Bernhard Ganter · Rudolf Wille

Formal Concept
Analysis
Mathematical natural stagnant

Foundations inland constant

Second Edition

temporary maritime

plash, sea,
puddle lagoon

channel mere, pond,

reservoir
trickle, river, rivulet,
brook, beck, runnel, tarn, lake,
rill, stream, torrent canal pool
Formal Concept Analysis
Bernhard Ganter • Rudolf Wille

Formal Concept Analysis

Mathematical Foundations

Second Edition
Bernhard Ganter Rudolf Wille (1937-2017)
Institute of Algebra Fachbereich Mathematik
Technische Universität Dresden Technische Universität Darmstadt
Dresden, Germany Darmstadt, Germany

ISBN 978-3-031-63421-5 ISBN 978-3-031-63422-2 (eBook)

https://doi.org/10.1007/978-3-031-63422-2
The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG
1999, 2024
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the
whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, expressed or implied, with respect to the material contained herein or for any errors or
omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

If disposing of this product, please recycle the paper.

Preface to the first edition

Formal Concept Analysis is a field of applied mathematics based on the math-

ematization of concept and conceptual hierarchy. It thereby activates mathematical
thinking for conceptual data analysis and knowledge processing. The underly-
ing notion of “concept” evolved early in the philosophical theory of concepts and
still has effects today. For example, it has left its mark in the German standards
DIN 2330 and DIN 2331. In mathematics it played a special role during the emer-
gence of mathematical logic in the 19th century. Subsequently, however, it had
virtually no impact on mathematical thinking. It was not until 1979 that the topic
was revisited and treated more thoroughly. Since then, through a large number
of contributions, Formal Concept Analysis has obtained such breadth that a sys-
tematic presentation is urgently needed, but can no longer be realized in one
volume.
Therefore, the present book focuses on the mathematical foundations of For-
mal Concept Analysis, which can be regarded chiefly as a branch of applied lat-
tice theory. A series of examples serves to demonstrate the utility of the mathe-
matical definitions and results; in particular, to show how Formal Concept Ana-
lysis can be used for the conceptual unfolding of data contexts. These examples
do not play the role of case studies in data analysis. A separate volume is in-
tended for a comprehensive treatment of methods of conceptual data and knowl-
edge processing. The general foundations of Formal Concept Analysis will also
be treated separately.
It is perfectly possible to use Formal Concept Analysis when examining hu-
man conceptual thinking. However, this would be an application of the mathe-
matical method and a matter for the experts in the respective science, for example
psychology. The adjective “formal” in the name of the theory has a delimiting
effect: we are dealing with a mathematical field of work, that derives its com-
prehensibility and meaning from its connection with well-established notions of
“concept”, but which does not strive to explain conceptual thinking in turn.
The mathematical foundations of Formal Concept Analysis are treated in
seven chapters. By way of introduction, elements of mathematical order and lat-
tice theory which will be used in the following chapters have been compiled in a
chapter “zero”. However, all difficult notation and results from this chapter will
be introduced anew later on. A reader who knows what is understood by a lattice
in mathematics may skip this chapter.
The first chapter describes the basic step in the formalization: An elementary
form of the representation of data (the “cross table”) is defined mathematically
(“formal context”). A formal concept of such a data context is then explained.
The totality of all such concepts of a context in their hierarchy can be interpreted
as a mathematical structure (“concept lattice”). It is also possible to allow more
complex data types (“many-valued contexts”). These are then reduced to the
basic type by a method of interpretation called “conceptual scaling”.
The second chapter examines the question of how all concepts of a data context
can be determined and represented in an easily readable diagram. In addition,

v
vi Preface to the first edition

implications and dependencies between attributes are dealt with. The third chap-
ter supplies the basic notions of a structure theory for concept lattices, namely
part- and factor structures as well as tolerance relations. In each case the extent
to which these can be elaborated directly within the contexts is studied.
These mathematical tools are then used in the fourth and fifth chapter, in or-
der to describe more complex concept lattices by means of decomposition and
construction methods. Thus, the concept lattice can be split up into (possibly
overlapping) parts, but it is also possible to use the direct product of lattices or
of contexts as a decomposition principle. A further approach is that of substitu-
tion. In accordance with the same principles, it is possible to construct contexts
and concept lattices. As an additional construction principle, we shall describe a
method of doubling parts of a concept lattice.
The structural properties examined in mathematical lattice theory, for exam-
ple the distributive law and its generalizations or notions of dimension, play a
role in Formal Concept Analysis as well. This shall be treated in the sixth chap-
ter. The seventh chapter finally deals with structure-comparing maps, examining
various kinds of morphisms. Particular attention is given to the scale measures,
occuring in the context of conceptual scaling.
We limit ourselves to a concise presentation of ideas for reasons of space.
Therefore, we endeavour to give a complete reference to further results and the
respective literature at the end of each chapter. However, we have only taken into
account such contributions closely connected with the topic of the book, i.e., with
the mathematical foundations of Formal Concept Analysis. The index contains
all technical terms defined in this book, and in addition some particularly im-
portant keywords. The bibliography also serves as an author index.
The genesis of this book has been aided by the numerous lectures and activ-
ities of the “Forschungsgruppe Begriffsanalyse” (Research Group on Concept
Analysis) at Darmstadt University of Technology. It is difficult to state in detail
which kind of support was due to whom. Therefore, we can here only express
our gratitude to all people who contributed to the work presented in this book.
Two years after the German edition, this English translation has been finished.
In its content there are only a few minor changes. Although there is vivid activity
in the field, the mathematical foundations of Formal Concept Analysis have been
stable over the last years.
The authors are extremely grateful to Cornelia Franzke for her precise and
cooperative work when translating the book. They would also like to thank K.A.
Baker, P. Eklund and R.J. Cole, M.F. Janowitz, and D. Petroff for their careful
proofreading.
Preface to the second edition

For December 13, 1979, the minute book for the almost daily after lunch semi-
nar („Mittagsseminar“) of Rudolf Wille’s research group at TU Darmstadt notes
his lecture on “Lattice Theory as Algebra of Concepts”, in which he explained
how the relation between concept extents and concept intents can be understood
as a Galois connection and how therefore complete lattices can be interpreted
as algebraic structures of concept hierarchies. This was preceded by months of
lively discussion in another seminar, dealing with the meaning and significance
of mathematical order theory. Two authors had a decisive influence on this sem-
inar and thus on the genesis of the Formal Concept Analysis:
• the mathematician Garrett Birkhoff with his application-oriented view of lat-
tice theory,1 and
• the educationalist Hartmut von Hentig with his critical yet constructive under-
standing of science.2
It was in this seminar that Wille presented his definition of formal concepts for
the first time, according to K. E. Wolff [423], who reports in more detail on this
development (another source is [418]). Years of intense productivity followed,
during which Wille worked out the theory of Formal Concept Analysis with his
collaborators and students.
Sadly, Rudolf Wille, the founder and idea provider of Formal Concept Analy-
sis, passed away in January 2017. A new edition of this book had already been
planned in 2003, when Rudolf was still in good health. The book had proven it-
self, so we wanted to get by with only a few changes. These included Contextual
Attribute Logic, which later was incorporated as the first section of a new eighth
chapter. However, it took twenty years before work on a new edition began in
earnest. In the meantime, the first edition (published in German 1996 [161],
in English 1999 [163], in Chinese 2005 [164]) had been cited several thousand
times.
The variety of publications on Formal Concept Analysis, including quite a
number of books, is almost impossible to keep track of. A first, but by no means
complete, impression is given by the proceedings of the conference series ICFCA
(International Conference on Formal Concept Analysis) and CLA (Concept Lattices
and their Applications). Many authors have dealt with applications and with the
algorithms necessary for them, i.e., with topics that are only marginally men-
tioned in this book. But even if only those contributions that concern the “math-
ematical foundations” are considered, there are far too many to be included in
the book.

1 G. Birkhoff: Lattice Theory. Amer. Math. Soc. Providence, 1st edition 1940, 2nd (revised) edi-
tion 1948, 3rd (new) edition 1967
2 H. von Hentig: Magier oder Magister? Über die Einheit der Wissenschaft im Verständigungsprozeß.

Klett, Stuttgart 1972

vii
viii Preface to the second edition

The now immense number of contributions to Formal Concept Analysis al-

lows us to say with all modesty that the development of this area has been suc-
cessful, at least more successful than we could have imagined in the beginning.
We have noticed time and again that this success has surprised, even alienated,
some of our experienced and respected scientific colleagues. And we observed
that many attempts to approach the same mathematical topic differently were
not very successful. Here, in the preface to the second edition, in a place which
experience has shown to be skipped by even the most careful readers, is an op-
portunity to mention some of the reasons for this success from the author’s sub-
jective point of view.
◦ Formal Concept Analysis arose out of the question of the meaning of mathe-
matics, especially order and lattice theory, and not just its applicability. The ex-
plicit and repeatedly stated goal was to support and strengthen human rational
thinking and cognitive autonomy with mathematical means. The fact that such
ideas are considered unmathematical by many colleagues was seen as a mod-
ern narrowing of mathematics by Rudolf Wille’s group. They were by no means
alone in pursuing such goals. What was unusual, however, was that specialists
from one of the most abstract corners of mathematics, who valued tackling hard
mathematical problems, opened up to these tasks and pursued them systemati-
cally and energetically.
◦ The methods are developed for contextual knowledge processing. The word is
meant to express that the focus is on real data, which may come, e.g., from obser-
vations or from knowledge collections. Data analysis and knowledge processing
should make such data better and more reliably understandable and thus acces-
sible to human judgment. Abstract laws of mathematics, logic and stochastics are
important for this, but they have to prove themselves on concrete data by con-
tributing to their understanding. It was frowned upon in the “Research Group
on Concept Analysis” to present invented sample data. The effectiveness of new
methods should always be demonstrated on real data, or at least on excerpts
from such data sets.
◦ And consequently, it had to be possible to answer real questions about the
data. The research group was therefore transdisciplinary and philosophically ac-
tive (which some mathematicians found difficult). It was not unusual for a psy-
chologist or a linguist to speak at the (almost daily) seminar, and writings by
C.S. Peirce, K.-O. Apel, J. Habermas, and many others were read in the study
groups. In this respect, too, the group was not isolated. V. Duquenne introduced
them to the logic of the 17th century Port Royal school, S.O. Kuznetsov to the
Moscow approaches, which in turn referred to the English philosopher J.S. Mill.
◦ Consequently, Formal Concept Analysis has developed its own technical terms
and notations. There are occasional attempts to reformulate results in usual math-
ematical language, pointing out that the core is Birkhoff’s result of 1940, on which
the Basic Theorem is based. In fact, this theorem is central to the theory. It is a
pity that it received so little attention for forty years.
Acknowledgements

When it became clear that the second edition of this book was only a few
weeks away from completion, I asked potential readers for a critical prelimi-
nary reading. An invitation was distributed via [email protected],
the standard mailing list for Formal Concept Analysis. To my delight, I have
been contacted by many interested readers from a wide variety of backgrounds,
and the responses that came from these contacts have noticeably improved the
manuscript. All the comments were friendly, constructive, and tried to encourage
the author, which was inspiring. Some have reported how they have benefited
from the first edition, others have suggested additional topics that should be in-
cluded or at least cited. However, this was mostly in vain, as the manuscript is
already extensive enough. And more than a few have honored the author (and
future readers) by spending a great deal of their personal time pointing out er-
rors, typos, and misleading wording.
My heartfelt thanks go to all those involved, namely Alexandre Albano, Jaume
Baixeries, Mayukh Bagchi, Radim Belohlavek, Daniel Borchmann, Dmitry Ig-
natov, René Jansen, Francesco Kriegel, Sergei O. Kuznetsov, Léonard Kwuida,
Thomas Morgenstern, Amedeo Napoli, Sergei Obiedkov, Vladimir Parkhomenko,
Uta Priss, Sebastian Rudolph, Baris Sertkaya, Gerd Stumme, Francisco Valverde-
Albacete, Karl Erich Wolff.
It is certainly no easy task to be the editor of a publishing house who over-
sees the creation of a book like this one. It takes a great deal of expertise and
sensitivity to support the often stubborn authors in their creative work while
maintaining the quality standards set by the publisher. Alexandru Ciolan, the
editor in charge at Springer, has done an exemplary job and deserves a special
word of thanks.
Finally, I thank Uta Wille for allowing her late father’s name to continue as
author. I am grateful for her trust in my work on the second edition.

Darmstadt, April 2024

Bernhard Ganter

ix
Contents

Preface
to the first edition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
to the second edition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

0 Order-theoretic foundations 1
0.1 Ordered sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
0.2 Complete lattices . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
0.3 Closure operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
0.4 Galois connections . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
0.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 20

1 Concept lattices of formal contexts 23

1.1 Formal context and formal concept . . . . . . . . . . . . . . . . . . 23
1.2 Formal context and concept lattice . . . . . . . . . . . . . . . . . . 29
1.3 Context constructions . . . . . . . . . . . . . . . . . . . . . . . . . . 41
1.4 Conceptual scaling of many-valued contexts . . . . . . . . . . . . 59
1.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 67

2 Determination and representation 77

2.1 All formal concepts of a formal context . . . . . . . . . . . . . . . . 77
2.2 Diagrams of concept lattices . . . . . . . . . . . . . . . . . . . . . . 83
2.3 Implications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
2.4 Clauses and dependencies . . . . . . . . . . . . . . . . . . . . . . . 104
2.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 112

3 Parts, factors, and bonds 117

3.1 Subcontexts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.2 Complete congruences and tolerances . . . . . . . . . . . . . . . . 124
3.3 Closed subrelations . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
3.4 Bonds and connections . . . . . . . . . . . . . . . . . . . . . . . . . 147
3.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 160

4 Decompositions of concept lattices 163

4.1 Subdirect decompositions . . . . . . . . . . . . . . . . . . . . . . . 163
4.2 Atlas decompositions . . . . . . . . . . . . . . . . . . . . . . . . . . 170

xi
xii Contents

4.3 Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184

4.4 Tensorial decompositions . . . . . . . . . . . . . . . . . . . . . . . 197
4.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 214

5 Constructions of concept lattices 215

5.1 Subdirect product constructions . . . . . . . . . . . . . . . . . . . . 216
5.2 Gluings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
5.3 Local doubling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
5.4 Tensorial constructions . . . . . . . . . . . . . . . . . . . . . . . . . 237
5.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 247

6 Properties of concept lattices 249

6.1 Size and breadth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
6.2 Distributivity, modularity and semimodularity . . . . . . . . . . . 252
6.3 Semidistributivity and local distributivity . . . . . . . . . . . . . . 262
6.4 Dimension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
6.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 279

7 Context comparison and conceptual measurement 283

7.1 Automorphisms and foldings . . . . . . . . . . . . . . . . . . . . . 283
7.2 Morphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
7.3 Scale measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
7.4 Measurability theorems . . . . . . . . . . . . . . . . . . . . . . . . 303
7.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 308

8 Contextual concept logic 311

8.1 Contextual attribute logic . . . . . . . . . . . . . . . . . . . . . . . 312
8.2 Pre-, proto-, and semiconcepts . . . . . . . . . . . . . . . . . . . . . 317
8.3 Concept algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
8.4 Contextual logic with relations . . . . . . . . . . . . . . . . . . . . 326
8.5 Notes, references, and trends . . . . . . . . . . . . . . . . . . . . . 332

References 335

Formal contexts and concept lattices in this book 361

Index 363
Chapter 0
Order-theoretic foundations

Formal Concept Analysis is based on mathematical order theory, in particular

on the theory of complete lattices. The reader is not required to be familiar with
these areas. The mathematical foundations are surveyed in this chapter. How-
ever, there was no room for a comprehensive introduction to order theory. For
this, we refer to the textbooks listed at the end of this chapter. In general, the
reader is supposed to have experience with mathematical texts: we use the tech-
nical language of mathematics, in particular of set theory, without further expla-
nation.
The first section introduces ordered sets, the second complete lattices. These
two sections constitute the basis for the following chapters. The third section,
dealing with closure systems, and the fourth on Galois connections may be
skipped on first reading. Much of what they contain will be introduced again
later under a different name. These two sections show how the basic notions of
Formal Concept Analysis have their roots in order and lattice theory. In most
aspects we follow the “classical” presentation by G. Birkhoff.

0.1 Ordered sets

Definition 1 A binary relation R between two sets M and N is a set of pairs

(m, n) with m ∈ M and n ∈ N , i.e., a subset of the set M × N of all such pairs.
Instead of (m, n) ∈ R we often write m R n. If N = M , we speak of a binary
relation on the set M . R−1 denotes the inverse relation to R, that is, the relation
between N and M with n R−1 m :⇔ m R n. ♦

The symbol ♦ indicates the end of a definition, > the end of an example. The “end-of-proof”
symbol  is also used in this book if no proof is given.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 1

B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_1
2 0 Order-theoretic foundations

Definition 2 A binary relation R on a set M is called an order relation (or

shortly an order), if it satisfies the following conditions for all elements x, y, z ∈
M:

1. x R x (reflexivity)
2. x R y and x 6= y together imply not y R x (antisymmetry)
3. x R y and y R z together imply x R z (transitivity)
For an order relation, we usually use the symbol ≤ instead of R (and ≥ instead
of R−1 ), and we write x < y for x ≤ y and x 6= y. As usual, we read x ≤ y as “x
is less than or equal to y”, etc. An ordered set is a pair (M, ≤), with M being a set
and ≤ an order relation on M .1 ♦
Examples of ordered sets are numerous in this book. The following are certainly
familiar to the reader: The real numbers R with the usual ≤-relation, but also the
space Rn with

(x1 , x2 , . . . , xn ) ≤ (y1 , y2 , . . . , yn ) : ⇔ xi ≤ yi for i = 1, 2, . . . , n,

the natural numbers N with the divisibility relation | ; the powerset P(X) of all
subsets of any set X with set inclusion. Even the equality relation = is a (trivial)
example of an order. >

Definition 3 An element a of an ordered set is called a lower neighbor of b, if

a < b and there is no element c fulfilling a < c < b. In this case, b is an upper
neighbor of a, and we write a ≺ b. ♦
Every finite ordered set (M, ≤) can be represented by a line diagram (also
called a Hasse diagram by many authors). The elements of M are depicted by
small circles in the plane. If x, y ∈ M with x ≺ y, the circle corresponding to y is
depicted above the circle corresponding to x (permitting sideways shifts), and
the two circles are joined by a line segment. From such a diagram we can read off
the order relation as follows: x < y if and only if the circle representing y can be
reached by an ascending path from the circle representing x. Figure 0.1 presents
line diagrams for all ordered sets with up to four elements.
Definition 4 Two elements x, y of an ordered set (M, ≤) are called comparable
if x ≤ y or y ≤ x, otherwise they are incomparable. A subset of M in which
any two elements are comparable is called a chain; a subset in which any two
distinct elements are incomparable is called an antichain. If M itself is a chain,
then ≤ is a linear order. The width of a finite ordered set (M, ≤) is defined to be
the maximal size of an antichain in (M, ≤), for an arbitrary ordered set (M, ≤)
it is defined to be the supremum of the sizes of antichains in (M, ≤). Similarly,
the length is defined to be the supremum of the sizes of chains in (M, ≤), minus
one. ♦
1Instead of ordered sets, many authors speak of partially ordered sets and use the abbreviation
poset.
0.1 Ordered sets 3

d d
d d d d d d d
d
d d d d d Jd d Jd d d d d d
J J

d d d d d d d d
d Jd Jd d dJ d d d d d d d d d
J J J

Jd d d Jd d d d d @d d d
J J @

d d d d d d d d d d
d d

A A

A A

d
d A d A d
d d d
d A d A d
d d d d d d

Figure 0.1 Line diagrams of all ordered sets with up to four elements.

Definition 5 If (M, ≤) is an ordered set and a, b, c, d are elements of M with

b ≤ c, we define the interval

[b, c] := {x ∈ M | b ≤ x ≤ c}.

The set c c c
(a] := {x ∈ M | x ≤ a} d

is called a principal ideal and

ac cb
[d) := {x ∈ M | x ≥ d}

is called a principal filter.

Thus, a ≺ b is equivalent to a < b and [a, b] = {a, b}. ♦

Definition 6 A map ϕ : M → N between two ordered sets (M, ≤) and (N, ≤) is

called order-preserving (or, synonymously, monotone), if

x ≤ y ⇒ ϕx ≤ ϕy

for all x, y ∈ M . If ϕ furthermore fulfills the converse implication

x ≤ y ⇐ ϕx ≤ ϕy,

ϕ is called an order-embedding. In this case, ϕ is necessarily injective. A bijective

order-embedding is called an (order-) isomorphism. (M, ≤) ∼ = (N, ≤) expresses
that (M, ≤) and (N, ≤) are order-isomorphic. If ϕ satisfies
4 0 Order-theoretic foundations

x ≤ y ⇒ ϕx ≥ ϕy

for all x, y ∈ M , then ϕ is order-reversing or antitone. ♦

c
Not every bijective order-preserving map is an c c
order-isomorphism, as the example shows. In c @@c
c
order to prove that a certain order-preserving @@c
map ϕ is an isomorphism, it is usually shown c
that the inverse map ϕ−1 exists and is also order-
Bijective, order preserving,
preserving.
but not an isomorphism.

Definition 7 The (direct) product of two ordered sets (M1 , ≤) and (M2 , ≤) is
defined to be the ordered set (M1 × M2 , ≤) with

(x1 , x2 ) ≤ (y1 , y2 ) :⇔ x1 ≤ y1 and x2 ≤ y2 . ♦

The definition of the product can be extended to any number of factors: If T is

an index set and (Mt , ≤), t ∈ T , are ordered sets, then

× (M , ≤) := × M , ≤
with
t∈T
t
t∈T
t

(xt )t∈T ≤ (yt )t∈T :⇔ xt ≤ yt for all t ∈ T.

c c c
H 
c c @c c c c c @ c H c c c @ c
@ @ HH @
∼

× =
HHHHHH     
c c @ cHH c
Hc
@c@@c
@ @ @ @ @ H @
H
H 
HH 
@ c
@
H
H

Figure 0.2 An example of a product of two ordered sets.

Definition 8 To be able to define the cardinal sum or disjoint union of two or-
dered sets, we first introduce the notation

Ṁt := {t} × Mt .

The sets Ṁ1 and Ṁ2 will then be disjoint copies of M1 and M2 . We define

(M1 , ≤) + (M2 , ≤) := (Ṁ1 ∪ Ṁ2 , ≤),

the order relation being specified as follows:

0.2 Complete lattices 5

(s, a) ≤ (t, b) : ⇔ s = t and a ≤ b in Ms .

This definition is also easily generalized in the case of any number of sum-
mands. ♦
The Duality Principle for ordered sets. The inverse relation ≥ of an order rela-
tion ≤ is also an order relation. It is called the order which is dual to ≤. A line
diagram of the dual ordered set (M, ≤)d := (M, ≥) can be obtained from the line
diagram of (M, ≤) by a horizontal reflection. If (M ≤) = ∼ (N, ≤)d , the two orders
are called dually isomorphic.
One obtains the dual statement Ad of an order-theoretic statement A (which,
apart from purely logical components, contains only the symbol ≤), by replacing
all occurences of the symbol ≤ in A by ≥. A holds for an ordered set if and only
if Ad holds for the dual ordered set. This Duality Principle is used to simplify
definitions and proofs. If a theorem asserts two statements that are dual to each
other, one usually proves only one of them, the other one follows “dually”, i.e.,
with the same proof for the dual order.
Definition 9 Let (M, ≤) be an ordered set and A a subset of M . A lower bound
of A is an element s of M with s ≤ a for all a ∈ A. An upper bound of A is
defined dually. If there is a largest element in the set ofVall lower bounds of A, it
is called the infimum of A and is denoted by inf A orW A; dually, a least upper
bound is called supremum and denoted by sup A or A. If A = {x, y}, we also
write x ∧ y for inf A and x ∨ y for sup A. Infimum and supremum are frequently
also called meet and join. ♦
If an order R on M is a subset of a second, say S, then S is called an order
extension of R. If S is even a linear order, then it is called a linear extension of
R. Each order is the intersection of its linear extensions, and the smallest num-
ber of linear extensions needed to represent the order as an intersection is its
order dimension. This dimension is discussed in detail in Section 6.4, although a
different, yet equivalent definition is used there.

0.2 Complete lattices

Definition 10 An ordered set L := (L, ≤) is a lattice if for every two elements

x and y in L both the supremum x ∨Wy and the infimum x ∧Vy exist. L is called
a complete lattice if the supremum X and the infimum WX exist for every
subset X of L. Every complete lattice L has a largest element, L, called the
Vunit
element of the lattice, denoted by 1L . Dually, the smallest element 0L := L is
called the zero element.2 ♦
The definition of a complete lattice requires that supremum and infimum exist
for every subset X, in particular for X = ∅. We have ∅ = 1L and ∅ = 0L ,
V W

2 Other names for these elements are top and bottom, other symbols are > and ⊥.
6 0 Order-theoretic foundations

d d d d
@ @
d d d @d d d
@d
@
@ @ @
d d d @d @d d @d d 
@ @ @ 
d @d d @d @ d

Figure 0.3 Line diagrams of the lattices with five elements.

from which it follows that L 6= ∅ for every complete lattice. Every non-empty
finite lattice is a complete lattice.
The order relation can be reconstructed from the lattice operations infimum
and supremum by
x ≤ y ⇔ x = x ∧ y ⇔ x ∨ y = y.
If T isWan index set and X :=V{xt | t ∈ T } aVsubset of L, instead of X we also
W
write t∈T xt and instead of X we write t∈T xt . The operations of the supre-
mum and infimum, respectively, are associative. The familiar particular case of
the associative laws, i.e., x∧(y∧z) = (x∧y)∧z, respectively x∨(y∨z) = (x∨y)∨z,
can be generalized as follows: If {Xt | t ∈ T } is a set of subsets of L, then
_  _  _ [  ^  ^  ^ [ 
Xt = Xt and dually Xt = Xt .
t∈T t∈T t∈T t∈T

The Duality Principle for lattices. The definitions of a lattice and a complete
lattice, respectively, are self-dual: If (L, ≤) is a (complete) lattice, then so is
(L, ≤)d = (L, ≥). Therefore, the Duality Principle for ordered sets carries over
to lattices: We obtain the dual statement of an order-theoretic statement when
we replace the symbols ≤, ∨, ∧, , , 0L , 1L etc. by ≥, ∧, ∨, , , 1L , 0L etc.
W V V W

Proposition 1 An ordered set in which the infimum exists for every subset is a complete
lattice.
Proof Let X be any subset of the ordered set. We have to prove that the supre-
mum of X exists. The set S of all upper bounds of X has an infimum s (even if
S is empty). Every element of X is a lower bound of S, i.e., is less than or equal
to s. Hence s itself is an upper bound of X and consequently the supremum. 

Examples of lattices.
1) For every set M the powerset P(M ), i.e., the set of all subsets of M , is ordered
by set inclusion ⊆ and (P(M ), ⊆) is a complete lattice. In this case the lattice
operations supremum and infimum are set union and intersection.
0.2 Complete lattices 7

2) Every closed real interval [a, b] in its natural order forms a complete lattice
([a, b], ≤) with the usual infimum and supremum, respectively, as lattice opera-
tions. The ordered set (R, ≤), on the other hand, is a lattice, but it is not complete:
It lacks a greatest and a least element.
We will give further examples of complete lattices from mathematics in the
next section and throughout this book. >
Definition 11 For an element v of a complete lattice L := (L, ≤) we define
_
v∗ := {x ∈ L | x < v}
^
and v ∗ := {x ∈ L | v < x}.

We call v -irreducible3 if v 6= v∗ , i.e., if v cannot be represented as the supre-

W
mum of strictly smaller Velements. In this case, v∗ is the unique lower neighbor
of v. Dually we call v -irreducible4 if v 6=V v ∗ . J(L) denotes the set of all -
W
irreducible elements and M (L) the set of all -irreducible elements. A set X ⊆ L
is called supremum-dense in L if every element from L can be represented as the
supremum of a subset of X and, dually, infimum-dense if v = {x ∈ X | v ≤ x}
V
for all v ∈ L. ♦

Proposition 2 An element v of aVcomplete lattice is -irreducible if and only if it has

W
exactly one lower neighbor, and -irreducible if andWonly if it has exactly one upper
neighbor. Every supremum-dense subset contains all -irreducible elements and every
infimum-dense subset contains all -irreducible elements. Conversely, in a finite lattice
V
L the set J(L) is supremum-dense and M (L) is infimum-dense.

Proof v is -irreducible if and only if v∗ 6= v. This, on the other hand, is equiva-

W
lent to the fact that v∗ is the largest
V element less than v, that is, it is in particular
the only lower neighbor of v. For -irreducible elements we conclude dually. The
second statement of the proposition is trivial, the third is proved inductively:
Every element v which is not -irreducible W itself, is the supremum of strictly
W
smaller elements. If those are suprema of -irreducible elements, so is v. 
It is easy to Vstate examples of complete lattices which contain neither -
W
irreducible nor -irreducible elements, as for instance the real intervalW [0, 1] in its
natural order. The upper neighbors of the zero element are always -irreducible
(if they exist). They are called the atoms of the
V lattice. The coatoms, i.e., the lower
neighbors of the unit element, are always -irreducible. A complete lattice in
which every element is the supremum of atoms is called atomistic. Figure 1.11
on p. 40 shows how this property relates to similar ones.
Definition 12 A subset U of a complete lattice L which is closed under suprema,
i.e., for which holds

3 Read: supremum-irreducible
4 Read: infimum-irreducible
8 0 Order-theoretic foundations
_
T ⊆U ⇒ T ∈ U,

isVa -subsemilattice of L. Dually, a subset which is closed under infima is called

W
a -subsemilattice. A subset which is closed under both suprema and infima is
called a complete sublattice. ♦
Note that a complete Wsublattice must always contain the largest element ∅ and
V
the smallest element ∅ of L. Thus, even in the finite case, not every sublattice
is a complete sublattice.
Definition 13 A map ϕ : L1 −→ L2 between two complete lattices L1 and L2 is
called supremum-preserving, if 5
_ _
ϕ X= ϕ(X)

of L1 . Another name is -morphism, and dually: in-

holds for every subset X V
W
fimum-preserving map, -morphism. If ϕ is both supremum-preserving and
infimum-preserving, then ϕ is a complete lattice homomorphism or complete
homomorphism. ♦
Every supremum-preserving map, in particular every complete homomor-
phism, is order-preserving. Conversely, every order-isomorphism between com-
plete lattices is automatically a lattice isomorphism, i.e., a bijective complete ho-
momorphism.
Lattice theory explores the structural regularities of lattices. At this point we
mention only the most important one, distributivity. This and other properties
will be discussed in more detail in Chapter 6.
Definition 14 A complete lattice L := (L, ≤) is called
• distributive if the following distributive laws hold for all x, y, z ∈ L:

(D∧ ) x ∧ (y ∨ z) = (x ∧ y) ∨ (x ∧ z)
(D∨ ) x ∨ (y ∧ z) = (x ∨ y) ∧ (x ∨ z)

• completely distributive if the following generalization of the two distributive

laws for arbitrary infima and suprema holds for all index sets S, T 6= ∅:
^ n_ o
(DW V ) {xs,t | t ∈ T } | s ∈ S =
_ n^ o
{xs,ϕs | s ∈ S} | ϕ : S → T .

♦
Distributivity and complete distributivity are self-dual properties: if they
hold for a lattice L, they also hold for Ld . The above-mentioned properties trans-
fer to complete sublattices. Powerset lattices are completely distributive.
5 ϕ(X) here stands for {ϕx | x ∈ X}.
0.3 Closure operators 9

For the special case S := {0, 1}, x0,t := x and x1,t := xt for all t ∈ T , the
property of complete distributivity yields the weaker law
_ _
(D∧ W ) x∧ xt = (x ∧ xt ).
t∈T t∈T

The dual law (D∨ V ) holds in the lattice of the closed sets of any given topolog-
ical space, and (D∧ W ) holds in the lattice of all open sets. Those lattices are not
completely distributive in general.

Remark This book deals almost exclusively with complete lattices. Unlike lat-
tices, however, they are not algebraic structures with fundamental operations of
fixed arity. The complete lattices therefore do not form an equational class in the
sense of Universal Algebra. Certain algebraic constructions, such as free algebras,
are therefore only available with restrictions, if at all.

0.3 Closure operators

Definition 15 A closure system6 on a set G is a set of subsets which contains

G and is closed under intersections. Formally: A ⊆ P(G) is a closure system if
G ∈ A and \
X⊆A⇒ X ∈ A.
The first condition (that G ∈ A) follows from the second as the case that X = ∅,
since G = ∅ ∈ A. The elements of the closure system are called closed sets.
T
A closure operator ϕ on G is a map assigning a closure ϕX ⊆ G to each subset
X ⊆ G under the following conditions:
1. X ⊆ Y ⇒ ϕX ⊆ ϕY (monotony)
2. X ⊆ ϕX (extensity)
3. ϕϕX = ϕX (idempotency)
♦

Proposition 3 An operator ϕ : P(G) → P(G) is a closure operator if and only if it is

extensive and satisfies the following condition for all X, Y ⊆ G:

ϕ(X ∪ Y ) = ϕ(X ∪ ϕY ).

Proof When ϕ is extensive (i.e., satisfies condition 2 above) then Y ⊆ ϕY and

thus X ∪ Y ⊆ X ∪ ϕY . If ϕ also is monotone, then this implies ϕ(X ∪ Y ) ⊆ ϕ(X ∪
ϕY ). Extensity yields X ⊆ ϕ(X ∪ Y ) and monotonicity gives ϕY ⊆ ϕ(X ∪ Y ),

6 Also called a Moore family.

10 0 Order-theoretic foundations

which adds up to X ∪ ϕY ⊆ ϕ(X ∪ Y ). With monotony and idempotency this

yields ϕ(X ∪ ϕY ) ⊆ ϕϕ(X ∪ Y ) = ϕ(X ∪ Y ).
Conversely assume ϕ(X ∪Y ) = ϕ(X ∪ϕY ), and that ϕ is extensive. For X = Y
it follows that ϕX = ϕ(X ∪ ϕX), which yields idempotency. And since ϕY ⊆
X ∪ ϕY ⊆ ϕ(X ∪ ϕY ) = ϕ(X ∪ Y ), one gets ϕY ⊆ ϕX from Y ⊆ X. 
Closure system and closure operator are closely related, as shown by the follow-
ing theorem:
Theorem 1 If A is a closure system on G then
\
ϕA X := {A ∈ A | X ⊆ A}

defines a closure operator on G. Conversely, the set

Aϕ := {ϕX | X ⊆ G}

of all closures of a closure operator ϕ is always a closure system, and

ϕA ϕ = ϕ as well as AϕA = A.

Proof We break the proof into steps:

• ϕA is a closure operator: From X ⊆ Y it follows that

{A ∈ A | X ⊆ A} ⊇ {A ∈ A | Y ⊆ A}.

Since set intersection is monotone, this implies

\ \
ϕA X = {A ∈ A | X ⊆ A} ⊆ {A ∈ A | Y ⊆ A} = ϕA Y.

Extensity is trivial. Idempotency: According to the definition of ϕA , each ele-

ment of A which contains X also contains ϕA X, and vice versa.
• A
Tϕ is a closure system: Let X ⊆ Aϕ . On account of the extensity of ϕ we have
Because of monotony and idempotency, from
T X ∈ X it always
T
X ⊆ ϕ( X). T
follows that ϕ( X) ⊆ ϕX = X, which implies ϕ( X) ⊆ X.
T

• X ∈ A ⇔ X = {A ∈ A | X ⊆ A} ⇔ ϕA X = X ⇔ X ∈ AϕA .
T

• For A ∈ Aϕ , X ⊆ A is equivalent to ϕX ⊆ A. Hence

\ \
ϕAϕ X = {A ∈ Aϕ | X ⊆ A} = {A ∈ Aϕ | ϕX ⊆ A} = ϕX,

since ϕX ∈ Aϕ . 

Every closure system A can be understood as the set of all closures of a closure
operator. Therefore, the terms “closure” and “closed set” are used as synonyms.
0.3 Closure operators 11

W 4 If S A is a closure system, then (A, ⊆) is a complete lattice with X =

V
Proposition
X and X = ϕA X for all X ⊆ A. Conversely, every complete lattice is isomorphic
T
to the lattice of all closures of a closure system.
Proof ItSis obvious that X is the infimum and thus (compare Proposition 1)
T
that ϕA X is the supremum of X. If LTis a complete V lattice, then the system
{(x] | x ∈ L} is a closure system, since y∈T (y] = ( T ] holds for each subset
T ⊆ L. 
However, a system of sets A ⊆ P(G) for which (A, ⊆) is a complete lattice is
not necessarily a closure system. Rather, such families of sets are precisely the
image sets of monotonous, idempotent operators.

Remark The definition of a closure operator as an extensive, order-preserving

and idempotent map can be formulated for arbitrary ordered sets. If the ordered
set is a complete lattice, then the equivalence of closure operators and closure
systems also generalizes, the latter then being the -subsemilattices. However,
V
this generalization does not provide anything structurally new, since it follows
from Proposition 4 that any closure system in a complete lattice is isomorphic to
a closure system in a powerset.

Examples For many mathematical structures, the system of substructures is

a closure system. The powerset evidently is a closure system. Other important
examples are:
(1) subspaces: For any vector space V , the system U(V ) of all subspaces is a
closure system. The complete lattice (U(V ), ⊆) is called the subspace lat-
tice of V ; in this lattice U1 ∨ U2 = U1 + U2 and more generally
_
X = {u1 + u2 + . . . + un | there are U1 , . . . , Un ∈ X
with ui ∈ Ui for i ∈ {1, . . . , n}}.

(2) subgroups: For any group G, the set U(G) of all subgroups of G is a clo-
sure system. The complete lattice (U(G), ⊆) is called subgroup lattice of G.
Provided that G is commutative, U1 ∨ U2 = U1 + U2 , and more generally,
_
X = {u1 + u2 + . . . + un | there are U1 , . . . , Un ∈ X
with ui ∈ Ui for i ∈ {1, . . . , n}}.

(3) closed sets: For a topological space T (for example, for Rn ), the set A(T )
of all closed sets of T is a closure system. In the complete lattice (A(T ), ⊆)
the supremum is equivalent to the topological closure of the union, i.e.,
_ [
X= X.
12 0 Order-theoretic foundations

(4) convex sets: For Rn the set C(Rn ) of all convex subsets is a closure system,
i.e., (C(Rn ), ⊆) is a complete lattice and in this lattice the supremum is the
convex closure of the union.
(5) the faces of polyhedra: For a polyhedron P , the set S(P ) of all faces of P is
a closure system. The complete lattice (S(P ), ⊆) is called the face lattice of
P ; for those lattices there is no general “good” description of the suprema.
(6) equivalence relations: For a set M , the set E(M ) of all equivalence rela-
tions on M is a closure system on M × M . The complete lattice (E(M ), ⊆)
is called the lattice of equivalence relations of M ; in this lattice
_
X = {(a, b) ∈ M × M | there are R1 , . . . , Rn ∈ X and
(xi , xi+1 ) ∈ Ri for i ∈ {1, . . . , n}
with a = x1 and b = xn+1 }.

>

An algorithm for finding closed sets

Theorem 1 can be used to determine the system of all closures of a given closure
operator. However, that method is cumbersome. For practical computations we
introduce the Next Closure algorithm, see Algorithm 1. It assumes that the

Algorithm 1 Next Closure(A, G, ϕ)

Input: A closure operator X 7→ ϕX on a finite linearly ordered set G and a subset A ⊆ G.
Output: The lectically next closed set after A if it exists; ⊥, otherwise.

for all g ∈ G in reverse order do

if g ∈ A then
A := A \ {g}
else
B := ϕ(A ∪ {g})
if B \ A contains no element < g then
return B
return ⊥

base set G is finite and linearly ordered (in an arbitrary way). The closures are
created one after the other. The previously generated closure is input for the
generation of the next one. It starts with the closure of the empty set, i.e. with ϕ∅.
The closures are generated in lectic order, defined as follows: A subset A ⊆ G
is lectically smaller than a subset B ⊆ G if A 6= B and the smallest element in
which A and B differ belongs to B.
0.3 Closure operators 13

More about the Next Closure algorithm will be said in Section 2.1, notably in
Theorem 5 (p. 81). The refinements presented in the Theorems 6 and 60 (p. 285)
generally work for closure systems.
Example To demonstrate how the algorithm works, we apply it to the closure
system of all closure systems on a fixed base set {1, . . . , m}, see Algorithm 2.
Each such closure system is a subset F ⊆ P({1, . . . , m}) of the powerset of
{1, . . . , m}. For the linear order, as required by Algorithm 1, we choose one which
extends the ⊇-order, i.e., with A ⊇ B ⇒ A ≤ B. A simple example of such an
order is obtained by encoding the subsets of {1, . . . , m} by the bit representations
of the integers 0, . . . , 2m −1 and then using the reverse order 0 > 1 > . . . > 2m −1.
Moreover, the closure system must be equipped with its generator map L : F → F
defined as follows: L[A] is the smallest set in F such that
\
A = {B ∈ F | B ≤ L[A], A ⊆ B}.

Algorithm 2 Next Closure System(m, ≤, F , L)

Input: m ∈ N and a ⊇-extending linear order ≤ on P({1, . . . , m}), and a closure system F ⊆
P({1, . . . , m}) with its generator map L.
Output: The lectically next closure system after F if it exists, together with its generator map;
⊥, otherwise.

if F 6= P({1, . . . , m}) then

A := max F {A is the largest subset not in F }
Add A to F
L[A] := A
for all B with B > A do
if L[B] > A then
Remove B from F
else if A ∩ B ∈
/ F or L[A ∩ B] > A then
Add A ∩ B to F
L[A ∩ B] := A
for all B < A do
if B ∈ F and A ∩ B ∈ / F then
Add A ∩ B to F
L[A ∩ B] := A
return F , L
else
return ⊥

Theoretically, this algorithm can be (and has been) used for counting: just
start with the trivial 1-element closure system and repeat. However, the number
of closure systems grows so quickly that a complete listing is out of reach already
for m = 7, see [80]. >
14 0 Order-theoretic foundations

Generators

If ϕ is a closure operator on G, then T ⊆ G is a generator of C ⊆ G if ϕT =

C holds. Being a generator is not a special property, because every subset is a
generator, namely of its closure. The question about the number numgen(C) of
generators of a given closed set C is not so easy to answer.7 Each of the 2|C|
subsets of C generates a closed set D ⊆ C. Therefore,
X
2|C| = numgen(D).
D⊆C,D=ϕD

An explicit formula is obtained from this using Möbius inversion (see Stan-
ley [349]): X
numgen(C) = 2|D| µ(D, C),
D⊆C,D=ϕD

where µ is the Möbius function of the closure system.

Definition 16 Let ϕ be a closure operator on a set G. A subset T ⊆ G is a minimal
generator8 if
ϕ(T \ {t}) 6= ϕ(T )
holds for all elements t ∈ T . ♦

Proposition 5 Each subset of a minimal generator is a minimal generator.

Proof Let S ⊆ T and assume that S is not a minimal generator. Then there is
some proper subset of S, say R, such that ϕR = ϕS. With Proposition 3 it follows
that ϕT = ϕ((T \ S) ∪ S) = ϕ((T \ S) ∪ ϕS) = ϕ((T \ S) ∪ ϕR) = ϕ((T \ S) ∪ R).
Thus (T \ S) ∪ R is a proper subset of T generating the same closure as T . So T
is no minimal generator either. 
Example Consider two closure systems on {a, b, c},

A1 := {∅, {c}, {b}, {b, c}, {a, b}, {a, b, c}}

A2 := {∅, {c}, {a}, {a, b}, {a, b, c}}.

The set of minimal generators in both cases is

{∅, {c}, {b}, {b, c}, {a}, {a, c}}. >

Some generalized finiteness condition is needed to ensure that each closed set
has a minimal generator, such as the condition of chain finiteness, which requires
that each chain of closed sets is finite, see p. 41. If each closed set has a minimal
7It has been proved to be a #P -complete problem (Kuznetsov [244]).
8“Minimal” in the sense of “non-redundant”, but not necessarily of “minimum size”. A closed
set may have several minimal generators of different cardinalities.
0.4 Galois connections 15

generator, then the number of closed sets is obviously less than or equal to the
number of minimal generators. This is exploited in Section 6.1 for the proof of
an upper bound on the number of formal concepts. The (chain finite) closure
systems, where each closed set has a unique minimal generator, are characterized
in Theorem 52 on p. 265.
Proposition 5 states that the set of minimal generators of a closure system on
G is always an order ideal9 in the powerset (P(G), ⊆) and thus a concept extent
of the free distributive lattice FCD(G), see Construction (4) on p. 48. Conversely,
for finite G each non-empty order ideal of (P(G), ⊆) is the set of minimal gen-
erators of a closure system (take as closed sets G together with all non-maximal
elements of the order ideal). This shows that the above example does not present
an exceptional case: There are 61 closure systems on {a, b, c} (see Figure 7.4 on
p. 291), but only 19 non-empty order ideals in (P({a, b, c}), ⊆).

Morphisms between closure systems

Two closure systems, say A on G and B on H, are isomorphic if there is a bijective

map α : G → H such that

C ∈ A ⇔ α(C) ∈ B.

This notion is more restrictive than that of lattice isomorphism: Each closure sys-
tem, when ordered by set inclusion ⊆, forms a complete lattice, and isomorphic
closure systems have isomorphic lattices. But the converse is not true. It hap-
pens that non-isomorphic closure systems give isomorphic lattices. This will be
explored in more detail when it comes to clarifying and reducing formal contexts
in Section 1.2.
In addition to isomorphisms, there are other useful mappings to compare clo-
sure systems. A mapping α : G → H is called closure continuous if the preimage
of every closed set ist closed, formally, if α−1 (C) ∈ A for every C ∈ B. The scale
measures, which play a central role in Chapter 7 of this book, are closure contin-
uous maps.

0.4 Galois connections

Definition 17 A Galois connection between two complete lattices10 L1 and L2

is a pair (ϕ, ψ) of maps ϕ : L1 → L2 , ψ : L2 → L1 , satisfying for all x ∈ L1 and
all y ∈ L2 the Galois condition

x ≤1 ψy ⇔ y ≤2 ϕx. ♦

9 Here we use definitions that will be explained later in Section 1.3.

10 Throughout this book L is an abbreviation for (L, ≤). Accordingly, Lt stands for (Lt , ≤t ).
16 0 Order-theoretic foundations

If this condition is met, then both mappings ϕ and ψ are order-reversing and their
compositions ψ ◦ ϕ and ϕ ◦ ψ are closure operators on L1 and L2 , respectively,
with dually isomorphic lattices of closed sets (comp. Propositions 6, 7, and 10
below).

Remark The name “Galois connection” comes from Galois theory, a branch of
algebra named after the mathematician Évariste Galois (1811–1832). There an
extension field E (satisfying certain conditions) of a field F is examined on the
basis of its Galois group, which consists of those automorphisms of E that leave F
pointwise fixed. The lattice L1 then is the lattice of intermediate fields, i.e., subfields
of E containing F . L2 is the subgroup lattice of the Galois group. ϕ maps an
intermediate field to its pointwise stabilizer, ψ maps a subgroup to the set of its
fixed points, which turns out to be an intermediate field. (The corresponding
context construction can be found on page 58).
In Galois theory a simplified notation is common, where the same symbol is
used for the two mappings ϕ and ψ, namely a simple “prime”, as it is also used
for the derivative of a function.
In fact, this Galois connection can be traced back to a single relation between
the field E and the Galois group G, namely the relation “is fixed by”. For an
intermediate field H and a subgroup S of G one has

H 0 = {α ∈ G | αh = h for all h ∈ H}
S 0 = {e ∈ E | αe = e for all α ∈ S}.

The condition in Definition 17 uses neither infima nor suprema and can there-
fore be generalized to arbitrary ordered sets without modification. An even fur-
ther generalization will be discussed in Section 3.4.
Definition 18 A pair (ϕ, ψ) of maps ϕ : P → Q and ψ : Q → P between two
ordered sets (P, ≤1 ) and (Q, ≤2 ) is called a Galois connection if it satisfies the
Galois condition p ≤1 ψq ⇔ q ≤2 ϕp for all p ∈ P and all q ∈ Q. The two maps
then are called dually adjoint to each other.11 ♦
See Figure 0.4 for an example.

Proposition 6 A pair (ϕ, ψ) of maps is a Galois connection if and only if all three of the
following conditions hold:
1. p1 ≤ p2 ⇒ ϕp1 ≥ ϕp2
2. q1 ≤ q2 ⇒ ψq1 ≥ ψq2
3. p ≤ ψϕp and q ≤ ϕψq .

Proof p ≤ ψq yields ϕp ≥ ϕψq because of 1., and from this we get ϕp ≥ q using
3., that is, one direction of the Galois condition. The other follows symmetrically.
11 In the following, we mostly use the same symbol ≤ for both ≤1 and ≤2 .
0.4 Galois connections 17

Figure 0.4 A pair of dually adjoint maps between two ordered sets.

Conversely, by the Galois condition it follows from ϕp ≤ ϕp that p ≤ ψϕp, i.e., 3.

Thus, from p1 ≤ p2 we can deduce that p1 ≤ ψϕp2 , which by the Galois condition
yields ϕp2 ≤ ϕp1 . 

Proposition 7 For every Galois connection (ϕ, ψ) we have

ϕ = ϕψϕ and ψ = ψϕψ.

Proof With q := ϕp we obtain by Condition 3. ϕp ≤ ϕψϕp and from p ≤ ψϕp by

1. ϕp ≥ ϕψϕp. 
The question under which conditions a given map ϕ can be extended to a
Galois connection is answered by the following proposition.
Proposition 8 A map
ϕ : (M, ≤) −→ (N, ≤)
between two ordered sets has a dual adjoint, if and only if the pre-image of each principal
filter is a principal ideal. The dual adjoint is unique.
Proof For any y ∈ N the pre-image of [y) equals {x ∈ M | y ≤ ϕx}. If ψ is dually
adjoint to ϕ, then by the Galois condition

{x ∈ M | y ≤ ϕx} = {x ∈ M | x ≤ ψy} = (ψy].

This also proves the uniqueness of ψ. Conversely, we can define ψ by

(ψy] := ϕ−1 ([y)) = {x ∈ M | y ≤ ϕx}

and thereby obtain x ≤ ψy ⇔ y ≤ ϕx, which according to the proposition is

characteristic of Galois connections. 
In the case that the two ordered sets are complete lattices, the characterization
can be further improved:
18 0 Order-theoretic foundations

Proposition 9 A map
ϕ : L1 −→ L2
between complete lattices has a dual adjoint if and only if
_ ^
ϕ xt = ϕxt
t∈T t∈T

holds for xt ∈ L1 .
Proof If ψ is dually adjoint to ϕ, then by the Galois condition
^
y≤ ϕxt ⇔ y ≤ ϕxt for all t ∈ T
t∈T
⇔ xt ≤ ψy for all t ∈ T
_
⇔ xt ≤ ψy
t∈T
_
⇔y≤ϕ xt .
t∈T

If, conversely, ϕ t∈T xt = t∈T W ϕxt , then ϕ definitely fulfills Condition 1) in

W V
Proposition 6. Defining ψy := {x ∈ L1 | y ≤ ϕx} for y ∈ L2 we immediately
obtain Condition V first part of 3). For y ∈ L2 it follows that
2) as well as the
ϕψy = ϕ {x ∈ L1 | y ≤ ϕx} = {ϕx | y ≤ ϕx} ≥ y, i.e., 3). Hence (ϕ, ψ) is a
W
Galois connection. 
We are particularly interested in the special case of a Galois connection be-
tween two powerset lattices. If M and N are two sets and ϕ : P(M ) → P(N ) is
a map (assigning a subset ϕA of N to each subset A of M ) and ψ is a map from
P(N ) to P(M ) such that conditions 1), 2) and 3) of Proposition 6 are fulfilled
(the order is set inclusion ⊆), then this is briefly called a Galois connection be-
tween M and N . The connection with the closure operators is emphasized by
the following proposition.
Proposition 10 The map A 7→ ψϕA is a closure operator on M and the map B 7→ ϕψB
is a closure operator on N . The maps ϕ and ψ, respectively, define dual isomorphisms
between the corresponding closure systems.
Proof Monotony and extensity of the maps follow immediately from Proposi-
tion 6, and idempotency follows from Proposition 7. We can also see from this
proposition that the closures in M are precisely the sets of the form ψB, B ⊆ N ,
and those in N are precisely the sets of the form ϕA, A ⊆ M . The maps ψB 7→
ϕψB and ϕA 7→ ψϕA, respectively, are order-reversing and by Proposition 7 in-
verse to each other, i.e., bijective. 
Galois connections between powerset lattices and binary relations between
their ground sets are closely interrelated. This is shown by the next theorem. In
preparation, we introduce some new notation:
0.4 Galois connections 19

Definition 19 If R ⊆ M × N is a relation, we write

X R := {y ∈ N | xRy for all x ∈ X} for X ⊆ M

and Y R := {x ∈ M | xRy for all y ∈ Y } for Y ⊆ N.

Since we have not presupposed that M and N are disjoint, this notation allows
ambiguous formulations, which, however, can easily be avoided.

Theorem 2 For every binary relation R ⊆ M × N , a Galois connection (ϕR , ψR ) be-

tween M and N is defined by

ϕR X := X R (= {y ∈ N | xRy for all x ∈ X})

and ψR Y := Y R
(= {x ∈ M | xRy for all y ∈ Y }).

If, conversely, (ϕ, ψ) is a Galois connection between M and N , then

R(ϕ,ψ) := {(x, y) ∈ M × N | x ∈ ψ{y}}

= {(x, y) ∈ M × N | y ∈ ϕ{x}}

is a binary relation between M and N , ϕR(ϕ,ψ) = ϕ, ψR(ϕ,ψ) = ψ and R(ϕR ,ψR ) = R.

Proof The Galois condition is easily verified. It implies that (ϕR , ψR ) is a Galois
connection and that the two sets used in order to define R(ϕ,ψ) are equal. Ac-
cording to this definition (x, y) ∈ R(ϕ,ψ) ⇔ y ∈ ϕ{x} and thus, by Proposition 9,
\
ϕX = ϕ{x}
x∈X
\
= ϕR(ϕ,ψ) {x}
x∈X
= ϕR(ϕ,ψ) X,

i.e., ϕR(ϕ,ψ) = ϕ and correspondingly ψR(ϕ,ψ) = ψ. The last statement R(ϕR ,ψR ) =
R follows immediately from the equivalence x ∈ ψR {y} ⇔ xRy. 
The use of the term “Galois connection” is not uniform. Some authors prefer
to replace one of the ordered sets by its dual. We prefer to call such pairs of maps
adjunctions. In the case of complete lattices we obtain:

V 11 To every -preserving map ϕ : L1 −→ L2 between complete lattices

W
Proposition
there is a -preserving map ψ : L2 −→ L1 with

ϕ(x) ≤ y ⇔ x ≤ ψ(y).

The maps ϕ and ψ uniquely determine each other: From ψ we obtain ϕ by

^
ϕ(x) = {y | x ≤ ψ(y)},
20 0 Order-theoretic foundations

and, conversely, ψ results from ϕ by

_
ψ(y) = {x | ϕ(x) ≤ y}.

In this case, ϕ is called a residuated map, ψ is called the residual map, and the maps
are adjoint to each other. If one of the maps is injective, the other one is surjective, and
vice versa.
Proof This is an immediate consequence of Proposition 9 and the preceding
propositions: If we replace L2 by the dual lattice Ld2 , ϕ and ψ form a Galois con-
nection. The relation between injectivity and surjectivity can be inferred from
Proposition 7. 

Remark: For mathematical Category Theory, the category of complete lattices

with the residuated mappings as morphisms is interesting because it has a spe-
cial property (it is ∗ -autonomous). This carries over to the category of formal con-
texts with the bonds as morphisms, which will be introduced in Section 3.4. More
about this can be found in the work of H. Mori [288] and of S. Borgwardt [64].

0.5 Notes, references, and trends

The idea of order is ubiquitous in mathematics. Its mathematical specification is

an achievement of the 19th century, at the end of which the notion of a lattice was
elaborated in the works of Dedekind, Peirce and Schröder. Much more can be
found in the detailed book Die Entstehung der Verbandstheorie by Mehrtens [280].
Duffus and Rival [109] give details on the history of the structural theory of
ordered sets.
A standard monograph on lattices and ordered sets is still Birkhoff’s poineer-
ing and influential book Lattice Theory [49, 50], which founded lattice theory in
1940. A popular modern introduction is offered by Davey and Priestley: Introduc-
tion to Lattices and Order [84]. The books of G. Grätzer, in especially his General
Lattice Theory [175], represent a comprehensive account in particular of the alge-
braic theory of lattices.
Most of what has been presented in Sections 0.3 and 0.4 originated in the 1940s
and was prompted by G. Birkhoff’s [49] discovery that from any binary relation
a complete lattice can be constructed. This was deepened by O. Ore [296] and
C. J. Everett [120]. For example, our Theorem 2 is covered almost verbatim by
Ore’s paper.
Many facts about Galois connections and residuated maps can be found in the
book Residuation Theory by Blyth and Janowitz [52]. Kuznetsov [246] discusses
early applications in data analysis.
The number of closure systems on an n-element base set is known up to n = 7,
thanks to contributions by Colomb et al. [80], Habib & Nourine [181]. See also
[62]. The values can be found in oeis.org as sequences A102896 and A193674.
0.5 Notes, references, and trends 21

Rudolph [324, 325] uses the lower and upper bounds given by Burosch et al. [71]
to investigate the possibilities of succintly representing closure systems. A dia-
gram of the closure systems on a three-element set is shown in Figure 7.4 on
p. 291.
Kuznetsov [247] (see also the earlier papers cited there) introduces a notion
of stability of closed sets based on the relative number of their generators. This
is closely related to the Möbius function. Hermann & Sertkaya [189] study com-
plexity issues connected with minimal generators.
Chapter 1
Concept lattices of formal contexts

The basic notions of Formal Concept Analysis are those of a formal context and a
formal concept. The adjective “formal” is meant to emphasize that we are dealing
with mathematical notions, which only reflect some aspects of the meaning of
context and concept in standard language. However, an overly frequent mention
of the suffix “formal” occasionally disrupts the flow of language. Thus, it shall
be understood that where we write context or concept we actually mean a formal
context or a formal concept, respectively.

1.1 Formal context and formal concept

Definition 20 A formal context K := (G, M, I) consists of two sets G and M

and a relation I between G and M . The elements of G are called the objects and
the elements of M are called the attributes of the formal context1 . In order to
express that an object g is in a relation I with an attribute m, we write g I m or
(g, m) ∈ I and read it as “ the object g has the attribute m ”.
The relation I is called the incidence relation of the formal context. Instead
of (g, m) 6∈ I we sometimes write g r I m. ♦
Example The formal context in Figure 1.1 was used when a Hungarian edu-
cational film was planned on the subject “Living Beings and Water”. Here the
objects are the living beings mentioned in the film and the attributes are the
properties which the film emphasizes.
A small context can be easily represented by a cross ta- m
ble2 , i.e., by a rectangular table the rows of which are ..
.
headed by the object names and the columns headed g ··· ··· ×
by the attribute names. A cross in row g and column m
means that the object g has the attribute m.
1 Strictly speaking: “formal objects” and “formal attributes”.
2 Not to be confused with a contingency table.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 23

B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_2
24 1 Concept lattices of formal contexts

a b c d e f g h i
1 | leech × × ×
2 | bream × × × ×
3 | frog × × × × ×
4 | dog × × × × ×
5 | waterweed × × × ×
6 | reed × × × × ×
7 | bean × × × ×
8 | maize × × × ×

Figure 1.1 Formal context of an educational film “Living Beings and Water”. The at-
tributes are: a: needs water to live, b: lives in water, c: lives on land, d: needs chlorophyll to
produce food, e: two seed leaves, f: one seed leaf, g: can move around, h: has limbs, i: suckles its
offspring.

Definition 21 For a set A ⊆ G of objects we define

A0 := {m ∈ M | g I m for all g ∈ A}

(the set of attributes common to the objects in A). Correspondingly, for a set B
of attributes we define

B 0 := {g ∈ G | g I m for all m ∈ B}

(the set of objects which have all attributes in B).3 ♦

Definition 22 A formal concept of the formal context (G, M, I) is a pair (A, B)
with
A ⊆ G, B ⊆ M, A0 = B and B 0 = A.
We call A the extent and B the intent of the formal concept (A, B). B(G, M, I)
denotes the set of all concepts of the formal context (G, M, I). ♦
Examples of formal concepts of the context in Figure 1.1 B
will be given after Definition 23. The extent A and the in-
×××××
tent B of a formal concept (A, B) are closely connected by A ×××××
×××××
the relation I. Each of the two parts determines the other
and thereby the concept, since B 0 = A and A0 = B, respec-
tively. The next proposition states further simple rules of
this interaction:
Proposition 12 If (G, M, I) is a formal context, A, A1 , A2 ⊆ G are sets of objects and
B, B1 , B2 are sets of attributes, then
3The notation introduced here is convenient but sometimes insufficient. In order to improve
comprehensibility it can be helpful to choose notations like A↑ , B ↓ to distinguish the derivation
operators, or AI , AJ to distinguish different incidence relations.
1.1 Formal context and formal concept 25

1) A1 ⊆ A2 ⇒ A02 ⊆ A10 1’) B1 ⊆ B2 ⇒ B20 ⊆ B10

2) A ⊆ A00 2’) B ⊆ B 00
3) A0 = A000 3’) B 0 = B 000
4) A ⊆ B 0 ⇔ B ⊆ A0 ⇔ A × B ⊆ I.

Proof 1) If m ∈ A02 , then g I m for all g ∈ A2 , i.e., in particular g I m for all g ∈ A1 ,

if A1 ⊆ A2 and thus m ∈ A01 . 2) If g ∈ A, then g I m for all m ∈ A0 , which implies
g ∈ A00 . 3) A0 ⊆ A000 follows immediately from 2’), and A ⊆ A00 together with 1)
yields A000 ⊆ A0 . 4) follows directly from the definition. 

The proposition shows that the two derivation operators form a Galois con-
nection between the powerset lattices P(G) and P(M ) (see Section 0.4). Hence
we obtain (by Proposition 10) two closure systems on G and M , which are dually
order-isomorphic to each other:
For every set A ⊆ G, A0 is an intent of some concept, since (A00 , A0 ) is always a
concept. A00 is the smallest extent containing A. Consequently, a set A ⊆ G is an
extent if and only if A = A00 . The same applies to intents. The union of extents
generally does not result in an extent. On the other hand, the intersection of any
number of extents (respectively intents) is always an extent (intent), as is proved
by the following proposition:
Proposition 13 If T is an index set and, for every t ∈ T , At ⊆ G is a set of objects,
then
[
0 \
At = A0t .
t∈T t∈T

The same holds for sets of attributes.

Proof We have
[
0 [
m∈ At ⇔ g I m for all g ∈ At
t∈T t∈T
⇔ g I m for all g ∈ At for all t ∈ T
⇔ m ∈ A0t for all t ∈ T
\
⇔ m∈ A0t .
t∈T

The set of all extents of (G, M, I) is sometimes denoted by Ext(G, M, I). For
the set of all intents we write Int(G, M, I).
Definition 23 If (A1 , B1 ) and (A2 , B2 ) are concepts of a formal context, then
(A1 , B1 ) is a subconcept of (A2 , B2 ) if A1 ⊆ A2 holds (which is equivalent to
B2 ⊆ B1 ). In this case, (A2 , B2 ) is a superconcept of (A1 , B1 ), and we write
(A1 , B1 ) ≤ (A2 , B2 ). The relation ≤ is called the hierarchical order (or simply
order) of the concepts. The set of all formal concepts of (G, M, I) ordered in this
26 1 Concept lattices of formal contexts

way is denoted by B(G, M, I) and is called the concept lattice of the formal con-
text (G, M, I). ♦
Example The formal context in Figure 1.1 has 19 concepts. The line diagram in
Figure 1.2 represents the concept lattice of this formal context. >

123
45678
ag ac ab ad
1234 5678
34 12
agh 678 356 adf
234 568
abg acd
acgh abc abdf
34 123 678 56
36
abgh acdf
23 68
acghi abcgh acde abcdf
4 3 7 6

abcdef
ghi

Figure 1.2 Concept lattice for the formal context of Figure 1.1. A much better way of la-
beling is shown in Figure 1.4.

Theorem 3 (The Basic Theorem on Concept Lattices)

The concept lattice B(G, M, I) is a complete lattice in which infimum and supremum
are given by:  \
^ [
00 
(At , Bt ) = At , Bt
t∈T t∈T t∈T
_  [
00 \ 
(At , Bt ) = At , Bt .
t∈T t∈T t∈T

A complete lattice L := (L, ≤) is isomorphic to B(G, M, I) if and only if there are

mappings γ̃ : G → L and µ̃ : M → L such that γ̃(G) is supremum-dense in L, µ̃(M )
is infimum-dense in L, and g I m is equivalent to γ̃g ≤ µ̃m for all g ∈ G and all m ∈ M .
In particular, L ∼
= B(L, L, ≤).

Proof of the Basic Theorem. First, we will explain the formula for the infimum.
Since At = Bt0 for each t ∈ T ,
1.1 Formal context and formal concept 27
 \ [
00 
At , Bt
t∈T t∈T

by Proposition 13 can be transformed into

 [
0 [
00 
Bt , Bt ,
t∈T t∈T

i.e., it has the form (X 0 , X 00 ) and is therefore certainly a concept. That this can
only be the infimum, i.e., the largest common subconcept of the concepts (At , Bt ),
follows immediately from the fact that the extent of this concept is exactly the
intersection of the extents of (At , Bt ). The formula for the supremum is sub-
stantiated correspondingly. Thus, we have proven that B(G, M, I) is a complete
lattice.
Now we prove, first for the special case (L, ≤) = B(G, M, I), the existence of
mappings γ̃ and µ̃ with the required properties. We set

γ̃g := ({g}00 , {g}0 ) for g ∈ G

and µ̃m := ({m}0 , {m}00 ) for m ∈ M.

As claimed, we have γ̃g ≤ µ̃m ⇔ {g}00 ⊆ {m}0 ⇔ {g}0 ⊇ {m} ⇔ m ∈ {g}0 ⇔
g I m. Furthermore, on account of the formulas proved above,
_ ^
({g}00 , {g}0 ) = (A, B) = ({m}0 , {m}00 )
g∈A m∈B

holds for every formal concept (A, B), i.e., γ̃(G) is supremum-dense and µ̃(M )
is infimum-dense in B(G, M, I). More generally, if ϕ : B(G, M, I) → (L, ≤) is an
isomorphism, we define γ̃ and µ̃ by

γ̃g := ϕ({g}00 , {g}0 ) for g ∈ G

and µ̃m := ϕ({m}0 , {m}00 ) for m ∈ M.

The properties claimed for these mappings are proved in a similar fashion.
Conversely, if (L, ≤) is a complete lattice and

γ̃ : G → L, µ̃ : M → L

are mappings with the above properties, then we define ϕ : B(G, M, I) → L by

_
ϕ(A, B) := {γ̃(g) | g ∈ A}.

Evidently, ϕ is order-preserving. In order to prove that ϕ is an isomorphism, we

need to show that ϕ−1 exists and is also order-preserving. Therefore, we define

ψx := ({g ∈ G | γ̃g ≤ x}, {m ∈ M | x ≤ µ̃m}),

28 1 Concept lattices of formal contexts

for x ∈ L and demonstrate that ψx is a formal concept of (G, M, I):

h ∈ {g ∈ G | γ̃g ≤ x} ⇔ γ̃h ≤ x
⇔ γ̃h ≤ µ̃n for all n ∈ {m ∈ M | x ≤ µ̃m}
⇔ h I n for all n ∈ {m ∈ M | x ≤ µ̃m}
⇔ h ∈ {m ∈ M | x ≤ µ̃m}0 .

The second condition follows correspondingly. We have defined a map ψ : L →

B(G, M, I), and we can read off directly from the definition that ψ is order-
preserving. Now we prove that ϕ = ψ −1 . We have
_
ϕψx = {γ̃g | g ∈ G, γ̃g ≤ x} = x,

since γ̃(G) is supremum-dense in (L, ≤). On the other hand, ϕ(A, B) =

V
{µ̃m |
m ∈ B}, since µ̃(M ) is infimum-dense in (L, ≤), and consequently
^
ψϕ(A, B) = ψ {µ̃m | m ∈ B}
^
= ({g ∈ G | γ̃g ≤ {µ̃m | m ∈ B}}, {. . .}0 )
= ({g ∈ G | γ̃g ≤ µ̃m for all m ∈ B}, {. . .}0 )
= ({g ∈ G | g I m for all m ∈ B}, {. . .}0 )
= (B 0 , B 00 ) = (A, B).

If we choose for a complete lattice (L, ≤) specifically G := L, M := L, I :=≤ and

γ̃ as well as µ̃ to be the identity of L, we obtain (L, ≤) ∼
= B(G, M, I). 
The Duality Principle for Concept Lattices. Let (G, M, I) be a formal context.
Then (M, G, I −1 ) is also a formal context, in fact,

B(M, G, I −1 ) ∼
= B(G, M, I)d ,

and
(B, A) 7→ (A, B)
is an isomorphism.
In other words: if we exchange the roles of objects and attributes, we obtain
the dual concept lattice. Thus, the Duality Principle extends to concept lattices.

The mappings γ̃ and µ̃ which appear in the Basic Theorem indicate how the
context can be identified in the concept lattice. This is further deepened by the
following definition.
Definition 24 For an object g ∈ G we write g 0 instead of {g}0 for the object intent
{m ∈ M | g I m} of the object g. Correspondingly, m0 := {g ∈ G | g I m} is
the attribute extent of the attribute m. Retaining the symbols used in the Basic
1.2 Formal context and concept lattice 29

Theorem, we write γg for the object concept (g 00 , g 0 ) and µm for the attribute
concept (m0 , m00 ). ♦
The line diagram in Figure 1.2 shows the intent and the extent of every formal
concept. The labelling can be simplified considerably by entering each object and
each attribute only once, namely to the corresponding object or attribute concept
(see Figure 1.3). One can then still read the formal context as well as all concept
extents and concept intents from the line diagram: If you look for the concept
extent of a formal concept represented by a small circle, this consists of the ob-
jects located at this circle or at the circles which can be reached by descending
line paths from this circle. Correspondingly, the intent can be found by follow-
ing all line paths going upwards from the circle and noting down the attributes
assigned to these circles.

g c b d

h f

1 5

2 8
i e
4 3 7 6

Figure 1.3 Line diagram with reduced labelling.

The sparing, reduced labelling enables us to enter the full names of the objects
and attributes of the formal context in Figure 1.1 into the diagram. This improves
the readability of the diagram, as can be seen in Figure 1.4.

1.2 Formal context and concept lattice

A formal context (G, M, I) can be easily reconstructed from the system of all
its formal concepts. G and M appear as the extent and the intent of the triv-
30 1 Concept lattices of formal contexts

needs water to live

can move around lives on lives in needs chlorophyll

land water

has limbs one seed leaf

leech waterweed

suckles its bream maize

offspring two seed leaves
dog frog bean reed

Figure 1.4 Concept lattice for the educational film “Living beings and water”.

ial boundary concepts: The set of all objects is the extent of the largest concept,
(∅0 , ∅00 ) = (G, G0 ). Dually, M is the intent of the least concept, (∅00 , ∅0 ) = (M 0 , M ).
The incidence relation I is given by
[
I = {A × B | (A, B) ∈ B(G, M, I)}.

It is even easier to read off the context from the concept lattice, as the Basic The-
orem shows. On the other hand, concept lattices of different contexts can well
be isomorphic to each other. The context manipulations which do not alter the
structure of the concept lattice include the merging of objects with the same in-
tents and attributes with the same extents, respectively:
Definition 25 A formal context (G, M, I) is called clarified, if for any objects
g, h ∈ G it always follows from g 0 = h0 that g = h and, correspondingly, m0 = n0
implies m = n for all m, n ∈ M . ♦
Example Figure 1.5 shows a context representing the service offers by an office
supply shop. Below is the clarified context. >
Also not affecting the structure of the concept lattice are attributes that can
be written as combinations of other attributes. More precisely: If m ∈ M is an
attribute and X ⊆ M is a set of attributes with m 6∈ X but m0 = X 0 , then the
attribute concept µm is the infimum of the attribute concepts µx, x ∈ X, i.e., the
set µ(M \ {m}) is also infimum-dense in B(G, M, I), and according to the Basic
Theorem
B(G, M, I) ∼= B(G, M \ {m}, I ∩ (G × (M \ {m}))).
1.2 Formal context and concept lattice 31

Furniture Computers Copy Type- Specialized

machines writers machines
Consulting × × × × ×
Planning × ×
Assembly and × × × × ×
installation
Instruction × × × ×
Training, ×
workshops
Original spare × × × × ×
parts and
accessories
Repairs × × × × ×
Service contracts × × ×

Furniture Computers Copy machines Specialized

and typewriters machines
Consulting, × × × ×
assembly and
installation,
original spare
parts and
accessories,
repairs
Planning × ×
Instruction × × ×
Training, ×
workshops
Service contracts × ×

Figure 1.5 Context and clarified context.

This removal of reducible attributes, i.e., attributesWwith -reducible attribute

V
concepts, and of reducible objects, i.e., objects with -reducible object concepts,
is what we call reducing the formal context. Full rows and full columns are
always reducible; thereby we mean objects g with g 0 = M and attributes m with
m0 = G, respectively.
Definition 26 W A clarified context (G, M, I) is called object reduced if every ob-
ject concept is -irreducible, and attribute reduced if every attribute concept is
-irreducible. A formal context that is both object reduced and attribute reduced
V
is reduced. ♦
It is easy to find infinite formal contexts in which all attributes and all objects
are reducible (see page 7). As a rule this means that we cannot simultaneously
omit all reducible objects and attributes. This is no problem, however, in the case
32 1 Concept lattices of formal contexts

Computers
e
Training,
Workshops @
@ Copy machines,
@ typewriters
@@e
Service contracts
Furniture e
Planning J
J eSpecialized machines
J Instruction
J
J
e
JJ
Consulting,
Assembly and installation,
Original spare parts and accessories,
Repairs

Figure 1.6 The concept lattice for the context of Figure 1.5.

of finite concept lattices, since

V in a finite lattice each element is the join of -
W
irreducible and the meet of -irreducible elements (see Proposition 2, p. 7).

Proposition 14 For every finite4 lattice L there is -up to isomorphism5 - a unique

reduced formal context K(L) with L ∼
= B(K(L)), namely

K(L) := (J(L), M (L), ≤).



This is called the standard context of the lattice L. For practical work with for-
mal contexts, the proposition has the following consequences: Every finite for-
mal context can be brought into a reduced form without changing the structure
of the concept lattice, and the latter is unique. We first clarify the context, i.e., we
merge objects with the same intents and attributes with the same extents. Then
we delete all objects, the intent of which can be represented as the intersection
of other object intents, and correspondingly all attributes, the extent of which is
the intersection of other attribute extents.
It is easy to reconstruct the concepts of the original formal context from those
of the reduced context, if one has kept a record of the reduction process. If
4See also Proposition 16.c).
5Two formal contexts (G1 , M1 , I1 ) and (G2 , M2 , I2 ) are called isomorphic if there are bijective
mappings α : G1 → G2 , β : M1 → M2 with g I1 m ⇔ (αg) I2 (βm) for all g ∈ G1 , m ∈ M1 , see
Definition 97 on p. 283.
1.2 Formal context and concept lattice 33

c
c c × × c
c × c × c @c × c
c c × @c c

c c c
× × c @c × × c × × c @c
× @c × × c @c × × c @c
× c @c × @c

c c
c  @c
× × c c@ c × × × c
c c @c
@
× × c@ c@ c × ×
\ c
× × @c × @ @ c × \c
c c
c @@c c @
× × × × @c
× × c@
@c c × × c@
@c
× × @c
@ × @c
@

Figure 1.7 Reduced formal contexts with up to three objects, and their concept lattices.
The context (∅, ∅, ∅) was omitted, as was its (one-element) concept lattice.

we denote, for a finite clarified context (G, M, I), the set of its irreducible ob-
jects by Girr and the set of irreducible attributes by Mirr , the reduced context is
(Girr , Mirr , I ∩ (Girr × Mirr )), and each concept (A, B) of (G, M, I) corresponds to
the formal concept (A ∩ Girr , B ∩ Mirr ) of (Girr , Mirr , I ∩ (Girr × Mirr )). For every
object g ∈ G and every extent A of (G, M, I)

g ∈ A ⇔ g 00 ∩ Girr ⊆ A ∩ Girr ,

holds dually for the attributes. If we note down the set g 00 ∩ Girr for every re-
ducible object g and the set m00 ∩ Mirr for every reducible attribute m, it is easy
to obtain the concepts of (G, M, I) from those of (Girr , Mirr , I ∩ (Girr × Mirr )).
There is another way to carry out the reduction of a clarified context, namely
with the help of arrow relations, which we will define next and which will be-
come important for other reasons later on. These relations can conveniently be
entered into the cross table, since they only apply to “empty cells”, i.e., to object-
attribute pairs which do not stand in the relation I.
34 1 Concept lattices of formal contexts

Non-Aligned

Non-Aligned
Group 77

Group 77
OACPS

OACPS
OPEC

OPEC
SIDS

SIDS
LDC

LDC
Afghanistan × × × East Timor × × × ×
Algeria × × × Ecuador × ×
Angola × × × × × Egypt × ×
Antigua and Barbuda × × × × El Salvador ×
Argentina × Equatorial Guinea × × × ×
Armenia Eritrea × × × ×
Azerbaijan × × Eswatini × × ×
Bahamas × × × × Ethiopia × × × ×
Bahrain × × Fiji × × × ×
Bangladesh × × × Gabon × × × ×
Barbados × × × × Gambia × × × ×
Belize × × × × Ghana × × ×
Benin × × × × Grenada × × × ×
Bhutan × × × Guam
Bolivia × × Guatemala × ×
Botswana × × × Guinea × × × ×
Brazil × Guinea-Bissau × × × × ×
Brunei × × Guyana × × × ×
Burkina Faso × × × × Haiti × × × × ×
Burundi × × × × Honduras × ×
Cambodia × × × India × ×
Cameroon × × × Indonesia × ×
Cape Verde × × × × Iran × × ×
Central African Republic × × × × Iraq × × ×
Chad × × × × Ivory Coast × × ×
Chile × × Jamaica × × × ×
China × Jordan × ×
Colombia × × Kazakhstan
Comoros × × × × × Kenya × × ×
DR Congo × × × × Kiribati × × × ×
Rep. Congo × × × × Kuwait × × ×
Cook Islands × × Kyrgyzstan
Costa Rica × Laos × × ×
Cuba × × × × Lebanon × ×
Djibouti × × × × Lesotho × × × ×
Dominica × × × × Liberia × × × ×
Dominican Republic × × × × Libya × × ×

Figure 1.8 Membership of developing countries in supranational groups. (Part 1). LDC =
Least Developed Countries, SIDS = Small Island Developing Countries, OPEC = Petrol Ex-
porting Countries, OACPS = African, Caribbean and Pacific States.
1.2 Formal context and concept lattice 35

Non-Aligned

Non-Aligned
Group 77

Group 77
OACPS

OACPS
OPEC

OPEC
SIDS

SIDS
LDC

LDC
Madagascar × × × × Samoa × × ×
Malawi × × × × São Tomé + Príncipe × × × × ×
Malaysia × × Saudi Arabia × × ×
Maldives × × × × Senegal × × × ×
Mali × × × × Seychelles × × × ×
Marshall Islands × × × Sierra Leone × × × ×
Mauritania × × × × Singapore × × ×
Mauritius × × × × Solomon Islands × × × ×
Micronesia × × × Somalia × × × ×
Moldova South Africa × × ×
Mongolia × × South Sudan × × ×
Morocco × × Sri Lanka × ×
Mozambique × × × × Sudan × × × ×
Myanmar × × × Suriname × × × ×
Namibia × × × Syria × ×
Nauru × × Tajikistan ×
Nepal × × × Tanzania × × × ×
Nicaragua × × Thailand × ×
Niger × × × × Togo × × × ×
Nigeria × × × × Tonga × × ×
Niue × × Trinidad + Tobago × × × ×
North Korea × × Tunisia × ×
North Macedonia Turkmenistan × ×
Oman × × Tuvalu × × ×
Pakistan × × Uganda × × × ×
Palau × × Un. Arab Emirates × × ×
Palestine × × Uruguay ×
Panama × × Uzbekistan ×
Papua New Guinea × × × × Vanuatu × × × ×
Paraguay × Venezuela × × ×
Peru × × Vietnam × ×
Philippines × × Yemen × × ×
Qatar × × Zambia × × × ×
Rwanda × × × × Zimbabwe × × ×
Saint Kitts and Nevis × × × ×
Saint Lucia × × × ×
St. Vincent + Grenadines × × × ×

Figure 1.8 Membership of developing countries in supranational groups. (Part 2). For data
sources, see Section 1.5.
36 1 Concept lattices of formal contexts

Argentina, Brazil, China, Costa Rica, El Armenia, Guam, Kazakhstan,

Salvador, Paraguay, Tajikistan, Uruguay Kyrgyzstan, Moldova, North
Macedonia
Azerbaijan, Bahrain, Boli-
via, Brunei, Chile, Colombia,
Ecuador, Egypt, Guatemala, Group of 77 LDC SIDS
Honduras, India, Indonesia,
Jordan, Lebanon, Malaysia,
Mongolia, Morocco, Nica-
ragua, North Korea, Oman, Non-
Pakistan, Palestine, Panama, aligned OACPS
Peru, Philippines, Qatar,
Sri Lanka, Syria, Thailand, Uzbekistan
Tunisia, Turkmenistan,
Vietnam
Algeria, Iran, Iraq,
Kuwait, Libya, Saudi Cook
Islands,
Arabia, United Arab Nauru,
Emirates, Venezuela Niue,
Palau

OPEC

Singapore South Tuvalu

Sudan
Afghanistan, Bangla- Marshall Is-
desh, Bhutan, Cam- lands, Mi-
bodia, Laos, Myan- croesia, Sa-
mar, Nepal, Yemen moa, Tonga

Botswana, Cameroon, Eswatini, Kiribati, Solomon

Ghana, Ivory Coast, Kenya, Na- Islands
mibia, Zimbabwe, South Africa
East Timor
Equatorial Guinea, Gabon,
Nigeria, Rep. Congo
Antigua+Barbuda, Bahamas,
Benin, Burkina Faso, Burundi, Barbados, Belize, Cape Verde,
Central African Republic, Chad, Cuba, Dominica, Dominican
Djibouti, DR Congo, Eritrea, Republic, Fiji, Grenada,
Angola
Ethiopia, Gambia, Guinea, Lesotho, Guyana, Jamaica, Maldives,
Liberia, Madagascar, Malawi, Mali, Mauritius, Papua New
Comoros,
Mauritania, Mozambique, Niger, Guinea- Guinea, Saint Kitts and Nevis,
Bissau, Saint Lucia, Saint Vincent +
Zambia, Rwanda, Senegal, Sierra Grenadines, Seychelles,
Haiti, São
Leone, Somalia, Sudan, Tanzania, Tomé + Suriname, Trinidad + Tobago,
Togo, Uganda Príncipe Vanuatu

Figure 1.9 Concept lattice of the formal context of developing countries, shown in Figure 1.8.
1.2 Formal context and concept lattice 37

Definition 27 If (G, M, I) is a formal context, g ∈ G an object, and m ∈ M an

attribute, we write

/ I and

(g, m) ∈
g . m :⇔
if g 0 ⊆ h0 and g 0 6= h0 , then h I m,
/ I and

(g, m) ∈
g % m :⇔
if m0 ⊆ n0 and m0 6= n0 , then g I n,
g%. m : ⇔ g . m and g % m.

♦
Thus, g . m if and only if g 0 is maximal among all object intents which do
not contain m. In other words: g . m holds if and only if g does not have the
attribute m, but m is contained in the intent of every proper subconcept of γg. If
we now let _
(γg)∗ := {x ∈ B(G, M, I) | x < γg}
as in Definition 11, then (γg)∗ is a subconcept of γg and γg is -irreducible, if
W
and only if γg 6= (γg)∗ . This, on the other hand, is equivalent to the fact that there
is an attribute m in the intent of (γg)∗ which is not contained in the intent of γg,
i.e., to g . m for some m ∈ M . Therefore, we obtain

g . m ⇔ γg ∧ µm = (γg)∗ 6= γg
g % m ⇔ γg ∨ µm = (µm)∗ 6= µm.

Example Figure 1.10 shows the context from Figure 1.5 with the arrow relations;
beside it the reduced context.

× × × ×
× × .
% . × .
%.
.
% × × × % × ×
.
% × % % × %
.
% × × .
%

Figure 1.10 Context with arrow relations, and the reduced context.

The significance of the arrow relations for the reduction of a formal context is
shown by the next proposition:

Proposition 15 The following statements hold for every formal context:

a) γg is V -irreducible ⇔ There is an m ∈ M with g . m.
W
b) µm is -irreducible ⇔ There is a g ∈ G with g % m.
38 1 Concept lattices of formal contexts

Furthermore, the following statements hold for every finite6 formal context:
c) γg is V -irreducible ⇔ There is an m ∈ M with g %
W
. m.
d) µm is -irreducible ⇔ There is a g ∈ G with g % . m.

Proof This follows immediately from the above-mentioned observations togeth-

er with Proposition 2. If we choose m0 maximal with respect to g . m (in a finite
context this is certainly possible), then g % m, i.e., g %
. m. 
In order to reduce a finite clarified context, we therefore first enter the arrow
relations in the cross table and then delete all rows and columns not contain-
ing a double arrow. The condition of finiteness in Propositions 14 and 15 can be
weakened:
Definition 28 A formal context (G, M, I) is called doubly founded, if for every
object g ∈ G and every attribute m ∈ M with (g, m) ∈/ I there is an object h ∈ G
and an attribute n ∈ M with

g % n and m0 ⊆ n0 as well as h . m and g 0 ⊆ h0 .

A complete lattice (L, ≤) is called doubly founded, if for any two elements x < y
of L there are elements s, t ∈ L with:
s is minimal with respect to s ≤ y, s 6≤ x, as well as
t is maximal with respect to t ≥ x, t 6≥ y.
♦

By means of Proposition 15 we realize easily that the attribute n and the object
h that appear in Definition 28 must be irreducible. The same applies to the lattice
elements sVand t in the second part of the definition: s must be -irreducible and
W
t must be -irreducible. This means that the property “doubly founded” implies
the existence of “many” irreducible elements.

Proposition 16
a) Every finite formal context is doubly founded.
b) A formal context which does neither contain infinite chains g1 , g2 , . . . of objects
with g10 ⊂ g20 ⊂ . . . nor infinite chains m1 , m2 , . . . of attributes with m01 ⊂ m20 ⊂
. . . is doubly founded.
c) Each formal concept of a doublyV founded context is the supremum of -irreducible
W
concepts and the infimum of - irreducible concepts. Hence Proposition 14 also ap-
plies to concept lattices of doubly founded contexts.
d) If (G, M, I) is doubly founded and g ∈ G, m ∈ M , the following hold true: if
g . m, then there is an attribute n with g %. n, and if g % m, then there is an object
h with h % . m. Hence parts c) and d) of Proposition 15 also apply to doubly founded
contexts.

6 cf. also Proposition 16.d).

1.2 Formal context and concept lattice 39

Proof b) If (g, mi ) ∈ / I holds and g % mi does not hold, according to the

definition of the arrow relations there must be an attribute mi+1 6= mi with
(g, mi+1 ) ∈/ I and mi0 ⊂ mi+1
0
. Thus, if (g, m) ∈ / I, then starting from m1 := m we
obtain by this argument a chain of attributes with increasing attribute extents. By
assumption this chain must be finite and therefore end with an attribute mj =: n
with g % n. The dual argumentation shows that there also must be some h with
h . m and g 0 ⊆ h0 .
a) follows immediately from b).
c) Let (A, B) be a concept and (C, D) := {γx | x ∈ A, γx −irreducible}. We
W W
assume that (C, D) < (A, B). Then there is m ∈ D, g ∈ A with (g, m) ∈ / I. Hence
there is an h ∈ G with h . m and g 0
⊆ h 0
, i.e., h ∈ A. Because of h . m, γh is
-irreducible, i.e., h ∈ C and thus m ∈ h0 , which is a contradiction.
W
d) From g . m it follows that (g, m) ∈ / I, hence there is an n ∈ M with m0 ⊆ n0
and g % n. All together we obtain g % . n. 

Proposition 17 If B(G, M, I) is doubly founded, so is (G, M, I). If a complete lattice

(L, ≤) is not doubly founded, neither is the context (L, L, ≤).

Proof Let B(G, M, I) be doubly founded and g ∈ G, m ∈ M with (g, m) ∈ / I,

i.e., γg 6≤ µm. For x := µm and y := µm∨γg there is a concept t which is maximal
with
V respect to t ≥ x, t 6≥ y. On account of this property of maximality t must be
-irreducible. Hence there is an attribute n with t = µn. Thus we obtain g % n
and m0 ⊆ n0 . The second condition is obtained dually.
If (L, L, ≤) is doubly founded, so must be (L, ≤): If x < y in (L, ≤), then
certainly y 6≤ x, i.e., (x, y) ∈
/ I in (L, L, ≤). From the definition of the doubly
foundedness of (L, L, ≤) now follows the existence of an element s ∈ L with
s . x and y 0 ⊆ s0 , hence s is minimal with respect to s 6≤ x, s ≤ y. The second
condition can again be shown by means of the dual argument. 
Thus, a complete lattice (L, ≤) is doubly founded if and only if every context
(G, M, I) with (L, ≤) ∼ = B(G, M, I) is doubly founded. One should note, how-
ever, that the concept lattice of a doubly founded context does not necessarily
have to be doubly founded, as shown by the example (N, N, ≤).
It frequently occurs that a statement can be proved for all finite lattices but
not for all complete lattices. We will (when possible) replace the condition of
finiteness by “doubly foundedness”. This is not in every case the strongest pos-
sible relaxation. The restriction to “doubly founded” is adopted for reasons of
uniformity. Mathematical lattice theory uses numerous other conditions, some
of which are represented by Figure 1.11 in their hierarchical order. We will only
give a short explanation of the terminology used in this context: A complete lat-
tice (L, ≤) is supremum-founded, if for any two elements x < y from L there is
an element s ∈ L which is minimal with respect to s ≤ y, s 6≤ x. The dual prop-
erty is “infimum-founded”. A concept lattice B(G, M, I) is algebraic (dually:
co-algebraic), if for every subset A ⊆ G from

F ⊆ A ⇒ F 00 ⊆ A for every finite subset F

40 1 Concept lattices of formal contexts

e
@@
J(L) supremum-dense
@ eM (L) infimum-dense
@
e
coalgebraic
@ algebraic
e @ A@
@ e
supremum-founded @ @ einfimum-founded
@ @
e @e
@
@ @
@ @
e
@ @ e
A@ e e
@ @ @ A@
@
@ @ @
atomistic @ @ ecoatomistic
e @e e @@ e
@
@ A@
@
@ @
@ @ @ @
e
@ @ e
A@ e @ @ e
@ e Ae
@ @ @ A@
@ @
@ @ df @ @
@@ @e A@
@ e
@ @ e
@ @A@
@e
@ @ e
@
B(N, PE (N), ∈)@ @ @ @ @ B(N, PE (N), ∈)d
@ e
A@ e
@ @ @ e
@
@ A@
@ e @ @ e
@
@ Ae
@ @ @ @
A@ e
@ @ e e
@ @ e A e

@ @ @A@
@ @ @
e@
@ @ @ @
B(C[0,1] ) @@ @ e
@ cf @
@ B(C[0,1] )d
A@
@e @ e Ae
@ @ f e@ @@ A
A@ e
@ @ e@ @@ e

@ @ @
@ @ @
e@@ @ e
@ A e@@ e
e@ A
@ @ @e
B(I3 )
@ @ e
J B( I 3 ) d
@
@
B(S) @@ e
J
B( S ) d
@ eJ
B(N, N, 6=) J B(N, N, =)

The abbreviations stand for: df := doubly founded, cf := chain-finite, f := finite. PE (N)

denotes the set of all finite subsets of the natural numbers. Furthermore, let S be the
formal context arising as a subposition (see Section 1.4) of the contexts (N, PE (N), ∈)
and (PE (N), PE (N), =). C[0,1] is the convex-ordinal scale for the real unit interval
([0, 1], ≤), as defined in Section 1.4.

Figure 1.11 Foundedness compared to related conditions.

1.3 Context constructions 41

it follows that A = A00 . An ordered set is chain finite, if every chain contained
by it is finite.
The lattice presented in Figure 1.11 is the result of an attribute exploration in
accordance with Section 2.3, i.e., the represented implications between the prop-
erties are really provable. We omit the proofs.

1.3 Context constructions

Formal contexts are very simple mathematical objects. It is not surprising that
they are easy to construct and that there is an almost inexhaustible variety of
constructions. Some of them are collected in this section. We start with simple
and frequently used composition methods, with sums, products and unions of
formal contexts, etc. We usually formulate only the composition of two formal
contexts, but mostly the definitions can be generalized to arbitrary many. The
additional statements on the concept lattices of the resulting contexts carry over.
This part is followed by context constructions from mathematical structures, es-
pecially ordered sets. Our list is diverse, but far from complete. Further construc-
tions are described in the later chapters of this book, but of course also in the
literature.
What is presented here are, for the time being, purely mathematical construc-
tions. Their interpretation and their meaning for conceptual knowledge process-
ing is not yet addressed here. The subsequent section gives first hints.
Some of the context compositions require the parts to be disjoint. We will use
the abbreviations Ġj := {j} × Gj , Ṁj := {j} × Mj and I˙j := {((j, g), (j, m)) |
(g, m) ∈ Ij } for j ∈ {1, 2} in the following definition.
Definition 29 Let K := (G, M, I) and Kt := (Gt , Mt , It ) for i ∈ {1, 2} be formal
contexts. It is:
Kc := (G, M, (G × M ) \ I) the complementary context to K,
Kd := (M, G, I −1 ) the dual context to K,
and, if G = G1 = G2 ,
K1 K2 := (G, Ṁ1 ∪ Ṁ2 , I˙1 ∪ I˙2 ) the apposition of K1 and K2 ,
as well as dually, if M = M1 = M2 ,
K1
:= (Ġ1 ∪ Ġ2 , M, I˙1 ∪ I˙2 ) the subposition of K1 and K2 .
K2
Kcd is called the contrary context to K,
K | Kc is the dichotomization of K, and
K1 ∪˙ K2 := (Ġ1 ∪ Ġ2 , Ṁ1 ∪ Ṁ2 , I˙1 ∪ I˙2 ) the disjoint union of K1 and K2 .
♦
42 1 Concept lattices of formal contexts

Remark: By using Ġi for {i} × Gi and Ṁi , respectively, we intend to make sure
that the sets are disjoint. However, strictly speaking, apposition and subposition
under this definition become non-associative. We will overlook this fact and tac-
itly identify the contexts

( K1 K2 ) K3 and K1 ( K2 K3 ) .

The same applies to the subposition, even to hybrid forms of the two operations.
We do not distinguish between

K1 | K2 K1 K2
and .
K3 | K4 K3 K4

The two abbreviations

× := (G, M, G × M )
∅ := (G, M, ∅)
are occasionally used without further describing the sets G and M , if they are
evident from the context. For example

K1 ×
∅ K2

denotes the context (Ġ1 ∪ Ġ2 , Ṁ1 ∪ Ṁ2 , I˙1 ∪ I˙2 ∪ (Ġ1 × Ṁ2 )), the concept lattice of
which is isomorphic to the vertical sum of the concept lattices B(K1 ) and B(K2 )
(provided that K1 does not contain a full column and K2 does not contain a full
row, cf. Section 4.3).
Each extent of K1 ∪˙ K2 , apart from the extent Ġ1 ∪ Ġ2 , is entirely contained
in one of the sets Ġi . The corresponding applies to the intents. Therefore, the
concept lattice L := B(K1 ∪˙ K2 ) is a horizontal sum, i.e., it is the union L =
(L1 , ≤) ∪ (L2 , ≤) of two sublattices which only overlap in the smallest and the
largest element: L1 ∩ L2 = {0L , 1L }. Provided that there are no full rows or
columns in K1 and K2 , we have (Li , ≤) ∼ = B(Ki ) or, more generally, (Li , ≤) =
˙
B(Ġ1 ∪ Ġ2 , Ṁ1 ∪ Ṁ2 , Ii ).

Definition 30 The direct sum of two formal contexts Ki :=

(Gi , Mi , Ii ) is defined by7
K1 ×
K1 + K2 := × K2

(Ġ1 ∪ Ġ2 , Ṁ1 ∪ Ṁ2 , I˙1 ∪ I˙2 ∪ (Ġ1 × Ṁ2 ) ∪ (Ġ2 × Ṁ1 )). Context sum
♦

7 For the “dot” notation see the remarks before Definition 29. A more general definition is given
in Section 5.1.
1.3 Context constructions 43

The concept lattice of a context sum is isomorphic to the product of its concept
lattices. In the case of two contexts we therefore obtain

B(K1 + K2 ) ∼
= B(K1 ) × B(K2 ),

since (A, B) is a concept of K1 + K2 if and only if (A ∩ Ġi , B ∩ Ṁi ) is a concept

of K̇i := (Ġi , Ṁi , I˙i ), for i ∈ {1, 2}. This means that the isomorphism is given by
(A, B) 7→ ((A ∩ Ġ1 , B ∩ Ṁ1 ), (A ∩ Ġ2 , B ∩ Ṁ2 )).
Definition 31 The semiproduct will be particularly important for conceptual
scaling, see Sections 2.4 and 7.3. It is defined by
`
K1 a
K2 := (G1 × G2 , Ṁ1 ∪ Ṁ2 , ∇)

with
(g1 , g2 )∇(j, m) : ⇔ gj Ij m for j ∈ {1, 2}. ♦

× ×
× `
× ∼ × ×
a
= →
× × ××
× ×

× ×
× ×
×
× `
× ××
← × ∼
=
a

× ××
××
× ×
×××

Figure 1.12 Two semiproducts and their concept lattices

The extents of the semiproduct are precisely the sets of the form A1 × A2 , each
set Aj `being an extent of Kj . This also yields the structure of the concept lattice
B(K1 a K2 ): Essentially, the concept lattice is the product of the concept lattices
of the factor contexts, though there is a modification regarding the zero elements.
Precisely, the instruction for the construction reads as follows: Provided the the
extent of the zero element of B(Kj ) is empty, remove that zero element from
B(Kj ) (for j ∈ {1, 2}). Then take the direct product of these ordered sets and,
if an element was removed earlier, add a new zero element to make a complete
lattice. This lattice is then isomorphic to the concept lattice of the semiproduct.
Figure 1.12 shows a small example.
44 1 Concept lattices of formal contexts

Definition 32 The direct product is given by

K1 × K2 := (G1 × G2 , M1 × M2 , ∇)

with
(g1 , g2 ) ∇ (m1 , m2 ) : ⇔ g1 I1 m1 or g2 I2 m2 . ♦
The concept lattice of the direct product is called the tensor product of the con-
cept lattices of the factor contexts. The tensor product will be discussed in more
detail later (Sections 4.4, 5.4). The cross table of the direct product is obtained by
replacing each empty cell in the table of K1 by a copy of K2 and each cross by a
rectangle full of crosses of the size of K2 . For an example see Figures 4.19 (page
198) and 4.20. An easy conclusion is that

(K1 × K2 )d = Kd1 × Kd2 .

One instance of the direct product deserves special mention, namely the case
where K2 is the standard context of the three-element chain, i.e.,

×
K2 = .

The direct product of an arbitrary K with this context yields

K ×
,
K K

and it can be shown that the concept lattice of this is isomorphic to the order
relation of B(K) (which, understood as an ordered set of pairs with component-
wise order, is itself a complete lattice). Compare the context recursion for the
free completely distributive lattices below on page 49.
Definition 33 Forming the Boolean matrix product of two contexts assumes
that the attribute set of the first is equal to the object set of the second. Let
(G, F, IGF ) and (F, M, IF M ) be such formal contexts. The Boolean matrix prod-
uct then is
(G, F, IGF ) · (F, M, IF M ) := (G, M, I)
with
(g, m) ∈ I : ⇔ ∃f ∈F (g, f ) ∈ IGF and (f, m) ∈ IF M .
♦
This product is seen less as a construction method. Rather, the task is to factorize
a given formal context K, i.e., to find an (ideally small) set F and relations IGF
and IF M to form a factorization of K. We will go into this in more detail in the
last part of Section 6.4.
Another context construction, the so-called substitution sum, where a context
is inserted into another context, will be described in Section 4.3. The sum and
1.3 Context constructions 45

the product of reduced contexts are reduced (cf. Corollary 3, p. 200). Reducible
objects or attributes with empty intents or extents may occur in the case of the dis-
joint union. Semiproducts of reduced contexts are reduced if the factors (allow-
ing for one exception at most) are atomistic, i.e., if they satisfy g 0 ⊆ h0 ⇒ g = h.
It is easy to state numerous simple arithmetical rules for context constructions,
which are useful for some proofs. In particular, the direct product is (up to iso-
morphism) commutative and associative; it is distributive over the direct sum,
the apposition and the subposition. We note down one of these results for later:

Proposition 18 We have

(K1 + K2 ) × K3 = (K1 × K3 ) + (K2 × K3 ).

Proof We may assume that the three contexts Ki =: (Gi , Mi , Ii ), i ∈ {1, 2, 3},
have disjoint object sets and disjoint attribute sets. By

(G1 ∪ G2 ) × G3 = (G1 × G3 ) ∪ (G2 × G3 )

and
(M1 ∪ M2 ) × M3 = (M1 × M3 ) ∪ (M2 × M3 ),
the two contexts of the proposition have the same objects and attributes. For the
incidence we find the same on both sides as well, namely

 g ∈ G1 and m ∈ M2 or


 g ∈ G2 and m ∈ M1 or


(g, h) I (m, n) ⇔ h I3 n or
 g ∈ G1 , m ∈ M1 , and g I1 m or


g ∈ G2 , m ∈ M2 , and g I2 m.



A list of formal context families

We now present some interesting context families, but must explain in advance
why we often use the synonym “scale” instead of “formal context”. The next sec-
tion explains how Formal Concept Analysis deals with data that are not given in
the form of formal contexts. The method used for this is called Conceptual Scaling.
We call formal contexts that are particularly suitable for this purpose conceptual
scales, see Definition 35 (p. 60). The simplest variants, which we call “elemen-
tary standard scales”, will be discussed in more detail in Definition 37. Besides,
the formal contexts in this list serve as a reservoir of examples for mathematical
reasoning.
First, contexts are described that can be obtained naturally from arbitrary or-
dered sets ((1)–(7)). The examples that follow come from graphs, equivalence
46 1 Concept lattices of formal contexts

relations, and other simple standard mathematical notions. Finally, some con-
structions from advanced mathematics are briefly mentioned.
(1) For every set S the contranominal scale

NcS := (S, S, 6=)

is reduced. The concepts of this context are precisely the pairs (A, S \ A)
for A ⊆ S. The concept lattice is isomorphic to the powerset lattice of S,
and thus has 2|S| elements. For S = {1, 2, . . . , n} it is listed as elementary
standard scales (Def. 37) under the name Nnc .
(2) From an arbitrary ordered set (P, ≤) we obtain the general ordinal scale

O(P,≤) := (P, P, ≤).

Its concepts are precisely the pairs (X, Y ) with X, Y ⊆ P where X is the
set of all lower bounds of Y and Y is the set of all upper bounds of X.
This concept lattice is called the Dedekind-MacNeille completion of the
ordered set (P, ≤). It is the smallest complete lattice in which (P, ≤) can
be order-embedded, in the sense of the following theorem:

e
ce ce
@ @
e e
Q e ce e c
e
Q  @ @ @ @
e e
Q e
Q  @ @
@ @
@e
@
c
@@e
c
@@e

Figure 1.13 Example of an ordered set (P, ≤) and its completion B(P, P, ≤).

Theorem 4 (Dedekind’s Completion Theorem) For an ordered set (P, ≤)

ιx := ((x] , [x)) for x ∈ P

defines an embedding ι of (P, ≤) in B(P, P, ≤); moreover, ι X = ιX or ι X =

W W V
ιX if the supremum or infimum of X, respectively, exists in (P, ≤). If κ is an arbitrary
V
embedding of (P, ≤) in a complete lattice L, then there is always also an embedding λ of
the ordered set B(P, P, ≤) in L with κ = λ ◦ ι.
1.3 Context constructions 47

Proof Evidently, the concepts of (P, P, ≤) are precisely the pairs (A, B) with
A, B ⊆ P and

A = B ↓ := {x ∈ P | x ≤ y for all y ∈ B},

B = A↑ := {y ∈ P | x ≤ y for all x ∈ A};

in particular, all pairs ((x] , [x)) with x ∈ P are concepts of (P, P, ≤), which con-
firms ι as an embedding. If the supremum of X exists in (P, ≤), then
h_  \
X = [x) ,
x∈X

i.e., ι
W
X=
_ i h_   \
↓ \  _ _
= X , X = [x) , [x) = ((x] , [x)) = ιX.
x∈X x∈X

The equation for existing infima is shown dually. With respect to the missing
part of the proof we refer to Proposition 39 (p. 119). 
We mention two familiar instances of this construction.
• What are the formal concepts of O(Q,≤) = (Q, Q, ≤), where (Q, ≤) is the set of
rational numbers in their natural order? Apart from the boundary concepts
(∅, Q) and (Q, ∅), there is exactly one concept for each real number r ∈ R,
namely
((r], [r)),
where
(r] := {q ∈ Q | q ≤ r} and [r) := {q ∈ Q | r ≤ q}.
These Dedekind cuts were used by R. Dedekind to construct the real numbers
from the rational ones. Indeed,

B(O(Q,≤) ) ∼
= (R ∪ {−∞, ∞}, ≤).

• Let (N, | ) be the set of non-negative integers (including zero), ordered by

divisibility. The ordinal scale O(N, | ) has exactly one formal concept for each
n ∈ N, namely (Div(n), n · N), where Div(n) denotes the set of divisors of n.
Indeed, B(O(N, | ) ) ∼
= (N, | ), since (N, | ) is already a complete lattice with zero
as the largest element.
(3) From an arbitrary ordered set (P, ≤) we furthermore obtain the reduced
context
Ocd
(P,≤) := (P, P, 6≥),

which is called the contraordinal scale for (P, ≤). In this case, the formal
concepts are precisely the pairs (X, Y ) with the following properties:
• X ∪ Y = P and X ∩ Y = ∅,
48 1 Concept lattices of formal contexts

• X is an order ideal in (P, ≤), i.e., x ∈ X and z ≤ x always imply

z ∈ X.
Because of X ∪ Y = P and X ∩ Y = ∅ this is equivalent to:
• Y is an order filter in (P, ≤), i.e., y ∈ Y and y ≤ z always imply z ∈ Y .

The formal context (P, P, 6≥) is doubly founded, since

x.y ⇔ x%y ⇔ x=y

holds for x, y ∈ P . Hence if x is an object and y is an attribute with (x, y) ∈ / I

(i.e., x ≥ y), then x % x and x0 = P \ [x) ⊃ P \ [y) = y 0 , hold for the attribute x,
as required by Definition 28.
The concept lattice B(P, P, 6≥) is isomorphic to the lattice of the order ideals
of (P, ≤). A look at (1) shows that all concepts of the contraordinal scale are con-
cepts of the contranominal scale NcP as well. We will prove later (Theorem 17, p.
140) that for this reason B(P, P, 6≥) is a complete sublattice of B(P, P, 6=), which
means that these lattices are completely distributive. Birkhoff’s theorem (Theo-
rem 46, p. 254) shows that the lattices constructed in this way are precisely the
doubly founded completely distributive lattices. In particular, every finite dis-
tributive lattice is isomorphic to the concept lattice of a contraordinal scale. The
dual lattice, i.e., B(P, P, 6≤), is often denoted by 2(P,≤) , because it is isomorphic
to the lattice of the order-preserving maps of (P, ≤) to the two-element lattice.

d
a b c d e f d @d
e d df a × × × × × @d
b × × × × ×
c dZ dd −→ d @d
Z
c × × × −→
d × × × @d
a dZ db
Z
e ×
f × d @d
(P, ≤) @d
(P, P, 6≥)
B(P, P, 6≥)

Figure 1.14 An ordered set (P, ≤), the corresponding contraordinal scale and its concept
lattice, i.e., the ideal lattice of (P, ≤).

(4) Free distributive lattices We obtain an interesting special case of (3) by

choosing the powerset of a set S as our ordered set (P, ≤), i.e., by consid-
ering the context
(P(S), P(S), 6⊇).
Because of A 6⊇ B ⇔ B ∩ (S \ A) 6= ∅, this context is isomorphic to
1.3 Context constructions 49

(P(S), P(S), ∆) with X ∆ Y : ⇔ (X ∩ Y ) 6= ∅.

The concept lattice is the free completely distributive lattice FCD(S). If

for S := {1, 2, . . . , n} we denote the context (P(S), P(S), 6⊇) by An , we
can state an easy recursion rule for the generation of these contexts:

An ×
A0 = ∅ and An+1 = .
An An

×
Note that An+1 = An × , as in Definition 32.
The construction can be generalized by taking an ordered set (S, ≤) as
the base set, the set OI(S, ≤) of the order ideals of (S, ≤) as the object set
and the set OF(S, ≤) of the order filters of (S, ≤) as the attribute set. The
concept lattice

FCD(S, ≤) := (OI(S, ≤), OF(S, ≤), ∆)

is called the free completely distributive lattice over the ordered set
(S, ≤).
(5) For an arbitrary ordered set (P, ≤), we define a filter to be a subset of P
which is an order filter and in which furthermore any two elements have a
common lower bound. Hence F ⊆ P is a filter if and only if the following
two conditions are satisfied:
1. From x ∈ F and y ≥ x it follows that y ∈ F ,
2. for any two elements x, y ∈ F there is an u ∈ F with u ≤ x and u ≤ y.
Dually, an ideal is defined to be a subset of P which is an order ideal and
contains a common upper bound for any two elements contained in it.
Filters in this sense are among other things the principal filters. Dually,
each principal ideal is an ideal. The set of all filters is denoted by F (P, ≤),
the set of all ideals by I(P, ≤). We obtain the doubly founded context

F(P,≤) := (F (P, ≤), I(P, ≤), ∆),

where again
F ∆ I : ⇔ F ∩ I 6= ∅.
(6) Again from an ordered set (P, ≤) we obtain the general interordinal scale

I(P,≤) := (P, P, ≤) | (P, P, ≥),

the concept system of which we explain by means of the extents: the at-
tribute extents are precisely the principal ideals and the principal filters
of (P, ≤), the concept extents are all intersections of those sets. These in-
50 1 Concept lattices of formal contexts

Figure 1.15 A nested line diagram of the free distributive lattice FCD(4). Such diagrams are
introduced in 2.2. The one shown here is due to S. Thiele [369]. The method that led to it is
explained in Example 17.
1.3 Context constructions 51

clude all intervals8 of (P, ≤). In general, these are all sets which constitute
intersections of intervals.
(7) By analogy with (6) we obtain the convex-ordinal scale

C(P,≤) := (P, P, 6≥) | (P, P, 6≤).

In this case, the extents are precisely the convex subsets of (P, ≤), i.e.,
those subsets which contain with any two elements a and b all elements
c with a ≤ c ≤ b.

1,a 1,b 1,c 1,d 1,e 1,f 2,a 2,b 2,c 2,d 2,e 2,f
a × × × × × ×
b × × × × × ×
c × × × × × ×
d × × × × × ×
e × × × × × ×
f × × × × × ×

Figure 1.16 The convex-ordinal scale of the ordered set from Figure 1.14.

If the ordered sets in the above definitions are composable, for example as
a cardinal sum or as a direct product, then it is to be expected that the formal
contexts defined from them also can be split up. This is true, even if in different
ways, as exemplified by the following rules:

Proposition 19 We have

O(P1 ,≤)+(P2 ,≤) = O(P1 ,≤) ∪˙ O(P2 ,≤)

I(P1 ,≤)+(P2 ,≤) = I(P1 ,≤) ∪˙ I(P2 ,≤)
Ocd cd cd
(P1 ,≤)+(P2 ,≤) = O(P1 ,≤) + O(P2 ,≤)

C(P1 ,≤)+(P2 ,≤) = C(P1 ,≤) + C(P2 ,≤)

Ocd cd cd
(P1 ,≤)×(P2 ,≤) = O(P1 ,≤) × O(P2 ,≤)

C(P1 ,≤)×(P2 ,≤) = Ocd cd c c

(P1 ,≤) × O(P2 ,≤) | O(P1 ,≤) × O(P2 ,≤)


8 in the sense of Definition 5 (p. 3), i.e., only the “closed” intervals.
52 1 Concept lattices of formal contexts

a f
b e
c d

Figure 1.17 The concept lattice of the convex-ordinal scale from Figure 1.16.

List of formal context families, continued

(8) Graph adjacency context. If R is a symmetric relation on S (easily visu-

alized by the edges of an undirected graph with or without loops) then
with
(S, S, R)
we obtain a formal context, the concepts of which are precisely the pairs
(A, B), A ⊆ S, B ⊆ S, which are maximal with respect to the property
that each element of A is in the relation R with each element of B (in
the visualization these are maximal complete bipartite edge sets). Thus,
together with (A, B), (B, A) is also a concept, and the map

(A, B) 7→ (B, A)

is a polarity, i.e., an order-reversing bijection which is inverse to itself

(another term for this is involutory antiautomorphism). Conversely, every
complete polarity lattice (i.e., every complete lattice with a polarity) is
isomorphic to the concept lattice of a formal context (S, S, R) with a sym-
metric relation R.
1.3 Context constructions 53

If the relation R is irreflexive, the extent and the intent of each concept
must be disjoint and we have

(A, B) ∧ (B, A) = (∅, ∅0 )

and (A, B) ∨ (B, A) = (∅0 , ∅),

i.e., (A, B) and (B, A) are complementary to each other: Their infimum is
the smallest, their supremum the largest element of the concept lattice. A
lattice with this kind of polarity is called an ortholattice; the complete or-
tholattices are (up to isomorphism) precisely the concept lattices of con-
texts with an irreflexive, symmetric relation.
There are many examples of such contexts in this book. They can be easily
recognized if the cross table is represented symmetric to the main diag-
onal. The context K(2,3) in Figure 1.21 is the context of a polarity lattice
but not of an ortholattice. The same applies to the context in Figure 5.8 on
p. 235, although this only becomes clear after an adroit reassembly of the
cross table.
(9) Partitions of a set Let S be a set and s ∈ S an arbitrary element. If we
now choose G to be the set of all two-element subsets of S and M to be
the set of all subsets of S \ {s}, by the definition

{x, y}  X :⇔ |{x, y} ∩ X| 6= 1
 
|S|
we obtain a context (G, M, ) with objects and 2|S|−1 attributes,
2
which is reduced except for one full column. Every extent of this context
is a set of two-element subsets of S, i.e., it can be understood as a sym-
metric reflexive relation on S; actually, the relations occurring are pre-
cisely the equivalence relations on S. Hence the concept lattice B(G, M, )
is isomorphic to the lattice E(S) of equivalence relations. We can give a
mnemonic rule for this context series as well. We get P1 := (∅, {∗}, ∅) and
obtain the n+1-st context of this series, Pn+1 , from the n-th as follows: We
form the apposition of Pn with the cross table Prev n , which is identical to
Pn , apart from the fact that the columns are written down in the reversed
order.

Pn Prev
n

2n − 1 . . . 2n−1 2n−1 − 1 . . . 0

We add n further rows, which we fill with crosses such that the columns
of this subcontext look like the binary representations of the numbers
2n − 1, . . . , 0. An example is given in Figure 1.18.
54 1 Concept lattices of formal contexts

× ×× ×
×× ××
10011001
× × × ×
32100123 −→
××××
76543210
×× ××
× × × ×

Figure 1.18 Context P4 for the lattice of equivalence relations on a 4-element set.

(10) Integer partitions. A partition of a positive integer n is Pa nonincreasing

n
sequence a1 ≥ a2 ≥ . . . ≥ an−1 ≥ an ≥ 0 with n = i=1 ai . Such a
partition is written down as n = a1 + a2 + . . . , with zeros omitted at the
end. Brylawski [67] has shown that the set of all partitions of n forms
a lattice when ordered as follows: For partitions a := (a1 , . . . , an ) and
b := (b1 , . . . , bn ) define
k
X k
X
a≤b :⇔ ai ≤ bi for k = 1, . . . , n.
i=1 i=1

Figure 1.19 shows the lattice for n = 8. Brylawski also proved that these
lattices admit a polarity (in the sense defined in (8) above), given by a 7→
a∗ with
a∗i := |{j | aj ≥ i}|, i ∈ {1, . . . , n}.
It can be deduced that there are precisely
 
(n + 1) · (n + 2)
−1
6

join-irreducible partitions of n, and that these are of the form

(q + 1, . . . , q + 1, q, . . . , q , 1, . . . , 1),
| {z } | {z } | {z }
r m−r s

where

n − s = q · m + r, s ≥ 0, m > r ≥ 0, q ≥ 2, and q ≥ 3 if s 6= 0 6= r.

Since a is meet-irreducible if and only if a∗ is join-irreducible, these in-

formations provide all that is necessary for writing down the standard
context.
(12) Permutations The set of all permutations of the set {1, . . . , n} can be
given a lattice order in a natural way. For this purpose we call a pair
1.3 Context constructions 55

7+1

6+2

6+1+1 5+3

5+1+1 4+4

4+3+1
5+1+1+1
4+2+2

4+2+1+1 3+3+2

3+3+1+1
4+1+1+1+1
3+2+2+1

3+2+1+1+1 2+2+2+2

3+1+1+1+1+1 2+2+2+1+1

2+2+1+1+1+1

2+1+1+1+1+1+1

1+1+1+1+1+1+1+1

Figure 1.19 The lattice of partitions of the integer 8.

(ϕi, ϕj) an inversion of the permutation ϕ if i < j but ϕi > ϕj. If we

order the permutations by

σ ≤ τ : ⇔ every inversion of σ is also an inversion of τ ,

we obtain, as proved by Yanagimoto and Okamoto [426], a lattice Σn .

There is a simple recursion rule for the description of the context: Putting

K0 := L0 := × and

Ln+1 := ∅ Ln , Kn+1 :=
Kn Kn
,
Ln Ln Kn Ln
then we obtain
Σn ∼
= B(Kn ).
The contexts Kn are reduced except for the full rows and full columns. Σ4
is presented in Figure 1.20.
56 1 Concept lattices of formal contexts

e4321
 HH

e3421 HH
 H
  HH HH

e e3241 HH e
HH
 4231

HH 4312
H

  HH e
 e2431  e3214 
    HH 3412
   
e e e e
  HH
4213 2341
 
 4132
 
H 3142
H HH
H  H

e e e He
H
HH 2314

2413
 
 1432 
 H  
    3124
H
    
e H e e
 H    
2143 HH  4123 
 1342 

  
e e e
H  
2134
HH
HH H
H 1423  1324
H HH e 
1243 
HH
H 
HH
H e
1234

Figure 1.20 The lattice Σ4 of the permutations of {1, 2, 3, 4}.

(13) Closure systems The intersection of any number of closure systems on

the same base set S results in a closure system again. Therefore, the clo-
sure systems on S themselves form a closure system (on P(S)) and there-
fore also a complete lattice. It is not difficult to give the standard contexts.
The objects are all proper subsets of S, while the attributes are all pairs
(A, b) with A ⊆ S, b ∈ S \ A. (Readers already familiar with Section 2.3
of this book will guess that such pairs (A, b) can also be understood as
implications A → b with one-element conclusions.) A subset X ⊆ S is
incident with a pair (A, b) if and only if it respects the implication A → b,
i.e. if A 6⊆ X or b ∈ X. The standard context for the lattice of all closure
systems on S is therefore

(P(S) \ {S}, {(A, b) | A ∪ {b} ⊆ S, b ∈

/ A}, |=),

where
X |= (A, b) : ⇔ (A ⊆ X ⇒ b ∈ X).
The extents of this contexts are precisely the closure systems on S, while
the intents, when read as sets of implications, describe the corresponding
closure operators. The arrow relations are also easy to specify, since one
gets
1.3 Context constructions 57

X . (A, b) ⇔ A ⊆ X, b ∈
/ X,
X % (A, b) ⇔ X %
. (A, b) ⇔ X = A.
It will be shown in Chapter 6 of this book that many structural properties
of these lattices can immediately be deduced from this characterization
of the arrows.

Formal contexts from mathematical structures

In advanced mathematics, relations that lead to interesting Galois connections

are ubiquitous. A detailed list would go beyond the scope of this book. We there-
fore limit ourselves to a few particularly typical examples and for this purpose
draw on an inspiring collection compiled by C. Hillman [192].
(14) If V is a finite dimensional vector space and V ∗ is the dual space of V ,
then
(V, V ∗ , ⊥) with a ⊥ ϕ : ⇔ ϕa = 0
is a doubly founded context, the extents of which are precisely the sub-
spaces of V . An example is given in Figure 1.21.

××××××××
× × × ×
×× ××
× ×× × K(d,2) K(d,2)
×××× K(d+1,2) = .
× × × × K(d,2) Kc(d,2)
×× ××
× × ××

Figure 1.21 K(3,2) := (GF(2)3 , GF(2)3 , ⊥), a formal context derived from the 3-dimen-
sional vector space over the two-element field. The help lines are intended to illustrate
the recursive construction given on the right. The first row and column are reducible. The
remaining 7 × 7-table gives the point-line incidence of the projective plane of order 2.

(15) Similarly, if H is a Hilbert space and ⊥ is the orthogonality relation de-

fined by x ⊥ y : ⇔ hx, yi = 0, then the concept lattice of the context

(H, H, ⊥)

is isomorphic to the (orthomodular) lattice of the closed subspaces of H,

since (U, U ⊥ ) is a concept for each such subspace U .
(16) The Pontryagin duality connects locally compact abelian group G with
its character group G
b through the formal context
58 1 Concept lattices of formal contexts

(G, G,
b I), where g I χ : ⇔ χ(g) = 1.

The extents of the formal concepts are the closed subgroups, the intents
are their annihilators.
(17) Galois theory Let E be a field, F be a subfield and let AutF (E) denote
the group of those automorphisms of E which pointwise fix F. Consider
the formal context

(E, AutF (E), I), where e I ϕ : ⇔ ϕ(e) = e.

The concept extents are intermediate fields between F and E the concept
intents are subgroups of AutF (E), but not necessarily all of them. It re-
quires additional conditions (“finite dimensional Galois extension”) to
make sure that every intermediate subfield is a concept extent and every
subgroup is an intent.
(18) Define for a group G the formal context

(G, G, I), where x I g : ⇔ x = g −1 xg ( ⇔ gx = xg).

The first definition specifies that x is incident to g exactly when x is a fixed

point under conjugation with g, similar to (17). The second, equivalent
definition also works for semigroups and therefore leads up to the oper-
ators of functional analysis. It also shows that the incidence is symmetric,
which makes it an instance of (8). The set of concept extents is thus equal
to that of concept intents. For S ⊆ G, S 0 is always a subgroup, called the
centralizer or commutant of S, but not all subgroups are necessarily of
that form.
(19) Pol-Inv has been nicknamed “the most basic Galois connection in alge-
bra”.9 Fix a base set A and denote by OA the set of all operations and by
RA the set of all relations on A (both finitary, but not nullary). Consider
the formal context
(OA , RA , .),
where f . r expresses that the operation f preserves the relation r. The
concept extents and intents are the (locally closed) clones of operations
and relations, respectively. The derivation operators are denoted by Pol
(for “polymorphisms”) and Inv (for “invariant relations”). A powerful
tool results from the fact that the two closure operators Pol Inv and Inv
Pol can also be calculated directly, i.e. without using the incidence.

9 Our presentation follows Kerkhoff et al. [219], where precise definitions can be found, but
also a reference to the quotation.
1.4 Conceptual scaling of many-valued contexts 59

1.4 Conceptual scaling of many-valued contexts

In standard language the word “attribute” is not only used for properties which
an object may or may not have. Attributes such as “color”, “weight”, “sex”,
“grade” have values. We call them many-valued attributes, in contrast to the one-
valued attributes considered so far. Data with many-valued attributes is usually
found in a form vaguely described as a data table. This is cast here into a mathe-
matical definition under the name many-valued context.
Definition 34 A many-valued context (G, M, W, I) consists of sets G, M and W
and a ternary relation I between G, M and W (i.e., I ⊆ G × M × W ) for which
it holds that

(g, m, w) ∈ I and (g, m, v) ∈ I together imply w = v.

The elements of G are called objects, those of M (many-valued) attributes and

those of W attribute values. (g, m, w) ∈ I we read as “the attribute m has the
value w for the object g”.
The many-valued attributes can be regarded as partial maps from G in W .
Therefore, it seems reasonable to write m(g) = w instead of (g, m, w) ∈ I.
m(G) := {m(g) | g ∈ G} then is the set of all values of the attribute m in this
many-valued context. The domain of an attribute m is defined to be

dom(m) := {g ∈ G | (g, m, w) ∈ I for some w ∈ W }.

The attribute m is called complete, if dom(m) = G. A many-valued context is

complete, if all its attributes are complete. ♦
Like the one-valued contexts treated so far, many-valued contexts can be rep-
resented by tables, the rows of which are labelled by the objects and the columns
labelled by the attributes. The entry in row g and column m then represents the
attribute value m(g). If the attribute m does not have a value for the object g,
there will be no entry.10

Conventional Front-wheel Rear-wheel Mid-engine All-wheel

Figure 1.22 Drive concepts for motorcars.11

10Further information on the role of the “empty cells” in a context will be given in the notes at
the end of the chapter.
60 1 Concept lattices of formal contexts

Example The many-valued context represented in the upper part of Figure 1.23
shows a comparison of the different possibilities of arranging the engine and the
drive chain of a motorcar (cf. Figure 1.22). >
How can we assign concepts to a many-valued context? We do this in the fol-
lowing way: The many-valued context is transformed into a one-valued one, in
accordance with certain rules, which will be explained below. The concepts of
this derived one-valued context are then interpreted as the concepts of the many-
valued context. However, this interpretive process, called conceptual scaling, is
not at all unique. The concept system of a many-valued context depends on the
scaling, and scaling is an interpretive decision. This may at first be confusing, but
has proved to be an excellent instrument for a purposeful evaluation of data.
In the process of scaling, first of all each attribute of a many-valued context is
interpreted by means of a formal context. This context is called a conceptual scale.
Definition 35 A scale for the attribute m of a many-valued context is a (one-
valued) context Sm := (Gm , Mm , Im ) with m(G) ⊆ Gm . The objects of a scale
are called scale values, the attributes are called scale attributes. ♦
Every formal context can be used as a scale. Formally there is no difference
between a scale and a context. However, we will use the term “scale” only for
contexts which have a clear conceptual structure and which bear meaning. Some
particularly simple contexts are used as scales time and again. A summary (in
tabular form) of the most important ones can be found in Figure 1.28 at the end
of this section.
As already mentioned, the choice of the scale for the attribute m is not math-
ematically compelling, it is a matter of interpretation. The same is true for the
second step in the process of scaling, the joining together of the scales to make
a one-valued context. In the simplest case, this can be achieved by putting to-
gether the individual scales without connecting them. This is described below
as plain scaling. Particularly when dealing with numerical scales this may well be
unsatisfactory. In this case we need the scaling by means of a composition operator.
In the case of plain scaling the derived one-valued context is obtained from
the many-valued context (G, M, W, I) and the scale contexts Sm , m ∈ M as fol-
lows: The object set G remains unchanged, every many-valued attribute m is
replaced by the scale attributes of the scale Sm . If we imagine a many-valued
context as represented by a table, we can visualize plain scaling as follows: Ev-
ery attribute value m(g) is replaced by the row of the scale context Sm which
belongs to m(g). A detailed description will be given in the following definition,
for which we first introduce an abbreviation: The attribute set of the derived con-
text is the disjoint union of the attribute sets of the scales involved. In order to
make sure that the sets are disjoint, we replace the attribute set of the scale Sm
by
Ṁm := {m} × Mm .
as in Definition 29.
11 Source: Schlag nach! 100 000 Tatsachen aus allen Wissensgebieten. BI-Verlag 1982.
 Drive Drive Road Self Space Construc- Maintain-
efficiency efficiency holding steering economy ability
empty loaded tion cost
Conventional poor good good understeering good medium excellent
Front-wheel good poor excellent understeering excellent very low good
Rear-wheel excellent excellent very poor oversteering poor low very poor
Mid-engine excellent excellent good neutral very poor low very poor
All-wheel excellent excellent good underst./neut. good high poor

Drive Drive Road Self Space Construc- Maintain-

efficiency efficiency holding steering economy
empty loaded tion cost ability
++ + – ++ + – ++ + – – u o n u/n ++ + – – – vl l m h ++ + – – –
Conventional × × × × × × × ×
Front-wheel × × × × × × × ×× ×
1.4 Conceptual scaling of many-valued contexts

Rear-wheel × × × × × × × × ×
Mid-engine × × × × × × × × × × ×
All-wheel × × × × × × × × ×

++ := excellent; + := good; – := poor; – – := very poor; u := understeering; o := oversteering;

n := neutral; vl := very low; l := low, m := medium; h := high.

Figure 1.23 A many-valued context: Drive concepts for motorcars. Below a derived one-valued context.
61
62 1 Concept lattices of formal contexts

Definition 36 If (G, M, W, I) is a many-valued context and Sm , m ∈ M are scale

contexts, then the derived context with respect to plain scaling is the context
(G, N, J) with [
N := Ṁm ,
m∈M

and
g J (m, n) : ⇔ m(g) = w and w Im n. ♦

f
@
 @
f
Dl+ @
@ @
@ @
fM+@
@
@ @
R+ f @ @ fDe+
aa  @ @ ! !
aa f
 aa  @ ! @! @
 ! @ 
 Dl++ @ fCl
! @
E+ f f aa@f!! @ @f
De
 @ ++
a a
 @ aaaaaa
@
 @ @ !!@ !! @!!!
 aa aa  ! !@ !! @ 
M-
@ @
!@
f @f aa f a f!
a  !
f!
! @f @f  E-
aa Su aa
@@
! ! !!
@ aa aa  @ Cvl @  @! !! 
@  aa Ch a  a @ R++@ ! !!!@ ! 
@f aa f @  R- -,So
E++ fDl- fE- -,M- - ! f
M++,Cm,De-   aa @ !@ ! ! !
aa Su/n
Standard aa
@  Sn !!
! Rear-wheel

aa @   !AA
All-wheel Mid-engine
!
a@f
a !!
Front-wheel


Figure 1.24 Concept lattice for the context of drive concepts.

Example We obtain the one-valued context in Figure 1.23 as the derived context
of the many-valued context presented above it, if we use the following scales.
The abbreviations are: De := drive efficiency when empty, Dl := drive efficiency
when loaded, R := road holding/handling properties, S := self steering effect,
E := economy of space, C := cost of construction, M := maintainability.
++ + − ++ + −−
SDe := SDl := ++ × × SR := ++ × ×
+ × + ×
− × −− ×
1.4 Conceptual scaling of many-valued contexts 63

u o n u/n ++ + − −− vl l m h
u × ++ × × vl × ×
SS := o × SE := SM := + × SC := l ×
n × − × m ×
u/n × −− × × h ×

The concept lattice is shown in Figure 1.24. If we had used the scale SE for the
attributes De, Dl and R as well, the derived context would have only turned out
slightly different. >
The mathematical definition of formal contexts allows to turn relations orig-
inating from arbitrary domains into formal contexts and to investigate their
concept lattices, i.e., even those where an interpretation of the sets G and M as
“objects” or “attributes” seems artificial. This is the case for many formal con-
texts from mathematics, whose concept lattices then have structural properties
that are highly rare in empirical data. Nevertheless, such formal contexts are of
great importance for data analysis. They are suitable, for example, as ideal struc-
tures or as scales for the conceptual scaling of many-valued contexts, introduced
above.

Standard scales

The simplest and most commonly used scales are now introduced in standard-
ized form. The elementary ones are discussed in Definition 37, a more general
selection is compiled in Figure 1.28. That these scales are standardized means the
following: The definition of scaling requires that the value set of a many-valued
attribute is a subset of the object set of the scale used. For the standardized scales
we frequently use n := {1, 2, . . . , n} as the object set. It may then be necessary to
rename the objects before scaling.

Definition 37 (Elementary standard scales)

Nominal scales Nn := (n, n, =).
N4 1 2 3 4
Nominal scales are used to scale attributes, the values of
which mutually exclude each other. If an attribute for exam- 1 ×
ple has the values {masculine, feminine, neuter}, the use of a 2 ×
nominal scale suggests itself. We thereby obtain a partition of 3 ×
the objects into extents. In this case, the classes correspond 4 ×
to the values of the attribute.
Contranominal scales Ncn := (n, n, 6=).
The concepts of these formal contexts are precisely the pairs Nc4 1 2 3 4
(A, n \ A) for A ⊆ n. The concept lattice is isomorphic to the 1 × × ×
powerset lattice of n, and thus has 2n elements. Not frequently 2 × × ×
used for conceptual scaling, but of general significance, see [2] 3 × × ×
for example. We sometimes use the symbol 6= in context con- 4 × × ×
structions.
64 1 Concept lattices of formal contexts

(One-dimensional) ordinal scales On := (n, n, ≤).

O4 1 2 3 4
These scale many-valued attributes, the values of which are 1 × × × ×
ordered and where each value implies the weaker ones. If an 2 × × ×
attribute has for instance the values {loud, very loud, extremely 3 × ×
loud}, ordinal scaling suggests itself. The attribute values then
4 ×
result in a chain of extents, interpreted as a hierarchy.

(One-dimensional) interordinal scales In := (n, n, ≤) (n, n, ≥) .

Questionnaires often offer opposite
pairs as possible answers, as for ex- I4 ≤ 1 ≤ 2 ≤ 3 ≤ 4 ≥ 1 ≥ 2 ≥ 3 ≥ 4
ample active–passive, talkative–taciturn 1 × × × × ×
etc., allowing a choice of intermediate 2 × × × × ×
values. In this case, we have a bipolar 3 × × × × ×
ordering of the values. Such attributes 4 × × × × ×
lend themselves to scaling by means
of an interordinal scale. The extents of the interordinal scale are precisely the
intervals of values, in this way, the betweenness relation is reflected conceptually.

However, bipolar attributes often also lend themselves to biordinal scaling:

Biordinal scales Mn,m := (n, n, ≤) ∪˙ (m, m, ≥).
In common usage we often use opposite
pairs not in the sense of an interordinal M4,2 ≤ 1 ≤ 2 ≤ 3 ≤ 4 ≥ 5 ≥ 6
scale, but simpler: each object is assigned 1 × × × ×
one of the two poles, allowing gradua- 2 × × ×
tions. The values {very low, low, loud, very 3 × ×
loud} for example suggest this way of scal- 4 ×
ing: loud and low mutually exclude each 5 ×
other, very loud implies loud, very low im- 6 × ×
plies low. We also find this kind of partition
with a hierarchy in the names of the school marks: An excellent performance ob-
viously is also very good, good, and satisfactory, but not unsatisfactory or a fail.

The dichotomic scale D := ({0, 1}, {0, 1}, =). The dichotomic 0 1
scale constitutes a special case, since it is isomorphic to the scales 0 ×
N2 and M1,1 and closely related to I2 . It is frequently used to scale 1 ×
attributes with values of the kind {yes, no}.

A more comprehensive list of standard scales is compiled in Figure 1.28. ♦

A special case of plain scaling which frequently occurs is the case that all
many-valued attributes can be interpreted with respect to the same scale or fam-
ily of scales. Thus we speak of a nominally scaled context, if all scales Sm are
nominal scales etc. We call a many-valued context nominal, if the nature of the
1.4 Conceptual scaling of many-valued contexts 65

c ≤4 ≥1
c c≤ 4
c ≤4 4
J
c 4 J c≥ 5
J ≤ 3 c @ c≥ 2 c≤ 3
≤3c 5
 B@ 3 ≤ 2 c @ c @ c≥ 3 3
≤2c c≤ 2
c1 c2 c3 c4
 BB @
@ 2 c≥ 6 ≤ 1 c @ c @ c @ c≥ 4 2
≤1c 6 1 2 3  4
@ B  1 Q
Q A

 c≤ 1
@c 1
@@B c
B @ QQA c

A

Figure 1.25 The concept lattices of the elementary scales are named after the scales. The
figure shows a nominal lattice, B(N4 ), a biordinal lattice, B(M4,2 ), an interordinal lat-
tice, B(I4 ), and an ordinal lattice, B(O4 ). The ordinal lattice B(On ) is isomorphic to the
n-element chain Cn .

Forum Romanum Baedecker Les Guides Bleus Michelin Polyglott

1 Arch of Septimus Severus ∗ ∗ ∗∗ ∗
2 Arch of Titus ∗ ∗∗ ∗∗
3 Basilica Julia ∗
4 Basilica of Maxentius ∗
5 Phocas column ∗ ∗∗
6 Curia ∗
7 House of the Vestals ∗
8 Portico of Twelve Gods ∗ ∗ ∗
9 Tempel of Antonius and Fausta ∗ ∗ ∗∗∗ ∗
10 Temple of Castor and Pollux ∗ ∗∗ ∗∗∗ ∗
11 Temple of Romulus ∗
12 Temple of Saturn ∗∗ ∗
13 Temple of Vespasian ∗∗
14 Temple of Vesta ∗∗ ∗∗ ∗

Figure 1.26 Example of an ordinal context: Ratings of monuments on the Forum Ro-
manum in different travel guides. The context becomes ordinal through the number of
stars awarded.

data suggests nominal scaling; a many-valued context is called an ordinal con-

text if for each attribute the set of values is ordered in a natural way. An example
is presented in Figure 1.26, see also Figure 1.27.
In general, it is not uncommon for the way the data are presented to suggest
a particular scale. We then speak of a pre-scaled many-valued context, without
defining this precisely. It is often clear what values the individual attributes can
take, i.e., sets W (m), m ∈ M , are given such that {m(g) | g ∈ G} ⊆ W (m). Then
the many-valued context is stratified. In addition, relations on these sets W (m)
may be given, e.g., an order.
66 1 Concept lattices of formal contexts

e

AQ

 Q

A QQ

Polyglott:∗  M:∗

A
A eGB:∗ Q Q eBaedecker:∗
Q
e e

Curia Temple Basilica of
of Maxentius
!! @
Basilica Julia !
!
Romulus
@
House of the Vestals @
M:∗∗ @
e e @e

Temple of 
@ @
Vespasian @ @
@ @
e @@ e @ eColumn
@
Temple of Saturn of Phocas
@ @
@ @
Portico of the
@ @
GB:∗∗ @
Twelve Gods
@@ e e @e
@ @
@ @
@ @
Temple of Vesta
e @
@ e @e
@
Arch of Titus
M:
@
∗∗∗
Arch of @,
, e
Septimus Severus @
PP
Temple of Antoninus
@e
@ and Faustina
Temple of Castor
and Pollux

Figure 1.27 The concept lattice of the ordinal context from Figure 1.26. M := Michelin,
GB := Les Guides Bleus.

By definition, the number of attributes of a derived context is the sum of the

numbers of scale attributes used, and thus tends to be large. It is therefore com-
mon to use selected subsets of these derived attributes, or attribute combinations.
This leads to the notion of a view:
Definition 38 A view of a formal context (G, M, I) is a formal context (G, N, J),
where for each n ∈ N there is a set An ⊆ M such that

g J n : ⇔ An ⊆ g I .

A contextual view of a many-valued context T is a view of a derived context of

T; the concept lattice of such a contextual view is a conceptual view of T.
♦
1.5 Notes, references, and trends 67

Symbol Definition Name Basic meaning

OP (P, P, ≤) general ordinal scale hierarchy

On (n, n, ≤) one-dimensional rank order

ordinal scale

Nn (n, n, =) nominal scale partition

Mn1 ,...,nk On1 +···+nk multiordinal partition with

scale rank orders

Mm,n Om+n biordinal scale two-class

rank orders

Bn (P(n), P(n), ⊆) n-dimensional dependency of

Boolean scale attributes
` `
Gn1 ,...,nk O n1 a
· · · a O nk k-dimensional multiple
grid scale ordering

Ocd
P (P, P, 6≥) contraordinal scale hierarchy and
independence

Ncn (n, n, 6=) contranominal partition and

scale independence

D ({0, 1}, {0, 1}, =) dichotomic scale dichotomy

` `
Dk D a
··· D a
k-dimensional multiple
dichotomic scale dichotomy
| {z }
k−times

IP OP OdP general betweenness

interordinal scale relation

In On Odn one-dimensional linear between-

interordinal scale ness relation

CP Ocd c
P OP convex-ordinal convex ordering
scale

Figure 1.28 Standardized scales of ordinal type.

1.5 Notes, references, and trends

1.1

Formal Concept Analysis has been developed from the end of the 1970s at the
Faculty of Mathematics of Darmstadt University of Technology. The first pro-
grammatic publication on Formal Concept Analysis was
68 1 Concept lattices of formal contexts

Rudolf Wille: Restructuring lattice theory: An approach based on hierarchies of

concepts. [389]
This paper contains already many of the ideas described in this book, including
the proof of the Basic Theorem on Concept Lattices.
The mathematical foundations for the Basic Theorem were already known in
the early 1940s, as reported at the end of the previous chapter (p. 20). Quite cer-
tainly the mathematical substance of that theorem can be mainly attributed to
Birkhoff [49], even if the second part has not been formulated there. This can
be found – in an order-theoretic version – in J. Schmidt [337] and Banaschewski
[20]. The general version presented in this book first appeared in [389]. It is not
quite easy to attribute the intermediate steps to particular authors. The fact that
a finite lattice is determined by its irreducibles was well known to lattice theo-
rists. One source is Markowsky [275]. The proof of the Basic Theorem has been
machine-checked, see Sertkaya and Oğuztüzün [341].
The example in Figure 1.1 is taken from an educational study, cf. Takács [366].
There had been earlier proposals to make use of the mathematical possibili-
ties offered by these results for data analysis. The interpretation of the incidence
relation as an object-attribute-relation was expressly mentioned in the first edi-
tion of Birkhoff’s book on lattice theory. Remarkable approaches can be found in
the work of Riguet [319] and Barbut [23], see Barbut & Monjardet [24]. Some
French authors therefore employ the term treillis de Galois, which was used in
their works for “concept lattice” (German Begriffsverband).
The Darmstadt group was presumably the first who systematically elaborated
these possibilities into a formal theory of conceptual knowledge, which then was
tested and further developed in many applications. The decisive factor in the
success of this work was among other things the formalization of “context” and
the interpretation of “concept” as a unity of extension and intension.
The understanding of “concept” which is formalized here has ramified and
deep-reaching roots in philosophy, which are described in more detail else-
where [410]. This tradition of thought finds expression even in the standards12
DIN 2330 and DIN 2331, which in turn were discussed by the Darmstadt group
at the beginning of the development. Further information on the origin of For-
mal Concept Analysis and its intellectual background can be found in [408] and
[105].

1.2

There is a simple way to assign a clarified context to every context K := (G, M, I):

K◦ := (G/ ker γ, M/ ker µ, I ◦ ),

12DIN stands for “Deutsche Industrienorm” and is characteristic for standards issued by the
German National Bureau of Standards.
1.5 Notes, references, and trends 69

the symbols having the following meanings: ker γ is the equivalence relation on
G with
(g, h) ∈ ker γ : ⇔ γg = γh.
ker µ is defined correspondingly. The equivalence classes of ker γ are the objects
of K◦ , those of ker µ the attributes. The incidence is defined by

([g] ker γ, [m] ker µ) ∈ I ◦ : ⇔ gIm.

The number of reduced contexts with four objects is 126. It is 13596 for five
objects, and 108096891 for six. The latter value was found by Klaus Hoffmann,
see sequence A047684 in https://oeis.org/.
Even without the additional
|G|\|M | 1 2 3 4 5 6
condition “reduced”, it is not
1 2 3 4 5 6 7
easy to determine the numbers.
2 3 7 13 22 34 50
The figures in the adjacent table
3 4 13 36 87 190 386
were calculated by M. Wiesend
4 5 22 87 317 1053 3250
in the Bayreuth group. Our
5 6 34 190 1053 5624 28576
source are lecture notes by
6 7 50 386 3250 28578 251610
A. Kerber and W. Lex [218].
7 8 70 734 9343 136758 2141733
Further values can again be
8 9 95 1324 25207 613894 17256831
found in https://oeis.org/,
see e.g., sequences A002623, A002727, and A006148.
Neither is it easy to determine the maximal possible number f (n) of attributes
in a reduced context with n objects. For small n we obtain

n 1 2 3 4 5 6
,
f (n) 1 2 4 7 13 24

where f (6) = 24 was recently found by Ignatov [197]. The question was men-
tioned on StackExchange as #3241237, and a constructive lower bound was given
there by Peter Taylor,
X  i 
f (n) ≥ i .
0≤i<n 2

Taylor’s construction provides the following attribute extents:

jik
{S ∪ {i + 2, . . . , n} | S ⊆ {1, . . . , i}, |S| = , i < n}.
2
An easy upper bound is f (n + 1) ≤ 2 · f (n). A better estimate, namely f (n) ≤
n
/n, was proved by Kleitman [221] under weaker assumptions.

n
n
b2c + 2
Several authors have addressed the problem of generating formal contexts
or closure systems at random, see Kerber & Lex [218], Emilion & Lévy [115],
Boley et al. [58], Ganter [144], Felde et al. [125], Borchmann & Hanika [63],
Sakurai [330].
70 1 Concept lattices of formal contexts

The arrow relations have been introduced in [390] following the example
of the weak perspectivities in congruence theory (cf. [175]). There were numer-
ous forerunners. For example Day [86] already used the double arrow relation
(“relation ρ”) in order to characterize semidistributivity as well as a “relation
C”, which is closely related to the arrow relations, in order to describe the con-
gruences of finite lattices. Doubly founded lattices have first been mentioned in
[396]. Geyer [169] has examined possible configurations of the arrow relations.
Algebraic lattices in Formal Concept Analysis were studied by Hitzler et al. [193],
see also [236].
The data for Figure 1.8 were compiled from the following sources:
List of developing countries: DAC list 2023 of the German government,
LDC, SIDS: https://www.un.org/ohrlls/,
OPEC: https://www.opec.org/opec_web/en/about_us/25.htm,
OACPS: https://www.oacps.org/.

1.3

Context constructions are treated in many papers, including [134]. The comple-
mentation has been comprehensively examined in Deiters [91], [90], but see also
Berry & Sigayret [46]. Instead of (g, m) ∈
/ I we sometimes write g r
I m, as in the
following simple, but useful proposition:
Proposition 20 (Krötzsch [236]) g ∈ hII ⇔ h ∈ gr
IrI
. 
Weinheimer [384] introduces the product apposition as a further construction.
For Xia’s context product see Definition 98 on p. 284. Doerfel [99] proves that
products of doubly founded complete lattices are doubly founded. The concept
lattices of the semiproduct powers of the dichotomic scale are precisely the “full
concept lattices” in Lex [262].
There is another definition of a product of contexts that suggests itself,

K1 & K2 := (G1 × G2 , M1 × M2 , &)

with
(g1 , g2 ) & (m1 , m2 ) : ⇔ g1 I1 m1 and g2 I2 m2 .
This has been considered by various authors (Schaffert [332], Reuter [315], Erné
[118]), but does not have the importance of the direct product in mathematical
literature. The extents of K1 & K2 are besides G1 × G2 precisely the sets U1 × U2
with
Ext(Ki ) if Gi is an attribute intent of Ki ,

Ui ∈
Ext(Ki ) \ {Gi } if not.
The concept lattice is therefore closely related to those of the context sum and
the semiproduct.
1.5 Notes, references, and trends 71

The Dedekind Completion Theorem (which generalizes Dedekind’s construc-

tion of the real numbers from the rational numbers) can be found in an order-
theoretic version already in MacNeille [272] and J. Schmidt [337]. Compare also
Banaschewski and Bruns [21].
Further interesting contexts can be obtained from an ordered set (P, ≤). For
example, the concept lattice of the context (P, P, 6>) can be interpreted as the lat-
tice of the maximal antichains of (P, ≤), see [395] and Reuter [317]. Ignatov [198]
used this to determine the number of maximal antichains in a Boolean lattice
with 7 atoms, see sequences A326359 and A348260 of https://oeis.org.
Free distributive lattices have been examined using the methods of Formal
Concept Analysis in [404] and Bartenschlager [26], [25] and using closely re-
lated methods even earlier by Markowsky [274]. Compare also Luksch [264].
The extensive literature on this subject can be looked up in [26], [25]. The ∆-
relation has been defined in [395].
Symmetric contexts have been treated by B. Schmidt [336] and Schaffert [332].
There are other connections to Graph Theory. Of course, every formal context
can be read as a bipartite graph (and conversely), with the formal concepts corre-
sponding to the maximal bicliques (i.e., maximal complete bipartite subgraphs).
T. Hanika’s thesis [182] contains a wealth of information and links on this topic.
Another connection results from the fact that the minimal separators of a graph
always form a complete lattice, see Berry & Sigayret [45]. Lehmann [258] gives
families of graphs for which standard contexts for the latttice of minimal sepa-
rators are exponentially larger than the graphs.
Integer partitions as formal concepts were studied in [146] and by Behrisch et
al. [32]. Almazaydeh et al. [3] have characterized the arrow relations for these
formal contexts.
Flath ([132], [131]) has generalized the description of the irreducible ele-
ments of the lattice Σn of permutations by Bennett and Birkhoff [51] to mul-
tipermutations and used it among other things in order to determine the order
dimension of those lattices with the methods of concept analysis. Meschke [282]
studies lattices of orders, with the permutation lattices as a special case.
The lattice of closure systems was studied in detail by Caspard and Mon-
jardet [72], but see also Reppe’s dissertation [312] and Grötzsch [178].

1.4

Many-valued contexts have been introduced already in [389]; of the numerous

related models we would like to mention the relational data bases of Codd [77] but
also the information systems of Pawlak [298] as well as the Chu spaces (cf. [303]).
Their use in conceptual file and knowledge systems has been discussed in Vogt,
Wachter & Wille [379], in Scheich, Skorsky, Vogt, Wachter & Wille [333] and
in [407]. With respect to conceptual scaling of many-valued contexts see [159].
The term “scale” has been chosen in order to emphasize the connection with
mathematical Theory of Measurement (cf. [229]), although the approaches dif-
72 1 Concept lattices of formal contexts

fer considerably. Whereas in Measurement Theory a scale is usually understood

to be a map to the real numbers, i.e., to a fixed structure, it has proved to be ex-
traordinarily useful for conceptual scaling to be able to choose different scales for
different many-valued attributes in accordance with their conceptual structure,
even if the value set remains the same. Therefore, there are many ordinal scales
in Concept Analysis, in contrast to Measurement Theory.
Interpretations of scales have also been treated by Spangenberg and Wolff
[347].
“Empty cells” of a one-valued context (i.e., pairs (g, m) with (g, m) 6∈ I) are
considered to be not concept forming. If we want to use the negation for the for-
mation of concepts, we have to dichotomize the respective attribute m, i.e., we
have to introduce an additional attribute ¬m with (g, ¬m) ∈ I : ⇔ (g, m) 6∈ I.
This corresponds to interpreting the attribute as two-valued and scaling it with
the dichotomic scale.
“Empty cells” of a many-valued context (i.e., pairs (g, m) with (g, m, w) 6∈
I for all w ∈ W ) in the case of plain scaling usually result in empty cells in the
derived context. If it is useful in terms of content, they can also be interpreted as
values and be included into the scaling, see e.g. Figure 1.26.
The contraordinal scales are of central importance when treating distributive
concept lattices [396]. Scaling by means of those scales has been carefully ex-
amined by Strahringer [355]. On this basis, Strahringer and Wille [356], [357]
develop an ordinal data analysis.
Strahringer has also worked on convex-ordinal scaling [354]. Strahringer and
Wille show in [358] that this kind of scaling lends itself to formulating a gener-
alized cluster analysis. This has been further elaborated by Leonhard and Winter-
berg [261]. Formal Concept Analysis has also proved useful for the classification
of ordinal cluster methods; compare Janowitz & Wille [201].

Research developments

Generalized Formal Concept Analysis

The question of whether and how Formal Concept Analysis can be generalized
was also intensively pursued. This seems obvious, since the theory starts with
an extremely simple form of data, the formal context, and one can investigate
whether the same approach will lead to success with other, more sophisticated
data. However, it should not be overlooked that our approach is intentionally
simple and yet universal. The Basic Theorem (Theorem 3) states that every com-
plete lattice is isomorphic to a concept lattice. Any “generalization” that ulti-
mately leads to a complete lattice can therefore also be formulated with the help
of formal contexts. Nevertheless, such alternative paths to the same goal can be
very valuable because they provide a different language framework that may be
more intuitive or practical for certain applications.
1.5 Notes, references, and trends 73

It should also be recalled that the reason for the success of Formal Con-
cept Analysis is that it provides mathematics which supports human conceptual
thinking. Mathematical methods, even very abstract ones, are welcome if they
have the potential to facilitate human-oriented knowledge processing. Mere gen-
eralization without reference to real meaning is not our aim.
The most important addendum to the model presented in Sections 1.1 and 1.2
of this book in our view is the inclusion of many-valued contexts by means of
conceptual scaling as introduced in Section 1.4.

Generalizations of Formal Concept Analysis have been discussed in several

variants:

Triadic Concept Analysis

Lehmann and Wille [257] have outlined a triadic concept analysis, where the
incidence relation is ternary and the concepts consist of three sets. An algorithm
for finding the triadic concepts was given in [235]. Several studies have shown
that this approach can be used for applications, see e.g. Jäschke et al. [202], Be-
lohlavek & Vychodil [41], Glodeanu [172], Missaoui & Kwuida [283], Ignatov
et al [199], Rudolph et al. [326], or Felde & Stumme [127]. The mathematical
theory, however, seems to be demanding. The “Basic Theorem” for the obtained
trilattices was proved in [409] and was later generalized to the n-adic case by
Voutsadakis [382]. Biedermann [48] has presented a first systematic study. But
many fundamental questions have so far remained unanswered. Remarkable re-
cent results by Bazin and his coauthors [29, 30, 31] suggest that the answers may
be different than thought.

Fuzzy concepts

Early on, the question of how to combine Formal Concept Analysis with aspects
of Fuzzy Logic was discussed. There are related elements in the work of Pawlak
[297] and Burusco Juandeaburre & Fuentes-Gonzales [204], and Umbreit [371]
has presented a first comprehensive study. Wolff [422] draws a comparison to
conceptual scaling.
The topic was then taken up by Belohlavek, Vychodil and their collaborators
at Palacky University, Olomouc. Their careful research resulted in a large num-
ber of publications which cannot be adequately discussed here, see [34, 42] for
example, but also, for a broader setting, [33]. Much of the work of this research
group is documented in the Proceedings of the CLA conference series, available
at cla.inf.upol.cz. The “Basic Theorem” proved in [34] was later simplified by
Krupka [240].
74 1 Concept lattices of formal contexts

A group at the University of Malaga is studying a different approach to fuzzy-

fication called multi-adjoint concept lattices, see Medina et al. [278]. S. Krajči [228]
contributed an early paper in this direction.
Not working with fuzzy sets, but somewhat related is the approach by Val-
verde-Albacete and Peláez-Moreno [373], where the incidence relation takes val-
ues in an idempotent semifield.

Faulty data

The question of how Formal Concept Analysis deals with unreliable data was
also repeatedly raised. Again, it is worth recalling the fundamental aim of this
theory development, which is to strengthen cognitive autonomy. This requires
that simplifying assumptions are made transparent, including their impact on
the results. By no means should an opaque mathematical formalism hide arti-
facts that cannot be justified rationally. It would be presumptuous to claim that
Formal Concept Analysis can simply redeem these claims. But they remain es-
sential.
It may first be necessary to specify what exactly is meant. Various authors have
dealt with incomplete (e.g., Burmeister and Holzer [70], Dubois and Prade [107],
Krupka and Lastovicka [241], [154], Zhi and Li [427]), faulty (e.g., Pensa and
Boulicaut [299], Cellier et al. [74], Borchmann [61]), and with probabilistic data
(e.g., Vityaeva et al. [375], Kriegel [230]). It seems unclear whether these ap-
proaches can interact and be unified.
Fuzzy concept analysis was already metioned as one possible approach. Some
authors have also explored the links to rough set theory, see Kent [214] for an
early contribution, Poelmans et al. [301] for a survey, and Meschke [281] for a
theoretical analysis. The Darmstadt group has focused on “formal contexts with
question marks” to handle incomplete data.

Logical information systems and pattern structures

A differently formulated approach to concept lattices is offered by pattern struc-

tures, which are studied by the group around S. O. Kuznetsov, see [151] for an
early publication and [248] and Kaytoue et al. [213] for comprehensive presenta-
tions. Similar approaches underly the logical information systems of Ferre and
Ridoux [128, 129], but also the abstract concept lattices of Soldano and Ven-
tos [345].
These methods build on the idea of admitting descriptions13 like graphs, in-
tervals, logical formulas etc. as attributes of the formal context. Of course, this is
also possible in the standard model, and therefore these approaches can also be
traced back to it. However, in some situations it is more convenient and natural

13 which then must be equipped with the order of a complete semilattice

1.5 Notes, references, and trends 75

to be able to use the syntax of logical expressions. The Generalized Formal Concept
Analysis by Chaudron and Maille [75] is also similar.

Structurally similar theories

There are at least two elaborated theories that are structurally closely related to
Formal Concept Analysis but have distinctly different application goals and use
different terminology. One is Barwise and Seligman’s theory of distributed sys-
tems, and is documented in [28]. The other is the theory of knowledge spaces
(and later of learning spaces) developed by Doignon and Falmagne [101, 122].
The latter provides a theory of human learning with extensive practical appli-
cations. One difference of the mathematical theory is that the knowledge states
(which roughly correspond to concept extents), are not necessarily closed under
intersections, but under set unions.
Not so close, but still related, are approaches by Diday [93] and Marty [277].
A similar attempt at restructuring with respect to mathematical logic is made by
[411].

Additional structure

Formal contexts with additional structure, e.g., with an additional operation,

have also been studied. The respective concept lattices in this case carry addi-
tional structural properties as well. Examples are the polarity lattices and ortho-
lattices introduced in 1.3. Generalizations can be found in Hoch [194]. Contexts
with an algebraic structure have been examined by Vogt [376], [378], [377], [380]
and by U. Wille [420], [421], contexts with a topological structure by Hartung
[185], [188], [186] and Săcărea [329]. K. E. Wolff [424] drafts a Temporal Concept
Analysis, see also Neouchi et al. [292].
Relational Concept Analysis is discussed in the section on Contextual Logic
on page 333, but see also Section 8.4.
Chapter 2
Determination and representation

How do you calculate the concept lattice and a good diagram for a given for-
mal context? In principle, this is simple and can be done by hand. We introduce
this in the first section of this chapter. However, the work becomes tedious very
quickly, and computer assistance is desired. Since the introduction of Formal
Concept Analysis, so many algorithms and implementations have emerged that
we cannot discuss them here and instead refer to [154]. These computer pro-
grams are impressively fast. A concept lattice with a billion elements can easily
be computed, and diagrams are also provided. However, computer-generated
lattice diagrams are rarely satisfactory. This is not so much because of the pro-
gramming as because it is not really understood what actually makes a good
diagram. Nested line diagrams, which we introduce alongside, allow readable
representations of systems consisting of a few hundred concepts. However, di-
agrams of larger concept lattices are usually no longer readable. For this pur-
pose, we will develop techniques in later chapters that allow a decomposition
into readable parts.
Another area in which there has been extensive development is the theory of
implications between attributes, in particular the knowledge acquisition method
of attribute exploration. This is also extensively represented in the already cited
book [154], so that we can limit ourselves here to the basic idea.

2.1 All formal concepts of a formal context

In principle, it is not difficult to find all the concepts of a formal context. The
following proposition summarizes some naive methods for doing so:
Proposition 21 Each formal concept of a formal context (G, M, I) has the form (X 00 , X 0 )
for some subset X ⊆ G and also the form (Y 0 , Y 00 ) for some subset Y ⊆ M . Conversely,
all such pairs are concepts.
Every concept extent is the intersection of attribute extents and every concept intent
is the intersection of object intents. 

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 77

B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_3
78 2 Determination and representation

However, this proposition does not immediately lead to a workable method.

Only for a very small formal context (G, M, I) will one take the trouble to deter-
mine the expression (X 00 , X 0 ) for each of the 2|G| subsets of G in order to generate
all formal concepts. The second part of the proposition at least yields a method
to determine all concepts of a small formal context by hand. To do this, first cre-
ate a list of concept extents as follows: At the beginning, the list is empty. Then
proceed as described below:
First step. The extent G is entered into the list.
Then one carries out the following for each attribute m ∈ M (the attributes are
processed in an arbitrary order):
Step m. For each set A, entered into the list in an earlier step, form the set

A ∩ m0

and include it into the list, provided that it is not yet contained within it.
It is easy to see that this list finally contains exactly those sets which can be writ-
ten as an intersection of attribute extents, i.e. according to the proposition exactly
the concept extents. Then, using the formal context, we can find the concept in-
tent A0 for each such extent A. Thus we obtain a list of all formal concepts (A, A0 )
of the context.

Step Extents Step Extents Step Extents

1 {1, . . . , 8} d {5, 6, 7, 8}, {5, 6} g {1, 2, 3, 4}, {1, 2, 3}
a {6, 7, 8}, {6} {3, 4}, {3}
b {1, 2, 3, 5, 6} e {7}, ∅ h {2, 3, 4}, {2, 3}
c {3, 4, 6, 7, 8}, {3, 6} f {5, 6, 8}, {6, 8} i {4}

Figure 2.1 List of concept extents for the formal context in Figure 1.1.

Determining the concepts in this way is usually easier if one starts drawing a
diagram of the concept lattice at the same time. We show this, staying with the
example of the “living beings and water” context from Figure 1.1. The interme-
diate steps are presented in Figure 2.2. First, draw a small circle for the largest
concept in the context. If there are attributes whose attribute extent contains all of
G, then these attribute names are entered above that circle. In our example, this
would be “a”. Then we determine the attributes the extents of which are maxi-
mal among the remaining attribute extents. In our example, we obtain b, c, d and
g. The attribute concept of each of these attributes is represented by a small circle
below the circle already drawn. These circles are then linked to the circle of the
largest concept and the names of the new attributes are entered above the respec-
tive circles. Now we systematically form the infima of the already represented
2.1 All formal concepts of a formal context 79

ba −→ ba −→ ba
  HH  H
 J H  J HH
g g
b b JJ bb HH bd
 c
b cb JJ bb HH bd

J
JJ b
3,6

−→ ba −→ ba
  HH   HH
 J H  J H
g g
b c b JJ bb HH bd b c b JJ bb HH bd
 
H H   HH   HH
J H HH J
H J J J
JJ b HH b HH JJ b  JJ  JJ HHH HHH
6,7,8
b3,4
b b
1,2,3  H 6,7,8
b JJ b
5,6 5,6
3,6 H  3,6 H J
H
HH J J  HH
HJJ b JJ b JJ b
H
6 3 H  6

HH 
HH b

−→ ba −→ ba
 J H
HH   H
J HH
 
g g
b b JJ bb HH bd b b JJ bb HH bd
 c  c
hb J 
H HH J bf hb J 

 JH
H   H 
 JH
H  HH J bf

H H
b J JJ b JJ b HH b HH Jb b
J JJ b JJ b HH b HH Jb
 
3,4 1,2,3  H 6,7,8 5,6
J J b2,3  3,6
H J b 6,8 J Jb  H Jb
J  H J   HH
iqb JJ b eb HHJb ibq JJ b eb HHJb
4 3 H 7  6 H  
HH  HH 
HH qb HH qb

Figure 2.2 Intermediate steps in the drawing of Figure 1.3 (p. 29). The last step is to insert
the object names, which also helps verifying whether the diagram is correct.

concepts and represent the newly generated concepts by small circles with their
respective connecting lines. In our example, this procedure is first applied to the
concepts for b and c, then to those for b, c and d, and finally to those for b, c, d
and g, making use of our knowledge about concepts already determined (if nec-
essary one can note down the extent intersections temporarily at the respective
circles in order to remember them for later). If we have drawn the line diagram
for all concepts determined by this stage, we look for the attributes the extents of
which are maximal among the attribute extents not already used. In our exam-
ple, we obtain e, f and h. As above, we represent the attribute concepts and all
new intersections of the concepts now available. In our example, we would have
to go through this procedure one more time, that is for the attribute i. If we fi-
nally have worked our way through all the attributes, the resulting line diagram
80 2 Determination and representation

should be a correct representation of the concept lattice. In order to check this,

we first delete the extents entered provisionally and then attach the object names
(from below) to the concept circles, such that the equivalence γg ≤ µm ⇔ g I m
of the Basic Theorem is satisfied. If this is not possible for every concept, we have
committed errors, and these may easily happen. In our experience, these errors
are easy to correct. Going over the line diagram again to make it more readable
is usually helpful.

The algorithm described above for determining all concepts becomes cum-
bersome for larger contexts, since the created list must be consulted again and
again. There are faster algorithms for the generation of all concept extents, and
we have already given one in Section 0.3: “Next closure” (Algorithm 1, p. 12).
In fact, as a direct consequence of Proposition 12, it had already been stated that
the concept extents form a closure system and that the associated closure opera-
tor is the “extent closure”, i.e., the mapping X 7→ X 00 . Therefore, the algorithm
can be applied, and Algorithm 3 below is identical to Algorithm 1 except for an
adjustment to the symbols. (In the same way, of course, it applies to the closure
system of the concept intents (Next intent), since the intents of K are exactly
the extents of Kd . But it is not necessary to do both, because the intents can di-
rectly be computed from the extents.) To compute all concept extents, start with
the “first extent” A := ∅00 and then invoke Algorithm 3 until ⊥ is returned. The
previous output is used as input A in each such step.

Algorithm 3 Next Extent(A, K)

Input: A formal context K := (G, M, I) with a finite linearly ordered set G and a subset A ⊆ G.
Output: The lectically next concept extent after A if it exists; ⊥, otherwise.

for all g ∈ G in reverse order do

if g ∈ A then
A := A \ {g}
else
B := (A ∪ {g})00
if B \ A contains no element < g then
return B
return ⊥

The algorithm assumes a linear order of the set of objects G. However, this is
arbitrary and can be freely chosen, which occasionally results in additional op-
timization possibilities. For simplicity, we assume G = {1 < 2 < . . . < n}.
We consider the set of all subsets of G to be “in lexicographic order”, which
is defined next. A subset A ⊆ G is called lectically smaller than a subset B 6= A,
if the smallest element distinguishing A and B from each other belongs to B.
Formally:

A<B :⇔ ∃i∈B\A A ∩ {1, 2, . . . , i − 1} = B ∩ {1, 2, . . . , i − 1}.

2.1 All formal concepts of a formal context 81

This defines a linear strict order for the powerset P(G), i.e., for subsets A 6= B
always A < B or B < A holds. The algorithm finds for an arbitrary given set
A ⊆ G the extent that is smallest after A with respect to this lectic order. The
lectically smallest concept extent is ∅00 . The other extents are found incrementally
by determining the one which is lectically next to the last extent found. In the end,
we obtain the lectically largest extent, namely G. Figure 2.3 shows the result in
the case of the “living beings and water” context.
To make this precise, we define for A, B ⊆ G, i ∈ G,
A <i B :⇔ i ∈ B \ A and A ∩ {1, 2, . . . , i − 1} = B ∩ {1, 2, . . . , i − 1}.
A ⊕ i := ((A ∩ {1, 2, . . . , i − 1}) ∪ {i})00 .
It is easy to verify the following statements:
(1) A < B ⇔ A <i B for one i ∈ G.
(2) A <i B and A <j C with i < j ⇒ C <i B.
(3) i 6∈ A ⇒ A < A ⊕ i.
(4) A <i B and B extent ⇒ A ⊕ i ⊆ B, thus A ⊕ i ≤ B.
(5) A <i B and B extent ⇒ A <i A ⊕ i.
Theorem 5 The smallest concept extent larger than a given set A ⊂ G (with respect to
the lectic order) is
A ⊕ i,
i being the largest element of G with A <i A ⊕ i.
Proof Let A+ be the smallest extent after A with respect to the lectic order. On
account of A < A+ , we get A <i A+ for some i ∈ G by (1) and thus A <i A⊕i by
(5). By (4) it follows that A ⊕ i ≤ A+ , i.e., A ⊕ i = A+ because of A < A ⊕ i. The
fact that i is the largest element with A <i A⊕i results from (2), since A <j A⊕j
with j 6= i on account of A ⊕ i = A+ < A ⊕ j by (2) yields j < i. 

No. 1 2 3 4 5 6 7 8 i No. 1 2 3 4 5 6 7 8 i No. 1 2 3 4 5 6 7 8 i

1 ∅ 7 8 ×××× 4 15 × ×× 1
2 × 6 9 × 3 16 × × × 5
3 × 8 10 × 6 17 × × × ×× 4
4 × × 7 11 × × 4 18 × × ×× 5
5 ××× 5 12 ×× 6 19 × × ××××××
6 ×× 8 13 ×× ××× 2
7 ×× × 7 14 ×× 4

Figure 2.3 List of the extents for the context in Figure 1.1 in lectic order. Behind each
extent A, the element i with A+ = A ⊕ i is stated.

Note that Algorithm 3 implements exactly what Theorem 5 specifies, as does

Algorithm 1 on p. 12.
82 2 Determination and representation

It is not difficult to modify the Next extent algorithm so that it also generates
the neighborhood relation of the concept lattice, which is needed for the concept
lattice diagram. However, this requires that a list of concepts is created. See [154],
Algorithms 11 and 12.
Because the algorithm is so simple, it allows numerous variations and exten-
sions. We briefly mention an example and then formulate a theorem that extends
Theorem 5. Many other results can be found in [154].
We can take advantage of the fact that the extents are generated in lectic order.
For example, if

C := {1, 2, . . . , c}, D := {c + 1, c + 2, . . . , d} ⊆ G,

then we obtain as the lectic successors of C first all sets which contain C and are
disjoint to D. For arbitrary disjoint subsets C, D ⊆ G, one can find all concept
extents E of (G, M, I) with C ⊆ E, E ∩ D = ∅ by suitably adjusting the linear
order of G.
The premises of the algorithm can be weakened in some respects, which al-
lows generalizations. Without substantial modifications of the proof we obtain
the following theorem.
Theorem 6 If F is a family of extents of the context (G, M, I) with the property

A ∈ F and i ∈ G ⇒ (A ∩ {1, . . . , i − 1})00 ∈ F ,

we obtain for an arbitrary subset A ⊆ G the set A+ which is the lectically next in F –if
it exists– by
A+ = A ⊕ i,

i being the largest element of G for which A <i A ⊕ i and simultaneously A ⊕ i ∈ F .

To gain an understanding of possible applications of this theorem, consider
the following task: How can all partitions of a seven-element set be generated
where none of the classes has more than three elements? We can use the context
for the lattice of equivalence relations from 1.3.(8) (p. 53). The family F of parti-
tions with the property specified is an order ideal and thus satisfies the condition
in Theorem 6. It can therefore be scanned with the modified algorithm.
Note that the theorem also allows only large concept extents to be , i.e. ex-
tents E ⊆ G for which the ratio |E|/|G| exceeds a given parameter (which is
often called Minsupp): The corresponding frequent concept intents E 0 satisfy
the conditions of Theorem 6 and can therefore be generated by a suitably modi-
fied version of the Next intent algorithm. The large extents are generated along
the way.
2.2 Diagrams of concept lattices 83

2.2 Diagrams of concept lattices

The best and most versatile form of representation for a concept lattice is a well
drawn line diagram. It is however tedious to draw such a diagram by hand and
one would wish an automatic generation by means of a computer. Both the avail-
able algorithms and the implementations are gradually getting better and better,
but there is still no program that provides a general satisfactory solution. It is by
no means clear which qualities make up a good diagram. It should be transparent,
easily readable and should facilitate the interpretation of the data represented.
How this can be achieved in each individual case depends however on the aim of
the interpretation and on the structure of the lattice. Simple optimization crite-
ria (minimization of the number of edge crossings, drawing in layers, etc.) often
lead to unsatisfactory results. Nevertheless, automatically generated diagrams
are a great help: they can serve as the starting point for drawing by hand. There-
fore, we will describe simple methods of generating and manipulating line dia-
grams by means of a computer, later we suggest even better procedures with the
aid of the structure theory for concept lattices.
b: not equilateral
a: equilateral

c: isosceles
d: oblique

f: obtuse
e: acute

g: right

T1 : (0, 0), (6, 0), (3, 1) × × × ×

T2 : (0, 0), (1, 0), (0, 1) × × ×
T3 : (0, 0), (4, 0), (1, 2) × × ×
√
T4 : (0, 0), (2, 0), (1, 3) × × × ×
T5 : (0, 0), (5, 0), (2, 1) × × ×
T6 : (0, 0), (2, 0), (1, 3) × × × ×
T7 : (0, 0), (2, 0), (0, 1) × ×

Figure 2.4 A formal context for triangles.

As an illustration, we will use the formal context in Figure 2.4, in which trian-
gles are classified according to properties such as right-angled, equilateral, etc. The
choice of the triangles is not coincidental: the context is the result of an attribute
exploration, a technique to be discussed in the next section. But for the moment
we are only concerned with the question of how to obtain a line diagram for this
context.
We can use a computer program to obtain the concepts of the formal context
and the edges of the line diagram. The successor list is displayed on the right. We
84 2 Determination and representation

can read from it that the context has 18 concepts. These are denoted by the serial
numbers 1,. . . ,18. Behind the colon follow the upper neighbors of each concept.
In the line diagram, an edge must be drawn to each of the upper neighbors, and
those are all edges.
Obviously concept no. 1 is the unit element of
the concept lattice (since it has no upper neigh-
bor) and no. 18 is the zero element (since 18 does
not occur as an upper neighbor). 1: −
2: 1 −
As a graph, the line diagram is already com- 3: 2 −
pletely determined by this list. It can be used to 4: 1 −
sketch a diagram “from bottom to top”: first of all 5: 2 4 −
we draw the smallest element (concept no. 18), 6: 3 5 −
above it the upper neighbors (13, 15, 16, 17), then 7: 1 −
8: 7 −
their upper neighbors (6, 8, 10, 11, 12, 14), and so 9: 2 7 −
on. It is still open how points are arranged on pa- 10 : 9 −
per. This can be done “intuitively” but will then 11 : 3 9 −
require various iterations to develop a satisfactory 12 : 4 7 −
diagram. 13 : 8 12 −
14 : 5 9 12 −
There is however an efficacious method to sup- 15 : 10 14 −
port the generation of a line diagram. This geo- 16 : 6 11 14 −
metrical method is based on first understanding 17 : 6 −
the lattice-theoretic structure through a geometri- 18 : 13 15 16 17 −
cal representation of the concept lattice and then
to find the best possible arrangement for the line
diagram. This means that we draw as an interme-
diate step –by hand or with the aid of a computer– an auxiliary picture, which is
then used to draw the actual line diagram. This auxiliary picture is called the ge-
ometrical diagram. Intuitively, we think of this diagram in the following way: we
imagine that the lattice is realized by means of a three-dimensional line diagram
and look down on the lattice from its highest point, i.e., from the unit element.
From the top, we first see the lower neighbors of the unit element. In the ge-
ometrical diagram they are represented by unconcealed circles into which we
write the names of the respective elements. We continue to draw the geometrical
diagram in accordance with the following rules:
1. An element with exactly one upper neighbor is represented by a circle which
is partly covered by the upper neighbor.
2. An element with exactly two upper neighbors is represented by a connecting
line segment between the two upper neighbors. The name of the element is
written into a circle which is partly covered by this connecting line.
3. An element with exactly three upper neighbors is represented by a connecting
triangle between the upper neighbors. The name of the element is written into
the triangle.
2.2 Diagrams of concept lattices 85

Elements with n > 3 upper neighbors are represented analogously by an n-

simplex connecting the upper neighbors. The largest and the smallest element
of the lattice are omitted.

Figure 2.5 A geometrical diagram.

In this way we obtain the geometrical diagram in Figure 2.5. The individual
steps are noted down in the following table. The necessary information has been
taken from the above successor list.
2 lies immediately below 1: therefore a circle for 2.
3 immediately below 2: therefore a circle for 3, partly covered by the 2-circle.
4 immediately below 1: therefore a circle for 4.
5 immediately below 2 and 4: therefore a line segment for 5 between the 2-circle and the
4-circle.
6 immediately below 3 and 5: therefore a line segment for 6 between the 3-circle and the
5-line-segment.
7 immediately below 1: therefore a circle for 7.
8 immediately below 7: therefore a circle for 8, partly covered by the 7-circle.
9 immediately below 2 and 7: therefore a line segment for 9 between the 2-circle and the
7-circle.
10 immediately below 9: therefore a circle for 10, partly covered by the 9-circle.
11 immediately below 3 and 9: therefore a line segment for 11 between the 3-circle and the
9-line-segment.
12 immediately below 4 and 7: therefore a line segment for 12 between the 4-circle and the
7-circle.
13 immediately below 8 and 12: therefore a line segment for 13 between the 8-circle and the
12-line-segment.
14 immediately below 5, 9 and 12: therefore a triangle for 14 between the 5-circle, the 9-
circle and the 12-line-segment.
86 2 Determination and representation

15 immediately below 10 and 14: therefore a line segment for 15 between the 10-circle and
the 14-triangle.
16 immediately below 6, 11 and 14: therefore a triangle for 16 between the 6-line-segment,
the 11-line-segment and the 14-triangle.
17 immediately below 6: therefore a circle for 17, partly covered by the 6-line-segment.

d1
2d d4@ d7
@

d 5@ d9@ d12@
@ @ @
3d
@ @ @ @ @
6 d @ d11@ d14 @d @ @d
10 8
@ @ @
17 d
@d @d @d
16 15 13
@ B 
@ B
@ 
B 
@ B
@B 
@B d
18

Figure 2.6 A line diagram for the lattice of triangle concepts.

It still remains to be said how a good line diagram can be obtained from the
geometrical diagram. The derived line diagram for the concept lattice of the tri-
angles is presented in Figure 2.6. If one already has some experience with the
geometrical method, one can see from Figure 2.5 that the most striking substruc-
ture of the lattice consists of two Boolean cubes. But even without this experience,
one can soon reach this conclusion by proceeding systematically. As a rule, one
should start with the lower neighbors of the unit element being represented by
unconcealed circles. In Figure 2.5 these are the 2-, 4- and 7-circle. These circles are
connected pairwise by the line segments 5, 9 and 12, which in turn are connected
by the 14-triangle. This shows that the concepts 1, 2, 4, 5, 7, 9, 12 and 14 form a
Boolean sublattice. The question is how these eight elements can best be arranged
within the line diagram. After having drawn the unit element, it seems advisable
to put the co-atoms 2 and 7 on the outside and 4 between them, since below both
2 and 7 there is another “point”, which needs some space. The concepts 5, 7, 9
and 12 will be best placed in accordance with the rule of parallelograms, which says
that one should (if possible) place an element in a way that it makes up a paral-
lelogram together with three elements already represented and their connecting
line segments. The resulting picture of the Boolean sublattice represents a cube
standing on one of its corners. After the explanation given so far, it should not
be difficult to recognize the second Boolean sublattice, consisting of the concepts
2, 3, 5, 6, 9, 11, 14 and 16. Since the cube representing it shares the elements 2,
2.2 Diagrams of concept lattices 87

5, 9 and 14 with the first cube, it is obvious how to continue the drawing. How-
ever, one further rule should be observed, the so-called rule of lines, according to
which a line to a new “point” should be arranged in such a way that it continues
some line segments already drawn. If we observe the rule of lines and the rule of
parallelograms for the remaining elements 8, 10, 13, 15 and 17, we obtain from
the geometrical diagram a satisfactory line diagram, to which we only have to
add the zero element (no. 18) (cf. Figure 2.6). For the labelling with object and
attribute names, additional information is required, which a computer program
could supply by means of an assignment list (see Fig. 2.7).

Concept : Object Concept : Attribute

8 : T7 2 : oblique
10 : T5 3 : acute
11 : T3 4 : isosceles
13 : T2 7 : not equilateral
15 : T1 8 : right
16 : T6 10 : obtuse
17 : T4 17 : equilateral

Figure 2.7 The assignment to the concepts.

d isosceles
oblique
d d%@ dnot equilateral
@ %

acute
d d @ d @ d @ obtuse
@ @ @

d @ d @ d @ d @ @ dright
@ @ @ @ @

equilateral @ @ @
d @d @d @d
@ B 
@ B
@ 
B 
@ B
@B 
@B d

Figure 2.8 The labelled line diagram.

In general, it is advisable to draw the geometrical diagram as quickly as pos-

sible using the successor list. When doing so, one should not be afraid to draw
segments of lines and surfaces rather boldly. Experience shows that this kind of
88 2 Determination and representation

diagrams can still be used as instructions for drawing good line diagrams. It is
helpful to observe geometrical patterns and their respective realizations in the
line diagrams. In some (relatively rare) cases, it is advisable to construct the line
diagram from bottom to top; in this case one should use the so-called predecessor
list.
Both of the procedures described above make use of the computer in order
to obtain information necessary for a diagram. We will now explain a method
where a computer generates a diagram and offers the possibility of improving it
interactively. Programming details are irrelevant in this context. We will there-
fore only give a positioning rule which assigns points in the plane to the ele-
ments of a given ordered set (P, ≤). If a and b are elements of P with a < b,
the point assigned to a must be lower than the point assigned to b (i.e., it must
have a smaller y-coordinate). This is guaranteed by our method. We will leave
the computation of the edges and the checking for undesired coincidences to the
programming. We do not even guarantee that our positioning is injective (which
of course is necessary for a correct line diagram). This must also be checked if
necessary.
Definition 39 A set representation of an ordered set (P, ≤) is an order embed-
ding of (P, ≤) in the powerset of a set X, i.e., a map

rep : P → P(X)

with the property

x ≤ y ⇔ rep x ⊆ rep y. ♦
An example of a set representation for an arbitrary ordered set (P, ≤) is the
assignment
X := P, a 7→ (a].
In the case of a concept lattice

X := G, (A, B) 7→ A

resp. X := M, (A, B) 7→ M \ B
are representations which can be combined to

X := G ∪˙ M, (A, B) 7→ A ∪ (M \ B).

It is sufficient to limit oneself to the irreducible objects and attributes.1

For an additive line diagram of an ordered set (P, ≤) we need a set represen-
tation rep : P → P(X) as well as a grid projection

vec : X → R2 ,

assigning a real vector with a positive y-coordinate to each element of X. By

1 For set representation see also Section 6.4.
2.2 Diagrams of concept lattices 89
X
pos p := n + vec x
x∈rep p

we obtain a positioning of the elements of P in the plane. Here, n is a vector which

can be chosen arbitrarily in order to shift the entire diagram. By only allowing
positive y-coordinates for the grid projection we make sure that no element p is
positioned below an element q with q < p.

large small
natural artificial

flowing stagnant

river brook canal ditch lake slough pond basin

Figure 2.9 An additive line diagram of the concept lattice of a lexical field “waters”. The set
representation is based on the irreducible attributes, i.e. the positioning of the attribute
concepts determines that of all remaining concepts. If we interpret the line segments be-
tween the unit element and the attribute concepts as vectors, we obtain the position of an
arbitrary concept by the sum of the vectors belonging to attributes of its concept intent
starting from the unit element. Other diagrams for the same lattice can be found in Figure
2.10.

Every finite line diagram can be interpreted as an additive diagram with re-
spect to an appropriate set representation. For concept lattices we usually use
the representation by means of the irreducible objects and/or attributes. The re-
sulting diagrams are characterized by a great number of parallel edges, which
90 2 Determination and representation

improves their readability. Besides, it is particularly easy to manipulate these di-

agrams.
If we change –the set representation being fixed– the grid projection for an
element x ∈ X, this means that all images of the order filter {p ∈ P | x ∈ rep p}
are shifted by the same distance and that all other images remain in the same
position. In the case of the set representation by means of the irreducibles these
order filters are precisely principal filters or complements of principal ideals, re-
spectively. This means that we can manipulate the diagram by shifting principal
filters or principal ideals, respectively, and leaving all other elements in position.
Experience shows that the set representation by means of the irreducible at-
tributes is most likely to result in an easily interpretable diagram.
Occasionally, it can be convenient to represent a lattice as a part of a larger
order. For this purpose, we draw a line diagram of the order but represent only
those elements of the lattice by small circles which we actually mean. An example
is shown in Figure 5.1.
Even carefully constructed line diagrams loose their readability from a certain
size up, as a rule from around 50 elements up. One gets considerably further with
the nested line diagrams which will be introduced next. However, these diagrams
do not only serve to represent larger concept lattices. They offer the possibility
to visualize how the concept lattice changes if we add further attributes.
The basic idea of the nested line diagram consists of delimiting parts of an
ordinary diagram and replacing bundles of parallel lines between these parts by
one line each. Thus, a nested line diagram consists of framed boxes, which con-
tain parts of the ordinary line diagram and which can be connected by lines. In
the simplest case two boxes which are connected by a simple line are congruent.
Here, the line indicates that circles which coincide if one box is put on top of
the other are connected in the ordinary line diagram. A double line between two
boxes means that each element of the upper box is larger than each element of
the lower box. Figure 2.10 shows the concept lattice from Figure 2.9, once as an
ordinary line diagram and once as a nested diagram. For reasons of comprehen-
sibility we have left out the object and attribute names.
Furthermore, we allow that two boxes connected by a single line do not nec-
essarily have to be congruent, but they may each contain a part of two congruent
figures. In this case, the two congruent figures are drawn in the boxes as a “back-
ground structure”, but the elements are only marked by circles if they are part
of the respective substructures. The line connecting the two boxes then indicates
that the respective pairs of elements of the background shall be connected with
each other.
Nested line diagrams originate from partitions of the set of attributes. The
basis is the following Theorem:

Theorem 7 Let (G, M, I) be a formal context and M = M1 ∪ M2 . The map

(A, B) 7→ (((B ∩ M1 )0 , B ∩ M1 ) , ((B ∩ M2 )0 , B ∩ M2 ))

is a -preserving order embedding of B(G, M, I) into the direct product of

W
2.2 Diagrams of concept lattices 91

Figure 2.10 Line diagram and nested line diagram. Less trivial examples of nested line
diagrams can be found in Figures 1.15 on p. 50 and 2.17 on p. 103.

B(G, M1 , I ∩ G × M1 ) and B(G, M2 , I ∩ G × M2 ).

The component maps

(A, B) 7→ ((B ∩ Mi )0 , B ∩ Mi )
are surjective on B(G, Mi , I ∩ G × Mi ).

Proof If (A, B) is a concept of (G, M, I), then B ∩ Mi is the set of all attributes
common to the objects of A in the context (G, Mi , I ∩ G × Mi ), i.e., it is an intent
of this context. Hence the above-mentioned assignment is really a map into the
product. The union of the intents B ∩ M1 and B ∩ M2 again yields B, i.e., the
92 2 Determination and representation

map is injective. The fact that it is furthermore -preserving (and thus an order-
W
embedding) can again be seen from the concept intents. It remains to be shown
that the component maps are surjective. Let C be an intent of (G, Mi , I ∩G×Mi ).
Then B := C II is an intent of (G, M, I) with B ∩ Mi = C, i.e., the image of the
concept (B 0 , B) of (G, M, I) under the ith component map is the concept with
the intent C. 
In order to sketch a nested line diagram, we proceed as follows: First of all we
split up the attribute set: M = M1 ∪ M2 . This splitting up does not have to be
disjoint. More important for interpretation purposes is the idea that the sets Mi
bear meaning. Now, we draw line diagrams of the subcontexts Ki := (G, Mi , I ∩
G × Mi ), i ∈ {1, 2} and label them with the names of the objects and attributes,
as usual. Then we sketch a nested diagram of the product of the concept lattices
B(Ki ) as an auxiliary structure. For this purpose we draw a large copy of the
diagram of B(K1 ), representing the lattice elements not by small circles but by
congruent rectangular boxes, each of which contains a diagram of B(K2 ).
WBy Theorem 7 the concept lattice B(G, M, I) is embedded in this product as
a -semilattice. If a list of the elements of B(G, M, I) is available, we can enter
them into the product according to their intents. If not, we enter the object con-
cepts the intents of which can be read off directly from the context, and form all
suprema.
This at the same time provides us with a further, quite practicable method of
determining a concept lattice by hand: split up the attribute set as appropriate,
determine the (small) concept lattices of the subcontexts, draw their product
in form of a nested line diagram, enter the object concepts and close it against
suprema. This method is particularly advisable in order to arrive at a useful di-
agram quickly.

2.3 Implications

A fictitious example will serve as an introduction to the problem: imagine a man-

ufacturer of computer hardware, whose products can be combined in various
ways, but not arbitrarily. In order to structure the (reasonable) configurations
conceptually, we would have to examine a formal context, the objects of which
are the combinations and whose attributes are the components. If a list of these
combinations is not available, we have to create it, for which we can use our
knowledge of the rules of combinability. The starting point of concept analysis
in this case is not an explicitly given formal context. Rather, we derive the context
and at the same time the concept system from attribute logic, i.e., from the rules
governing the combinations of attributes.
This method is useful not only in the example mentioned above. It is often nec-
essary to classify a large number of objects in terms of a relatively small number
of attributes, and then it is often impractical to write down the entire formal con-
text and apply the procedures for determining the concept system that have been
2.3 Implications 93

described in the previous section. In such cases, the conceptual structure can be
inferred from the implications between the attributes, by which we mean statements
of the following kind:
“Every object with the attributes a, b, c, . . . also has the attributes x, y, z, . . .”
Formally, an implication between attributes (over M ) is a pair of subsets of the
attribute set M . It is denoted2 by A → B. The set A is the premise and B is the
conclusion.
In this section, we examine the set of all attribute implications that hold in
a given formal context. It turns out that these implications contain enough in-
formation to reconstruct the structure of the concept lattice. Conversely, the at-
tribute implications of a formal context can also be read from the concept lattice.
However, the system of all implications between attributes holding in a given
context tends to be large and to contain many trivial implications. Therefore, we
look for subsystems which suffice to describe the concept lattice. We start with
some simple definitions:
Definition 40 A subset T ⊆ M respects an implication A → B if A 6⊆ T or
B ⊆ T . T respects a set L of implications if T respects every single implication
in L. A → B holds in a set {T1 , T2 , . . .} of subsets if each of the subsets Ti respects
the implication A → B. A → B holds in a formal context (G, M, I) if it holds in
the system of object intents. In this case, we also say that A → B is an implication
of the context (G, M, I) or, equivalently, that within the context (G, M, I), A is a
premise of B. ♦
Proposition 22 An implication A → B holds in (G, M, I) if and only if B ⊆ A00 . It
then automatically holds in the set of all concept intents as well. 
How can we read off an implication from a concept lattice diagram? It is suf-
ficient to describe this for implications of the form A → m, since A → B holds
if and only if A → m holds for eachVm ∈ B. Now A → m holds if and only
if (m0 , m00 ) ≥ (A0 , A00 ), i.e., if µm ≥ {µn | n ∈ A}. This means that we have
to check in the diagram whether the concept labeled by m is located above the
infimum of all concepts labeled by an n from A.
It can occasionally be useful for the determination of the implications to re-
place the original context (G, M, I) by its complementary context (G, M, (G ×
M ) \ I), in particular if the latter has considerably fewer concepts as (G, M, I).
For m ∈ M and T A 0⊆ M the following S hold:
equivalences m ∈ A00 ⇔ {m} ⊆ A00
⇔ A ⊆ m ⇔ {n | n ∈ A} ⊆ m ⇔ G \ m ⊆ {G \ n | n ∈ A}. Thus, A → m
0 0 0 0 0

holds in the context (G, M, I) if and only if in the complementary context every
object with the attribute m has at least one attribute n from A.
Proposition 23 If L is a set of implications over M ,

Mod(L) := {X ⊆ M | X respects L}
2 When the sets are small, we shall omit brackets (as we have done earlier), i.e., we shall write
A → m instead of A → {m}, etc.
94 2 Determination and representation

is a closure system on M . If L is the set of all implications of a formal context, Mod(L)

is the system of all concept intents. 
The proof is trivial. The respective closure operator can be described as follows:
For a set X ⊆ M , let
[
X L := X ∪ {B | A → B ∈ L, A ⊆ X}.

We form the sets X L , X LL , X LLL , . . . until we finally obtain a set L(X) := X L...L
with L(X)L = L(X) (in the case of infinite contexts it can be necessary to con-
tinue this process transfinitely). L(X) then is the closure of X with respect to the
closure system Mod(L) which we have been looking for.
By means of the closure system Mod(L) it is also possible to construct a con-
text for every given set L of implications, the intents of which are precisely the
sets respecting L: (Mod(L), M, 3) has this property. In addition to L in this con-
text all implications hold which follow from L in the sense of the following def-
inition:
Definition 41 An implication A → B follows (semantically) from a set L of
implications between attributes if each subset of M respecting L also respects
A → B. A family of implications L is called closed if every implication following
from L is already contained in L.
A set L of implications of a formal context (G, M, I) is called complete if every
implication of (G, M, I) follows from L.3 ♦
In other words: An implication follows semantically from L if it holds in every
system of sets in which L holds as well. This is the case if and only if Mod(L) =
Mod(L ∪ {A → B}). More about this in the following section.
There is also a syntactic characterization of closed implication sets known as
the Armstrong rules [5]:

Proposition 24 A set L of implications on M is closed if and only if the following con-

ditions are satisfied for all W, X, Y, Z ⊆ M :
1. X → X ∈ L,
2. If X → Y ∈ L, then X ∪ Z → Y ∈ L,
3. if X → Y ∈ L and Y ∪ Z → W ∈ L, then S X ∪ Z → W ∈ L, and
4. if X → Yt ∈ L for all t ∈ T , then X → t∈T Yt ∈ L.


To efficiently decide whether an implication A → B follows from a list L of

implications, one checks whether

B ⊆ L(A)

holds, using a fast algorithm for computing L(A), e.g., Algorithm 15 in [154].

3 Note that a complete set must always consist of implications which hold in (G, M, I).
2.3 Implications 95

It should be noted that sometimes the inference rules can change, namely
when additional “background knowledge” about the attributes is known. A typ-
ical instance is that one attribute is known to be the “negation” of another. This
will be discussed in Section 8.1.

Implications with proper premise

In order to demonstrate that a set L of implications of a context is complete, we

have to show that every subset T ⊆ M respecting L is an intent.
A first attempt to find a manageable complete set of implications consists in
leaving out those implications which follow trivially from others or those which
hold in any context. For instance, A → B holds whenever B ⊆ A, and from A →
B and C ⊆ B it always followsS that A → SC. Correspondingly, from Aj → Bj for
j ∈ J it always follows that j∈J Aj → j∈J Bj . If we eliminate the implications
arising like that, there remain certain implications with a proper premise:
Definition 42 For an attribute set A ⊆ M of a formal context (G, M, I) we denote
by [
A• := A00 \ (A ∪ (A \ {n})00 )
n∈A

the set of those attributes contained in A00 but not in A or in the closure of any
proper subset of A. We call A a proper premise if A• 6= ∅, i.e., if
[
A00 6= A ∪ (A \ {n})00 .
n∈A

In particular, ∅ is a proper premise if ∅00 6= ∅. ♦

Proposition 25 If T is a finite subset of M , then

[
T 00 = T ∪ {A• | A is a proper premise with A ⊆ T }.

The set of all implications of the form

A → A• , where A is a proper premise,

of a formal context with a finite attribute set is complete.

Proof If T = T 00 the assertion is trivial, thus let m ∈ T 00 \T . A subset A of T which

is minimal with respect to the property m ∈ A00 has to be a proper premise, i.e.,
there is an implication A → A• with m ∈ A• . Since m had been chosen arbitrarily,
the first assertion follows. If T respects all implications of the form A → A• and
A is a proper premise, from what we have just proved it follows that T 00 = T , i.e.,
that T is an intent. 
96 2 Determination and representation

In certain respects, the set of proper premises is canonical with respect to the
property of being iteration-free as described in Proposition 25. In order to state
this more precisely, we first introduce a further term. A family of implications
can be simplified by merging implications with the same premise. We call a fam-
ily of implications contracted if there are no premises which occur more than
once. If L is any contracted family of implications satisfying the condition of the
proposition, i.e., with
[
T 00 = T ∪ {B | A → B ∈ L, A ⊆ T } for all T ⊆ M,

then L contains an implication E → F with E • ⊆ F for every proper premise E,

as can be seen easily, if we replace T by E in the condition.
In order to determine the proper premises of a doubly founded context
(G, M, I) we can use the arrow relation .. Following Definition 40 we call an
attribute set P a proper premise of an attribute m if P is a proper premise and
m ∈ P • holds.
Proposition 26 P is a premise of m if and only if

(M \ g 0 ) ∩ P 6= ∅

holds for all g ∈ G with g . m. P is a proper premise for m if and only if m 6∈ P and P
is minimal with respect to the property that (M \ g 0 ) ∩ P 6= ∅ holds for all g ∈ G with
g . m.
Proof For g ∈ G and P ⊆ M we have the equivalences

(M \ g 0 ) ∩ P 6= ∅ ⇔ P 6⊆ g 0 ⇔ g 6∈ P 0 .

Since P 0 6⊆ m0 is equivalent to the fact that there is an object g ∈ P 0 with g . m,

the first assertion follows. The property of minimality of the proper premises
yields the second assertion. 
According to the proposition, we obtain the proper premises by determining
for every attribute m the minimal attribute sets P with (M \ g 0 ) ∩ P 6= ∅ for all
g . m.

Pseudo-intents and the canonical basis

Even the set of implications described in Proposition 25 is in general still redun-

dant.
Definition 43 A set L of implications of a context is called non-redundant if
none of the implications follows from the others. ♦
Guigues and Duquenne [179] have shown that for every formal context with
a finite attribute set M there is a natural complete and non-redundant set of im-
2.3 Implications 97

plications. For the following results, we make the general assumption that the
attribute set M which occurs is finite. This permits a recursive definition of the
basic notion of a pseudo-intent (which takes the place of the proper premise):
Definition 44 P ⊆ M is called a pseudo-intent of (G, M, I) if and only if P 6= P 00
and Q00 ⊆ P holds for every pseudo-intent Q ⊆ P , Q 6= P . ♦

Theorem 8 ([179]) The set of implications

L := {P → P 00 | P pseudo-intent}

is non-redundant and complete.

Proof Evidently, L holds in (G, M, I). In order to show that L is complete, we

again have to show that every set T ⊆ M respecting L is an intent. Each such set
in particular respects all implications Q → Q00 where Q is a pseudo-intent and
Q ⊆ T . If we assume that T 6= T 00 , T itself satisfies the definition of a pseudo-
intent and the implication T → T 00 is in L but is not respected by T , a contradic-
tion.
In order to show that L is non-redundant, we consider an arbitrary pseudo-
intent P and show that P respects the set L \ {P → P 00 }. In fact, if Q → Q00 is an
implication in L \ {P → P 00 } with Q ⊆ P , then Q00 ⊆ P must hold, since P is a
pseudo-intent. 
In practice, the implications are not stated in the form P → P 00 but in the form
P → (P 00 \ P ). We call this the canonical basis4 of the attribute implications.
In the case of the developing countries context (Figure 1.8 on p. 35) this basis
consists of four implications (see Figure 2.11).

OPEC → Group 77, Non-Aligned

LDC, SIDS → OACPS
Group 77, Non-Aligned, LDC, OPEC → OACPS
Group 77, Non-Aligned, SIDS, OPEC → LDC, OACPS

Figure 2.11 Canonical basis for the context of developing countries.

The canonical basis is indeed canonical with respect to the properties stated.
For this we start with a simple proposition:
Proposition 27 If P and Q are intents or pseudo-intents with P 6⊆ Q and Q 6⊆ P , then
P ∩ Q is an intent.
Proof P as well as Q and thus also P ∩ Q respect all implications in L with the
possible exception of P → P 00 and Q → Q00 . If P 6= P ∩ Q 6= Q, then P ∩ Q also
respects these implications, i.e., it is an intent. 
4 Other names are “Duquenne-Guigues basis” or “stem base”.
98 2 Determination and representation

The next proposition shows among other things that there can be no complete
set which contains fewer implications than there are pseudo-intents:
Proposition 28 Every complete set Σ of implications contains, for each pseudo-intent
P , an implication A → B with A ⊆ P and A00 = P 00 .
Proof A pseudo-intent P is always not equal P 00 . Therefore, provided that Σ is
complete, there must be at least one implication A → B in Σ which leads out of
P , i.e., with A ⊆ P and B 6⊆ P . On account of B ⊆ A00 , we get A00 6⊆ P , and thus
A00 ∩ P cannot be a concept intent. By Proposition 27 this yields P ⊆ A00 and thus
P 00 = A00 . 
The recursive definition of the pseudo-intents provides us with a first, al-
though inefficient, algorithm for generating them. In the following we will de-
velop a more practicable procedure. As an immediate consequence of Proposi-
tion 27 we obtain:

Proposition 29 The set of all subsets of M which are intents or pseudo-intents of

(G, M, I) is a closure system. 

The closure operator for this closure system is obtained by a modification of

the operator L. Starting from a set X, we successively form
∗ [
X L := X ∪ {B | A → B ∈ L, A ⊆ X, A 6= X}
∗
L∗ ∗ ∗ ∗
[
XL := X L ∪ {B | A → B ∈ L, A ⊆ X L , A 6= X L }
∗
and so on, until we finally obtain a set L∗ (X) with L∗ (X) = L∗ (X)L . This set
is the pseudo-intent or intent which we have been looking for. We should bear
in mind that when using this method for generating a pseudo-intent P , we only
need implications A → B, the premise of which is a proper subset of P . This
permits a recursive generation of the pseudo-intents by means of the next intent
algorithm described in Section 2.1. Algorithm 4 below is from [154]. It computes
the canonical basis L. The premises of the implications in L are precisely the
pseudo-intents.
Remark: According to Proposition 28, no complete list can contain fewer im-
plications than the canonical basis. Nevertheless, even this basis can occasionally
still be simplified and thereby made more readable. Details can be found in [145],
Figure 2.16 shows an example.

Attribute exploration

We now return to the initial question of this section: how can we determine the
concept intents using the implications? We have seen that we do not need all
implications for this, but that a small subset of them is sufficient. So far, we have
2.3 Implications 99

Algorithm 4 Canonical Basis(M , 00 )

Input: A closure operator X 7→ X 00 on a finite set M , for example, given by a formal context
(G, M, I).
Output: The canonical basis for the closure operator.

A := ∅
L := ∅
while A 6= M do
if A 6= A00 then
L := L ∪ {A → A00 }
A := Next Closure(A, M, L)
return L

only explained how to obtain these implications from an existing formal context.
However, using the tools now at our disposal, we can also develop a method to
develop implication lists that are free of redundancy, even if the formal context
is only partially available.
This procedure, called attribute exploration, has been proven in many appli-
cations. It is documented in a separate monograph [154]. In practice, a computer
is used to manage the implication sets and to calculate which information is still
missing. The implications are interactively determined, i.e. in cooperation with
the user.
The algorithm for determining the pseudo-intents permits a modification re-
sulting in an interactive program: it is possible to modify the formal context by
adding new objects, even while the generation of the list L of the implications is
in progress. If the intents of these objects respect all implications determined so
far, the computation for the new context can be continued with the results so far
obtained. This is the content of the following proposition:
Proposition 30 Let K be a formal context and let P1 , P2 , . . . , Pn be the first n pseudo-
intents of K with respect to the lectic order. If K is extended by an object g the object
intent g 0 of which respects the implications Pi → Pi00 , i ∈ {1, . . . , n}, then P1 , P2 , . . . ,
Pn are also the lectically first n pseudo-intents of the extended context.
This can be proved for example by induction on n. 
Thus, if we have found a new pseudo-intent P , we can stop the algorithm
and ask, whether the implication P → P 00 should be added to L. The user can
answer this question in the affirmative or add a counter-example, which must not
contradict the implications which were confirmed so far. In the extreme case, the
procedure can be started with a context the object set of which is empty. In this
case, the user will have to enter all counter-examples, thereby creating a concept
system with a given “attribute logic”.
A detailed description of the procedure can be found in [154]. We limit our-
selves here to an example. From a book on measurement theory [321] we take
a list of properties of binary relations, which are used there in order to define
different types of relations, see Figure 2.12.
100 2 Determination and representation

Property Definition
r reflexive xRx for all x ∈ S
i irreflexive ¬xRx for all x ∈ S
s symmetric xRy ⇒ yRx for all x, y ∈ S
as asymmetric xRy ⇒ ¬yRx for all x, y ∈ S
an antisymmetric xRy and yRx ⇒ x = y for all x, y ∈ S
t transitive xRy and yRz ⇒ xRz for all x, y, z ∈ S
nt negatively transitive ¬xRy and ¬yRz ⇒ ¬xRz for all x, y, z ∈ S
c connex xRy or yRx for all x 6= y ∈ S
sc strictly connex xRy or yRx for all x, y ∈ S

Figure 2.12 Properties of binary relations.

Which implications exist between those properties? For every such implica-
tion it is easy to determine whether it holds for all binary relations, and there can
only be finitely many (since we are dealing with a finite number of attributes),
but very likely there are more than we would like to list. Our algorithm should
allow us to find only “good” implications right away. Implications which are not
valid for all binary relations have to be refuted by stating counterexamples.
First, we equip ourselves with a small sample set, for which we take all re-
lations on a one- or two-element set. Up to isomorphism there are twelve such
relations (Figure 2.13).

S R r i s as an t nt c sc
{0} ∅ × × × × × × ×
{0} {(0, 0)} × × × × × × ×
{0, 1} ∅ × × × × × ×
{0, 1} {(0, 0)} × × × ←
{0, 1} {(0, 0), (1, 1)} × × × ×
{0, 1} {(0, 0), (0, 1)} × × × × ←
{0, 1} {(0, 0), (1, 0)} × × × × ←
{0, 1} S × S \ {(0, 1)} × × × × × ×
{0, 1} S × S \ {(0, 0)} × × × ←
{0, 1} {(0, 1)} × × × × × ×
{0, 1} {(0, 1), (1, 0)} ×× × ×
{0, 1} S×S × × × × × ×

Figure 2.13 Examples of binary relations.

Now we have a context to start with (generally, this context can even be
empty). Of course, only implications which hold in this context can hold for all
binary relations, but not vice versa. Please note that the four objects marked by
2.3 Implications 101

No. S R
1 {0} ∅
2 {0} {(0, 0)}
3 {0, 1} ∅
4 {0, 1} {(0, 0), (1, 1)}
5 {0, 1} S × S \ {(0, 1)}
6 {0, 1} {(0, 1)}
7 {0, 1} {(0, 1), (1, 0)}
8 {0, 1} S×S
9 {0, 1, 2} S × S \ {(0, 1), (1, 2), (2, 0)}
10 {0, 1, 2} {(0, 1), (1, 2), (2, 0)}
11 {0, 1, 2} {(0, 1)}
12 {0, 1, 2} {(0, 1), (1, 0)}
13 {0, 1, 2} S × S \ {(0, 1)}
14 {0, 1, 2} S × S \ {(0, 1), (1, 0)}

Figure 2.14 A complete list of examples.

a ← are superfluous, since their intents are the intersections of other object in-
tents and therefore respect all implications respected by the other objects. We
will leave them out in the following. Now we use the algorithm in order to cal-
culate the first pseudo-intent. The lectically smallest pseudo-intent in this context
is {sc}, with {sc}00 = {r, t, nt, c, sc}. In other words, the implication

{sc} → {r, t, nt, c, sc}

holds in all examples stated so far. Does it hold for binary relations in gen-
eral? Of course not. A counter-example is for instance S = {0, 1, 2}, R =
S × S \ {(0, 1), (1, 2), (2, 0)}. This relation is reflexive, antisymmetric, connex
and strongly connex and has none of the other properties. We add this exam-
ple to our formal context and again ask for the smallest pseudo-intent. It is still
{sc}, but now {sc}00 = {r, c, sc}, and we have to check whether the implication
{sc} → {r, c, sc}, which we abbreviate by

sc → r, c,

holds for all binary relations. As a matter of fact, every strongly connex relation
is reflexive and connex. Therefore, we can add this implication to the list of im-
plications L.
The next pseudo-intent is {t, c} with {t, c}00 = {t, nt, c}. This suggests the im-
plication
t, c → nt,
which in fact holds for binary relations and therefore is added to the list, as well
as the following one
102 2 Determination and representation

r i s as an t nt c sc
1 × × × × × × ×
2 × × × × × × ×
3 × × × × × ×
4 × × × ×
5 × × × × × ×
6 × × × × × ×
7 × × × ×
8 × × × × × ×
9 × × × ×
10 × × × ×
11 × × × ×
12 × ×
13 × × × ×
14 × ×

Figure 2.15 The context of the examples.

strictly connex → reflexive, connex

transitive, connex → negatively transitive
antisymmetric, negatively transitive → transitive
asymmetric → irreflexive, antisymmetric
symmetric, connex → negatively transitive
symmetric, antisymmetric → transitive
irreflexive, transitive → asymmetric
irreflexive, antisymmetric → asymmetric
symmetric, asymmetric, transitive → negatively transitive
reflexive, connex → strictly connex
reflexive, negatively transitive → strictly connex
symmetric, negatively transitive, strictly connex → transitive
reflexive, irreflexive → all properties

Figure 2.16 A complete and non-redundant list of implications, given by the canonical
basis of the formal context in Figure 2.15, slightly simplified for better readability.
2.3 Implications 103

symmetric antisymmetric

A  e.Q

A

reflexive `  Q
` eA e

AA e
Q e irreflexive
 Q
Q
Q

A A Q
A
e
A e A
Q e
QA e
14 A

Q Z
Q AZ

A

Q A
12
Q 
Q 
Q
@
transitive negatively transitive @
connex
@
e e e
 Q
 
AQ
Q 
Q
AQ
e e
A e Q  e
A Q e e
A eQ e

AQ 
  Q 
QQ Q Q

A AQ
Q
A
AAQ
Q A A
Q
A
e
A eA e
Q
QA e
A A
Q e
Q
A e A
Q
QA e
Q
A
 Q
 A 
 Q A
11

 Q A
  Q
QA
A 9
QQ A
10
e
AA

QA
A

Q
A
A

Q
A

4 QQ
 Q  Q 
 Q  Q 
Q Q Q
@ @
@ @
@ @
e e

AQ  AQ
 Q
 e
A e Q e
 e
A Q e
 Q

 Q 13  Q Q

A AQ

Q
A AQ
Q

A A e
Q
QA e
A A
Q e
7Q
Q

A 
A
  A
A
  Q QA
A
  Q
QA e
QA

A

A

Q
A

Q  3
Q Q 
Q 
Q Q
@
@
@
e
 Q

e e
A e Q

AQ
 Q Q

A A
Q
A
e
5A eA e
Q
QA e
Q
8 AA

 Q QA
6
e QA e

2 QQ  1
Q e

Figure 2.17 The concept lattice for the context of the binary relations. The numbers refer to the
list of relations in Figure 2.14.
104 2 Determination and representation

an, nt → t,
which results from the pseudo-intent {an, nt} with

{an, nt}00 = {an, t, nt}.

After that, we obtain the pseudo-intent {as}, for which

{as}00 = {i, as, an, t, nt}

holds in the context of the examples. But the implication

as → i, an, t, nt

does not hold generally, as the following example shows: S := {0, 1, 2}, R :=
{(0, 1), (1, 2), (2, 0)}. This relation has the attributes i, as, an, nt, and we add it
to the context. Since it obviously respects all implications accepted so far, it has
no consequences for the pseudo-intents found up to then (cf. Proposition 30).
In the following, we first confirm the implications as → i, an and s, c → nt
as well as s, an → t, then we state a counter-example for i, t → as, an, nt etc.
The complete result is presented in Figures 2.14 to 2.17. We point out that the
premises of the implications in Figure 2.16 are precisely the pseudo-intents of
the context in Figure 2.15.
The procedure does not guarantee that the resulting formal context is reduced
(as in the example). Newly entered objects can make previously entered objects
dispensable. It is possible to “row-reduce” the context during the process (i.e.,
to delete dispensable objects). This has no effect on the implications.

2.4 Clauses and dependencies

How can the theory of implications be applied to the case of many-valued con-
texts? There are several answers to this question.
One is to study one-valued contexts that are derived from a many-valued con-
text through conceptual scaling, as introduced in Section 1.4. Basically, that leads
to implications between the individual attribute values, at least as long as we
keep to plain scaling. A typical such implication might say that
“Downtown real estate is expensive”.
However, the theory needs modification because the conceptual scales may pro-
vide additional background knowledge, resulting in a modified, stronger impli-
cation inference. This will be discussed in the second subsection, after we have
compiled some facts about clauses, which are a useful tool.
In colloquial language the term dependency of many-valued attributes is used
as exemplified by the following sentence
2.4 Clauses and dependencies 105

“The price of a real-estate depends on situation and size”.

This is meant to express a simultaneous dependency of attribute values, perhaps
even a gradual one, in the sense of “the larger the more expensive”. This is what
the third subsection is about.

Clauses

When reading diagrams of concept lattices, it is sometimes irritating that some

concept nodes carry no labels at all. As an example, consider the diagram in
Figure 1.4 on p. 30. The second circle seen from the left is without a label. It
represents the formal concept with the extent {dog, frog} and the concept intent

{needs water to live, has limbs, can move around, lives on land}.

The fact that no object label is attached to this concept means that no object in
the formal context has these four attributes, but no others. This concept intent
consists of the common attributes of dog and frog, but is not represented by a
single object from G. And every single attribute that dog and frog have in com-
mon, they share with another living being from G. Only the attribute combination
is characteristic for dog and frog. If, for example, “lizard” were added as a new
(reducible) object, then the dog-frog-lizard concept would get an object label as
well.
According to Proposition 23 (p. 93), the implications of a formal context K de-
termine the closure system of its concept intents and thus the concept lattice up
to isomorphism. However, they do not determine the formal context, not even up
to isomorphism, since adding reducible objects to K does not change its impli-
cations. The implications do not provide any information about which concept
intents are also object intents. This is only clear for -irreducible concept intents,
W
which must be object intents.
A formal concept without an object label leads to an expression similar to an
implication:
Every object in this concept extent has an attribute
which is not in the intent.
In the example discussed above, this specializes to
Every living being from G that lives on land and can move around
suckles its offspring or also lives in water.
The decisive
V difference to an implication is the word “or”. Such expressions of
the form A → S, called clauses over M , are dealt with in Section 8.1.
W

W 45 A subset T ⊆ M respects a clause A → S if A 6⊆5 T or T ∩S 6= ∅.

V W
Definition
A → S is a clause of K if every object intent of K respects it.
V
♦
5 Other names are “is all-extensional in K” or “holds in K”.
106 2 Determination and representation

Clauses are much more expressive than implications. Theorem 67 (p. 314)
will show that every formal context is determined by its clauses up to clarifica-
tion and isomorphism. But clauses are also much more difficult to handle than
implications, both from an algorithmic point of view and for human intuition.
Nevertheless, clauses can be very useful. Among other things, they provide ad-
ditional options for dealing with implications. We start with a special case.
Suppose you know that two attributes m and n are negations of each other.
On the one hand, this means that no object can have both attributes, but it also
means that every object must have one of them. If the implications A ∪ {m} → s
and A ∪ {n} → s both hold for some A ⊆ M and s ∈ M , then A → s must
also hold, obviously. But that can not be concluded using only the Armstrong
rules (see page 94). If two attributes are negations of each other, this results in
additional inference rules for implications. The same works for every clause.
Proposition 31 The following are equivalent:
1. A → S is a clause of K.
V W
2. The set L of implications of K is closed under the inference rule
If A ⊆ X and X ∪ {s} → B ∈ L for all s ∈ S, then


R∧A→∨S
X → B ∈ L.
Proof The inference rule R∧A→∨S does not hold for the implications of K iff
there is some attribute m ∈ M and some subset X ⊆ M containing A such that
all implications X ∪ {s} → m are in L, but X → m is not. X → m is not in L iff
there is some object g of K such
V that W
g 0 does not respect X → m, i.e., with X ⊆ g 0
and m ∈ / g . But g respects A → S if and only if A ⊆ g 0 enforces s ∈ g 0 for
0 0

some s ∈ S. 

Remark: The inference rules R∧A→∨S are introduced here as a tool of theory.
For the algorithmic implementation of implication inference with background
knowledge, we refer to Section 5.3 of [154].

Theorem 9 Let K := (G, M, I) be a formal context with finite attribute set M . A set L
of implications over M is the set of implications of some subcontext (H, M, I ∩(H ×M ))
if
Vand onlyW if L is closed under the Armstrong rules and under all rules R∧A→∨S , where
A → S is a clause of K.

Proof The clauses of K are inherited by the subcontexts, and, according to

Proposition 31, the associated inference rules are inherited as well. The Arm-
strong rules apply to all contexts. Therefore, one direction of the assertion is clear.
It remains to be shown that every set L of implications over M that is closed un-
der the inference rules mentioned is the set of implications of some subcontext.
Consider an arbitrary subset A ⊆ M . If A is not an object intent of K, we can
choose a (possibly empty) transversal S of

{g 0 \ A | A ⊆ g 0 , g ∈ G}.
2.4 Clauses and dependencies 107

Then A → S is a clause of K and the corresponding inference rule R∧A→∨S

V W
holds for L. For every m ∈ M we therefore have

if m ∈ L(A ∪ {s}) for all s ∈ S, then m ∈ L(A),

which simplifies to \
L(A ∪ {s}) = L(A).
s∈S

For any -irreducible L-closed set this is impossible. The -irreducible L-closed
T T
elements therefore all must be object intents of K. Since M is assumed to be finite,
the system of L-closed sets is generated by its irreducibles, and L is the set of
implications of the subcontext induced by these objects. 

Implications derived from many-valued contexts

Theorem 9 can be applied when examining the implications of formal contexts

derived from many-valued contexts using plain scaling. When applying the at-
tribute exploration method to scaled many-valued contexts, then the algorithm
should not ask any unnecessary questions. Every possibility to answer a ques-
tion automatically must therefore be utilized. Thus if additional inference rules
are available, they should be used. And indeed, Theorem 9 provides such rules.
The good news is that, under realistic conditions, implication inference remains
fast even in the many-valued case [150].
Recall the definition of plain scaling (Def. 36 on p. 62). For a given many-
valued context (G, M, W, I) and suitable scale contexts Sm := (Gm , Mm , Im ),
m ∈ M , a derived context is defined. The derived context then has the object set
G and the attribute set m∈M Ṁm , with g I (m, n) : ⇔ m(g) Im n.
S

Definition 46 (generalizing Definition 31 on p. 43) The semiproduct of formal

contexts Sm := (Gm , Mm , Im ), m ∈ M , is defined as

×
` [

a
Sm = Gm , {m} × Mm , ∇ ,
m∈M m∈M m∈M

with
(ga | a ∈ M ) ∇ (m, n) ⇔ gm Im n.
The formal contexts Sm (Gm , {m} × Mm , Im ) with
• •

•
gm Im (m, n) : ⇔ gm Im n

are called the factors of the semiproduct. ♦

108 2 Determination and representation

Proposition 32 Each context derived from a complete6 many-valued context via plain
scaling is isomorphic to a subcontext of the semiproduct of the scales used.

Proof An embedding is obtained by mapping each object g to the tuple

(m(g) | m ∈ M )

of its values in the many-valued context. 

Combining this proposition with Theorem 9 shows that for derived contexts
additional implication inference rules apply, and that these are determined by
the clauses of the semiproduct of the conceptual WNote that when AV1 ⊆ A2
V scales.
and S 1 ⊆ S 2 , then every subset that respects A 1 → S1 also respects A2 →
S2 , i.e., the second clause follows semantically from the first one. Moreover,
W
the inference rule corresponding to the second clause follows from the one for
the first clause. It therefore suffices (for finite M ) to know the minimal clauses of
a context, called prime implicants in Propositional Logic. For a semiproduct, these
are easy to determine:

Proposition 33 The minimal clauses of a semiproduct are clauses of its factors.

Proof Consider a semiproduct as in Definition 46, and let A → S be a clause

V W
of this semiproduct. Fix some m ∈ M and let Am := A ∩ Mm V, Sm := W S ∩ Mm .
Should Sm ∩ v 0 6= ∅ for each v ∈ Gm for which Am ⊆ v 0 , then Am → Sm is a
clause of Sm (even if there is no such v), confirming the proposition. But if that
•

happens for no m ∈ M , then from each Gm an object vm can be selected such that
Am ⊆ v m 0
and S ∩ vm
0
= ∅. CombiningS these 0results in an object (vm | m ∈ M ) of
the semiproduct with object intent m∈M vm ,V
which contains A and is disjoint
from S, i.e., which does not respect the clause A → S.
W

The combination of the last two propositions with Theorem 9 now finally clari-
fies which rules apply to the implication inference of derived contexts: the Arm-
strong rules and all rules R∧A→∨S of factor scale clauses.

Example We illustrate the effect of these results with the “smallest non-trivial
example”. It lists the possible outcomes of a test that consists of two parts as
follows:
A natural interpretation is to interpret the attribute values pass and fail as nega-
tions of each other. This suggests scaling all three many-valued attributes with
the dichotomic scale. Object and attribute names of the abstract scale must there-
fore be adapted to this specific case. Plain conceptual scaling with these scales
leads to the one-valued context shown in Figure 2.20.
The logic of the implications of this example is obvious even without any the-
ory: The overall result is “pass” if and only if both parts are passed. Surprisingly,
however, the canonical basis of the derived context consists of eight implications,
not two.
6 The completeness assumption can be avoided by modifying the scales, see [138].
2.4 Clauses and dependencies 109

part 1 part 2 overall

outcome 1 fail fail fail
outcome 2 fail pass fail
outcome 3 pass fail fail
outcome 4 pass pass pass

Figure 2.18 Possible outcomes of a two-part test.

D 0 1 S fail pass • part 2:

Spart 2
0 × fail × fail pass
1 × pass × fail ×
pass ×

Figure 2.19 Abstract, concrete and factor scale (using “part 2” as an example).

part 1: part 2: overall:

fail pass fail pass fail pass
outcome 1 × × ×
outcome 2 × × ×
outcome 3 × × ×
outcome 4 × × ×

Figure 2.20 The derived formal context.

The reason is that the background knowledge is not used for the computation
of the canonical base.
W The clauses of the (abstract) dichotomic scale are ∅ →
V
0 ∨ 1 and 0 ∧ 1 → ∅. >

Dependencies of many-valued attributes

Several notions of dependencies come from relational database theory. Here we

discuss the case of functional dependency, the (stronger) one of ordinal dependency
and will indicate generalizations. For reasons of simplicity, we will first concen-
trate on complete many-valued contexts.
Definition 47 If X ⊆ M and Y ⊆ M are sets of attributes of a complete many-
valued context (G, M, W, I), then we say that Y is functionally dependent on X
110 2 Determination and representation

if the following holds for every pair of objects g, h ∈ G:

(∀m∈X m(g) = m(h)) ⇒ (∀n∈Y n(g) = n(h)). ♦

That is to say, if two objects have the same values with respect to all attributes
from X the same must be true for the attributes from Y . The term “functional”
can be explained as follows: Y is functionally dependent on X if and only if there
is a map f : W X → W Y with

f (m(g) | m ∈ X) = (n(g) | n ∈ Y ) for all g ∈ G.

In the case of ordinal dependency, we consider an ordinal context, i.e., we have

for each attribute m ∈ M an order ≤m on the set m(G) of the values of m. (We
obtain the special case of functional dependency if we take the equality relation
for each of those orders.)
Definition 48 (see also Def. 49) Let (G, M, W, I) be a complete many-valued
context and let ≤m be an order relation on the set m(G) of the values of m for
every attribute m ∈ M . If X ⊆ M and Y ⊆ M are sets of attributes, we call Y
ordinally dependent on X if the following holds for each pair of objects g, h ∈ G:

(∀m∈X m(g) ≤m m(h)) ⇒ (∀n∈Y n(g) ≤n n(h)). ♦

Irrespective of which orders ≤m we have chosen, ordinal dependency always

implies functional dependency, since from m(g) = m(h) it follows that m(g) ≤m
m(h) as well as m(h) ≤m m(g), and vice versa. Thus, one would expect that
ordinal dependency is a kind of “order-preserving functional dependency”. In-
tuitively, this is quite correct, but it is difficult to formulate, since the condition
of being order-preserving is only required for the tuples (m(g) | m ∈ X) that
appear in the context. Not every map of this kind can be extended to form an
order-preserving map of W X to W Y .
The ordinal dependencies (and as a special case within them the functional
dependencies) of many-valued contexts can be expressed elegantly by implica-
tions of appropriate one-valued contexts. By means of the rule

(g, h) IO m : ⇔ m(g) ≤m m(h),

we define a one-valued context

KO := (G × G, M, IO )

for a complete many-valued context (G, M, W, I) with orders ≤m on the values.

For the functional dependencies the context can be simplified further: It is pos-
sible to take advantage of the symmetry of the equality relation and to define

KN := (P2 (G), M, IN )
2.4 Clauses and dependencies 111

by
{g, h} IN m : ⇔ m(g) = m(h).
Then,
P2 (G) := {{g, h} | g, h ∈ G, g 6= h}.
The contexts defined in this way exactly fit the above-mentioned definitions
of the dependencies and it is easy to prove the following proposition:
Proposition 34 In (G, M, W, I) the attribute set Y is functionally dependent on X if
and only if the implication X → Y holds in the context KN . In (G, M, W, I) the attribute
set Y is ordinally dependent on X if and only if the implication X → Y holds in the
context KO . 
Hereby we have traced back the theory of functional and ordinal dependen-
cies completely to the theory of implications. In particular, the algorithm men-
tioned in the previous section can also be used for the creation of a basis for the
functional or ordinal dependencies, respectively.
The translation works even if the many-valued context (G, M, W, I) is not com-
plete. In this connection, first of all we observe that Y is ordinally dependent on
X if and only if this is true for every single attribute in Y , i.e., if {n} is ordinally
dependent on X for every n ∈ Y . This means that it is sufficient to state in which
cases a single attribute is dependent on an attribute set. For the general case, this
can be formulated as follows:
Definition 49 (Def. 48 without the completeness assumption)
Let (G, M, W, I) be a many-valued context with an order relation ≤m on the set
m(G) of values for each attribute m ∈ M . If X ⊆ M is a set of attributes and
n ∈ M is an attribute, we say that n is ordinally dependent on X if

∀m∈X dom(n) ⊆ dom(m)

and n(g) 6≤ n(h) always implies that there exists an attribute m ∈ X with m(g) 6≤
m(h). ♦
In order to adapt Proposition 34, we have to modify the definitions of the
contexts KN and KO . We introduce a copy m̂ for every attribute m ∈ M which
is not complete. These new attributes have to be different from each other and
must not belong to M . We add the set

M̂ := {m̂ | dom(m) 6= G}

to the attribute set of the one-valued context. In the case of complete contexts,
we have M̂ = ∅, in general

KN := (P2 (G), M ∪˙ M̂ , IN ) and KO := (G × G, M ∪˙ M̂ , IO ),

with
{g, h} IN m̂ : ⇔ (g, h) IO m̂ : ⇔ g ∈ dom(m) and h ∈ dom(m)
112 2 Determination and representation

and, as above,

(g, h) IO m : ⇔ m(g) ≤m m(h), {g, h} IN m : ⇔ m(g) = m(h).

Proposition 34 can now be generalized as follows:

Proposition 35 The attribute n is functionally (resp. ordinally) dependent on X if and
only if the implications {n̂} ∪ X → n and n̂ → X̂ hold in the context KN (or in the
context KO , respectively). 
Do the approaches presented above extend to notions of dependency other
than those of functional and ordinal dependency? For which cases is it possible
to represent the dependencies of a many-valued context by means of the impli-
cations of an appropriate one-valued context?
Contributions on this topic can be found in the work of J. Baixeries and his co-
authors, see the notes on “database dependencies” in Section 2.5 below. Another
obvious generalization can be obtained by considering (complete) many-valued
contexts (G, M, W, I) with a given relation Θm on W for every attribute m ∈ M .
We abbreviate the sequence of these relations by Θ := (Θm | m ∈ M ) and define
an attribute set Y ⊆ M to be Θ-dependent on a set X ⊆ M if the following holds
for each pair of objects g, h ∈ G:

(∀m∈X m(g)Θm(h)) ⇒ (∀n∈Y n(g)Θn(h)).

A possible interpretation of these kinds of dependency consists in viewing the

Θm as tolerances or fuzziness. Then, a Θ-dependency describes a “functional
dependency with a tolerance”.
Proposition 34 can be applied to this case without problems. The Θ-dependen-
cies of (G, M, W, I) are precisely the implications of the context

KΘ := (G × G, M, IΘ ) with (g, h) IΘ m : ⇔ m(g) Θm m(h).

2.5 Notes, references, and trends

2.1

The algorithm in Theorem 5 has been taken from [143], see also [133]. For a dis-
cussion of other algorithms for finding formal concepts, see Section 2.4.4 of [154].
Further developments can be found in Ganter and Reuter [155], [135], Krolak–
Schwerdt, Orlik and Ganter [235]. With respect to complexity see Skorsky [344].
Schütt [339] gives an estimate of the number of concepts depending on |I|:

3 √|I|+1
|B(G, M, I)| ≤ ·2 −1 (for |I| > 2).
2
2.5 Notes, references, and trends 113

Other bounds are discussed in [1]. Finding the number of formal concepts is
known to be a #P -complete problem (Kuznetsov [245]).

2.2

The example of the waters from Figure 2.9 has been taken from the paper [220] by
Kipke and Wille. The automatic generation of diagrams has been discussed in de-
tail in the works of Skorsky, Luksch and Wille, see [344], [265] and [403] but also
Gepperth [167]. Besides, there are numerous implementations. Early ones were
Diagram for DOS by Frank Vogt and Conexp by Serhiy Yevtushenko. Toscana
[225] was a commercially available program system which facilitates and im-
proves the access to databases by means of elaborate nested line diagrams. See
also [407], [406] as well as Kühn & Ries [243]. The geometrical method has been
described in [400] and in [363] and has been supported by a computer program
by Kark [209]. Figure 2.5 is hand-drawn by R. Wille. Skorsky [343] has examined
the rule of parallelograms. Methods for automatic drawing were proposed, e.g.,
by Zschalig [430], in [139], and, more recently, by Dürrschnabel et al. [112].
[381] gives the “Basic Theorem” on line diagrams.
Other ways of representing contexts and concept lattices have been suggested,
which we shall not discuss here. See [400], Bokowski and Kollewe [57], Kollewe
[224], Lengnink [259], [260].

2.3

Attribute exploration which builds on the theory of implications has become so

extensive that it has its own monograph [154].
Implications and dependencies between attributes have already been exam-
ined in [389]. The implication base with the pseudo-intents was introduced into
Formal Concept Analysis by Duquenne and Guigues [179], [110], Theorem 8 has
also been taken from their publication.
Similar questions have also been of importance in the theory of relational
databases. In this context see Maier [273], Ch. 5. Further investigations can be
found in Wild [386].
Proper premises were introduced in [158], see also Rusch and Wille [327].
An implication A → B only holds in a context if every object having all the
attributes from A also has all the attributes from B. Various authors have tried
to weaken this condition. Burmeister [69] describes implications in the case of
incomplete knowledge by means of a three-valued Kleene-logic. This has also
been implemented in his afore-mentioned program ConImp. See also Burmeister
& Holzer [70]. Luxenburger [269], [270], [271] examines partial implications,
i.e., implications which only hold for part of the object set. These became known
later as association rules in data mining.
114 2 Determination and representation

The results of Duquenne and Guigues permit a more effective algorithmic im-
plementation of the attribute exploration process, which had already been sug-
gested earlier by Wille. This has been described in [143], [133]. A remarkably
early application of this technique was realized by Reeg and Weiss [311]. In the
case of their investigation, the attribute set consisted of 50 common properties of
finite lattices.
Stumme [359] allows exceptions and background implications.
The method of attribute exploration has been further developed in different
ways (cf. [402]): On the one hand into concept exploration [397] (see also Klotz
and Mann [222]), which instead of attributes uses concepts. A specialization of
concept exploration to the distributive case which has practical applications to
knowledge acquisition is presented by Stumme [361].
On the other hand, the attribute exploration can be developed into rule ex-
ploration where the implications are replaced by Horn clauses from predicate
logic. This has been investigated by Zickwolff [429].

2.4

The first two subsections are based on [138] and [137], see also Ganter and
Krauße [150]. In his thesis, Krauße has computed clause bases for many stan-
dard scales.
Most of the results on dependencies have been taken from [159]. A uniform
theory of the dependency of many-valued attributes has been sketched in [398],
compare also [158]. The Θ-dependencies can be looked up in Stöhr and Wille
[353]. Umbreit [371] furthermore examines implications and dependencies be-
tween fuzzy attributes. See below for more recent contributions.

Research developments

Algorithms for finding formal concepts

There are numerous algorithms for computing formal concepts and also several
comparative studies, see e.g. Section 2.4.4 in [154]. That book also describes 30
other Concept Analysis algorithms in detail. The reason we limit ourselves here
to the Next Closure algorithm, which is fast but not the theoretically fastest, is its
captivating simplicity and the versatility it offers. A better worst case complexity
is provided by the algorithm of Nourine and Raynaud [293].
2.5 Notes, references, and trends 115

Algorithms for lattices and diagrams

Similarly, there are a considerable number of programs that support working

with concept lattices. Many of them provide an editor for formal contexts, au-
tomatically draw a diagram of the concept lattice, calculate the arrow relations,
bases for the implications and many other parameters. There are also program
libraries in various currently popular programming languages. However, experi-
ence shows that the life expectancy of such implementations is shorter than that
of a book like this one. Therefore, we refrain from listing the current computer
programs for Formal Concept Analysis. Most of them are freely available on the
internet.

Implications

The theory of implications has been refined in numerous publications. In [154]

this is described in detail, but see also the survey of M. Wild [386]. Wild calls the
system of implications with proper premise the canonical direct basis.
Our Algorithm 4 for constructing the canonical basis looks awkwardly cum-
bersome, because it computes both intents and pseudo-intents, although only
the latter are needed. But it seems difficult to find something decisively better.
Complexity results by Kuznetsov & Obiedkov [250], Babin & Kuznetsov [12],
Distel [96], and Distel & Sertkaya [98] suggest, that neither detecting nor count-
ing pseudo-intents is straightforward. It seems to be possible to speed up Algo-
rithm 4. Borchmann & Kriegel [234] use their parallelized version of the Next
Closure algorithm for this. A different approach was chosen by Ryssel, Distel,
and Borchmann [328]. They first compute the proper premises and then trans-
form that result into the canonical basis. Szathmary et al. [365] follow a related
strategy.
Kriegel [231] finds a canonical basis for those implications of a context appo-
sition K1 | K2 that go from M1 to M2 , which is meaningful when examining the
influence of the attributes from K1 on the attributes of K2 . His result is actually
more general.

Database dependencies

Dependencies are a topic of relational database theory, where different variants

of them are investigated. Whether methods of Formal Concept Analysis can be
helpful for this has been investigated by various authors, in particular by J. Baix-
eries in his dissertation [13] and in a number of publications with coauthors (see
Balcazar & Baixeries [18] for an early contribution and Baixeries et al. [15] for a
more recent one). He has not only investigated functional dependencies, but also
others such as multivalued and (acyclic) join dependencies. An overview with
references to relevant publications is provided by Baixeries [14].
Chapter 3
Parts, factors, and bonds

If one wishes to examine parts of a rather complex concept system, it seems rea-
sonable to exclude some objects and/or attributes from the examination. We shall
describe the effects of this procedure on the concept lattice. The concept lattice
of a subcontext always has an order-embedding into that of the original context,
as will be shown in Section 3.1. Much more information can be obtained when
dealing with compatible subcontexts, which will be introduced later in this sec-
tion. It is easy to identify these particular subcontexts by means of the arrow
relations. One then obtains a factor lattice of the original concept lattice. The in-
terrelations between factor lattices, congruence relations and such subcontexts
will be described in the second section. Tolerance relations are generalized con-
gruence relations and also lead to a factor lattice. They correspond to certain
supersets of the incidence relation I, called block relations.
The complete sublattices of a concept lattice can also be described through parts
of the context, however not through subcontexts, but through certain subsets of
the incidence relation. Such closed relations will be defined in the third section.
In the fourth section we introduce bonds. These are relations that can connect
one formal context with a second one. They correspond in various ways to mor-
phisms between concept lattices and are used for this purpose in several places
in this book.

3.1 Subcontexts

Definition 50 If (G, M, I) is a formal context and if H ⊆ G and N ⊆ M , then

(H, N, I ∩ H×N ) is called a subcontext of (G, M, I).1 ♦

1We write I ∩ H × N for I ∩ (H × N ). Instead of (H, N, I ∩ H×N ) some authors simply use
(H, N ) or [H, N ]. We write K1 ≤ K to express that K1 is a subcontext of K.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 117
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_4
118 3 Parts, factors, and bonds

Subcontexts and order embeddings

We begin the section by asking how the concept system of a subcontext is related
to that of (G, M, I). If one merely omits attributes, that is, if one considers for a
subset N ⊆ M the context (G, N, I ∩ G × N ), then the answer is quite simple.
Any attribute extent of (G, N, I ∩ G × N ) is also an attribute extent of (G, M, I),
and since all concept extents are intersections of attribute extents, we obtain:

Proposition 36 If N ⊆ M , then every extent of (G, N, I ∩ G × N ) is an extent of

(G, M, I). 

This means that the omission of attributes is equivalent to a coarsening of

the closure system of the extents. The corresponding is true for the omission of
objects. We obtain a natural embedding of the concept lattice of (G, N, I ∩G×N )
into that of (G, M, I).

Proposition 37 For N ⊆ M , the map

B(G, N, I ∩ G × N ) → B(G, M, I), (A, B) 7→ (A, A0 )

is a -preserving order-embedding. Dually, for H ⊆ G, the map

V

B(H, M, I ∩ H × M ) → B(G, M, I), (A, B) 7→ (B 0 , B)

is a -preserving order-embedding.
W


× × × × ×
× × - × ×
× × × × ×

e e
@ @
e @e e u @e
@ @ - @ @
e @e @e e @e @e
HH 
 @
e @e
HH 

Figure 3.1 A -embedding of the concept lattice of a subcontext

V

An example is shown in Figure 3.1. If we combine the two parts of the propo-
sition, we obtain:
3.1 Subcontexts 119

Proposition 38 If H ⊆ G and N ⊆ M , the map

B(H, N, I ∩ H×N ) → B(G, M, I), (A, B) 7→ (A00 , A0 )

is an order-embedding, and so is the map

(A, B) 7→ (B 0 , B 00 ).

These order-embeddings are bijective if (H,
V N, I ∩ H×N ) is a dense subcon-
text, i.e., if γH is -dense and dually µN is -dense in B(G, M, I).2
W
If ϕ : B(G, M, I) → (L, ≤) is an order-preserving mapping, then α := ϕ ◦ γ
and β := ϕ ◦ µ are maps α : G → L, β : M → L with

g I m ⇒ αg ≤ βm.

If, conversely, (α, β) is a pair of maps satisfying this condition, then, for instance,
the map _
ϕ(A, B) := αg
g∈A

3
is order-preserving. A useful special case is considered in the next proposition:

Proposition 39 An order-embedding of B(G, M, I) in a given complete lattice (L, ≤)

exists if and only if there are maps α : G → L, β : M → L with

g I m ⇔ αg ≤ βm.

Proof If ϕ : B(G, M, I) → (L, ≤) is an order-embedding, then α := ϕ ◦ γ and

β := ϕ ◦ µ have the properties specified. If, conversely, (α, β) is a pair of maps
with g I m ⇔ αg ≤ βm, then the map ϕ(A, B) := g∈A αg is order-preserving.
W
We show that ϕ is, moreover, an order-embedding: If (A1 , B1 ) and (A2 , B2 ) are
concepts and if (A1 , B1 ) 6≤ (A2 , B2 ), then there exist an object h ∈ A1 and an
attribute n ∈ B2 with (h, n) 6∈ I, i.e., W αh 6≤ βn. On the other hand, αg ≤ βn
holds for all g ∈ A2 , and we have αh 6≤ {αg | g ∈ A2 }. Consequently, ϕ(A1 , B1 )
cannot be less than or equal to ϕ(A2 , B2 ). 

This means that the concept lattice of a subcontext is isomorphic to a suborder

of the entire concept lattice (which is not necessarily a sublattice). The derivation
operators with respect to a subcontext

(H, N, I ∩ H×N )

2 A clarified doubly founded formal context has, of course, a unique smallest dense subcontext,
consisting of its irreducible objects and attributes.
3 See also Section 7.2.
120 3 Parts, factors, and bonds

can be expressed in terms of those of (G, M, I): If A ⊆ H, then the set of common
attributes with respect to (H, N, I ∩ H×N ) is equal to A0 ∩ N . Dually, the extent
of (H, N, I ∩ H×N ) belonging to a set B ⊆ N is equal to B 0 ∩ H. However,
the concepts of a subcontext cannot simply be derived from those of (G, M, I)
by restricting their extent and intent to a subcontext. This can be done only for
compatible subcontexts, which will be examined next.

Compatible subcontexts

Definition 51 A subcontext (H, N, I ∩ H×N ) is called compatible if for every

concept (A, B) ∈ B(G, M, I) the pair (A ∩ H, B ∩ N ) is a concept of the subcon-
text. ♦

↓ ↓ ↓
→× ×
→××× ×
→ × ×
×××

Figure 3.2 Example of a compatible subcontext.

Restricting the concepts to a compatible subcontext yields a map between the

concept lattices, which necessarily has to be structure-preserving, as the follow-
ing proposition shows:

Proposition 40 A subcontext (H, N, I ∩ H×N ) of (G, M, I) is compatible if and only

if
ΠH,N (A, B) := (A ∩ H, B ∩ N ) for all (A, B) ∈ B(G, M, I)
defines a surjective complete homomorphism

ΠH,N : B(G, M, I) → B(H, N, I ∩ H×N ).

Proof According to Definition 51, (H, N, I ∩ H×N ) is compatible if and only if

ΠH,N is a map. The fact that this map must necessarily be infimum-preserving
can be recognized by examining the extents: The map A 7→ A ∩ H is evidently -
T
preserving, and the infimum of concepts is defined in terms of the intersection
of their extents (cf. Basic Theorem). Dually, we infer that ΠH,N is supremum-
preserving. The surjectivity can be seen as follows: If (C, C 0 ∩ N ) is a concept of
(H, N, I ∩ H×N ), then ΠH,N (C 00 , C 0 ) = (C 00 ∩ H, C 0 ∩ N ) is a concept with the
same intent, i.e., the same concept. 
3.1 Subcontexts 121

If there is a surjective complete homomorphism from a complete lattice (L1 , ≤)

onto a complete lattice (L2 , ≤), then (L2 , ≤) is sometimes also called a (complete)
homomorphic image of (L1 , ≤). Thus, the above proposition says that the con-
cept lattice of a compatible subcontext of (G, M, I) is always a homomorphic
image of B(G, M, I). For structure theory it is an important question whether
the converse is true as well, i.e., whether every homomorphic image originates
from a compatible subcontext. We shall defer this question until Section 3.2.
How can we recognize compatible subcontexts? We first give a technical con-
dition, which is often used in proofs. For algorithms, however, the characteriza-
tion by means of the arrow relations is more appropriate. We shall introduce it
right after.
Proposition 41 (H, N, I ∩ H×N ) is a compatible subcontext of (G, M, I) if and only
if:
a1) for every object h ∈ H and every attribute m ∈ M with (h, m) ∈
/ I there is some
attribute n ∈ N with (h, n) ∈
/ I and m0 ⊆ n0 ,
a2) for every attribute n ∈ N and every object g ∈ G with (g, n) ∈
/ I there is some
object h ∈ H with (h, n) ∈
/ I and g 0 ⊆ h0 .
The following are equivalent conditions:
b1) (A0 ∩ N )0 ∩ H ⊆ A00 for all A ⊆ G,
b2) (B 0 ∩ H)0 ∩ N ⊆ B 00 for all B ⊆ M .
Proof If (H, N, I ∩ H×N ) is compatible and m ∈ M , then (m0 ∩ H, m00 ∩ N ) has
to be a concept of the subcontext. If, therefore, h ∈ H is an object with (h, m) ∈
/ I,
there must be an attribute n ∈ m00 ∩ N with (h, n) ∈ / I. This is precisely condition
a1). a2) follows dually.
Now if a1) and a2) are satisfied, we show that b1) must hold: Assume that
A ⊆ G, h 6∈ A00 , h ∈ H. Then there exists some m ∈ A0 (i.e., m0 ⊇ A) with
(h, m) ∈ / I, i.e., by a1) some n ∈ A0 ∩N with n0 ⊇ A and (h, n) ∈/ I. Consequently,
h 6∈ (A0 ∩ N )0 and thus (A0 ∩ N )0 ∩ H ⊆ A00 . b2) follows correspondingly.
It remains to be shown that (H, N, I ∩ H×N ) is compatible if b1) and b2) are
satisfied. Let (A, B) be a concept of (G, M, I). Then (A∩H)0 ∩N ⊇ A0 ∩N = B∩N
and, by applying b2), (A ∩ H)0 ∩ N = (B 0 ∩ H)0 ∩ N ⊆ B 00 ∩ N = B ∩ N , i.e.,
(A∩H)0 ∩N = B ∩N and dually A∩H = (B ∩N )0 ∩H. Therefore, (A∩H, B ∩N )
is a concept of (H, N, I ∩ H×N ). 
In the case of doubly founded contexts, the compatible subcontexts can be
easily identified by means of the arrow relations.
Definition 52 A subcontext (H, N, I ∩ H×N ) of a clarified context (G, M, I) is
arrow-closed if the following holds: h % m and h ∈ H together imply m ∈ N ,
and g . n and n ∈ N together imply g ∈ H. ♦

Proposition 42 Every compatible subcontext is arrow-closed.

Every arrow-closed subcontext of a doubly founded context is compatible.
122 3 Parts, factors, and bonds

Proof If (H, N, I ∩ H×N ) is compatible and h ∈ H, m ∈ M are such that

h % m, then by Proposition 41.a1) there is an attribute n ∈ N with (h, n) ∈ / I
and m0 ⊆ n0 . Because of h % m, m0 is maximal with respect to (h, m) ∈ / I, i.e.,
m0 = n0 , i.e., m = n, i.e., m ∈ N . Dually, g . n and n ∈ N yield g ∈ H.
If, conversely, (H, N, I ∩ H×N ) is an arrow-closed subcontext of a doubly
founded context, we can prove 41.a1): Let h ∈ H be an object and let m ∈ M be an
attribute with (h, m) ∈ / I. By Definition 28 there exists an attribute n with m0 ⊆ n0
and h % n, i.e., n ∈ N , which was to be proved. 41.a2) follows correspondingly.
Thus, in the case of doubly founded contexts it is easy to determine the com-
patible subcontexts. We enter the arrow relations % and . into the context and
examine the directed graph (G ∪˙ M, % ∪ .). The compatible subcontexts
then correspond exactly to the arrow-closed components of the directed graph.
If we furthermore assume that (G, M, I) is reduced, we can elegantly describe

a b c d e c a b e d a b c d e .
.\ a b c d e
1 × . × .
% b b b b b 1 .
... .
. 1 × ×
2 × × × .
% × 2 .
. 2 × × × ×
b b b b
?
3 . × .
% × 3 .
..... 3 × ×
4 .
% .
% × × × 3 4 1 2 4 .
... 4 × × ×

Context example The graph of the arrows The relation .

. (G, M, .
.\ )

Figure 3.3 With reference to Proposition 43

the arrow-closed subcontexts in terms of the concepts of a context. For this pur-
pose we need the transitive closure of the arrow relations, as introduced in the
following definition:
Definition 53 For g ∈ G and m ∈ M we write g . . m if there are objects g =
g1 , g2 , . . . , gk ∈ G and attributes m1 , m2 , . . . , mk = m ∈ M with gi . mi for
i ∈ {1, . . . , k} and gj % mj−1 for j ∈ {2, . . . , k}. The complement (the negation)
of this relation is denoted by . \ , i.e., g .
. \ m ⇔ not g .
. . m. ♦

Proposition 43 Let (G, M, I) be a reduced doubly founded context. Then (H, N, I ∩

H×N ) is an arrow-closed subcontext if and only if (G\H, N ) is a concept of the context
(G, M, .
.
\ ).

Proof From the presuppositions doubly founded and reduced follows by Proposi-
tion 15 that for every object g there is an attribute m with g %
. m, and dually.
First of all, let (H, N, I ∩ H×N ) be arrow-closed. If g ∈ H and g % . m, then
m ∈ N must be true, i.e., g ∈ H holds if and only if there is an n ∈ N with
g .. n. Consequently, g ∈ G \ H if and only if g . \ n for all n ∈ N , i.e.,
.
G \ H = N .. \
. Now assume that m ∈ M \ N and g % . m. g ∈ H is impossible
because (H, N, I ∩ H×N ) is arrow-closed. Therefore, we get g ∈ G \ H, g ..m
3.1 Subcontexts 123

and thus m 6∈ (G\H).. \

. This shows (G\H)..
\
⊆ N , which means that (G\H, N )
is a concept of (G, M, .
.\ ).
For the converse we assume that (G \ H, N ) is a concept of (G, M, .
\ ). From
.
g . n and n ∈ N immediately follows g ∈ H; if we have h % m and h ∈ H,
it remains to be proved that m ∈ N . Assuming that m 6∈ N , there would be an
object g ∈ (G \ H) with g . . m, and because of h ∈ H an attribute n ∈ N with
h . . n. Taken together, g . . m, h % m and h . . n yield g .. n, which is
impossible. Thus, (H, N, I ∩ H×N ) is arrow-closed. 

Proposition 44 Every compatible subcontext of a clarified (resp. reduced, resp. doubly

founded) context is clarified (resp. reduced, resp. doubly founded).
The arrow relations are inherited by compatible subcontexts, i.e., g . m holds in
(H, N, I ∩ H×N ) if and only if g ∈ H, m ∈ N and g . m hold in (G, M, I), and the
corresponding is true for %.

Proof Several times in the proof we use the following argument: If h1 , h2 ∈ H

are objects with h01 ∩ N ⊆ h20 ∩ N , then h01 ⊆ h02 . This follows from Proposition
41: If m were an attribute with m ∈ h01 \ h20 , we should obtain by a1) an attribute
n ∈ N ∩ (h01 \ h20 ), which in the case of h01 ∩ N ⊆ h02 ∩ N is impossible. (Of
course, the corresponding applies to the attributes.) This immediately yields the
first assertion: h01 ∩ N = h20 ∩ N implies h01 = h02 , i.e., objects with the same object
intents in the subcontext have the same object intents in general.
If h ∈ H is irreducible in (G, M, I), then there exists an attribute m with
(h, m) ∈ / I and gI m for every g ∈ G with g 0 ⊇ h0 . By 41.a1) we find an n ∈ N with
(h, n) ∈ / I and n0 ⊇ m0 , i.e., with g I n for all g ∈ G with g 0 ⊇ h0 and particularly
g I n for all g ∈ H with g 0 ∩ N ⊇ h0 ∩ N . Together with the dual consideration
this shows that a compatible subcontext of a reduced context is reduced as well.
Now concerning the arrows: First of all we assume that h ∈ H and n ∈ N and
that h . n in (G, M, I). Then h0 is maximal with respect to n 6∈ h0 . According
to our previous considerations, in this case h0 ∩ N is maximal with respect to
n 6∈ h0 ∩ N . This means that h . n also holds in (H, N, I ∩ H×N ). Thus, all
arrows from (G, M, I) are being preserved in (H, N, I ∩ H×N ).
Is it possible, conversely, to infer h . n in (G, M, I) from h . n in (H, N, I ∩
H×N )? If not, there would have to be an object g ∈ G with (g, n) ∈ / I and g 0 ⊃ h0
and furthermore by Proposition 41.a2) an object h2 ∈ H with (h2 , n) ∈ / I and
h20 ⊇ g 0 , from which would follow h20 ⊃ h0 and (h2 , n) ∈ / I (the fact that h02 ∩ N =
h0 ∩ N is impossible, again follows from our first consideration). Thus, h0 would
not be maximal among the extents of (H, N, I ∩ H×N ) which do not contain n,
in contradiction to the precondition h . n.
We are still lacking the proof that the property of being doubly founded is
inherited by compatible subcontexts. Thus, assume that h ∈ H, n ∈ N and
(, n)h ∈ / I. If (G, M, I) is doubly founded there is an attribute m ∈ M with h % m
and n0 ⊆ m0 . We apply Proposition 41.a1) and obtain an attribute n2 ∈ N with
(h, n2 ∈ / I and m0 ⊆ n20 . It follows that m0 = n20 and thus that h % n2 , which
according to what we have just proved transfers to (H, N, I ∩ H×N ). One of
124 3 Parts, factors, and bonds

the conditions of doubly foundedness is proved thereby, the other one follows
dually. 
We should also mention that dense subcontexts are always compatible:

Proposition 45 For a subcontext (H, N, I ∩ H×N ) of (G, M, I) the following state-

ments are equivalent:
1. (H, N, I ∩ H×N ) is dense.
2. (H, N, I ∩ H×N ) is compatible and the map ΠH,N is injective.
3. for every concept (A, B) of (G, M, I),

(A ∩ H)00 = A and (B ∩ N )00 = B.

Proof 1) ⇔ 3): (H, N, I ∩ H×N ) is denseWif and only if both, for every object
g ∈ G there is a subset X ⊆ H with γg = x∈X γx, i.e., with g 0 = X 0 and thus
g ∈ X 00 , and the dual condition holds for the attributes. Because of γx ≤ γg for
all x ∈ X we have X ⊆ g 00 and thus the condition from 3) for the case A = g 00 .
The more general condition follows without difficulty.
3) ⇒2): In order to show that (H, N, I ∩ H×N ) is compatible, we prove the
conditions b) from Proposition 41: If A ⊆ G, then A0 is an intent and on account
of 3) it satisfies (A0 ∩ N )00 = A0 , which yields (A0 ∩ N )0 = A00 and thus b1). b2)
is dual. The injectivity of ΠH,N immediately follows from (A ∩ H)00 = A.
2) ⇒ 3): If ΠH,N is injective and (A, B) is a concept of (G, M, I), then (A ∩ H)00 =
A must hold, otherwise (A ∩ H)00 and A would be different extents having the
same intersection with H. 

3.2 Complete congruences and tolerances

In the preceding section we have seen that, for a compatible subcontext (H, N, I ∩
H×N ) of (G, M, I), the map ΠH,N is a surjective complete homomorphism of
B(G, M, I) onto B(H, N, I∩ H×N ), i.e., that the concept lattice of the subcontext
is always a homomorphic image of B(G, M, I). We shall now examine whether
the converse is true as well, i.e., whether every surjective complete homomor-
phism, every homomorphic image can be described in terms of a subcontext. In
the case of finite contexts this is true, in the case of infinite contexts not in general.
In order to clarify the situation, we require a notion from lattice theory, namely
that of the complete congruence relation.
Definition 54 A complete congruence relation of a complete lattice (L, ≤) is an
equivalence relation Θ on L satisfying:
^ ^ _ _
xt Θ yt for t ∈ T ⇒ ( xt ) Θ ( yt ) and ( xt ) Θ ( yt ).
t∈T t∈T t∈T t∈T
3.2 Complete congruences and tolerances 125

We define
[x]Θ := {y ∈ L | x Θ y},
which is the equivalence class of Θ containing x. The factor lattice

(L, ≤)/Θ := {[x]Θ | x ∈ L}

is ordered as follows:

[x]Θ ≤ [y] Θ:⇔ x Θ (x ∧ y) (⇔ (x ∨ y) Θ y).

In order to demonstrate that this is really an order relation we can, for instance,
argue as follows: If we define
^ ^
xΘ := {y ∈ (L, ≤) | y Θ x} = [x]Θ
_ _
and xΘ := {y ∈ (L, ≤) | y Θ x} = [x]Θ

for x ∈ (L, ≤), this immediately yields [x]Θ = xΘ , xΘ and [x]Θ ≤ [y]Θ ⇔ xΘ ≤
 

yΘ ⇔ xΘ ≤ y Θ . Thus, the congruence classes, i.e., the classes of a congruence

relation are intervals. They are ordered according to their smallest, or, which is
the same, according to their largest elements. We infer that
^ h^ i _ h_ i
[xt ]Θ = xt Θ and [xt ]Θ = xt Θ.
t∈T t∈T t∈T t∈T ♦

e
@
e @e e
@ @ @
e @e @e e @e
@ @ @
@ @ @e
@ @
@e
@ @e
@ e
@
@e

Figure 3.4 Congruence and factor lattice

If in the following we speak of congruence relations or congruences, we mean

complete congruence relations. The significance of the congruences for the prob-
lem we are concerned with is revealed by the following Homomorphism Theorem.
126 3 Parts, factors, and bonds

It asserts among other things that every homomorphic image of a complete lat-
tice can be found within the lattice itself, namely as a factor lattice.
Theorem 10 (Homomorphism Theorem) If Θ is a complete congruence relation of
a complete lattice (L, ≤), then x 7→ [x]Θ is a complete homomorphism of (L, ≤) onto
(L, ≤)/Θ. If, conversely, ϕ : (L1 , ≤) −→ (L2 , ≤) is a surjective complete homomor-
phism between complete lattices, then

ker ϕ := {(x, y) ∈ L1 × L1 | ϕ(x) = ϕ(y)}

is a complete congruence relation of (L1 , ≤); moreover,

[x] ker ϕ 7→ ϕ(x)

describes an isomorphism of (L1 , ≤)/ ker ϕ onto (L2 , ≤).

Proof The homomorphism properties of the map x 7→ [x]Θ have been proved
above. ker ϕ is evidently an equivalence relation. Moreover,

(xt , yt ) ∈ ker ϕ for all t ∈ T ⇔ ϕ(xt ) = ϕ(yt ) for all t ∈ T

^  ^
⇒ϕ xt = ϕ(xt )
t∈T t∈T
^ ^ 
= ϕ(yt ) = ϕ yt
t∈T t∈T
^  ^ 
⇒ xt ker ϕ yt ;
t∈T t∈T

the -compatibility follows by analogy. ([x] ker ϕ) 7→ ϕ(x) describes a bijection

W
which evidently satisfies the conditions for homomorphisms. 
The map x 7→ [x]Θ is occasionally denoted by πΘ and is called canonical pro-
jection onto the factor lattice.
We provide another result from lattice theory that we will need later. It de-
scribes which equivalence relations are congruences in a given lattice.
Theorem 11 (Characterization of Complete Congruence Relations)
An equivalence relation Θ on a complete lattice (L, ≤) is a complete congruence relation
of (L, ≤) if and
 only ifevery equivalence class of Θ is an interval of (L, ≤) (we then
write [x]Θ =: xΘ , xΘW for all x ∈ L), the lower bounds of these intervals being closed
under suprema (i.e., t∈T V xΘ =Θ yΘ )Θand the upper bounds of these intervals being
t

closed under infima (i.e., t∈T xt = z ).

Proof Let Θ be a complete congruence relation of (L, ≤) and x ∈ L. Then, xΘ :=
[x]Θ and xΘ := [x]Θ are elements of [x]Θ. If xΘ ≤ y ≤ xΘ , from yΘy and
V W
xΘ ΘxΘ it follows that

y = (xΘ ∨ y), (xΘ ∨ y) Θ (xΘ ∨ y), (xΘ ∨ y) = xΘ ,

3.2 Complete congruences and tolerances 127

and thus y ∈ [x]Θ. This means that [x]Θ = xΘ , xΘ . The maps x 7→ xΘ and
 

x 7→ xΘ form a Galois-connection between (L, ≤) and (L, ≤)d , since x ≤ y ⇒

x Θ ≤ yΘ , x ≥ y ⇒ W xΘ ≥ y Θ , x ≤ (xΘ )Θ , x ≥ (xΘ )Θ . By Proposition 9 it follows
that x 7→ xΘ is a -homomorphism and x 7→ xΘ is a -homomorphism of
V

Θ
(L, ≤) in itself. In particular t∈T xtΘ = t∈T x Θ and ,
W W t

V Θ
V
t∈T xt = t∈T xt
which was to be proved. Conversely, we assume that Θ is an equivalence relation
on (L, ≤) the equivalence classes of which are intervals with supremum-dense
lower bounds and infimum-dense upper bounds. Assume that xt Θyt for t ∈ T .
Then (xt )Θ = (yt )Θ and (xt )Θ = (yt )Θ for t ∈ T . Consequently,
_ _ _ _
(xt )Θ ≤ xt , yt ≤ (xt )Θ .
t∈T t∈T t∈T t∈T

Since, in general, a ≤ b implies aΘ ≤ bΘ (because a ≤ b yields a ≤ aΘ ∧ bΘ ≤

W since a ∧ b is an upper interval bound, a ∧ bΘ = a ), from
aΘ and thus, Θ Θ Θ Θ Θ

(xs )Θ ≤ t∈T (xt )Θ for every s ∈ T we furthermore obtain (xs ) = ((xs )Θ )Θ ≤

Θ
( t∈T (xt )Θ )Θ and thus t∈T (xt )Θ ≤ . Therefore:
W W W
t∈T (xt )Θ
" #
_ _ _ _
Θ
xt , yt ∈ (xt )Θ , ( (xt )Θ ) ,
t∈T t∈T t∈T t∈T

i.e., t∈T xt Θ t∈T yt . Dually, the -compatibility of Θ can be shown, which

W
W
V
proves that Θ is a complete congruence relation. 
In the light of the Homomorphism Theorem our question “What is the con-
nection between the compatible subcontexts and the congruences?” comes up
again. From the Homomorphism Theorem it immediately follows that the con-
cept lattice of a compatible subcontext (H, N, I ∩ H×N ) of (G, M, I) is always
isomorphic to a factor lattice of B(G, M, I): (H, N, I ∩ H×N ) induces a complete
congruence ΘH,N on B(G, M, I), namely the kernel of the complete homomor-
phism ΠH,N , and we get

B(H, N, I ∩ H×N ) ∼
= B(G, M, I)/ΘH,N ,

with

(A1 , B1 ) ΘH,N (A2 , B2 ) ⇔ A1 ∩ H = A2 ∩ H ⇔ B1 ∩ N = B2 ∩ N.

It is easy to identify the smallest and the largest elements of the congruence
classes. If (A, B) is a concept, the smallest element of the congruence class
[(A, B)]ΘH,N is the concept ((A ∩ H)00 , (A ∩ H)0 ) and the largest is the concept
((B ∩ N )0 , (B ∩ N )00 ).
We say that a complete congruence Θ is induced by a subcontext if there is a
compatible subcontext (H, N, I ∩ H × N ) with Θ = ΘH,N . Using the connection
between compatible subcontexts and congruences, we shall prove the following:
in the case of a doubly founded concept lattice every congruence is induced by
128 3 Parts, factors, and bonds

a subcontext. Provided that the context is reduced, this subcontext is uniquely

determined by the congruence. These restricting preconditions are not superflu-
ous, i.e., the general theory is somewhat more complicated.
First, we shall examine the problem of uniqueness. A congruence can be in-
duced by various subcontexts. These, however, only differ in their reducible ob-
jects and attributes. Among all the possible subcontexts there is always a largest
one.
Proposition 46 If a complete congruence Θ is induced by a subcontext (H, N, I ∩ H ×
N ), then

H ⊆ GΘ := {g ∈ G | γg is the smallest element of a Θ-class} and

N ⊆ MΘ := {m ∈ M | µm is the largest element of a Θ-class}.

In this case, Θ is also induced by the compatible subcontext

(GΘ , MΘ , I ∩ GΘ × MΘ ).

Proof If Θ is the congruence induced by (H, N, I ∩ H×N ) (i.e., Θ = ΘH,N ),

then according to the above-mentioned description of the smallest elements of
the ΘH,N -classes we have:

g ∈ GΘ ⇔ there is X ⊆ H with X 0 = g 0 .

This immediately yields H ⊆ GΘ and dually N ⊆ MΘ .

Why is (GΘ , MΘ , I ∩ GΘ × MΘ ) compatible? We use Proposition 41 and prove
condition a1): If g ∈ GΘ and m ∈ M with (g, m) ∈ / I, then there is a set X ⊆ H
with g 0 = X 0 , i.e., in particular some h ∈ H with (h, m) ∈ / I and g 0 ⊆ h0 . Since
(H, N, I ∩ H×N ) is compatible, there is an attribute n ∈ N with (h, n) ∈ / I and
m0 ⊆ n0 . This means that (g, n) ∈ / I, n ∈ N ⊆ MΘ and m0 ⊆ n0 hold as well.
Thus, condition a1) is satisfied, and a2) can be shown dually.
Finally, it remains to be proved that (GΘ , MΘ , I ∩ GΘ × MΘ ) induces the same
congruence as (H, N, I ∩ H×N ). In order to do so, it suffices to show that from
(A1 , B1 ) Θ (A2 , B2 ) it always follows that A1 ∩ GΘ = A2 ∩ GΘ ; the converse
implication immediately follows from H ⊆ GΘ and N ⊆ MΘ . We assume that
g ∈ A1 ∩ GΘ . Then there is some X ⊆ H with X 0 = g 0 , consequently X ⊇ B1 and
therefore X 00 ⊆ A1 , from which it follows that X = X ∩ H ⊆ A1 ∩ H = A2 ∩ H
and thus g ∈ A2 . 
Hence, it is possible to “saturate” a compatible subcontext by adding reducible
objects and attributes, without changing the corresponding congruence.
Definition 55 A subcontext (H, N, I ∩ H×N ) of (G, M, I) is called saturated if:

from g ∈ G, X ⊆ H and X 0 = g 0 it follows that g ∈ H and

from m ∈ M, Y ⊆ N and Y 0 = m0 it follows that m ∈ N. ♦

3.2 Complete congruences and tolerances 129

The preceding proposition together with this definition immediately yields:

Proposition 47 If a congruence Θ is induced by some subcontext, then it is also induced

by a saturated subcontext, which is then equal to

(GΘ , MΘ , I ∩ GΘ × MΘ ).

In a reduced context every subcontext is saturated. 

Now we turn to the second part of the question: Which congruences are in-
duced by subcontexts? On account of the propositions we know that H = GΘ
and N = MΘ can be chosen, if Θ is at all induced by a subcontext. It is easy
to state congruences which do not have this form, these examples are however
infinite. The following propositions provide an exact clarification.

Proposition 48 A complete congruence relation Θ is induced by a subcontext if and only

if {[γh]Θ | h ∈ GΘ } is supremum-dense and {[µn]Θ | n ∈ MΘ } is infimum-dense in
B(G, M, I)/Θ.

Proof If Θ is induced by a subcontext, then

B(H, N, I ∩ H×N ) ∼
= B(G, M, I)/Θ,

and the isomorphism (A, B) 7→ [(A, B)]Θ maps the supremum-dense set {γh |
h ∈ H} onto {[γh]Θ | h ∈ H}, i.e., this set is supremum-dense in B(G, M, I)/Θ,
and dually {[µn]Θ | n ∈ N } is infimum-dense. Because of H ⊆ GΘ and N ⊆ MΘ
thus the direction “⇒” of the assertion follows.
We begin the proof of the other direction by showing that (GΘ , MΘ , I ∩ GΘ ×
MΘ ) is compatible under the conditions specified. Assume that h ∈ GΘ and that
m ∈ M with (h, m) ∈ / I. Then [γh]Θ 6≤ [µm]Θ and, since {[µn]Θ | n ∈ MΘ } is
infimum-dense in B(G, M, I)/Θ, there is some n ∈ MΘ \ h0 with µn ≥ µm, i.e.,
n0 ⊇ m0 . This yields 41.a1) and dually 41.a2).
In order to show that Θ is induced by (GΘ , MΘ , I ∩ GΘ × MΘ ), we have to prove
that
(A, B) Θ (C, D) ⇔ A ∩ GΘ = C ∩ GΘ
for (A, B), (C, D) ∈ B(G, M, I). Let (A, B) be the smallest concept in the Θ-class
containing (A, B). For h ∈ GΘ we have γh ≤ (A, B)W⇔ γh ≤ (A, B), since γh also
is the smallest element of a Θ-class, and (A, B) = {γh | h ∈ A ∩ GΘ } because
{[γh]Θ | h ∈ GΘ } is supremum-dense. (A, B) and (C, D) are congruent if and
only if the classes [(A, B)]Θ and [(C, D)]Θ have the same smallest element, i.e.,
if A ∩ GΘ = C ∩ GΘ . 
If [v]Θ is -irreducible in B(G,
W M, I)/Θ, then the smallest element of the con-
W
gruence class [v]Θ must also be -irreducible and thus must be an W object concept
γg with g ∈ GΘ . Hence, the set {[γh]Θ | h ∈ GΘ } contains all -irreducible el-
ements, and likewise {[µn]Θ | n ∈ MΘ } contains all -irreducible elements of
V
B(G, M, I)/Θ. Thus, from Proposition 48 we can infer:
130 3 Parts, factors, and bonds

Proposition 49 If B(G, M, I)/Θ is doubly founded, then Θ is induced by a subcontext.


If we consider the concept lattice only up to isomorphism, then each con-
gruence is induced by a subcontext of a suitable context: Every complete lattice
(L, ≤) can be represented as a concept lattice B(L, L, ≤); and by Proposition 48,
in this representation every complete congruence is induced by a subcontext. In
the case of doubly founded contexts we can go even further. For this purpose we
first of all transfer Proposition 44:
Proposition 50 Every factor lattice of a doubly founded complete lattice is doubly
founded.
Proof Let [x]Θ < [y]Θ be two elements of B(G, M, I)/Θ and assume w.l.o.g. that
x is the largest and the smallest element of its class, i.e., that and
W
y x = [x]Θ
y = [y]Θ. Hence, in the doubly founded concept lattice B(G, M, I) we find an
V
element s that is minimal with respect to s ≤ y, s 6≤ x ∧ y. We claim that [s]Θ
has the corresponding property of minimality with respect to [x]Θ < [y]Θ. It is
certain that [s]Θ ≤ [y]Θ. On the other hand [s]Θ ≤ [x]Θ is impossible, since it
would yield s ∨ x ∈ [x]Θ and thus s ∨ x ≤ x. Consequently, we have [s]Θ V ≤
[y]Θ, [s]Θ 6≤ [x]Θ and shall prove the minimality: If [r]Θ < [s]Θ, r = [r]Θ
being the smallest element of its class, then r < s and, because of the property
of minimality of s, r ≤ x and hence [r]Θ ≤ [x]Θ. The second condition is proved
dually. 
If we combine Propositions 49 and 50, we obtain:
Theorem 12 If B(G, M, I) is doubly founded, then every complete congruence relation
is induced by a subcontext. 
At the end of this section, we shall use the above results in order to analyze
the system of all complete congruence relations of a concept lattice L. This set
of congruences is ordered by set inclusion ⊆; it even forms a closure system on
L × L and thus a complete lattice, the lattice C(L) of the complete congruence
relations of L.
If we suppose that L is doubly founded, then we may assume that L is the
concept lattice of a reduced, doubly founded context (G, M, I). This yields the
following simplifications: Every compatible subcontext of a reduced context is
saturated (cf. Definition 47), the compatible subcontexts are precisely the arrow-
closed subcontexts (Proposition 42), and every complete congruence is induced
by a subcontext (Theorem 12). Thus, in this case the arrow-closed subcontexts
correspond bijectively to the complete congruences.
The order of the congruence relations is also reflected by the subcontexts: If
Θ and Ψ are two congruences of L, then

Θ ⊆ Ψ ⇔ (A, B) Θ (C, D) ⇒ (A, B) Ψ (C, D)

for all (A, B), (C, D) ∈ L
⇔ A ∩ GΘ = C ∩ GΘ ⇒ A ∩ GΨ = C ∩ GΨ and
3.2 Complete congruences and tolerances 131

B ∩ MΘ = D ∩ M Θ ⇒ B ∩ MΨ = D ∩ MΨ
for all (A, B), (C, D) ∈ L
⇔ GΨ ⊆ GΘ and MΨ ⊆ MΘ .

Hence, if we order the subcontexts by

(H1 , N1 , I ∩ H1 × N1 ) ≤ (H2 , N2 , I ∩ H2 × N2 )

:⇔ H1 ⊆ H2 and N1 ⊆ N2 ,
under the preconditions specified, the ordered set of the arrow-closed subcon-
texts is dually isomorphic to the lattice of complete congruences. Now, how-
ever, the union as well as the intersection of arrow-closed subcontexts are arrow-
closed too. Therefore, the lattice of arrow-closed subcontexts is completely dis-
tributive. Thus, Proposition 43 makes it easy to state a context for the congruence
lattice as well.
Theorem 13 The congruence lattice of a doubly founded concept lattice B(G, M, I) is
isomorphic to the completely distributive lattice B(G, M, .
.
\ ).
Proof If (G, M, I) is reduced, then every congruence is induced by exactly one
subcontext (H, N, I ∩ H×N ). Furthermore we know by Proposition 43 that those
subcontexts correspond to the concepts of (G, M, . .\ ): (H, N, I ∩ H×N ) induces
a congruence if and only if (G \ H, N ) is such a concept.
The order of those subcontexts is dual to that of the concepts of (G, M, . .
\)
as well as to that of the congruences. This means that the latter two must be
isomorphic to each other.
For the structure of B(G, M, . .\ ), however, it is irrelevant whether (G, M, I) is
reduced, provided that B(G, M, I) is doubly founded. In this case, we can switch
to the reduced context (Girr , Mirr , I ∩ Girr × Mirr ) with Girr and Mirr being the
set of irreducible objects and attributes, respectively. The . .-relation is inherited
by this subcontext, since in Definition 53, apart from g and m, there only appear
irreducible objects and attributes. Therefore, B(G, M, . .\ ) and B(Girr , Mirr , .
.
\)
are isomorphic, every concept of (G, M, . \ ) is of the form
.
 
(G \ Girr ) ∪ A, B ∪ (A..\
∩ (M \ Mirr )) ,

(A, B) being a concept of (Girr , Mirr , .

.
\ ). 

Tolerances and block relations

Definition 56 Let L be a complete lattice. A complete tolerance relation on L is

a relation Θ ⊆ L×L which is reflexive, symmetric and compatible with suprema
and infima, i.e., for which holds
132 3 Parts, factors, and bonds

f
@
d f fc @ fe
@
@ @
f f @f
@
@ @
1 @ 3 @
@v
@ @ fa, b
@
4 @ 2
@f
@

Figure 3.5 The congruence lattice of the lattice from Figure 3.4 is at the same time the
lattice of arrow-closed subcontexts of Figure 3.3. The marked element corresponds to the
congruence from Figure 3.4 and the compatible subcontext in Figure 3.2.

^ ^ _ _
xt Θ yt for t ∈ T ⇒ ( xt ) Θ ( yt ) and ( xt ) Θ ( yt ).
t∈T t∈T t∈T t∈T

Hence, a complete tolerance relation is a congruence relation if it is transitive. ♦

Proposition 51 If Θ is a complete tolerance relation on L, then it follows from a Θ b

and x, y ∈ [a ∧ b, a ∨ b] that x Θ y.
Proof From a Θ b and a Θ a it follows that a Θ a ∧ b and correspondingly
b Θ a ∧ b. This yields a ∧ b Θ a ∨ b. It follows that x ∨ (a ∧ b) Θ x ∨ (a ∨ b), i.e.,
x Θ (a ∨ b) and correspondingly (a ∨ b) Θ y. Because of x, y ≤ a ∨ b we obtain
x Θ y. 
Definition 57 If Θ is a complete tolerance relation and a ∈ L, we define
^ _
aΘ := {x ∈ L | a Θ x} and aΘ := {x ∈ L | a Θ x}.

The intervals [a]Θ := [aΘ , (aΘ )Θ ], (a ∈ L) are called the blocks of Θ. ♦

aΘ is the smallest (and dually a is the largest) element related with a under
Θ

Θ . The dual definition [a]Θ := [(aΘ )Θ , aΘ ] also yields the blocks of Θ: Because of
((aΘ )Θ )Θ = aΘ and ((aΘ )Θ )Θ = aΘ we obtain [a]Θ = [aΘ ]Θ as well as [a]Θ = [aΘ ]Θ .
From aΘb and a ≤ b it follows that bΘ ≤ a ≤ b, i.e., a ∈ [b]Θ . Correspondingly,
from aΘb and a ≥ b it always follows that a ∈ [b]Θ . The blocks of a tolerance rela-
tion do not have to be disjoint, unless we are dealing with a congruence relation.
We have
[a]Θ ∩ [b]Θ 6= ∅ ⇔ aΘ ≤ (bΘ )Θ and bΘ ≤ (aΘ )Θ .
3.2 Complete congruences and tolerances 133

f f
@ @
@ @ 
f f
f @f f @ f
@ @@ @
@ @ @  @ 
f f @f f @f f f @ f f @f
@ @ 
@ 
@
@ @ @ @
@f @f @f @f
@ @
@ @
@f @f

Figure 3.6 The pairs which are linked together in the figure on the right (including the
pairs of neighboring elements) form part of a tolerance relation of the lattice on the left.

Proposition 52 The blocks of Θ are precisely the maximal subsets X of L with x Θ y

for all x, y ∈ X.
Proof Because of Proposition 51 we have x Θ y for all x, y ∈ [a]Θ . Now if z is an
arbitrary element with z Θ a and z Θ aΘ , we obtain z ≥ aΘ and z ≤ (aΘ )Θ , i.e.,
z ∈ [a]Θ . Hence, every block is maximal with regard to the property specified.
If X is an arbitrary maximal set of elements of L which areWpairwise related
under Θ, then from the compatibility it follows Vthat WX and X V areΘelements
V
ofWX. Hence, because of the maximality, and
W
X = [ X, X] ( X) = X,
( X)Θ = X, i.e., X is a block of Θ.
V

ϕ ψ
Proposition 53 The map x 7→ xΘ is a -morphism and the map x → xΘ is a -
W V
7
morphism. The two maps are adjoint to each other.

Proof We show that (ϕ, ψ) is a Galois connection between L and Ld . We have

x ≤ y ⇒ xΘ ≤ yΘ ⇒ xΘ ≥d yΘ , x ≤d y ⇒ x ≥ y ⇒ xΘ ≥ y Θ , x ≤ (xΘ )Θ and
x ≥ (xΘ )Θ ⇒ x ≤d (xΘ )Θ . Hence,
_ ^d _
ϕ( xt ) = ϕ(xt ) = ϕ(xt )

and correspondingly
^ _d ^
ψ( xt ) = ψ( xt ) = ψ(xt ). 

Definition 58 The set of all blocks of a complete tolerance relation of L is de-

noted by L/Θ and ordered by
134 3 Parts, factors, and bonds

' $
f
@
' f$
'
@
$
f @f
@ C@
' C$
 @ C @
f f @f f @f
& %
@ C @ 
@ C @ 
@C f @f

& &% %
@
@
@f
& %
Figure 3.7 The blocks of the tolerance relation from Figure 3.6.

^ ^ _ _
B1 ≤ B2 : ⇔ B1 ≤ B2 (⇔ B1 ≤ B2 ). ♦

The definition says that the smallest elements of the blocks are ordered in
the same way as their largest elements. This is correct because of xΘ ≤ yΘ ⇔
(xΘ )Θ ≤ (yΘ )Θ . In fact, even more is true: The set of the upper bounds of the
blocks is closed under infima, that of the lower bounds is closed under suprema,
in analogy to the case of the complete congruences (cf. Theorem 11, p. 126). This
is described by the following theorem:
Theorem 14 With the order described above, L/Θ is a complete lattice (the factor lat-
tice of L by Θ). The following equations hold for blocks Bt and for elements xt , t ∈ T ,
of L respectively:
_ _^ ^ ^_
Bt = [ Bt ]Θ resp. Bt = [ Bt ] Θ
t∈T t∈T t∈T t∈T
_ _ ^ ^
[xt ]Θ = [ xt ]Θ resp. Θ
[xt ] = [ xt ] Θ

Proof The proofs of the equations follow easily from Proposition 53. 
How can we describe complete tolerance relations of concept lattices in terms
of the contexts?
Definition 59 By a block relation of a context (G, M, I) we mean a relation J ⊆
G × M which satisfies the following conditions:
1. I ⊆ J,
2. for every object g ∈ G, g J is an intent of (G, M, I),
3.2 Complete congruences and tolerances 135

3. for every attribute m ∈ M , mJ is an extent of (G, M, I). ♦

We can use this definition as a starting point for some observations: If J is a
block relation of (G, M, I), then every extent of (G, M, J) is an extent of (G, M, I)
and every intent of (G, M, J) is an intent of (G, M, I). The intersection T of any
number of block relations of (G, M, I) is again a block relation, since g Jt
=
g , and the intersection of intents is always an intent, and dually. Hence, the
T Jt
block relations of (G, M, I) form a closure system and thus a complete lattice.

f
@
@
c f fd @ fe
@ 7 @
@ fg
@ @
a f b f ff
@f a b c d e f g
1 × • × • •
1 @ 2 5 6
@ 2 • × × • •
@ @ 3 × × × × × • •
@f @f 4 • • × × × × ×
3 @ 4 5 • • × × •
@ 6 • • × • ×
@f 7 • × •

Figure 3.8 The block relation J belonging to the tolerance from Figure 3.6 additionally
contains the pairs marked by dots.

Theorem 15 The lattice of all block relations of (G, M, I) is isomorphic to the lattice
of all complete tolerance relations of B(G, M, I). The map β assigning to any complete
tolerance relation Θ the block relation defined by

g β(Θ) m : ⇔ γg Θ (γg ∧ µm) ( ⇔ (γg ∨ µm) Θ µm)

is an isomorphism. Conversely,

(A, B) β −1 (J) (C, D) ⇔ A × D ∪ C × B ⊆ J

yields the tolerance corresponding to a block relation J.

Proof First, we show that J := β(Θ) is a block relation. Since Θ is reflexive,
I ⊆ J. According to the definition,

g J = {m ∈ M | γg Θ (γg ∧ µm)}.

We claim that this is an intent of (G, M, I). For this purpose we consider the
concept
136 3 Parts, factors, and bonds
^
{µm | γg Θ (γg ∧ µm)} = (g JI , g JII ).
If n is an attribute of this concept, we get
^
µn ≥ {µm | γg Θ (γg ∧ µm)},

and hence also

^
γg ∧ µn ≥ {γg ∧ µm | γg Θ (γg ∧ µm)}.

If we are aware that this infimum is in a Θ-relation with γg, we recognize that
(γg ∧ µn) Θ γg, i.e., that n ∈ g J . Hence, g J = g JII is an intent. Dually we prove
that every set of the form mJ is an extent of (G, M, I).
Now we start from a block relation J and define a relation τ (J) on B(G, M, I)
by
(A, B) τ (J) (C, D) : ⇔ A × D ∪ B × C ⊆ J.
Evidently τ (J) is reflexive and symmetric, and, if T is an index set and there are
concepts with (At , Bt ) Θ (Ct , Dt ) for t ∈ T , we argue as follows: For g ∈
T At we
have g J ⊇ Dt = T CtI and consequently g JI ⊆ CtII = Ct . Hence, for g ∈ t∈T At
we obtain g JI ⊆ t∈T Ct , i.e.,
!I
\
J JII
g =g ⊇ Ct ,
t∈T

which proves
!I
\ \
At × Ct ⊆ J.
t∈T t∈T

Analogously we show that

!I
\ \
Ct × At ⊆ J,
t∈T t∈T

and altogether we have proved

^ ^
(At , Bt ) τ (J) (Ct , Dt ).
t∈T t∈T

The dual argument proves, moreover, that τ (J) is compatible with suprema, i.e.,
that it is a complete tolerance relation.
Both maps β and τ are evidently order-preserving. In order to prove the the-
orem we furthermore have to show that they are inverse to each other. Let Θ be
a complete tolerance relation of B(G, M, I). We want to show that
3.2 Complete congruences and tolerances 137

(A, B) Θ (C, D) ⇔ (A, B) τ (β(Θ)) (C, D).

According to Proposition 51, we may limit ourselves to the special case (A, B) >
(C, D). We have

(A, B) Θ (C, D) ⇔ γg ∨ (C, D) Θ (C, D) for all g ∈ A

⇔ γg Θ γg ∧ (C, D) for all g ∈ A
⇔ γg Θ γg ∧ µm for all g ∈ A and m ∈ D
⇔ A × D ⊆ β(Θ)
⇔ (A, B) τ (β(Θ)) (C, D).

For the last part of the proof let J be a block relation of (G, M, I). Then

(g, m) ∈ J ⇔ g ∈ mJ
⇔ g II ⊆ g JJ ∩ mJ and
(g II ∩ mI )I = (g I ∪ mII )II ⊆ (g J ∪ mJJ )JJ = (g JJ ∩ mJ )J
⇔ g II × (g II ∩ mI )I ⊆ J
⇔ (γg) τ (J) (γg ∧ µm)
⇔ (g, m) ∈ β(τ (J)).

This proves that J = β(τ (J)). 

Corollary 1 If Θ is a tolerance relation on B(G, M, I) and J := β(Θ) is the corre-

sponding block relation, then

B(G, M, I)/Θ ∼
= B(G, M, J).

More precisely, we have:

1. (B I , B), (C, C I ) is a block of Θ if and only if (C, B) is a concept of (G, M, J).
 

2. The map  I
(B , B), (C, C I ) 7→ (C, B)


is an isomorphism of the lattice of the blocks of Θ onto the concept lattice of (G, M, J).
3. If (C, B) is a concept of (G, M, J), then
 I
(B , B), (C, C I ) = B(C, B, I ∩ C × B)


for the corresponding block of Θ

Proof According to Theorem 15, two concepts (A, B) ≤ (C, D) of B(G, M, I)

stand in the relation Θ to each other if and only if C × B ⊆ J, i.e., if B ⊆ C J and
C ⊆ B J . For an arbitrary concept (X, Y ) of (G, M, I) therefore

(X, Y )Θ = (X JI , X J ) and
Θ
(X, Y ) = (Y , YJ JI
) and consequently
138 3 Parts, factors, and bonds

((X, Y )Θ )Θ = (X JJ , X JJI ).

If we assume that B := X J and C := X JJ , then (C, B) is a concept of (G, M, J)

and the block [(X, Y )]Θ proves to be of the form we claimed:

[(X, Y )]Θ = (X, Y )Θ , ((X, Y )Θ )Θ = (B I , B), (C, C I ) .

   

Therefore, the map

[(X, Y )]Θ 7→ (X JJ , X J )
is an order isomorphism mapping the blocks of Θ onto the concepts of (G, M, J).
The third part of the assertion follows from Proposition 61. 

c, d, e
g
7@ c d e f g
a, b @ gf, g
g
@ 3×××
1,2 @ 5,6 4×××××
5 ××
@g
@ 6 × ×
3,4

Figure 3.9 The concept lattice B(G, M, J) of the block relation J is isomorphic to the fac-
tor lattice by the tolerance relation. As an example, we state the subcontext belonging to
the concept ({3, 4, 5, 6}, {c, d, e, f, g}) of J . Its concept lattice is isomorphic to the corre-
sponding block of the tolerance.

We close this section with two observations. The first deals with the question
which families of intervals form the systems of the blocks of a tolerance. This is
answered by the following theorem:
Theorem 16 Let L be a complete lattice, T an index set and

F := {[xt , xt ] | t ∈ T }

a family of intervals from L which are assumed to be pairwise distinct, i.e., s 6= t ⇒

[xs , xs ] 6= [xt , xt ]. Then the following conditions are equivalent:
1. F is the family of the blocks of a complete tolerance relation on a complete sublattice
of L.
2. a) The set {xt | t ∈ T } of the upper bounds of the intervals is W -closed.
V
b) The set {xt | t ∈ T } of the lower bounds of the intervals is -closed.
c) The upper and the lower bounds are ordered in the same way, i.e., xs ≤ xt ⇔
xs ≤ xt .
3.2 Complete congruences and tolerances 139

3. There is an order ≤ on T with respect to which (T, ≤) is a complete lattice and there
are maps
α : (T, ≤) → L, injective and -preserving,
W

α : (T, ≤) → L, injective and -preserving,

V

with α(s) ≤ α(t) ⇔ α(s) ≤ α(t), and F = {[α(t), α(t)] | t ∈ T }.

Proof 1 ⇒ 2: Every block of a complete tolerance relation Θ is of the form
[x]Θ = [(xΘ )Θ , xΘ ], i.e., the upper bounds of the blocks are precisely the elements
of the form xΘ . From Proposition 53 it follows that
^ ^
{xΘt | t ∈ T} = ( {xt | t ∈ T })Θ ,

i.e., the result is again an upper bound of a block. This proves a) and dually we
infer b). c) again follows from Proposition 53: xΘ ≤ y Θ ⇔ (xΘ )Θ ≤ (y Θ )Θ .
2 ⇒ 3: We order T by s ≤ t : ⇔ xs ≤ xt . Because of c) this is equivalent to
xs ≤ xt . Hence, the map defined by α(t) := xt is an order-isomorphism of (T, ≤)
on {xt | t ∈ T }, correspondingly α(t) := xt defines an order-isomorphism of
(T, ≤) on {xt | t ∈ T }. Therefore, according to a) and b), (T, ≤) is a complete
lattice.
3 ⇒ 1: We first show that under the conditions mentioned under 3, the set
[
U := {[α(t), α(t)] | t ∈ T }

is a complete sublattice of L. Hence, let xs , s ∈ S be a sequence of elements W from

U . For every s ∈ S there is a ts ∈ T with xs ∈ [α(ts ), α(ts )]. We W
claim that s∈S xs
also
W lies in such an interval, namely in [α(t), α(t)] with t := s∈S ts . Since α is
-preserving, α(t) ≤ s∈S xs is obvious, and since α is W order-preserving, for
W
every s ∈ S we have xs ≤ α(ts ) ≤ α( s∈S ts ) and thus s∈S xs ≤ α(t). Hence,
W
U is closed with regard to suprema as well as with respect to infima, as the dual
argument shows.
The relation
Θ := {(x, y) | ∃t∈T x, y ∈ [α(t), α(t)]}
evidently is reflexive and symmetric. Moreover, from (xs , ys ) ∈ Θ, s ∈ S it fol-
lows that xs , ys ∈ [α(ts ), α(ts )] holds for suitable ts ∈ T , i.e.,
_ _ _ _
{xs | s ∈ S}, {ys | s ∈ S} ∈ [α( ts ), α( ts )],

i.e., _ _
( {xs | s ∈ S}, {ys | s ∈ S}) ∈ Θ.
Hence, Θ is -compatible (and dually, of course, -compatible as well), i.e., it is
W V
a complete tolerance.
Finally we have to show that the intervals [α(t), α(t)] are in fact the blocks of Θ,
i.e., the maximal sets of elements which are pairwise related under the relation
Θ. If [u, v] is a block, then (u, v) ∈ Θ, i.e., u, v ∈ [α(t), α(t)] for some t, hence every
140 3 Parts, factors, and bonds

block of Θ is of this form. On the other hand, every [α(t), α(t)] is maximal too;
from [α(s), α(s)] ⊆ [α(t), α(t)] we can infer α(s) ≤ α(t), that is s ≤ t, as well as
α(t) ≤ α(s), that is t ≤ s. Together this yields s = t. 
The second observation establishes a link between compatible subcontexts and
block relations. Transitive tolerance relations are congruence relations, hence we
can describe a complete congruence relation (under suitable conditions, cf. The-
orem 12) in two ways: in terms of a compatible subcontext and in terms of a block
relation.
Proposition 54 Let Θ be a complete congruence relation of a doubly founded concept
lattice B(G, M, I), (GΘ , MΘ , I ∩ GΘ × MΘ ) the corresponding saturated compatible
subcontext and J = β(Θ) the block relation for Θ. Then:

(g, m) ∈ J ⇔ g 00 ∩ GΘ ⊆ m0
⇔ m00 ∩ MΘ ⊆ g 0
GΘ = {g ∈ G | g I = g J }
MΘ = {m ∈ M | mI = mJ }.

Proof According to Theorem 15, (g, m) ∈ J ⇔ (γg, γg ∧ µm) ∈ Θ. Two concepts

are congruent if and only if their extents have the same intersection with GΘ . In
the present case, this means g 00 ∩ GΘ = g 00 ∩ m0 ∩ GΘ which corresponds to the
first line of the assertion. The second line can be inferred dually.
According to the definition, g is in GΘ if and only if γg is the smallest element
of a Θ-class. This is equivalent to the fact that for every attribute m from (γg, γg ∧
µm) ∈ Θ it already follows that µm ≥ γg, i.e., that (g, m) ∈ J always implies
(g, m) ∈ I. 
The connection between congruence relations and block relations, according
to the theorem, can also be explained as follows: If ϕ is an homomorphism, Θ =
ker ϕ and J is the block relation belonging to Θ, then

(g, m) ∈ J ⇔ ϕγg ≤ ϕµm.

3.3 Closed subrelations

Definition 60 A relation J ⊆ I is called a closed relation of the context (G, M, I)

if every concept of the context (G, M, J) is also a concept of (G, M, I). ♦
Theorem 17 If J is a closed relation
S of (G, M, I), then B(G, M, J) is a complete sub-
lattice of B(G, M, I) with J = {A × B | (A, B) ∈ B(G, M, J)}. Conversely, for
every complete sublattice U of B(G, M, I) the relation
[
J := {A × B | (A, B) ∈ U }
3.3 Closed subrelations 141

is closed and B(G, M, J) = U .

Proof Let J be a closed relation of (G, M, I). According to the definition,

B(G, M, J) is a subset of B(G, M, I) containing (M I , M ) = (M J , M ) as well as
(G, GI ) = (G, GJ ).4 The characterization of the suprema and infima in the Ba-
sic Theorem Sshows that B(G, M, J) is a complete sublattice of B(G, M, I). The
relation J = {A × B | (A, B) ∈ B(G, M, J)} holds for every context (G, M, J).
Now, conversely, let U be a complete sublattice and
[
J := {A × B | (A, B) ∈ U }.

We have to show that J is a closed relation with U = B(G, M, J). It is evident that
U ⊆ B(G, M, J). Thus, it remains to be shown that every concept of (G, M, J)
belongs
T to U . We first prove this for the object concepts: Assume that g ∈ G and
D := {A | (A, B) ∈ U, g ∈ A}. D is an extent of (G, M, J), and consequently
g JJ ⊆ D. For every attribute m ∈ g J there exists a concept (A, B) ∈ U with
(g, m) ∈ A×B and because of D ⊆ A it follows that m ∈ DJ . Therefore, g J = DJ
and g JJ = D. This shows that for every g ∈ G the concept (g JJ , g J ) belongs to U .
Every concept of B(G, M, J) is however the supremum of such object concepts,
thus U ⊇ B(G, M, J), which remained to be proved. 

This means that the closed relations are in a 1-1 correspondence to the com-
plete sublattices. The map
[
C(U ) := {A × B | (A, B) ∈ U }

maps the set of complete sublattices bijectively onto the map of closed relations
of B(G, M, I).S It is furthermore order-preserving, U1 ⊆ U2 ⇔ J1 ⊆ J2 . However,
C is neither – nor –preserving. The intersection of closed relations does not
T
necessarily have to be closed, the closed relations in general do not form a closure
system. This is surprising in so far as the family of complete sublattices does
form a closure system: for every subset T of a complete lattice, the intersection
of all lattices containing T is also a sublattice (namely the complete sublattice
generated by T ).
If F is a family of closed relations and D := F, then there is, nonetheless,
T
always a largest closed relation in D, namely
[
J := {A × B | (A, B) concept, A × B ⊆ D}.

This is easily seen if we consider the following: If (A, B) is a concept of (G, M, I)

and J is a closed relation with A × B ⊆ J, then (A, B) is also a concept of
(G, M, J). Thus, the concepts (A, B) with A × B ⊆ D are precisely those which
are contained in each of the sublattices B(G, M, L), L ∈ F. This means that they

4In conformity with Definition 19 we write X I or X J instead of X 0 , in order to make clear

when we are referring to the context (G, M, I) or to (G, M, J), respectively.
 142 3 Parts, factors, and bonds

form precisely the intersection of those sublattices, i.e., they themselves are a
complete sublattice. These considerations yield the following proposition:
Proposition 55 For every set T ⊆ B(G, M, I) of concepts, there is a smallest closed
relation J of (G, M, I) containing all sets A × B with (A, B) ∈ T . B(G, M, J) is the
complete sublattice of B(G, M, I) generated by T . 

a b c d e f g h
1 × ×
2×××
3××⊗ ⊗
4 ⊗⊗⊗
5 ⊗⊗⊗⊗⊗
6 ⊗⊗ ⊗ ⊗
7 × ××
8 ××××
9 ⊗⊗⊗⊗⊗⊗
10 × × ×

Figure 3.10 Example of a closed relation in a context, from [136].

How can we recognize whether a relation is closed? A first clue is provided

by the next proposition.
Proposition 56 A subrelation J ⊆ I is closed if and only if

X JJ ⊇ X JI

holds for each subset X ⊆ G and for each subset X ⊆ M .

Proof (X JJ , X J ) is a concept of (G, M, I), if and only if X JJ = X JI and X JJI =
X J . If we set Y := X J the second condition can be rewritten as Y JI = Y JJ
because of X J = X JJJ . However, the inclusion Y JJ ⊆ Y JI holds for every
subrelation. 

The following characterization is somewhat more ambitious:

Proposition 57 The closed relations of a context (G, M, I) are precisely those subrela-
tions J ⊆ I which satisfy the following condition:
(C) (g, m) ∈ I \ J implies (h, m) 6∈ I for some h ∈ G with g J ⊆ hJ as well as (g, n) 6∈ I
for some n ∈ M with mJ ⊆ nJ .

Proof Let J be a closed relation of (G, M, I) and assume that (g, m) ∈ I \ J.

(g JJ , g J ) is a concept of (G, M, I), i.e., g J = g JJI . Since m ∈
6 g J , there is some
3.3 Closed subrelations 143

u
H g
e e
@He Hf
d e @H H ec
H H H
e @H e HH
HH HH HHH HH
HuH @eH
@ H HH H HH
H HHH
H b
H H
@e H Hu H
 uHHHe7HHeH
@ @ HHH@ HHH
H
H 4
8 HH@  HHHHu H
@H
H eH
H u H10 H ea
H
Hu
H@ H@ H H
H
@ H H
6 H
5 H He 2
 H
 HH H
h  3
eP Hu
H
1 PP 9
PP
Pu
PP

Figure 3.11 Diagram of the concept lattice for the context from Figure 3.10. The sublattice
consisting of the blackened elements, belongs to the above-mentioned closed relation.

h ∈ g JJ with m 6∈ hI , i.e., with (h, m) 6∈ I and g J ⊆ hJ . The second part of (C)

follows dually.
Conversely, let J ⊆ I be a relation satisfying (C) and let (A, B) be a concept of
(G, M, J). We have to show that (A, B) is a concept of (G, M, I), i.e., that A = B I
and B = AI . B ⊆ AI is trivial, we show B ⊇ AI . If we assume that there is
an attribute m ∈ AI which is not an element of B = AJ , then there should be
an object g ∈ A with (g, m) 6∈ J but (g, m) ∈ I. By means of condition (C) we
should find some h ∈ G with m 6∈ hI and hJ ⊇ g J ⊇ B. Because of hJ ⊇ B,
however, in this case we should obtain h ∈ A which would contradict m ∈ AI .
A = B I is proved dually, i.e., J is closed. 

Proposition 58 If J is a closed relation and

(H, N, I ∩ H × N )

is a compatible subcontext of (G, M, I), then J ∩H ×N is a closed relation of (H, N, I ∩

H×N ) and (H, N, J ∩ H × N ) is a compatible subcontext of (G, M, J).

Proof If (A, B) is a concept of (G, M, I), then (A ∩ H, B ∩ N ) is a concept of

(H, N, I ∩ H×N ). This holds in particular for the concepts of (G, M, J), in which
case (A ∩ H, B ∩ N ) is even a concept of (H, N, J ∩ H × N ). Each concept of
(H, N, J ∩ H × N ) originates in this way, i.e., J ∩ H × N is closed and (H, N, J ∩
H × N ) is compatible. 
The proposition has the following background: A homomorphism maps sub-
lattices onto sublattices. If (H, N, I ∩ H×N ) is a compatible subcontext and J is
a closed relation, then ΠH,N maps the sublattice B(G, M, J) onto the sublattice
B(H, N, J ∩ H × N ).
144 3 Parts, factors, and bonds

Proposition 58 contains as a special case the statement that a closed relation

remains closed if we omit reducible objects and attributes. By Proposition 44, a
dense subcontext is always compatible. In the following proposition we establish
a connection between closed relations and the arrow relations. First, we shall
explain an abbreviation used in this connection:

. ∪ %:= {(g, m) | g . m or g % m}.

In the following proposition this refers to the arrow relations in the context
(G, M, J):
Proposition 59 Let (G, M, J) be a doubly founded clarified context. Then the following
statement holds: J is a closed relation of (G, M, I) if and only if

J ⊆ I ⊆ G × M \ (. ∪ %).

Proof If J is a closed relation of (G, M, I) and (g, m) ∈ I \ J, then by Proposi-

tion 57 there exists some h with (h, m) 6∈ J and g J ⊆ hJ , i.e., g J ⊂ hJ and thus
(g, m) 6∈..
If, on the other hand, J ⊆ I ⊆ G × M \ (. ∪ %), then, according to Propo-
sition 56, it suffices to show for given X ⊆ G (and dually for X ⊆ M ) that
X JJ ⊇ X JI . Hence, assume that X ⊆ G, B := X J and g ∈ B I . If we had g 6∈ B J ,
then there would be an attribute m ∈ B with (g, m) 6∈ J and furthermore, on
account of the doubly-foundedness, an attribute n with g % n and nJ ⊇ mJ , i.e.,
in particular n ∈ B and consequently (g, n) ∈ I, in contradiction to g % n. 

Intervals

Full rows and full columns of a context belong to every closed relation and it is
sometimes awkward to have to carry them along. For simplification purposes,
we therefore occasionally use the notation

:= M 0 × M ∪ G × G0 .

The relation consists precisely of the trivial incidences in I. In the following

proposition we simply assume that = ∅ and give some simple examples of
closed relations.
Proposition 60 If (A, B) and (C, D) are concepts of a context (G, M, I) with G0 =
∅ = M 0 , then

I ∩ A × M, I ∩ G × D, I ∩ (A × M ∪ G × D)

and, if (A, B) ≤ (C, D), even I ∩ C × B are closed relations with

B(G, M, I ∩ A × M ) = {(G, ∅)} ∪ ((A, B)]
3.3 Closed subrelations 145

B0
C

C0

Figure 3.12 With reference to Proposition 61: Between (B 0 , B) and (C, C 0 ) lie precisely
the concepts of the context (C, B, I ∩ C × B).

B(G, M, I ∩ G × D) = {(∅, M )} ∪ [(C, D))

B(G, M, I ∩ (A × M ∪ G × D)) = ((A, B)] ∪ [(C, D))
B(G, M, I ∩ C × B) = {(∅, M ), (G, ∅)} ∪ [(A, B), (C, D)] .
Proof It suffices to undertake the proof for J := I ∩ C × B. It is clear that
(G, ∅) and (∅, M ) are concepts of B(G, M, J). Furthermore every concept (X, Y )
∈ B(G, M, I) with (A, B) ≤ (X, Y ) ≤ (C, D) is also a concept of (G, M, J), since
X × Y ⊆ I ∩ C × B ⊆ J. Hence, assume that (X, Y ) ∈ B(G, M, J). We may
assume that X ⊆ C and Y ⊆ B. Because of A × B ⊆ J, Y ⊆ B immedi-
ately yields A = B J ⊆ Y J = X, i.e., A ⊆ X and thus X I ⊆ AI = B. With
X J = X I ∩ B it follows that X I = X J . Dually, we recognize that Y I = Y J , i.e.,
that (X, Y ) ∈ B(G, M, I). 
As an immediate consequence we obtain:
Proposition 61 If (A, B) and (C, D) are concepts of (G, M, I) with (A, B) ≤ (C, D),
then
[(A, B), (C, D)] = B(C, B, I ∩ C × B).

Example We demonstrate this proposition by means of the concepts γ5 and µc
in the context from Figure 3.10. We have

γ5 = (500 , 50 ) = ({5, 9}, {c, d, e, f, g}),

µc = (c0 , c00 ) = (G \ {8}, {c}).

According to the proposition [γ5, µc] = B(C, B, I ∩ C × B) with

C := {1, 2, 3, 4, 5, 6, 7, 9, 10} and B := {c, d, e, f, g}.

146 3 Parts, factors, and bonds

It turns out that in this subcontext the objects 1, 2, 3, 5, 9 and 10 as well as the
attribute c are reducible. The reduced context is presented in Figure 3.13 together
with its concept lattice. It is an interval in the center of Figure 3.11.

c
d e f g
4 × × e c@ cg
@
6 × × dc @@c @ @ cf
7 × × 4 HH 6  7
H c

Figure 3.13 The concept lattice of this subcontext is isomorphic to an interval.

Further examples of closed relations are obtained from formal concepts in-
variant under an automorphism group and from dismantling:
If Γ is a group of automorphisms of the context (G, M, I), i.e., of pairs of maps
(α, β) with

α : G → G, β : M → M, g I m ⇔ α(g) I β(m),

then we obtain a closed relation IΓ by means of the definition

(g, m) ∈ IΓ : ⇔ g I β(m) for all (α, β) ∈ Γ

( ⇔ α(g) I m for all (α, β) ∈ Γ ),

as can be easily proved. The corresponding sublattice B(G, M, IΓ ) consists pre-

cisely of those concepts (A, B) of (G, M, I) which are invariant under Γ , i.e., for
which
(α(A), β(B)) = (A, B) holds for all (α, β) ∈ Γ.

Dismantling

It may happen that a closed relation differs very little from the incidence relation
I, in the extreme case only by one “cross”. This case corresponds to the disman-
tling of doubly irreducible elements. Therefore, we shall give a short description,
only sketching the order-theoretic results and referring to the corresponding lit-
erature for the proofs.
An element a of an ordered set shall be called doubly irreducible if a has ex-
actly one lower neighbor a∗ and exactly one upper neighbor a∗ and furthermore
the conditions
x < a ⇒ x ≤ a∗ , x > a ⇒ x ≥ a∗
3.4 Bonds and connections 147

(which are dispensable in the finite) are satisfied. According to Proposition 2, in

Wcomplete lattice the doubly
a V irreducible elements are precisely those which are
-irreducible as well as -irreducible.
We talk about the dismantling of a doubly irreducible element a in an ordered
set (P, ≤) if we mean the transition from (P, ≤) to the ordered set

(P \ {a}, (P \ {a})2 ∩ ≤).

In the following we shall write (P \ {a}, ≤) for this ordered set .

If we dismantle a doubly irreducible element a of a complete lattice (L, ≤),
then we obtain a complete sublattice L \ {a}. Obviously, the property of being
doubly irreducible is also necessary for this purpose. We get a further converse:
Proposition 62 If a is a doubly irreducible element of (P, ≤), then (P \ {a}, ≤) is a
complete lattice if and only if (P, ≤) is a complete lattice. 
We omit the (easy) proof and point to an application instead: If we want to
determine whether a given ordered set (P, ≤) is a complete lattice, we can first
remove doubly irreducible elements and then examine the remaining structure.
If (P, ≤) is finite we can gradually dismantle all doubly irreducible elements until
there finally remains a DI-kernel without doubly irreducible elements.
It can be shown by means of a simple argument that the DI-kernel is unique,
i.e., that it does not depend on the order in which the doubly irreducible elements
are being dismantled.
Dismantling an element corresponds to canceling a cross in the context:
Proposition 63 If a = γg = µm is a doubly irreducible concept of a clarified context
(G, M, I), then
B(G, M, I) \ {a} = B(G, M, I \ {(g, m)}).
Proof We have already noted that B(G, M, I) \ {a} is a complete sublattice. By
Theorem 17 the corresponding closed relation is given by
[
J := {A × B | (A, B) 6= a}.

Now, if (h, n) ∈ I is an arbitrary incident object-attribute-pair, then (h, n) ∈ h00 ×

h0 and (h, n) ∈ n0 × n00 . Hence, from (h, n) 6∈ J it follows that (h00 , h0 ) = a =
(n0 , n00 ), i.e., (since (G, M, I) is clarified) h = g and n = m. 

3.4 Bonds and connections

Homomorphisms, i.e., structure-preserving mappings, are an important tool

that allows concept lattices to be compared and connected. Structure-preserving
mappings between formal contexts are also of interest. They will be discussed
later in Section 7.2. However, it is also possible to connect formal contexts by
means of relations, and the first foundations for this will be laid here.
148 3 Parts, factors, and bonds

Bonds between formal contexts

We call bonds those relations between formal contexts that generate adjunctions
between the associated concept lattices. This connection is made precise by
Proposition 66. But before we clarify some basic definitions and facts.

Definition 61 A bond from (G, M, I) to (H, N, J) M N

is a relation R ⊆ G × N for which it holds that
G I R
• for every g ∈ G, g R is an intent of (H, N, J) and
• for every n ∈ N , nR is an extent of (G, M, I).
H J

♦
This is obviously equivalent to:
• Every extent of (G, N, R) is an extent of (G, M, I) and
• every intent of (G, N, R) is an intent of (H, N, J).
Since intersecting extents yields extents and intersecting intents gives intents,
the intersection of bonds is always a bond. From this follows the next proposi-
tion:
Proposition 64 The set of all bonds from (G, M, I) to (H, N, J) is a closure system on
G × N. 
The corresponding closure operator is easy to describe, since a subset of G ×
N is closed if and only if there is an intent of (H, N, J) in each “row” and an
extent of (G, M, I) in each “column”. Thus, the extent closure operator of the
first context is to be applied column-wise and the intent closure operator of the
second context is to be applied row-wise. This usually has to be repeated several
times alternately.
Bonds were introduced by R. Wille [393] to describe subdirect products of
concept lattices, see Chapter 5 for an introduction. Subdirect products are certain
complete sublattices of the direct product, and since direct products correspond
to context sums and complete sublattices to closed relations, the following is natural
(and an instance of Theorem 37 (p. 218)):
Proposition 65 R ⊆ G × N is a bond from (G, M, I)
to (H, N, J) if and only if M N

C := I ∪ R ∪ J ∪ (H × M ) G I R
is a closed relation of the context sum

(G, M, I) + (H, N, J) H @ J
@

(assuming G ∩ H = ∅ = M ∩ N ).
3.4 Bonds and connections 149

Proof The formal concepts of the sum (cf. Definition 30) are precisely the pairs
of the form
(A1 ∪ A2 , B1 ∪ B2 ),
where (A1 , B1 ) ∈ B(G, M, I) and (A2 , B2 ) ∈ B(H, N, J). Now let (U, V ) be a
concept of (G, N, R). Then (U ∪ V J , V ∪ U J ) is a concept of

(G ∪ H, M ∪ N, C).

In order for this to be a concept of the sum, it is necessary that U is an extent of

(G, M, I) and that V is an intent of (H, N, J). In other words, R must be a bond.
But conversely, if R is a bond, and if (A, B) is a concept of

(G ∪ H, M ∪ N, C),

then let U := A ∩ G and V := B ∩ N . Since

U = (B ∩ M )I ∩ V R ,

U must be an extent of (G, M, I) and B ∩ M = U I . Dually we get that V is

an intent of (H, N, J) and V J = A ∩ H. Thus (A, B) is a concept of the sum
(G, M, I) + (H, N, J). 
Remark: Proposition 65 does not claim that all closed relations of (G, M, I) +
(H, N, J) are of this form; this is usually not the case. It is not difficult to char-
acterize such closed relations. Since H × M ⊆ C holds, they must contain the
concept (M I ∪ H, M ∪ H J ), but also (as a closer look reveals) every subconcept
and every superconcept of this concept. This characterizes such closed relations.

Adjunctions between concept lattices were introduced on Page 19, see also
Proposition 11 there. They are in 1-1-correspondence to bonds, as the following
result shows.
Proposition 66 From every bond R ⊆ G × N between formal contexts (G, M, I) and
(H, N, J) we obtain an adjunction, i.e., a pair (ϕR , ψR ) of mappings

ϕR : B(G, M, I) → B(H, N, J), ψR : B(H, N, J) → B(G, M, I),

such that ϕR is a -morphism and ψR is a -morphism residual to ϕR , by

W V

ϕR (A, AI ) := (ARJ , AR ), ψR (B J , B) := (B R , B RI ).

Each adjunction is obtained in this way from exactly one bond, in fact,

R = {(g, n) | ϕγg ≤ µn}

= {(g, n) | γg ≤ ψµn}.

For the proof see that of Theorem 20 below, where Corollary 2 states just the dual
of Proposition 66. 
150 3 Parts, factors, and bonds

Adjunctions differ from Galois connections only in that the connecting map-
pings are order-preserving instead of order-reversing. This has an important con-
sequence: adjunctions can be concatenated.
The category of concept lattices with adjunctions as morphisms has remark-
able properties (∗ -autonomous, see Mori [288], Borgwardt [64]). Since (as stated
in Proposition 66) such adjunctions are represented by bonds, it is natural to ask
for a concatenation of bonds. However, the usual relation product does not come
into question for this, simply because bonds do not fit together appropriately;
they always go from objects to attributes. The following definition, on the other
hand, will work.
Definition 62 Let Ki := (Gi , Mi , Ii ) (i ∈ {1, 2, 3}) be formal contexts and let
J12 ⊆ G1 × M2 and J23 ⊆ G2 × M3 be relations. We define

J12 ◦ J23 := {(g, m) ∈ G1 × M3 | g J12 I2 ⊆ mJ23 }.

♦
Proposition 67 (cf. Figure 3.14) If J12 is a bond, then
[
J12 ◦ J23 = B J12 × AJ23 ,
(A,B)∈B(K2 )

Proof If (A, B) ∈ B(K2 ) and g ∈ B J12 then B ⊆ g J12 and thus A ⊇ g J12 I2 . Any
m ∈ AJ23 therefore satisfies g J12 I2 ⊆ mJ23 , which proves that B J12 × AJ23 ⊆
J12 ◦ J23 . For the other direction let g ∈ G1 , A := g J12 I2 and B := AI2 . Then
(A, B) ∈ B(K2 ), and, since g J12 is an intent of K2 , B = AI2 = g J12 I2 I2 = g J12 ,
which implies g ∈ B J12 . If mJ23 contains g J12 I2 for some m ∈ M3 , then mJ23
contains A (= g J12 I2 ) and thus m ∈ AJ23 . 

B J12
J12 J12 ◦ J23

I2 J23 A

B AJ23

Figure 3.14 With reference to the definition of J12 ◦ J23 .

3.4 Bonds and connections 151

Theorem 18 If J12 is a bond from K1 to K2 and J23 is a bond from K2 to K3 , then

J12 ◦ J23 is a bond from K1 to K3 .

M1 M2 M3 M1 M2 M3

g
G1 I1 J12 G1

G2
n
G2 I2 J23 g J12 I2
mJ23
|{z}

G3 I3 g J12 G3

|{z} m
g J12 I2 J23

Figure 3.15 With reference to the proof of Proposition 68.

To prove this theorem, we first need a proposition.

Proposition 68 If J12 is a bond from K1 to K2 and if J23 is a bond from K2 to K3 , then

for g ∈ G1 , m ∈ M3 the following holds:

m ∈ g J12 I2 J23 ⇔ mJ23 ⊇ g J12 I2 ⇔ g J12 ⊇ mJ23 I2 ⇔ g ∈ mJ23 I2 J12 .

Proof (cf. Figure 3.15) The equivalence in the middle

mJ23 ⊇ g J12 I2 ⇔ g J12 ⊇ mJ23 I2

follows immediately from Proposition 12 on p. 24, because g J12 is an intent and

mJ23 is an extent of K2 . The other two equivalences are dual to each other and
result easily from the definitions. For example,

m ∈ g J12 I2 J23 ⇔ mJ23 ⊇ g J12 I2

simply reflects the fact that m belongs to the J23 -intent of g J12 I2 iff the J23 -extent
of m contains g J12 I2 . 

Proof of Theorem 18: According to Definition 62 we have for fixed g ∈ G1 that

{m ∈ M3 | (g, m) ∈ J12 ◦ J23 } = {m ∈ M3 | g J12 I2 ⊆ mJ23 },

which by Proposition 68 this is equal to g J12 I2 J23 , the latter being an intent of K3 .
Dually one shows that for fixed m ∈ M3
152 3 Parts, factors, and bonds

{g ∈ G1 | (g, m) ∈ J12 ◦ J23 } = mJ23 I2 J12 ,

an extent of K1 . Thus J12 ◦ J23 is a bond. 

I1 J12 J13

I2 J23

I3

Figure 3.16 With reference to Proposition 69.

Proposition 69 Let Ki := (Gi , Mi , Ii ), i ∈ {1, 2, 3} be formal contexts, and let (see

Figure 3.16)
• J12 be a bond from K1 to K2 ,
• J23 be a bond from K2 to K3 , and
• J13 be a bond from K1 to K3 .
Moreover, let (ϕ12 , ψ12 ), (ϕ23 , ψ23 ), and (ϕ13 , ψ13 ) be the adjunctions corresponding
to5 these bonds according to Proposition 66. Then

ϕ13 = ϕ23 ◦ ϕ12 and ψ13 = ψ12 ◦ ψ23

if and only if
J13 = J12 ◦ J23 .
Proof (g, m) is an element of the bond corresponding to ϕ23 ◦ϕ12 iff ϕ23 ◦ϕ12 γg ≤
µm. The intent of ϕ23 ◦ϕ12 γg is g J12 I2 J23 . It was shown in the proof of Theorem 18
that this is equal to
{m ∈ M3 | (g, m) ∈ J12 ◦ J23 }.
So J13 is indeed the bond corresponding to ϕ23 ◦ ϕ12 and ψ12 ◦ ψ23 . 
Remark: The results outlined here can be expressed particularly clearly in the
language of mathematical Category Theory. The formal contexts as objects form
a category with the bonds as morphisms. This category is naturally equivalent
to the category of complete lattices with the residuated mappings and there-
fore inherits the special properties of that category (see Theorem 3.8 of Borg-
wardt [64]). The functors that establish this equivalence are K 7→ B(K) on the
one hand and (L, ≤) 7→ (L, L, ≤) on the other, with the actions on the morphisms
explained above.
5 Here ϕij is an abbreviation for ϕJij and ψij abbreviates ψJij .
3.4 Bonds and connections 153

The identical mapping is residuated for each concept lattice B(G, M, I). The
corresponding bond is the incidence relation I. Other self-bonds are, e.g., the
closed relations and the block relations (see Definitions 60 and 59).

Proposition 70 (from [64]) Let K := (G, M, I) be a formal context, J be a bond from

K to K, and ϕJ be the corresponding residuated map as in Proposition 66. Then
• ϕJ is monotone,
• ϕJ is extensive iff J ⊆ I,
• ϕJ is intensive6 iff I ⊆ J,
• ϕJ is idempotent iff B = B IJ holds for all intents B of (G, M, J), and
• ϕJ is a closure operator iff J is a closed relation.


Relational Galois connections

The basic construction for Formal Concept Analysis was described in Section 1.1:
From a relation I ⊆ G × M between two sets G and M a Galois connection be-
tween the powerset lattices P(G) and P(M ) is obtained by the derivation oper-
ators, and from it the concept lattice B(G, M, I).
Galois connections between complete lattices were introduced in Section 0.4
and there it was already mentioned that the definition can be extended to arbi-
trary ordered sets without modifications. Now we generalize even further and
introduce Galois connections between formal contexts. Here the connection is
established by a pair of relations rather than mappings.7 In the first condition of
Definition 63, which is illustrated in Fig. 3.17, g I hΨ is short for hΨ ⊆ g I , etc. The
second condition will be simplified in Proposition 71.
Definition 63 A (relational) Galois connection between (G, M, I) and (H, N, J)
is a pair (Φ, Ψ ) of relations

Φ⊆G×N and Ψ ⊆H ×M

satisfying
1. g I hΨ ⇔ h J g Φ for all g ∈ G, h ∈ H (the Galois condition),
2. Φ is the largest relation fulfilling the Galois condition for the given Ψ , and
conversely. ♦

Proposition 71 The second condition of Definition 63 can be reformulated as follows:

2’. g Φ is an intent of (H, N, J) and hΨ is an intent of (G, M, I).

6 i.e., ∀x ϕJ (x) ≤ x
7 See Definition 103 on p. 296 for the latter.
154 3 Parts, factors, and bonds

M N

G I Φ

H Ψ J

Figure 3.17 The Galois condition requires that the left hand part hΨ of a row from the
lower part is contained in a row g I of the upper part iff the right hand side g Φ of the latter
is contained in the former, in hJ .

Proof Fix Ψ ⊆ H × M . A relation Φ satisfies the direction “⇒” of the Galois

condition iff
g Φ ⊆ {h ∈ H | g I hΨ }J .
The implication “⇐” is equivalent to

g ΦJ ⊆ {h ∈ H | g I hΨ },

which implies
g Φ ⊆ {h ∈ H | g I hΨ }J = g ΦJJ .
Thus from Φ we obtain another relation satisfying the Galois condition for Ψ by
replacing each g Φ by its closure g ΦJJ . This then is the largest possible choice. The
dual argument works for Ψ . 
This proof also shows that one gets condition 2) of Definition 63 for free, so to
speak. If a pair of relations satisfies the Galois condition, then these relations can
be uniquely enlarged so that condition 2) is also satisfied. One only has to replace
g Φ and hΨ by the concept intents they generate.
The next proposition shows that Definition 63 in fact generalizes Definition 18.

Proposition 72 If (ϕ, ψ) is a Galois connection between ordered sets (P, ≤1 ) and

(Q, ≤2 ), then (Φ, Ψ ), given by

(p, q) ∈ Φ : ⇔ ϕp ≤2 q and (q, p) ∈ Ψ : ⇔ ψq ≤1 p,

is a relational Galois connection between (P, P, ≤1 ) and (Q, Q, ≤2 ).

Proof Observe that pΦ = [ϕp) and q Ψ = [ψq) are intents for p ∈ P and q ∈ Q. If
(ϕ, ψ) is a Galois connection between the ordered sets, then
3.4 Bonds and connections 155

p I q Ψ ⇔ p ≤1 ψ(q) ⇔ q ≤2 ϕ(q) ⇔ q J pΦ .

Now that the definition of a Galois connection has been generalized in several
steps (from complete lattices to arbitrary ordered sets and further to arbitrary
binary relations, from mapping pairs to relations), the next theorem shows that
we finally reach the starting point again.
Theorem 19 Whenever (ϕ, ψ) is a Galois connection between concept lattices
B(G, M, I) and B(H, N, J), then

g Φ = Y : ⇔ ϕ(g II , g I ) = (Y J , Y )

and
hΨ = X : ⇔ ψ(hJJ , hJ ) = (X I , X)
defines a relational Galois connection between (G, M, I) and (H, N, J).
Conversely we obtain from each relational Galois connection (Φ, Ψ ) between (G, M, I)
and (H, N, J) a Galois connection (ϕ, ψ) between the concept lattices B(G, M, I) and
B(H, N, J) by

ϕ(X, X I ) := (X ΦJ , X Φ ) and ψ(Y, Y J ) := (Y Ψ I , X Ψ ).

The two constructions are inverse to each other.

Proof The proof is straightforward. 

Dual bonds

The name “dual bond” is admittedly somewhat ambiguous. It denotes a bond

from one formal context to the dual one of the second. This is how it is specified
in Definition 64.
Definition 64 A dual bond from (G, M, I) to (H, N, J) is a relation R ⊆ G × H
for which it holds that
• for every g ∈ G, g R is an extent of (H, N, J) and
• for every h ∈ H, hR is an extent of (G, M, I). ♦

Theorem 20 From each relational Galois connection (Φ, Ψ ) between (G, M, I) and
(H, N, J) we obtain a dual bond R ⊆ G × H by means of

(g, h) ∈ R :⇔ g I hΨ ( ⇔ h J g Φ ).

Conversely, if R is a dual bond between (G, M, I) and (H, N, J), then

156 3 Parts, factors, and bonds

g Φ := g RJ for g ∈ G,
hΨ := hRI for h ∈ H

defines a relational Galois connection. The two constructions are inverse to each other.
Proof If R is defined as above from a relational Galois connection (Φ, Ψ ) we get

g R = g ΦJ and hR = hΨ I for g ∈ G, h ∈ H.

Then obviously R is a dual bond. Conversely, if R is a dual bond, then g R = g RJJ

and thus
h J g Φ ⇔ h ∈ g ΦJ = g RJJ = g R ⇔ g R h,
and analogously g I hΨ ⇔ g R h. This shows that (Φ, Ψ ) is a relational Galois
connection.
As a corollary, we obtain a generalized version of Birkhoff’s construction:
Corollary 2 For every dual bond R ⊆ G × H,

ϕ(X, X I ) := (X R , X RJ ), ψ(Y, Y J ) := (Y R , Y RI )

defines a Galois connection between B(G, M, I) and B(H, N, J). Conversely, for every
Galois connection (ϕ, ψ),

R := {(g, h) | (g II , g I ) ≤ ψ(hJJ , hJ )} = {(g, h) | (hJJ , hJ ) ≤ ϕ(g II , g I )}

is a dual bond. The two constructions are inverse to each other.

Theorem 20 has several nice consequences. First of all, it shows that relational
Galois connections lead to closure operators:
Proposition 73 If (Φ, Ψ ) is a relational Galois connection between (G, M, I) and
(H, N, J), and if R is the corresponding dual bond, then the mappings

A 7→ ARR for A ⊆ G
C 7→ C RR for C ⊆ H

are closure operators with dually isomorphic closure systems. 

Note that the closed sets of these closure operators are just the extents of the
images of the two mappings ϕ and ψ in Theorem 19.
Secondly, Theorem 20 shows us a simple way to construct the lattice of all
(relational) Galois connections. This comes from the fact that the dual bonds
form a closure system with a simple closure operator:
Proposition 74 A relation R ⊆ G × H is a dual bond from (G, M, I) to (H, N, J) if
and only if

A × {h} ⊆ R ⇒ AII × {h} ⊆ R and {g} × B ⊆ R ⇒ {g} × B JJ ⊆ R.

3.4 Bonds and connections 157

Proof This is immediate from Definition 64. 

The direct product of two formal contexts Kt := (Gt , Mt , It ), t = 1, 2 has been
defined in Definition 32 on p. 44 as

K1 × K2 := (G1 × G2 , M1 × M2 , ∇)

with
(g1 , g2 ) ∇ (m1 , m2 ) :⇔ g1 I1 m1 or g2 I2 m2 .

Proposition 75 All extents of the direct product are dual bonds.

Proof It suffices to prove this for attribute extents. We have

(m, n)∇ = mI1 × G2 ∪ G1 × nI2 .

Abbreviating R := (m, n)∇ , we find that R ⊆ G1 × G2 is a relation such that

- g R = G2 or g R = nI2 if g ∈ G1 and
- hR = g1 or hr = mI1 if h ∈ G2 .
In any case, the result is an extent of K1 or K2 , respectively. 
The converse does not hold. There are examples of dual bonds that are not
regular, i.e., which are not extents of the direct product. The regular dual bonds
are characterized in Theorem 21, for which we use the following notation: If R ⊆
G × H is a relation and A ⊆ G, then let

R(A) := {h ∈ H | ∃g∈A (g, h) ∈ R}.

If A has only one element, i.e., if A = {g}, we omit the brackets and write R(g)
instead of R({g}). Note that
[
R(A) = R(g).
g∈A

Theorem 21 (Krötzsch et al. [237]) A dual bond R ⊆ G × H from (G, M, I) to

(H, N, J) is regular iff \
R(mI )JJ
r
R(g) =
I
m∈gr

holds for all g ∈ G.

A proof of this theorem will be given below, but it requires some preparation.
Recall that the object set of the direct product (G, M, I) × (H, N, J) is G × H.
Subsets of this object set are the relations R ⊆ G×H, and applying the derivation
operator of the direct product leads to

R∇ := {(m, n) | (g, h)∇(m, n) for all (g, h) ∈ R}.

158 3 Parts, factors, and bonds

Proposition 76 Let R ⊆ G × H be a relation between the object sets of formal contexts

(G, M, I) and (H, N, J). For each attribute m ∈ M , the three sets
\
R∇ (m), R(mI )J , and R(g)J
r

I
g∈mr

are equal. Furthermore,

\
R∇∇ (g) = R∇ (g I )J = R(mI )JJ
r r

I
m∈gr

holds for each object g ∈ G.

Proof Note that

R∇ (m) = {n ∈ N | (g, h) ∇ (m, n) for all (g, h) ∈ R}

= {n ∈ N | g I m or h J n for all (g, h) ∈ R},
and R(mI )J = {n ∈ N | h J n for all h ∈ H with (g, h) ∈ R for some g r
r
I m}.

Now let n ∈ R∇ (m). If h ∈ H is such that (g, h) ∈ R for some g r I m, then

clearly h J n. Thus R∇ (m) ⊆ R(mr ) . For the converse we need to show that if
I J

n ∈ R(mr ) , then (m, n) ∈ R∇ . Consider some (g, h) ∈ R. If g r

I J
I m, then h J n
since n ∈ R(mr ) , so certainly (g, h) ∇ (m, n), i.e., n ∈ R∇ (m). This proves the
I J

equality of the first and the second set. The equality of the second and the third
is immediate from the definition of the derivation operators.
To verify the second part of the assertion, note that R∇ is a relation between
the object sets of the dual contexts. The first part, which has already been proven,
can be applied to this and confirms the assertion. 
Proof of Theorem 21. We show more than is claimed in the theorem, namely that
R ⊆ G × M is an extent of the direct product if and only if the condition stated
in the theorem is fulfilled. The assumption that R is a dual bond is not needed.
However, with Proposition 76 this is easy, because obviously R is an extent if
and only if R = R∇∇ , and the latter is the case if and only if R(g) = R∇∇ (g)
is satisfied for all g ∈ G. But because according to Proposition 76 R∇∇ (g) =
I JJ
r
, this is exactly the condition in the theorem.
T
m∈gr I R(m ) 

Dense subcontexts

When the product R ◦ S was defined in Definition 62, we did not require that
the factors were bonds. Now we describe another situation in which this product
occurs naturally. Let K := (G, M, I) be a formal context and let G1 ⊆ G and M1 ⊆
M define a dense subcontext (G1 , M1 , I1 ), where I1 := I ∩ G1 × M1 , meaning that
3.4 Bonds and connections 159

{γg | g ∈ G1 } is -dense and {µm | m ∈ M1 } is -dense in K. The subcontext

W V
naturally divides (G, M, I) into four parts, which are labeled as
follows in the adjacent graphic:
M1 M2
G2 := G \ G1 , M2 := M \ M1 ,
R := I ∩ G1 × M2 , S := I ∩ G2 × M1 , G1 I1 R
and T := I ∩ G2 × M2 .
All objects in G2 and all attributes in M2 must be
reducible, since (G1 , M1 , I1 ) is dense, and therefore G2 S T
every m , m ∈ M2 , must be an extent and every g ,
R S

g ∈ G2 , must be an intent of (G1 , M1 , I1 ).

What conditions on a relation from T ⊆ G2 × M2 can ensure that (G1 , M1 , I1 )
is a dense subcontext? The following proposition shows that there is only one
possible choice.

Proposition 77 Let (G1 , M1 , I1 ) be a subcontext of (G, M, I) and let G2 , M2 , R, S,

and T be defined as above. Then (G1 , M1 , I1 ) is a dense subcontext of (G, M, I) if and
only if the following conditions 1.–3. are all satisfied:
1. For each m ∈ M2 , mR is an extent of (G1 , M1 , I1 ),
2. for each g ∈ G2S
, gS is an intent of (G1 , M1 , I1 ),
3. T = R ◦ S := B S × AR | (A, B) ∈ B(G1 , M1 , I1 ) .
Condition 3 can equivalently be replaced by each of the following:
4. T = {(h, n) ∈ G2 × M2 | nRI1 ⊆ hS }.
5. T = {(h, n) ∈ G2 × M2 | hSI1 ⊆ nR }.

Proof Conditions 1. and 2. are obviously necessary for (G1 , M1 , I1 ) to be a dense

subcontext. To see that 3. also is necessary, pick an arbitrary h ∈ G2 and let

Ah := hSI1 ∩ hT R .

Both hSI1 and hT R are extents of (G1 , M1 , I1 ) (the latter according to 1.), and
therefore Ah is the extent of some concept (Ah , AIh1 ) ∈ B(G1 , M1 , I1 ). When
(G1 , M1 , I1 ) is dense, then A0h = h0 , which is short for AIh1 = hS and AR
h = h .
T

From h ∈ Ah we then get

I1 S

{h} × hT ⊆ AIh1 S × AR
h,

which proves that T is contained in the right hand side of 3., i.e., in R ◦ S.
For the other inclusion let (X, Y ) ∈ B(G1 , M1 , I1 ) such that (h, n) ∈ Y S × X R .
Then h ∈ Y S and thus Y ⊆ hS = AIh1 , which implies that Ah ⊆ X and therefore
X R ⊆ AR h . This together with n ∈ X forces n ∈ Ah = h , i.e., (h, n) ∈ T .
R R T

To see that 3. is sufficient when 1. and 2. are given, we must infer from 3. that
each h ∈ G2 is reducible. So let Y := hS , X := Y I1 . Then Y is an intent of
(G1 , M1 , I1 ), h ∈ Y S and from 3. we get hT ⊇ X R . Let n ∈ hT \ X R . Then there
160 3 Parts, factors, and bonds

is some (U, V ) ∈ B(G1 , M1 , I1 ) such that h ∈ V S and n ∈ U R . But h ∈ V S

yields V ⊆ hS = Y and consequently U = V I1 ⊇ Y I1 = X. We get X ⊆ U and
conclude U R ⊆ X R , which is a contradiction to n ∈ U R , n ∈ / X R.
To prove that 4. is equivalent to 3., assume first that (h, n) ∈ B S × AR for some
concept (A, B) ∈ B(G1 , M1 , I1 ). The B ⊆ hS and A ⊆ nR , thus B = AI1 ⊇ nRI1 ,
which implies nRI1 ⊆ hS . Conversely, if nRI1 ⊆ hS , then A := nR is an extent
with corresponding concept intent B := AI1 = nRI1 . We get hS ⊆ B and thus
(h, n) ∈ B S × AR . The proof for 5. is analogous. 

3.5 Notes, references, and trends

3.1

Compatible subcontexts as well as their characterization by means of the arrow

relations were introduced in [390]. Proposition 43 has been taken from a paper
by Knecht and Wille [223].

3.2

Congruence relations belong to the standard subjects of textbooks on abstract

algebra. Theorem 10 is “classical”. Books on lattice theory, however, usually ex-
amine lattice congruence (where the requirement is compatibility with suprema
and infima of finite sets only). These differ considerably from the complete con-
gruence relations. Of course, every complete congruence is a lattice congruence
in the weaker sense. But whereas the lattice congruences always form a distribu-
tive sublattice of the lattice of equivalence relations, the situation in the case of
the complete congruences is more complicated: The supremum of two complete
congruences does not have to coincide with the supremum as equivalence rela-
tions, and every complete lattice is isomorphic to the lattice of the complete con-
gruence relations of a suitable concept lattice (Teo [368], Grätzer [174]). There-
fore Theorem 11 cannot simply be derived from the corresponding theorems for
algebras, but follows [388]. With regard to the congruence theory for concept
lattices see also Reuter and Wille [318].
Czedli [83] had discovered that tolerance relations also yield a factor lattice.
Bandelt [22] examined this connection in more detail. The interrelation between
complete tolerance relations and block relations was first described in [394].
Wille has also suggested the use of tolerance relations in order to obtain counting
formulas by means of the Möbius function. In this context see also [314] and Vogt
[376]. Krupka [239] considers more general symmetric relations, which he calls
“incompatible tolerances”. He proposes to understand the totality of these as a
residuated lattice, whose elements can then be interpreted as truth values and
thus provide a way to simplify the concept lattice. See also Bartl and Krupka [27].
3.5 Notes, references, and trends 161

3.3

Closed relations were introduced in [397], in order to simplify the description of

subdirect products, which had been tackled in [391]. The concept lattice from
Figure 3.11 originally resulted from an analysis of biological data, see [136].
Kauer and Krupka [212] give an algorithm for finding the closed relation cor-
responding to a complete sublattice generated by given elements. For the case of
the (doubly founded, distributive) contranominal scale (P, P, 6≥), the complete
sublattices were described in [140] as the extents of

(B(P, P, 6≥), P × P, ◦), where (A, B) ◦ (p, q) :⇔ ¬(p ∈ B and q ∈ A).

Kästner [210] shows how retracts of concept lattices (i.e., sublattices that are im-
age of an idempotent complete endomorphism) can be described by compatible
subcontexts in closed relations.
The dismantling of doubly irreducible elements was examined by Duffus and
Rival [108]. They also proved the uniqueness of the DI-kernel. A very simple
proof was given by Farley [123]. Berry and Sigayret [46] study complementary
contexts. Felde and Koyda [126] generalize the dismantling construction. Instead
of removing only doubly irreducible elements, they allow to remove intervals
[u, v], provided u is supremum-prime in (v] and v is infimum-prime in [u). They
prove that there is again a unique kernel.
Distributive lattices that are generated by their doubly irreducible elements
were examined by Monjardet & Wille [287].
It is possible to describe in terms of structure what happens if we add “a cross”
to the incidence relation I. Here, we shall limit ourselves to cardinality: The con-
cept lattice can become larger but it can also become smaller. A simple estimate
by Skorsky shows
1 |B(G, M, I ∪ {(g, m)})| 3
≤ ≤ .
2 |B(G, M, I)| 2
See also Kauer and Krupka [211]. The concept lattice can only become smaller
if g %
. m. Assume w.l.o.g. that (G, M, I) is clarified. If neither g % m nor g .
m, then, by Proposition 59, B(G, M, I) is a complete sublattice of B(G, M, I ∪
{(g, m)}).

3.4

Much of what we present in 3.4 has been moved there from Section 5.1 of the first
edition. The notion of a bond was introduced in [399] (under the French name
“liaison”). Relational Galois connections are from [141], but see also Xia [425],
Domenach and Leclerc [103]. The sublattices for the closed relations correspond-
ing to bonds (Proposition 65) were characterized by Kaarli et al. [206]. Kaarli
& Radecezki [207] explore the algebraic meaning of self-bonds.
162 3 Parts, factors, and bonds

Examples of irregular bonds occur in [141]. Krötzsch & Malik [238] remark
that both the ordinal scale M3 and the biordinal scale M2,1 have irregular self-
bonds (the concept lattices of which are called M3 and N5 in algebraic lattice
theory). These authors show connections to distributivity. Among other things,
they prove that Galois connections from and to completely distributive complete
lattices are regular and that a complete lattice which only has regular Galois
connections to other complete lattices must be distributive.
Chapter 4
Decompositions of concept lattices

A complex concept lattice can possibly be split into simpler parts. For this, the
mathematical modeling must prove its worth by providing effective and versatile
decomposition methods. Any such decomposition principle can be reversed to
create a method of construction. Therefore some of the following topics will be
taken up again in the next chapter with this second focus.
If a lattice can be represented as a sublattice of a direct product, this is called
a subdirect decomposition. The theory described in the previous chapter allows an
elegant description of these decompositions by means of the formal context. This
is the subject of the first section.
The tolerance relations introduced in 3.2 result in coverings of the concept
lattice by overlapping intervals. This fact will be used as a principle of decompo-
sition in the second section.
A surprisingly versatile context operation consists in inserting one context into
another one. We shall explain this in more detail in the third section and describe
the corresponding lattice construction, the substitution product. We then put some
effort into proving a decomposition theorem for this product. (Theorem 30).
In the fourth section we shall finally introduce the tensor product of complete
lattices by means of the direct product of the contexts. Similarly as in the case of
the direct product of lattices (which corresponds to the context sum), tensorial
decomposability is rare. Therefore, we transfer the idea of the subdirect prod-
uct, which we explain in 4.1, to contexts and obtain the notion of the subtensorial
decomposition of concept lattices.

4.1 Subdirect decompositions

The direct1 product of ordered sets has already been introduced in Definition 7.
We shall repeat it here for the special case of complete lattices:

1 The word “direct” is often omitted.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 163
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_5
164 4 Decompositions of concept lattices

Definition 65 Let T be an arbitrary index set. For a family (Lt , ≤t )t∈T of com-
plete lattices, the product is defined to be

× ×
 
(Lt , ≤t ) := Lt , ≤
t∈T t∈T

with
(xt )t∈T ≤ (yt )t∈T : ⇔ xt ≤t yt for all t ∈ T.
The lattices (Lt , ≤t ), t ∈ T , are the factors of the product, and the maps

πs : ×L
t∈T
t −→ Ls with πs ((xt )t∈T ) := xs

defined for s ∈ T are the canonical projections. ♦

Without difficulty one proves:
Proposition 78 Every product of complete lattices is a complete lattice. The infimum
and the supremum can be formed componentwise. The canonical projections are surjec-
tive complete homomorphisms. 
The direct product of concept lattices corresponds to the direct sum of the
formal contexts, cf. Definition 30 (p. 42).
Definition 66 A (complete)2 subdirect product of complete lattices is a com-
plete sublattice of the direct product for which the canonical projection maps
onto the factors are all surjective.
A subdirect decomposition of a complete lattice (L, ≤) is a family Θt , t ∈ T ,
of complete congruence relations of (L, ≤) with
\
Θt = ∆,
t∈T

where ∆ denotes the trivial congruence ∆ := {(x, x) | x ∈ L}. The lattices

(L, ≤)/Θt , t ∈ T , are called the factors of the subdirect decomposition. ♦

Theorem 22 If (L, ≤) is a complete subdirect product of lattices (Lt , ≤t ), t ∈ T , then

the kernels of the canonical projections

{ker πt | t ∈ T }

form a subdirect decomposition of (L, ≤). Conversely, for every subdirect decomposition
Θt , t ∈ T of (L, ≤), by
ι(v) := ([v]Θt )t∈T
an isomorphism of (L, ≤) onto a subdirect product of the factor lattices (L, ≤)/Θt , t ∈ T ,
is defined.
2For stylistic reasons, we will frequently leave out the adjective “complete”, i.e., in the following
“subdirect” should be replaced by “complete subdirect” where necessary.
4.1 Subdirect decompositions 165

(1,1)
d
(1,4) d @ d(6,5)
@
1
(5,4) d @ d (6,3) @ d (4,5) d
@ @

1
(2,4) d @ d (3,3) @ d (4,3) 5 d @ d6 d
@ @ @

(2,2) d 2 d 3 d @ d4 4 d @ d5
@ @ @ @
@ @
@ d (0,3) 2 d @d3
@ @ @ @
@ @
@d @d @d
@ @ @

(0,0) 0 0

Figure 4.1 The lattice on the left is a subdirect product of the two lattices on the right.

Proof The kernels of the canonical projections are congruences. Hence, in order
to prove the first part, we only have to show that their intersection is the trivial
congruence ∆. Two elements (vt )t∈T and (wt )t∈T of the direct product are differ-
ent if there is some s ∈ T with vs 6= ws , i.e., with πs (v) 6= πs (w), which
T is equiv-
alent to (v, w) 6∈ ker πs . Hence, from v 6= w it follows that (v, w) 6∈ t∈T ker πt ,
which was to be proved.
Now we show that the map ι defined in the second part has the properties
claimed. ι is a complete homomorphism, since for an arbitrary family (vj )j∈J of
elements of (L, ≤) we have
 ^   ^  
ι vj = vj Θt
j∈J j∈J t∈T
^ 
= [vj ] Θt
j∈J t∈T
^
= ([vj ] Θt )t∈T
j∈J
^
= ι (vj ) .
j∈J

Likewise, we show that ι preserves suprema. ι is injective, since from v, w ∈ L,

v 6= w follows the existence of some t ∈ T with (v, w) 6∈ Θt . This means, however,
that [v]Θt 6= [w]Θt , i.e., that ι(v) 6= ι(w).
Hence, ι is an isomorphism of (L, ≤) onto the complete sublattice ι(L) of the
product × t∈T ((L, ≤)/Θt ). It remains to be shown that the canonical projections
166 4 Decompositions of concept lattices

d d
@ @
d @d d @d
@ @ @ @
d @d @d d @d @d
@ @ @ @
d @d @d d @d @d
@ @
d @ d @
@ @ @ @
@ @d @ @d
@ @
@d @d

Figure 4.2 The two congruences represented here form a subdirect decomposition of the
lattice in Figure 4.1. The factor lattices by these congruences are precisely the factors of
the subdirect product represented in the figure above.

πs : ι((L, ≤)) −→ Ls , s ∈ T,

are surjective. This follows from

πs (ι(L)) = {[v]Θs | v ∈ L} = (L, ≤)/Θs . 

In the definition of the subdirect decomposition, we have not excluded the

trivial case that one of the congruences is the trivial congruence ∆. The lattices
which only allow such decompositions are described by the following definition:
Definition 67 A complete lattice (L, ≤) is called (completely) subdirectly irre-
ducible if every subdirect decomposition of (L, ≤) contains the trivial congru-
ence ∆. ♦
This property can also be formulated as follows: If (L, ≤) is isomorphic to a
subdirect product of lattices (Lt , ≤t ), t ∈ T , then (L, ≤) is canonically isomor-
phic to one of the factors (Lt , ≤t ). (Canonically isomorphic here means that the
canonical projection πt is bijective and is therefore an isomorphism from (L, ≤)
to (Lt , ≤t )).
It is particularly easy to read off this property from the lattice of congruence
relations of (L, ≤), since (L, ≤) is subdirectly irreducible if and only if this lattice
has exactly one atom:
Proposition 79 A complete lattice (L, ≤) is subdirectly irreducible if and only if (L, ≤)
has a smallest non-trivial3 congruence, i.e., a complete congruence relation Θ with Θ 6=
∆ and Θ ≤ Ψ for all complete congruences Ψ 6= ∆.

3 We allow the total congruence L × L.

4.1 Subdirect decompositions 167

d
d @ @
@@
@
@

@
@

d @@ @d
@
@@ @d
@
@
@ @
@

@
@ @
@

d @@ @
@ @d
@
d @ @ @d
@ @d
@
@@ @
@ @
@

Q 
Q 
Q 
Q 

@
@
@d
@
@d
@

Figure 4.3 The nested line diagram helps to follow the definition of the subdirect product.

Proof If Θ is such a congruence and if ι is an isomorphism of (L, ≤) onto a sub-

direct product ι(L) of lattices (Lt , ≤t ), t ∈ T , then, because of
\
ker(πt ◦ ι) = ∆,
t∈T

it is not possible that

Θ ≤ ker(πt ◦ ι) for all t ∈ T.
Hence, ker(πt ◦ ι) = ∆ for at least one t ∈ T .
If, on the other hand, there is no such minimal congruence, then
168 4 Decompositions of concept lattices
\
{Θ | Θ ∈ C((L, ≤)), Θ 6= ∆} = ∆

for the family C((L, ≤)) of all congruences of (L, ≤). Hence, these congruences
form a proper subdirect decomposition of (L, ≤). 
The examination of subdirect decompositions can be carried out directly on
the context if we use the interplay between congruence relations, compatible and
arrow-closed subcontexts which we have developed in the preceding sections.
In order to do so, we must presuppose that the lattice (L, ≤) we examine is dou-
bly founded and thus isomorphic to the concept lattice of a reduced context K,
since in this case the congruences are in one-to-one correspondence to the arrow-
closed subcontexts of K.

Proposition 80 If (G, M, I) is a reduced context of a doubly founded concept lattice,

then the subdirect decompositions of B(G, M, I) correspond
S bijectively to the
S families of
arrow-closed subcontexts (Gt , Mt ,I ∩ Gt × Mt ) with t∈T Gt = G and t∈T Mt =
M.

Proof According to the observations preceding Theorem 13, t∈T ΘtS = ∆ holds
T
for a family Θt , t ∈ T of congruences if and only if t∈T Gt = G and t∈T Mt =
S
M holds for the corresponding arrow-closed subcontexts (Gt , Mt , I ∩Gt ×Mt ).
It is particularly easy to recognize the subcontexts belonging to subdirectly
irreducible factors of B(G, M, I). We must be aware that for every object g there
is always a smallest arrow-closed subcontext containing g. We shall call such a
subcontext a 1-generated arrow-closed subcontext. (The corresponding is true
for the attributes, but since in a reduced context every object is connected to an
attribute by a double arrow and vice versa, it suffices to concentrate on one of
the sets G or M , respectively).

Proposition 81 A doubly founded reduced context (G, M, I) is 1-generated if and only

if B(G, M, I) is subdirectly irreducible.

Proof If (G, M, I) is 1-generated, then for every family (Gt , Mt , I ∩ Gt × Mt ),

t ∈ T , of arrow-closed subcontexts there is some t ∈ T with Gt = G and Mt =
M . By means of Proposition 80 we recognize that this is equivalent to subdirect
irreducibility. 

Theorem 23 Every doubly founded complete lattice has a subdirect decomposition into
subdirectly irreducible factors.

Proof W.l.o.g. we may assume that (L, ≤) is the concept lattice of a reduced con-
text (G, M, I), and, since structural properties are preserved under isomorphism,
even assume that (L, ≤) = B(G, M, I). For g ∈ G let (Gg , Mg , I ∩ Gg × Mg )
denote the smallest arrow-closed subcontext of (G, M, I) containing g. Proposi-
tion 42 shows that this subcontext is reduced as well. According to Proposition
81 the corresponding concept lattice is subdirectly irreducible. Hence, together
with Proposition 80,
4.1 Subdirect decompositions 169

(Gg , Mg , I ∩ Gg × Mg ), g ∈ G,

provides a subdirect decomposition of (L, ≤) into subdirectly irreducible fac-

tors. 

d
fd
@ c
@d
b @ @ a b c d e f
d @d @ dd
1 × × %
. . × ×
5
a @ @ 2 .
% × × .
% ×
d @d @d
2 3 3 . .
% × × ×
e @ 4 × × × × %
. ×
d
1 @ 5 × × .
%
@ @
@ @d
@ 4
@d

Figure 4.4 Using the arrow relations in the context, we can examine which subdirect de-
compositions are possible for the concept lattice.

b a d c e f

KA
AK
 A
 6 6
A
A
A

A
A
AUA


AU
AU
? ?
3 2 1 4 5

Figure 4.5 We recognize five 1-generated subcontexts: one for each object.

To close this section, we apply the theory we have developed to the example
given in the beginning (Figure 4.1). A representation of the lattice as a concept
lattice is presented in Figure 4.4. The arrow relations of the context are shown
in Figure 4.5 as a graph. We can read off that there are exactly five 1-generated
subcontexts. In the main, there is only one subdirect decomposition of this lattice
into subdirectly irreducible factors. The respective subcontexts are generated by
the objects 2, 4 and 5. We can add the subcontexts generated by the objects 1 or
170 4 Decompositions of concept lattices

3, they are, however, contained in the subcontext generated by 2 and therefore

dispensable.
The subdirect decomposition shown in Figure 4.2 corresponds to the two sub-
contexts ({1, 2, 3}, {a, b, c, d}) and ({1, 4, 5}, {c, e, f }). We recognize that the sec-
ond factor can be chosen smaller: object 1 is superfluous, the second subcontext
can be replaced by ({4, 5}, {e, f }). Figure 4.6 shows the corresponding congru-
ence. It can replace the second congruence in Figure 4.2. The factor lattice ob-
tained from this congruence has three elements.

d
@
d @d
@ @
d @d @d
@ @
d @d @d
@
d @
@ @
@ @d
@
@d

Figure 4.6 A coarser congruence may be chosen for the second congruence in the subdi-
rect decomposition from Figure 4.2.

4.2 Atlas decompositions

A map which is meant to represent a larger area on a larger scale necessarily

becomes unwieldy. Usually, it is split up into an atlas: a collection of manage-
able part maps covering the desired area, together with a general map as well as
additional information, which show how the individual maps are related.
An analogous procedure for large unwieldy lattices can be introduced with
the help of the tolerance relations. According to Theorem 14 (p. 134), a tolerance
relation on a complete lattice L := (L, ≤) provides a decomposition of L into
intervals which are themselves elements of a complete lattice. Thus, L can be
understood as being constructed from a family of complete lattices, whose in-
dices in turn form a complete lattice. This means that the blocks of the tolerance
appear in the role of the part maps in the atlas and the factor lattice acts as the
“general map”.
4.2 Atlas decompositions 171

It is, however, in the case of the tolerances not quite as easy as in the case of
topographic maps to explain the interconnection between the part maps. In the
general case, this is done through a family of adjoint pairs of mappings. It is
easier in the case of glued tolerances: here those maps ensue automatically.
Definition 68 Let Q be a complete lattice and, furthermore, let Lq := (Lq , ≤) be
a complete lattice for every element q ∈ Q. Let

ϕqr : Lq → Lr and ψqr : Lr → Lq

be maps for every pair q ≤ r in Q. Such a family

(Lq | q ∈ Q)

is called a Q-atlas if the following conditions are satisfied:

0. Lq ⊆ Lr ⇒ q = r.
1. Lq ∩ Lr is an order filter of Lq and an order ideal of Lr if q ≤ S r.
2. {q ∈ Q | x ∈ Lq } is an interval [xmin , xmax ] in Q for every x ∈ q∈Q Lq .
3. ϕqq x = x = ψqq x for all x ∈ Lq .
4. ϕqr x ≤ y holds in Lr if and only if x ≤ ψqr y holds in Lq .
5. ϕrs ϕqr = ϕrq and ψqr ψrs = ψqr .
6. r∨s
ϕrq x = ϕq∨s x for all x ∈ Lq ∩ Lq∨s and ψst = ψs∧r
t∧r
for all y ∈ Lt ∩ Lt∧r .
The sum of the Q-atlas is defined as the pair
 [ 
Lq , v
q∈Q

with
x v y : ⇔ xmin ≤ ymin and ϕyxmin
min
x≤y
for all x, y ∈ Lq .
S
q∈Q ♦

Condition (4) says that for q ≤ r the two maps ϕrq and ψqr are mutually adjoint
W 11 on p. 19, as can easily
in the sense of Proposition be seen from Propositions
6 and 8. Hence, ϕrq is -preserving and ψqr is -preserving. If we denote the
V
boundary elements of (Lt , ≤) by 0t and 1t , respectively, we obtain in particular
ϕrq 0q = 0r and ψqr 1r = 1q .

Proposition 82
xmin ≤ ymin and ϕyxmin
min
x≤y
is equivalent to
xmax ≤ ymax and x ≤ ψxymax
max
y.

Proof Let xmin ≤ ymin and ϕxymin

min
x ≤ y. In the following we shall use the abbre-
viations
q := xmin , r := ymin , s := xmax , t := ymax .
172 4 Decompositions of concept lattices

By (6) we obtain
y ≥ ϕqr x = ϕsr∨s x.
Hence, ϕqr x is an element of (Lr , ≤) ∩ (Lr∨s , ≤), and by (1) it follows that y ∈
(Lr∨s , ≤), i.e., s ≤ t. Since r ≥ xmin , we have s∧r ∈ [xmin , xmax ] and x ∈ (Ls∧r , ≤)
follows from (2). Together with (6) and (3) we obtain

ϕs∧r
q x = x.

Thus
r
ϕrs∧r x = ϕs∧r ϕs∧r r
q x = ϕq x ≤ y

because of (5). With (4) and (6) we obtain

r
x ≤ ψs∧r y = ψst y.

The opposite direction of the equivalence follows dually. 

Theorem 24 The sum of a Q-atlas is a complete lattice L := (L, ≤), in which infima
and suprema can be described as follows:

t u
_ ^
xt = ϕqrt xt and xt = ψsqt xt
t∈T t∈T
t∈T t∈T

with xt ∈ (Lqt , ≤), r := t∈T qt and s := t∈T qt . The complete lattices (Lq , ≤),
W V
q ∈ Q are precisely the blocks of a complete tolerance relation Θ of L, and q 7→ (Lq , ≤)
describes an isomorphism of Q onto L/Q; furthermore, we have

ϕqr = x t 0r for all x ∈ (Lq , ≤)

and
ψqr y = y u 1q for all y ∈ (Lr , ≤).
In this way we obtain a bijective assignment between the complete tolerance relations on
a complete lattice and the representations of this lattice as the sum of a Q-atlas.

Proof Without difficulty we prove that v is an order which corresponds, more-

over, on each of the (Lq , ≤) to the order given there. Furthermore, we prove that
x v y always follows from ϕqr x = y: By (6) we first obtain ϕr∨x max
xmax x = y and
thus xmax ≤ ymax . With
ymax
ϕr∨x max
y = ϕyymax
max
y=y

we get ϕyxmax
max
x = y and, because of (4), x ≤ ψxymax max
y. Consequently, x v y holds
according to Proposition 82. Dually we infer that x = ψqr y always implies x v y.
We now show that the supremum has the form specified in the theorem: As-
sume that xt ∈ (Lqt , ≤) and xt v y for t ∈ T . By Proposition 82 andW(2) we
obtain y ∈ (Lymin ∨qt , ≤) for all t ∈ T and thus y ∈ (Lymin ∨r , ≤) for r := t∈T qt .
Therefore, from ϕxymin
min
xt ≤ y it follows that ϕqytmin ∨qt xt ≤ y because of (6) and
4.2 Atlas decompositions 173

min ∨r min ∨r
ϕqrt v ϕyqtmin ∨r xt = ϕyymin ymin ∨qt
∨qt ϕqt xt v ϕyymin ∨qt y = y

because of (5), (6) and (3). Thereby we have proved t∈T ϕqt xt v y. Since
r
W

t∈T ϕqt xt is an upper bound of each xt , t ∈ T , t∈T ϕqt xt is the supremum

r r
W W
of the xt , t ∈ T , in (L, v). Because of Proposition 82, the dual proof yields the
equation for the infimum. Thus, we have proved that the sum of a Q-atlas is al-
ways a complete lattice.
For x, y ∈ L now assume that

x Θ y : ⇔ x, y ∈ (Lq , ≤) for at least one q ∈ Q.

The proved description of the suprema and infima immediately yields that Θ is
a complete tolerance relation of L. We can use Proposition 52 to show that the
(Lq , ≤) are precisely the blocks of Θ. This requires however to prove the maxi-
mality of the (Lq , ≤). Assume therefore that q ∈ Q and that y is an element with

xΘy for all x ∈ (Lq , ≤).

In particular, we have y Θ 0q , which yields {0q , y} ⊆ (Lr , ≤) for some r ∈ Q and

because 0r ≤ 0q it follows that r ≤ q. Likewise, we obtain from y Θ 1q some
s ∈ Q with y ∈ (Ls , ≤) and s ≥ q. In all, we have q ∈ [ymin , ymax ] and thus by (2)
y ∈ (Lq , ≤), which together with (0) yields the assertion.
Evidently, q 7→ (Lq , ≤) describes an isomorphism of Q on L/Θ. For q ≤ r,
x ∈ (Lq , ≤) and y ∈ (Lr , ≤) we have

x t 0r = ϕqr x ∨ ϕrr 0r = ϕqr x

and dually y u 1q = ψqr y. If we define for an arbitrary complete tolerance relation

Ξ morphisms between its blocks through these equations, we obtain a L/Ξ-atlas
whose sum again is L. This shows the bijective assignment we claimed. 

The conditions become simpler in the case of tolerances with overlapping

neighborhoods.
Definition 69 A complete tolerance relation Θ of a lattice L := (L, ≤) has over-
lapping neighborhoods if

B1 ≺ B2 in L/Θ implies B1 ∩ B2 6= ∅.

Let Σ(L) denote the smallest tolerance relation comprising all pairs (x, y) with
x ≺ y in L.
In the case of doubly founded lattices, a tolerance with overlapping neigh-
borhoods is called glued, and Σ(L) is called the skeleton tolerance. The factor
lattice L/Σ(L) is then called the skeleton of L. ♦
The intersection of any number of tolerance relations is again a tolerance re-
lation. Therefore Σ(L) is well-defined.
174 4 Decompositions of concept lattices

Theorem 25 Σ(L) is the smallest tolerance relation of L with overlapping neighbor-

hoods. In particular, the skeleton tolerance is the smallest glued tolerance.
Proof First we show that Σ(L) has overlapping neighborhoods. For this pur-
pose, let B1 =: [a, b] and B2 =: [c, d] be blocks of Σ(L) with B1 ≺ B2 . We show
that the assumption B1 ∩ B2 = ∅ leads to a contradiction. It would imply that
b < b ∨ c and from b < x ≤ d it would follow that generally

B1 = [b]Σ(L) < [x]Σ(L) ≤ [d]Σ(L) = B2 ,

i.e., because of B1V≺ B2 , that [x]Σ(L) = B2 and thus x ≥ c. For this reason we
have that b ∨ c = {x | b < x ≤ d}, from which we can infer that b ≺ b ∨ c and
consequently (b, b ∨ c) ∈ Σ(L). Because of (b ∨ c)Σ(L) = c this yields c ≤ b ≤ d,
i.e., the desired contradiction b ∈ B2 .
It remains to be shown that Σ(L) is smallest among the tolerance relations
with overlapping neighborhoods. Hence, let Θ be any such tolerance relation
and x ≺ y a pair of neighboring elements of L. We must show that there is a
block of Θ containing x and y. This is certainly the case if x ∈ [y]Θ , i.e., we may
assume x 6∈ [y]Θ and in particular [x]Θ < [y]Θ . Now we consider an arbitrary
block [u, v] with [x]Θ ≤ [u, v] < [y]Θ . Because of u < yΘ ≤ y, y ∈ [u, v] would
immediately follow from y ≤ v. It would imply yΘu and thus a contradiction
u < yΘ . Hence, y 6≤ v and, because of x = (v ∧ y) Θ (v ∧ y Θ ) = v, we obtain xΘv
and thus x ∈ [u, v]. Hence, every block between [x]Θ and [y]Θ contains x. Since
the lower bounds of the blocks are closed under suprema, this also holds for
_
Bx := {B ∈ L/Θ | [x]Θ ≤ B < [y]Θ }.

Hence, this block must be a lower neighbor of [y]WΘ . Since Θ has overlapping
neighborhoods, it follows that Bx ∩ [y]Θ 6= ∅, i.e., Bx ∈ [y]Θ . If y 6∈ Bx , then
y ∧ Bx = x and thus x ∈ [y]Θ , contradicting [x]Θ < [y]Θ !
W

The corresponding block relation β(Σ(L)) can be described easily.

Theorem 26 Let (G, M, I) be a doubly founded context and let

Σ := Σ(B(G, M, I))

be the skeleton tolerance. Then the following statements hold for the corresponding block
relation J := β(Σ):
a) J is the smallest block relation of (G, M, I) containing all pairs (g, m) with g %
. m.
b) J contains all pairs (g, m) with g . m or g % m.
Proof b) From g . m it follows that γg ∧ µm = (γg)∗ ≺ γg, i.e., (g, m) ∈ J
according to the definition of β. Dually we show that J furthermore contains all
pairs (g, m) with g % m.
a) Let K be a block relation comprising all pairs (g, m) with g %
. m and let fur-
thermore (A, B) and (C, D) be concepts with (A, B) ≺ (C, D). We want to show
4.2 Atlas decompositions 175

that (A, B) and (C, D) are related under the tolerance relation β −1 (K) belonging
to K. For this purpose, we consider an object g ∈ C and an attribute m ∈ B with
(g, m) ∈/ I. Since the context is doubly founded, we find an object h with h0 ⊇ g 0
and h . m, i.e., in particular (h, m) ∈/ I. We again make use of the fact that the
context is doubly founded to get an attribute n with n0 ⊇ m0 and h % n and thus
h%. n (since µn ≥ µm ≥ (γh)∗ ). Consequently, (h, n) ∈ K, which is equiva-
lent to (γh, γh ∧ µn) ∈ β −1 (K), or shorter (γh, (γh)∗ ) ∈ β −1 (K). From that we
infer ((A, B) ∨ γh, (A, B) ∨ (γh)∗ ) ∈ β −1 (K), i.e., ((C, D), (A, B)) ∈ β −1 (K), as
claimed. 

Now we concentrate on the case of glued tolerances; in particular we presup-

pose that the lattices are doubly founded. In the case of finite concept lattices,
this condition is automatically fulfilled, and the intention to use the theory for
drawing line diagrams is primarily aimed at the finite case.
We describe the system of the blocks of a glued tolerance in abstract terms:
Definition 70 Let (Lq , ≤), q ∈ Q be a family of doubly founded complete lattices.
Let the index set Q be a lattice of finite length. We call ((Lq , ≤) | q ∈ Q) a Q-atlas
with overlapping neighbor maps, if for each two elements q, r ∈ Q the following
conditions are satisfied:
0. (Lq , ≤) ⊆ (Lr , ≤) ⇒ q = r.
1. If q ≤ r, then (Lq , ≤) ∩ (Lr , ≤) is an order filter in (Lq , ≤) and an order ideal
in (Lr , ≤).
2. If q is a lower neighbor of r, then (Lq , ≤) ∩ (Lr , ≤) 6= ∅.
3. The orders of (Lq , ≤) and (Lr , ≤) coincide on the intersection (Lq , ≤)∩(Lr , ≤).
4. (Lq , ≤) ∩ (Lr , ≤) ⊆ (Lq∧r , ≤) ∩ (Lq∨r , ≤).
5. q ≤ r ≤ s ⇒ (Lq , ≤) ∩ (Ls , ≤) ⊆ (Lr , ≤) ♦

This is compatible with Definition 68. The maps postulated there ensue canon-
ically in the glued case, as the following proposition shows.

Proposition 83 A Q-atlas with overlapping neighbor maps is a Q-atlas in the sense of

68 if we define the maps ϕqr and ψqr as follows:
1. ϕqq x := ψqq x := x for all x ∈ (Lq , ≤);
2. let
ϕrq x := x ∨ 0r (in (Lq , ≤)),
ψqr y := y ∧ 1q (in (Lr , ≤))
for q ≺ r;
3. let
ϕrq := ϕqqm
m−1
◦ · · · ◦ ϕqq10 ,
ψqr := ψqq01 ◦ · · · ◦ ψqqm−1
m

for q = q0 ≺ q1 ≺ · · · ≺ qm = r.
176 4 Decompositions of concept lattices

Proof Conditions 0) and 1) are the same as in Definition 68. Condition 2) of

Definition 68 follows like this: The set

{q ∈ Q | x ∈ (Lq , ≤)}

is by 5) convex and by 4) closed against ∨ and ∧, i.e., an interval in the lattice Q

(which is of finite length).
The remaining conditions refer to the maps ϕqr and ψqr . First, we have to show
that the maps in the manner specified are well-defined for all q ≤ r. In the case
that q ≺ r, then (Lq , ≤) ∩ (Lr , ≤) by 2) is not empty and, with 1), we find 0r ∈
(Lq , ≤). Hence, the supremum x ∨ 0r can be formed within (Lq , ≤) for all x ∈
(Lq , ≤) and by 1) lies in (Lr , ≤).
Since Q is of finite length, for any two elements q < r in Q there exists at least
one, but possibly several chains of neighbor elements between q and r. We have
to show that the definition of ϕrq is independent of the choice of such a chain
(the proof of ψqr then works analogously). Hence, let q0 ≺ q1 ≺ · · · ≺ qm and
r0 ≺ r1 ≺ · · · ≺ rn be chains with q0 = r0 and qm = rn . The proof works through
induction on the length of the interval [q0 , qm ].

1st case: q1 ∨ r1 = qm .
Since the smallest elements of (Lq1 , ≤) and (Lr1 , ≤) belong to (Lq0 , ≤), the supre-
mum 0q1 ∨ 0r1 can be formed in (Lq0 , ≤). This element belongs to (Lq1 , ≤) as well
as to (Lr1 , ≤); i.e., the two lattices are not disjoint. Because of 4) we have

(Lq1 , ≤) ∩ (Lr1 , ≤) ⊆ (Lq0 , ≤) ∩ (Lqm , ≤),

i.e., (Lq0 , ≤) and (Lqm , ≤) are not disjoint either, and 0qm has to be an element
of (Lq0 , ≤). Because of 5), 0qm therefore belongs to all lattices (Lqi , ≤), i ∈
{0, . . . , m} and to all lattices (Lrj , ≤), j ∈ {0, . . . , n}. The sets (Lq0 , ≤) ∩ (Lqi , ≤)
and (Lq0 , ≤) ∩ (Lrj , ≤) are therefore all nonempty. Since by 1) they are order
ideals, the elements 0qi and 0rj , respectively, must all belong to (Lq0 , ≤).
Hence, for x ∈ (Lq0 , ≤) it follows that

(ϕqqm−1
m
◦ · · · ◦ ϕqq10 )x = (. . . ((x ∨ 0q1 ) ∨ 0q2 ) ∨ . . .) ∨ 0qm ,

and all those suprema are being formed in (Lq0 , ≤). Therefore,

ϕqqm
0
x = x ∨ 0q1 ∨ · · · ∨ 0qm = x ∨ 0qm

and correspondingly

ϕrrn0 x = x ∨ 0r1 ∨ · · · ∨ 0rn = x ∨ 0rn ,

which on account of qm = rn yields the desired result.

4.2 Atlas decompositions 177

2nd case: q1 ∨ r1 < qm . Assume that

q 1 = s 1 ≺ s 2 ≺ · · · ≺ s j = q 1 ∨ r1
r 1 = t 1 ≺ t2 ≺ · · · ≺ tk = q 1 ∨ r 1

and sj ≺ sj+1 ≺ · · · ≺ sl = qm as well as tk+i = sj+i for i ∈ {1, . . . , l − j}. By the

induction hypothesis

ϕssjj−1 ◦ · · · ◦ ϕss12 ◦ ϕqq01 = ϕttkk−1 ◦ · · · ◦ ϕtt21 ◦ ϕrr01 ,

ϕqqm
m−1
◦ · · · ◦ ϕqq12 = ϕssl−1
l
◦ · · · ◦ ϕss12
and
t
ϕrrn−1
n
◦ · · · ◦ ϕrr21 = ϕtl−j
l−j−1
◦ · · · ◦ ϕtt12 .
Now, by insertion we obtain

ϕqqm
m−1
◦ · · · ◦ ϕqq10 = ϕrrnn−1 ◦ · · · ◦ ϕrr10 .

Thereby we have shown that the definition of ϕrq is independent of the choice of
the chain.
Furthermore, we have to prove (4), i.e., that the pairs of maps ϕqr , ψqr , q ≤ r
are adjunctions. If q ≺ r, then this follows immediately from the definition, since
for x ∈ (Lq , ≤), y ∈ (Lr , ≤) we have

ϕrq x ≤ y ⇔ x ∨ 0r ≤ y ⇔ x ≤ y

and dually
x ≤ ψqr y ⇔ x ≤ y ∧ 1q ⇔ x ≤ y.
Thus, we get

ϕqqm−1
m
◦ · · · ◦ ϕqq10 x ≤ y
⇔ ϕqqm−1
m−2
◦ · · · ◦ ϕqq10 x ≤ ψqqm−1
m
y
.
⇔ ..
⇔ x ≤ ψqq01 ◦ · · · ◦ ψqqm−1
m
y.

Condition (5) follows immediately from parts (1) and (3) of the definition. What
remains to be shown is condition 6) of Definition 68. In order to do so, we first
consider the case q ≺ r, for which from x ∈ (Lq , ≤) ∩ (Lq∨s , ≤) it follows that:
r∨s
ϕrq x = x ∨ 0r = x ∨ 0r ∨ 0q∨s = x ∨ 0r∨s = ϕq∨s x

(if q ∨ s ≺ r ∨ s does not hold, then the last of these equalities is inferred as
above for ϕqqm
0
). The general case is obtained by concatenation along a chain of
neighbors. 
178 4 Decompositions of concept lattices

Definition 71 The sum of a Q-atlas with overlapping neighbor maps given by

((Lq , ≤) | q ∈ Q) is described by
 [ 
Lq , ≤ ,
q∈Q

≤ being the transitive closure of the union of the orders on the summands. ♦

Theorem 27 The sum of a Q-atlas with overlapping neighbor maps is a complete lattice
L where the summands (Lq , ≤), q ∈ Q are precisely the blocks of a complete tolerance
relation Θ and where q 7→ (Lq , ≤) describes an isomorphism of Q onto L/Q.
Conversely, in a complete lattice L the blocks of a tolerance Θ with overlapping neigh-
borhoods, for which Q := L/Θ is of finite length, always form a Q-atlas with overlapping
neighbor maps whose sum is L.

Proof First of all, we shall prove that the order v of the Q-atlas, which is de-
scribed by Proposition 83, is equal to the transitive closure ≤ of the union of
the orders on the summands. According to the definition, v on the summands
(Lq , ≤) coincides with their respective orders, which is the reason why from
x ≤ y always follows x v y. If, conversely, x v y, i.e., xmin ≤ ymin and ϕyxmin
min
x ≤ y,
then for
xmin = q0 ≺ q1 ≺ · · · ≺ qm = ymin
as in Proposition 83 it follows that

ϕxymin
min
x = (. . . ((x ∨ 0q1 ) ∨ 0q2 ) ∨ . . .) ∨ 0qm ,

which yields x ≤ y. Thus we have proven that

[  [ 
(Lq , ≤), ≤ = (Lq , ≤), v .
q∈Q q∈Q

Therefore, Proposition 83 yields the assertions of the first part of Theorem 24.
Furthermore, we get that in a complete lattice L the blocks of a glued tolerance
Θ form a Q-atlas. It only remains to be shown that this Q-atlas is in fact a Q-atlas
with overlapping neighbor maps. Conditions 0), 1), 2) and 3) of Definition 70
are obviously satisfied. From 0q ≤ x ≤ 1q and 0r ≤ x ≤ 1r it follows that

0q∨r = 0q ∨ 0r ≤ x ≤ 1q ∧ 1r = 1q∧r ,

which proves 4). 5) can be seen from the fact that q ≤ r ≤ s and 0s ≤ x ≤ 1q ,
because of 0r ≤ 0s and 1q ≤ 1r , immediately yield 0r ≤ x ≤ 1r . 
Theorem 27 can be applied quite practically for the representation in dia-
grams, provided that the lattice which is to be represented has a tolerance with
overlapping neighborhoods. This can be checked by entering the arrow relations
into the context and enriching the relation
4.2 Atlas decompositions 179

J := I ∪ . ∪ %

in accordance with the conditions in Definition 59, until a block relation is ob-
tained (namely β(Σ)). According to Corollary 1, we obtain the blocks (“maps”)
as concept lattices of subcontexts. Diagrams are created from these. The overlaps
can be read off, even in the case of a reduced labelling of the individual maps,
since every concept is stated with its correct extent and intent.
According to Theorem 27, the lattice is uniquely described by this set of dia-
grams. The correctness of the atlas can be verified by means of the conditions in
Definition 70.
We shall demonstrate this using the example of a lattice of subgroups. Fig-
ure 4.7 shows a computer-generated diagram, which was taken from a book on
orthomodular lattices [205].

19

18 17 16

14 15 13 11 12

9 10 8 6 7

5 4 3 2

Figure 4.7 Computer-generated lattice diagram

180 4 Decompositions of concept lattices

10 14 15 16 17 18
2 .
% × × × × ×
3 × .
% × × %
. ×
4 × .
% × .
% × ×
5 × × × .
% .
% ×
6 × .
% × × ×
11 . × × .
%

Figure 4.8 The standard context for the lattice from Figure 4.7

The standard context (cf. Page 32) for this lattice, including the arrow rela-
tions, is presented in Figure 4.8. A short examination shows that J := I ∪ . ∪ %
is already a block relation, i.e., that it is equal to β(Σ).
The concept lattice of (G, M, J) is a three-element chain with formal concepts

({2, 3, 4, 5, 6, 11} , {14, 16, 17, 18}),

({2, 3, 4, 5, 6} , {14, 15, 16, 17, 18}), and
({2, 3, 4, 5} , {10, 14, 15, 16, 17, 18}).

The subcontexts belonging to the blocks of the relation are represented to-
gether with their concept lattices in Figure 4.9. In the case of the lattice presented
in the middle of this figure, it can be easily seen how the blocks overlap. The
smallest element of this block is the concept with the extent {2}. We discover it
in the lower lattice on the right side. The largest element of the lower lattice has
the intent {15, 18}. It can easily be found in the middle lattice.
Hence, the lower and the middle lattice have the five elements of the interval

[({2}, {14, 15, 16, 17, 18}), ({2, 3, 4, 5}, {15, 18})]

in common. Analogous are the middle and the upper lattice, which overlap in
the interval
[γ6, µ18],
also having five elements.
In the present case, it proves to be particularly convenient that we have chosen
congruent diagrams for the overlap areas. This makes it possible to superimpose
the individual part-diagrams (Figure 4.10) and thus to obtain a new diagram for
the subgroup lattice (Figure 4.11), which reflects the structure particularly well,
due to its construction method.
4.2 Atlas decompositions 181

d
@
@
14 16 17 18 18 d 17 d @ 16
@d
2 × × × × A@ @
3 × × A@ @
4 × ×
d A d14@ d
A @ @
5 × × @d
6 × × × × 4 @5 A 3 11
11 × × @ A
@A
@A d
2,6
18
d
A@
A@
15 d 17 d A d14@ d16
A @
14 15 16 17 18
2 × × × × × A@ @ A
3 × × × A@ @ A
4 × × × A @ @A
5 × × × d Ad @ d @A d
6 × × × × 4 @5 A 3 6
@ A
@A
@A d
2
15,18
d
A@
A@
10 d 17 d A d14@ d16
A @
10 14 15 16 17 18
A@ @ A
2 × × × × × A@ @ A
3 × × × × A @ @A
4 × × × × d Ad @ d @A d
5 × × × × 4 @5 A 3 2
@ A
@A
@A d

Figure 4.9 The blocks of the skeleton tolerance

182 4 Decompositions of concept lattices

e
@
@
@
18 17 @ e16
` e e @
`
``
A@ @
` ` A @ @
`
18 ` `
A @ @
e 4 e 5 A e14@
` `
@e
`
@e
@
` ` ` ` ` ` 3 11
A @` ` ` ` ` ` @` ` ` A
A@ @ A

` A` @` ` ` ` ` @ A
15 e 17 e` A e` @ @e @ @A e
` ` 14 16 `
` ` A@ @ A
` ` ` 2,6
`` A@ ``
@ A
` `
15,18 ` ` `
A @ @ A
e 4 e 5A e @ @e @ @A e`
` ` ` 3
A@
` `` @ ` ` ` A` ` ` 6
A ``
@
` ` @` A
`
` ` A` @` ` ` ` ` ` @ A
10 e 17 e A e @ e @ @ @A e
14 16
` `` 2
`
A@ @ A
A@ ``
@ A
A @ @A ` `
e Ae @ @e @@A e`
4 @ 5 A 3 2
@ A
@A
@A e
@

Figure 4.10 Atlas of the part diagrams. The dotted lines link equal concepts in the different
part diagrams. If the diagram is contracted along those lines, we obtain the diagram in Figure
4.11.
4.2 Atlas decompositions 183

f
@
@
@
@
18 f 17 f @@ f16
A@ @
A@ @
A @ @
A @ @
15 f f A f14@@f @@f
@
A @ A 11
A@ @ A
A @ @ A
A @ @A
10 f f A f @
@ f @A f
@
A@ @ A 6
A@ @ A
A @ @ A
A @ @A
f A f @
@ f @A f
@
4 @ 5 A 3 2
@ A
@ A
@A
@A f
@

Figure 4.11 The same lattice as in Figure 4.7, but with a diagram which better reflects the
structure.
184 4 Decompositions of concept lattices

4.3 Substitution

In the case of the substitution sum a context is inserted into another one at the
place of an “empty cell”, i.e., a non-incident object-attribute pair. We can visual-
ize the construction by imagining the respective row and column suitably mul-
tiplied, so that there is room for the context which is to be inserted. For reasons
of convenience we presuppose that the two contexts are disjoint and non-empty,
and thus we obtain the following definition.
Definition 72 Let K1 = (G1 , M1 , I1 ) and K2 = (G2 , M2 , I2 ) be contexts and
/ I1 be a non-incident object-attribute pair in K1 . We presuppose that
(g, m) ∈
6 M2 and (G1 \ {g}) ∩ G2 = ∅ = (M1 \ {m}) ∩ M2 . We define substitu-
G2 6= ∅ =
tion sum of K1 with K2 on (g, m) to be the context

K1 (g, m) K2 := (G, M, I)

with G := (G1 \ {g}) ∪˙ G2 , M := (M1 \ {m}) ∪˙ M2 and

I := {(h, n) ∈ I1 | h 6= g, n 6= m} ∪ G2 × g I1 ∪ mI1 × M2 ∪ I2 .

We speak of a proper substitution sum if G2I2 = ∅ = M2I2 holds4 . Then g I1 is

an intent and mI1 is an extent of (G, M, I). The corresponding concepts shall be
denoted by a and b. ♦

M2
m
g G2 K2

K2 -
K1

Figure 4.12 To form the substitution sum K1 (g, m) K2 , the context K2 is inserted into “the
empty cell” (g, m) of K1 . The hatchings in the resulting context are meant to indicate that
every object 6∈ G2 is incident either with all or with no element from M2 and, dually, that
all objects from G2 have the same intents with regard to the attributes 6∈ M2 .

From the definition it immediately follows that the substitution sum is restrict-
edly associative:
4The cases K2 ∼ = ({g}, {m}, ∅), when K1 (g, m) K2 = K1 , as well as K1 = ({g}, {m}, ∅) are
admitted. In the following we shall consider proper substitution sums only.
4.3 Substitution 185

Proposition 84

K1 (g, m) (K2 (h, n) K3 ) = (K1 (g, m) K2 ) (h, n) K3 ,

if g ∈ G1 , m ∈ M1 , h ∈ G2 , n ∈ M2 and (g, m) 6∈ I1 , (h, n) 6∈ I2 . 

The substitution sum generalizes several of the context operations we have
already introduced. The context sum, the disjoint union and the construction
from page 42 leading up to the vertical sum of the concept lattices can be obtained
as special cases of the form

(K0 (g, m) K1 ) (h, n) K2

if we choose the contexts

m n m n m n
g ×, g , resp. g ×
h × h h

for K0 , since as results we obtain

K1 ×, K1 ∅ and
K1 ×
× K2 ∅ K2 ∅ K2
We shall examine how the concept lattice of the substitution sum is related to
those of the summands. It turns out that the concept lattice of the second sum-
mand is “hung up” several times in the concept lattice of the first summand,
similarly to the sails of a ship in the rigging. (see Figure 4.13).

Rigging

Sails aa
J a
@ J aa
@ Ja J
aa J
aa Ja
b aa
J
J a
Ja J
aa J
aa J

Figure 4.13 Sails and rigging

The first proposition shows that we can rediscover B(K1 ) as a sublattice:

186 4 Decompositions of concept lattices

Proposition 85 The rigging of a proper substitution sum

(G, M, I) = K1 (g, m) K2 ,

defined as
U := B(G, M, I \ I2 ),
is a complete sublattice of B(K1 (g, m) K2 ) that is isomorphic to B(K1 ). U contains in
particular all concepts which are ≥ a or ≤ b.
Proof Proposition 57 shows that I \ I2 is closed. The context (G, M, I \ I2 ) is up
to clarification identical to K1 , i.e., the isomorphy follows from Theorem 17. A
concept (A, B) ≤ b satisfies A ⊆ mI1 and therefore A × B ∩ I2 = ∅, which implies
(A, B) ∈ U . 
The remaining concepts of K := (G, M, I), i.e., those
which do not belong to B(G, M, I \ I2 ), all contain “one
cross from I2 ” and thus are entirely contained in the M2 g 0
subcontext (G2 ∪mI1 , M2 ∪g I1 ). This subcontext, accord-
ing to the definition of K, is the sum of K2 and K3 := ∅
G2 K2
@
@
(m , g , I3 := I ∩ (m × g )), i.e., the concept lat-
I1 I1 I1 I1 @
m0 @ K3
tice of this subcontext is isomorphic to B(K2 ) × B(K3 ). @
@
We do not claim that we thereby obtain a sublattice,
∅
but by Proposition 38 we know that we find an order-
embedding of this concept lattice into B(K) by assign-
ing the concept (AII , AI ) of K to every concept (A, B)
of K2 + K3 . The proposition just mentioned suggests a
further order-embedding, namely (A, B) 7→ (B I , B II ). In the present case, how-
ever, this yields the same map, since AI = B or B I = A for every concept (A, B)
of K2 + K3 .
We denote the image of this mapping by P and summarize:
Proposition 86 The map

ϕ : B(K2 + K3 ) → B(K1 (g, m) K2 ),

(A, B) 7→ (AII , AI ) (= (B I , B II )),

is an order-embedding, mapping the concept with the extent G3 (= mI1 ) onto b and the
concept with the intent M3 (= g I1 ) onto a. The range P covers all concepts which do
not belong to U . 
We call P the sails of the substitution sum. The following theorem shows that
the two parts, rigging and sails, determine the structure of the concept lattice of
a substitution sum. However, we must indicate how U and P are joined together.
Before doing so, we shall introduce the corresponding lattice construction.
Definition 73 For complete lattices U and W , |W | > 1 and elements a, b ∈ U
with a 6≤ b, we define the substitution product U (a, b) W of U and W on (a, b)
to be the concept lattice of the (proper) substitution sum
4.3 Substitution 187

U (a, b) W := B ((U, U, ≤) (a, b) (W0 , W1 , ≤))

with W0 := W \ {0W } and W1 := W \ {1W }. ♦

Hence, according to this definition, the substitution product of U and W is the
concept lattice of the context (G, M, I) with

G = (U \ {a}) ∪ (W \ {0W }),

M = (U \ {b}) ∪ (W \ {1W }),

and
g ≤ m in U, if g ∈ U, m ∈ U

g ≤ b in U, if g

∈ U, m ∈ W

gIm ⇔
 a ≤ m in U,
 if g ∈ W, m ∈ U
g ≤ m in W, if g ∈ W, m ∈ W

The rigging of this substitution sum is naturally isomorphic to U ; therefore, we

also denote it by U . We also take over the names a and b for the corresponding
elements of the substitution product.

c c
 cHH  cHH
c BZ HH c BZ HH
 
c Z
Bc Z c
Hc c  Z Hc
C c c c
 Bc Z c
H bJ bJb  HH A C
 cH C b  HA cH C
 HH AA C J b bb C
b c c c c
Jb c c
Jb c c c ca

H a b  J JJ J b b b H
C H cH  C H cHb
H   H bb J J
  bbJb  
C AA cHH c c C AA cHHb c b
b Jb
J bJcJb
JcJ c
Cc Z c Cc Z c
HH Z B  c c c@c @ HH Z B  c
HHZB c  HHZB c 
U H c W@c
@
U (a, b)W H c

Figure 4.14 A substitution product.

Theorem 28 (Properties of the substitution product) The concept lattice L of the

substitution sum
L = B(K1 (g, m) K2 )
with rigging U and sails P has the properties (Subst 1) – (Subst 4), specified below.
a, b are as in the propositions and W := B(K2 ). Conversely, every complete lattice L
satisfying (Subst 1) – (Subst 4) is isomorphic to U (a, b) W . In particular,

B(K1 (g, m) K2 ) ∼
= B(K1 ) (γg, µm) B(K2 ).
188 4 Decompositions of concept lattices

(Subst 1) L = U ∪ P , P ∩ U ⊆ (b] ∪ [a), a, b ∈ U , a 6≤ b.

(Subst 2) U is a complete sublattice of L,
(Subst 3) W is order-isomorphic to P ∩ (a], and

p = (p ∧ a) ∨ (p ∧ b) as well as p = (p ∨ a) ∧ (p ∨ b)

for all p ∈ P ,
(Subst 4) If u ∈ U and p ∈ P \ U , then u ≤ p ⇒ u ≤ b and u ≥ p ⇒ u ≥ a.

Proof In Propositions 85 and 86 we have already shown that B(K1 (g, m)K2 )
has the properties (Subst 1) and (Subst 2), and furthermore that P is order-
isomorphic to the direct product of B(K2 ) and B(K3 ). This yields the first part
of (Subst 3). A concept p = (X, Y ) of K1 (g, m)K2 which does not belong to U
satisfies X ⊆ G2 ∪ mI1 and Y ∩ M2 6= ∅. Every subconcept of it which belongs to
U must therefore have an extent which is entirely contained in mI1 , i.e., it is ≤ b.
This proves (Subst 4). Now we can infer the second part of (Subst 3), since an
upper bound of (p ∧ a) ∨ (p ∧ b) which is less than or equal to p must be contained
in P and must therefore be equal to p.
We now assume, conversely, that L has the properties (Subst 1) – (Subst 4).
First we derive some information from the structure of P . From (Subst 3) we
infer that P ⊆ [a ∧ b, a ∨ b]. The maps p 7→ p ∨ b and q 7→ q ∧ a are isomorphisms
between P ∩ (a] and P ∩ [b) which are inverse to each other, since it holds for
p ≤ a that
(p ∨ b) ∧ a = (p ∨ b) ∧ (p ∨ a) = p,
and dually. In this way we do not only obtain the isomorphism σ : W → P ∩ (a]
postulated in (Subst 3) but a further isomorphism τ : W → P ∩ [b) by virtue of
τ (x) := σ(x) ∨ b. From (Subst 1) we can see that the elements of P ∩ (a], with
the exception of the boundary elements a and a ∧ b, do not belong to U . The
corresponding is true for P ∩ [b).
In order to prove the isomorphy of L with U (a, b) W , we use the Basic The-
orem on Concept Lattices. The context defining U (a, b) W has the object set
G = (U \ {a}) ∪ (W \ {0W }) and the attribute set M = (U \ {b}) ∪ (W \ {1W })
(cf. the explanation following Definition 73). The maps

γ̃ : G → L, µ̃ : M → L,

which are defined by

if x ∈ U if x ∈ U
 
x x
γ̃(x) := and µ̃(x) :=
σ(x) if x ∈ W τ (x) if x ∈ W ,

satisfy the conditions mentioned in the Basic Theorem, as we shall

V show. For this
purpose we have to prove that γ̃(G) is -dense and µ̃(M ) is -dense in L and
W
that gIm is equivalent to γ̃g ≤ µ̃m.
We have γ̃(G) = U ∪ (P ∩ (a]), since σ(0W ) = a ∧ b according to (Subst 2) is
an element of U . Likewise, µ̃(M ) = U ∪ (P ∩ [b)). Because of p = (p ∧ a) ∨ (p ∧ b),
4.3 Substitution 189

every element p 6∈ U is the supremum of some element of P ∩ (a] and some

element of (b], i.e., in any case of elements of γ̃(G). Dually, we show that µ̃(M ) is
infimum-dense.
In order to prove g I m ⇔ γ̃g ≤ µ̃m, we distinguish four different cases,
depending on whether g ∈ U or g ∈ W and whether m ∈ U or m ∈ W . If
both are contained in U , the assertion is obviously right. If g ∈ U , m ∈ W , then
by Definition 73 g I m is equivalent to g ≤ b in U . On the other hand, µ̃(m) =
τ (m) 6∈ U or µ̃(m) = b, i.e., by (Subst 4) γ̃g ≤ µ̃m ⇔ γ̃g ≤ b, and, because of
γ̃g = g the conditions are equivalent. The case g ∈ W, m ∈ U is treated dually.
Finally, we have to deal with the case g, m ∈ W : In this case we have g I m ⇔
g ≤ m ∈ W ⇔ σ(g) ≤ σ(m) ⇔ σ(g) ≤ σ(m) ∨ b (since from σ(g) ≤ σ(m) ∨ b it
follows that σ(g) = σ(g)∧a ≤ (σ(m)∨b)∧b = σ(m)), and because of γ̃(g) = σ(g)
as well as µ̃(m) = τ (m) = σ(m) ∨ b everything has been proved. 
It is easy to derive further information on the set P . In the following proposi-
tion we compile some information (without proof).

Proposition 87 (Further Properties) The concept lattice

B(K1 (g, m) K2 )

of a proper substitution sum has (using the notations in Theorem 28) the following prop-
erties:
(Subst 5) The sails P are isomorphic to a direct product

P ∼
= (P ∩ (a]) × (P ∩ (b]).

The elements of P ∩ (b] are precisely those of the form x = (x ∨ a) ∧ b.

(Subst 6) It holds for x ∈ P ∩ (a], y ∈ P ∩ [b) that

x≤y ⇔ x≤y∧a ⇔ x ∨ b ≤ y.

(Subst 7) Each element of P ∩ (a] is the supremum of object concepts in P ∩ (a], and
each element of P ∩ W[b) is the infimum of attribute concepts in P ∩ [b).V
(Subst 8) If 1W is -irreducible, then a is an object concept. If 0W is -irreducible,
then b is an attribute concept. 
190 4 Decompositions of concept lattices

The substitution product of a lattice L with the two-element

lattice is always isomorphic to L. Therefore, we call L substi-
tutionally indecomposable if L has more than two elements ×
×
and if from
L∼= L1 (a, b) L2 substitutionally
indecomposable
it always follows that |L1 | = 2 or |L2 | = 2. We conclude our
investigation by settling the question: under which circum-
stances does substitutional decomposability of a concept lat- ×
×
tice B(K) imply decomposability of the context K? This im-
plication does not apply generally, since due to the addition
(in special cases also the removal) of reducible objects and at- substitutionally
tributes a context can loose its property of being a substitution decomposable
sum.
In the following theorem (which we prepare by a proposition) we therefore
switch over to a dense subcontext.

Proposition 88 A context K is isomorphic to a proper substitution sum of contexts with

concept lattices isomorphic to U and W , if and only if there is a lattice isomorphism ψ of
B(K) onto a substitution product U (a, b)W with

ψγ(G) ⊆ U ∪ (a] and ψµ(M ) ⊆ U ∪ [b),

for which additionally the following special condition is satisfied:

• if 0W is W-irreducible, then b ∈ ψµ(M ),
V
• if 1W is -irreducible, then a ∈ ψγ(G).

Proof If K = K1 (g, m)K2 , then, by the preceding theorem,

B(K) ∼
= B(K1 ) (γg, µm) B(K2 ).

If h ∈ G is an object, then the object concept belongs either to the rigging U W or

it is contained in the object set G2 ; then, however, γh ≤ γg = a. If 1W is -
irreducible, then 1W is an object concept in K2 , this object then also belongs to K
and is mapped under ψγ on a. We argue dually for the attributes.
Now, conversely, let ψ be an isomorphism with the properties stated in the
proposition. Let P again be the sails of U (a, b) W , and furthermore assume that
ga 6∈ G and mb 6∈ M . We define a context K1 by

GU := {g ∈ G | ψγg ∈ U \ {a}}, G1 := GU ∪ {ga }

MU := {m ∈ M | ψµm ∈ U \ {b}}, M1 := MU ∪ {mb }

I1 := I ∩ (GU × MU ) ∪ {(ga , m) | ψµm ≥ a} ∪ {(g, mb ) | ψγg ≤ b}.
The concept
W lattice of this context K1 := (G1 , M1 , I1 ) is isomorphic to U , since
ψγ(G) is -dense in U (a, b) W , and by (Subst 4) (ψγ(G) ∩ U ) ∪ {a} then is -
W
dense in U . Dually, (ψµ(M ) ∩ U ) ∪ {b} is infimum-dense in U . Therefore, the
4.3 Substitution 191

maps γ1 : G1 → U and µ1 : M1 → U being defined by

ψγg, if g ∈ GU , falls m ∈ MU ,
 
ψµm,
γ1 (g) := and µ1 (m) :=
a, if g = ga , b, if m = mb ,

by the Basic Theorem are sufficient for the isomorphy of B(K1 ) with U , since
evidently g I m ⇔ γ1 g ≤ µ1 m.
The context K2 := (G2 , M2 , I2 ) is explained through

G2 := G \ GU , M2 := M \ MU , I2 := I ∩ G2 × M2 . 

We claim that B(K2 ) ∼

= W (∼
= P ∩ (a]) and argue again with the help of the Basic
Theorem, using the maps γ2 : G2 → P ∩ (a] and µ2 : M2 → P ∩ (a], which are
defined as follows:

γ2 g := ψγg, µ2 m := (ψµm) ∧ a.

With (Subst 6) we obtain

ψγg ≤ (ψµm) ∧ a ⇔ ψγg ≤ ψµm,

i.e., γ2 g ≤ µ2 m ⇔ (g, m) ∈ I2 . (Subst 7) says that the object concepts are -

W
dense in P ∩ (a]. By (Subst 1), however, P ∩ (a] does W not contain elements of U
with the exception of a and a ∧ b. Therefore, γ2 G2 is -dense in the lattice P ∩ (a],
since by (Subst 8) the largest element a is also a supremum of elements of γ2 G2 .
If we take into account that by (Subst 3) x V7→ x ∧ a is an isomorphism of P ∩ [b)
on P ∩ (a], we obtain that, dually, µ2 M2 is -dense in P ∩ (a]. Without difficulty
we verify that
K = K1 (ga , mb )K2 .


Theorem 29 If B(K) ∼ = U (a, b) W , then there is a dense subcontext K0 of K which

is a proper substitution sum of contexts with concept lattices isomorphic to U and W ,
provided that the isomorphism ψ : U (a, b) W → B(K) can be chosen such that the
following (necessary) extra condition is satisfied:
• If 1W is V-irreducible, then ψ(a) is an object concept γga of K.
W
• If 0W is -irreducible, then ψ(b) is an attribute concept µmb of K.

Proof If
ψ : B(K) → U (a, b) W
is an isomorphism, so that ψ −1 satisfies the extra condition, then we define sub-
context K0 := (G0 , M0 , I ∩ G0 × M0 ) through

G0 := {g | ψγg ∈ U ∪ (a]}
M0 := {m | ψµm ∈ U ∪ [b)}.
192 4 Decompositions of concept lattices

We only have to prove that K0 is a dense subcontext, since in this case Proposi-
tion 88 yields the rest of the assertion.
For this purpose we show that γG0 is -dense. The corresponding asser-
W
tion M0 isWproved dually. Since ψ is an isomorphism,
W we can prove instead that
ψγG0 is -dense in U (a, b) W . ψγG is certainly -dense, i.e., every element
s ∈ U (a, b) W can be represented as a supremum
_
s= X, X := (s] ∩ ψγG.

We distinguish four different cases:

s ∈ U, s 6≥ a By (Subst 4), X ⊆ U .
s≤a Then, trivially, X ⊆ (a].
s∈P By (Subst W∧ a) ∨ (s ∧ b) and s ∧ b ∈ U , s ∧ a ≤ a. Therefore,
3), s = (s
we have s = (X ∩ (a]) ∨ (X ∩ U ).
W
s≥a By (Subst 3) for every p ∈ P , a ∨ p = a ∨ (p ∧ b) with p ∧ b ∈ U . Hence,
each element of X which does not belong to U , can be replaced by elements
of (a] ∪ (b], and those in turn are suprema of elements of ψγG ∩ ((a] ∪ (b]).
This means that in the end those elements of ψγG are sufficient which are con-
tained in U ∪ (a], i.e., the images of G0 , as claimed. 
We shall use these results to prove a theorem on “unique prime factor de-
composition” for the substitution product of finite lattices. However, this does
not work quite smoothly, the extra condition makes itself felt and prevents a re-
sult without exceptions. The decisive technical aid is a refinement result, which
shows that two substitution products can only be isomorphic if they are made
up of the same factors.
Proposition 89 Let L1W, L2 , L3 , L4 be doubly foundedVcomplete lattices. Assume that
L3 and L4 each have a -reducible unit element and a -reducible zero element. If
∼ L2 (a2 , b2 ) L4
L1 (a1 , b1 ) L3 =

holds for suitable elements a1 , a2 , b1 , b2 , then there are lattices W1 , W2 , W3 , W4 and

elements c1 , . . . , c4 as well as d1 , . . . , d4 with
∼ W1 (c1 , d1 ) W2 ,
L1 = L2 ∼
= W1 (c2 , d2 ) W3 ,
∼ W3 (c3 , d3 ) W4 ,
L3 = ∼ W2 (c4 , d4 ) W4 .
L4 =

The proof of the proposition can be illustrated by Figure 4.15.

It represents a substitution sum K := K1 (g, m) K2 . The subcontext K2 is
drawn in, K∗1 =: (G∗1 , M1∗ , I1∗ ) denotes the context resulting from K1 by omission
of the object g and of the attribute m. We obtain a context which is isomorphic
to K1 by adding an arbitrary non-incident object-attribute pair from K2 . Hence,
except for the names of g and m, K1 is also given, and those names are irrelevant
for the substitution sum. Therefore, in this situation we write K = K1 ( , ) K2 as
an abbreviation for the fact that K = K1 (g, m) K2 holds for suitable g, m.
4.3 Substitution 193

M2

G2 K2

K∗1

K = K1 ( , ) K2 .

Figure 4.15 With reference to the proof of Proposition 89.

Proof. Assume that

L := L1 (c1 , d1 ) L3 ∼
= L2 (c2 , d2 ) L4

and that K is the (reduced) standard context for L. We can apply Theorem 29,
since the extra condition is irrelevant because of the additional preconditions.
Hence, K is in two ways a substitution sum:

K∼
= K1 (g1 , m1 ) K3 ∼
= K2 (g2 , m2 ) K4 ,

with
B(Ki ) ∼
= Li and Ki =: (Gi , Mi , Ii )
for i ∈ {1, . . . , 4}. The presuppositions of the proposition guarantee that K3 and
K4 have neither full nor empty rows or columns.
Hence, the isomorphy of the substitution products corre- W
sponds to the isomorphy of two substitution sums. In the fol- L1 1 L2
lowing, we shall make use of this circumstance. There are, how-
@
R
@
W2 W3
ever, some complications, since the result for substitution sums
which corresponds to Proposition 89 does not hold in general. L@@
R
4 L
W4 3
If, however, we manage to find contexts L1 , L2 , L3 , L4 with
K1 = L1 ( , ) L2 , K 2 = L1 ( , ) L3 , L1
K1 K2
K3 = L3 ( , ) L4 , K 4 = L2 ( , ) L4 , @
R
@
L2 L3
then the statement of the proposition results from Theorem 28.
K@
R
@
4 K3
This is true for some special cases. These will be dealt with first. L4

Case 1: G3 ⊂ G4 and M3 ⊂ M4 (see Figure 4.16).

We define a subcontext
K∗0 := (G∗0 , M0∗ , I0∗ )
through
G∗0 := G4 \ G3 and M0∗ := M4 \ M3 .
194 4 Decompositions of concept lattices

M4

K3
G4
K∗0

K∗2 M1∗

G∗1

Figure 4.16 With reference to the proof of Proposition 89.

As above we expand this context into a context K0 by adding a non-incident pair

from K3 . The latter cannot contain any full rows or full columns, because K4 does
not contain any either. Therefore, we have

K∗1 = K2 ( , ) K∗0 , K1 = K2 ( , ) K0 , K4 = K0 ( , ) K3

If, as a fourth context, we add the trivial context 2 := ({g}, {m}, ∅), we obtain
the desired refinement with

K1 = K2 ( , ) K0 K2 = K2 ( , ) 2
K3 = 2 ( , ) K3 K4 = K0 ( , ) K3

(cf. the left diagram of Figure 4.17). Of course, thereby we have also dealt with
the converse case G4 ⊂ G3 , M4 ⊂ M3 .

K3
K2 K0
K1 K2 K1 K2
K∗∗
@
2
R
@ @
R
@
K0 0 K4 K3
M1∗
K@ K@
2
R
@
4 K3 R
@
4 K3
K3 K4

G∗1

Figure 4.17 With reference to the proof of Proposition 89.

4.3 Substitution 195

Case 2: G3 ∩ G4 = ∅ = M3 ∩ M4 .
We define (middle figure in 4.17)

G0∗∗ := G \ (G3 ∪ G4 ),
M0∗∗ := M \ (M3 ∪ M4 ).

Let K0 be the subcontext with the object set G0 := G0∗∗ ∪ {g, h} and the attribute
set M0 := M0∗∗ ∪{m, n}, with (g, m) and (h, n) being non-incident object-attribute
pairs of K3 or K4 , respectively.
We recognize that K0 (g, m) K4 is equal to K1 . Furthermore, K2 = K0 (h, n) K3 .
Once again by using the trivial context 2 we obtain the refinement represented
in the diagram on the right, which yields the statement of the proposition.
The cases G3 ⊆ G4 , M4 ⊆ M3 (resp. dually) and G3 ∩ G4 6= ∅, M3 ∩ M4 = ∅
turn out to be trivial: If, for instance, G3 ⊆ G4 and m ∈ M3 \M4 , then m0 ∩G4 = ∅
or m0 ∩ G4 = G4 and consequently m0 ∩ G3 = ∅ or m0 ∩ G3 = G3 , contrary to our
presuppositions. If M3 ∩ M4 = ∅, then G3 ⊆ G4 is obviously impossible, since
otherwise K3 would have constant columns. So, if g ∈ G3 ∩ G4 , h ∈ G3 \ G4 and
m ∈ M4 , then from g I m it immediately follows that h I m and thus h I n for all
n ∈ M4 , which in turn necessitates g I n for all n ∈ M4 . Similarly, from (g, m) ∈
/I
it follows that g 0 ∩ M4 = ∅. This means that K4 would contain a full row or an
empty row, which is contrary to the presuppositions.
What remains is the case that the two subcontexts K3 and K4 intersect non-
trivially. We can proceed similarly as we have done so far and introduce contexts
L1 , . . . , L4 , as presented in Figure 4.18:

K3 L∗3
L1
L4 K1 K2
@
R
@
K4 L∗2 L2 L3

L∗1 K@
R
@
4 K3
L4

Figure 4.18 With reference to the proof of Proposition 89.

With H1∗ := G \ (G3 ∪ G4 ), N1∗ := M \ (M3 ∪ M4 ), H2∗ := G4 \ G3 , N2∗ :=

M4 \ M3 , H3∗ := G3 \ G4 , N3∗ := M3 \ M4 and H4 := G3 ∩ G4 , N4 := M3 ∩ M4
we define subcontexts L1 , . . . , L4 , which indeed yield the refinement indicated
in the diagram, provided that the substitution sums appearing in the process are
proper substitution sums. However, it may happen that L4 contains full rows or
full columns (for L2 and L3 this shall be excluded, as in the first case).
A simple trick helps us along: We extend the context K by an object g∞ and
an attribute m∞ with g∞0
:= g40 ∩ M1∗ and m0∞ := m04 ∩ G∗1 , g4 ∈ G4 and m4 ∈ M4
196 4 Decompositions of concept lattices

being chosen arbitrarily. The new object and the new attribute are reducible in
K as well as in the subcontexts K1 , . . . , K4 , i.e., the respective concept lattices are
isomorphic. If by L+ 4 we denote the context resulting from L4 , we obtain

K 1 = L1 ( , ) L2 K2 = L1 ( , ) L3
K3 = L3 ( , ) L+
4 K4 = L2 ( , ) L+
4,

and thus the assertion. 

By means of this proposition we can finally prove the result which we asserted
above. In this context, the three-element lattice C3 plays a special role, because
it is the only substitutionally indecomposable lattice which does not satisfy the
additional condition from Proposition 89.
Theorem 30 If no substitutional decomposition of the finite lattice L contains a factor
isomorphic to C3 , then any two substitutional decompositions of L into indecomposable
lattices have the same length and pairwise isomorphic factors.
Proof According to Proposition 84 each substitution decomposition can be put
into a left-bracketed form. Hence, let
∼ ((. . . (M1 ( , ) M2 ) . . .) ( , ) Mm−1 ) ( , ) Mm
L=

L∼
= ((. . . (N1 ( , ) N2 ) . . .) ( , ) Nn−1 ) ( , ) Nn
be two decompositions of L into indecomposable factors M1 , . . . , Mm or N1 , . . .,
Nn , respectively (the names of the elements in brackets are irrelevant for the
proof). Assume that n is the largest possible length which this kind of decom-
position of L can have. We proceed by induction on n.
According to Proposition 89 there are lattices W1 , . . . , W4 with

((. . . (M1 ( , ) M2 ) ( , ) M3 . . .) ( , ) Mm−1 ∼ = W1 ( , ) W2 ,

∼
((. . . (N1 ( , ) N2 ) ( , ) N3 . . .) ( , ) Nn−1 = W1 ( , ) W3 ,

Mm ∼
= W3 ( , ) W4 and Nn ∼
= W2 ( , ) W4 .
Since Mm and Nn are substitutionally indecomposable, |W2 | = |W3 | = 2 or
∼ W4 ∼
|W4 | = 2. In the first case Mm = = Nn and
∼ W1 ∼
(. . . (M1 ( , ) M2 ) . . .)Mm−1 = = (. . . (N1 ( , ) N2 ) . . .)Nn−1 ,

and the assertion follows by means of induction. If |W4 | = 2, then Mm ∼

= W3 and
Nn ∼
= W2 , and we have
∼L∼
(W1 ( , ) Nn ) ( , ) Mm = = (W1 ( , ) Mm ) ( , ) Nn

Since N1 is indecomposable, for n = 2 we obtain |W1 | = 2, and thus m = 2,

M1 ∼= N2 and N1 = ∼ M2 . If n > 2 we can infer that every substitutional decom-
position of W1 has at most n − 2 factors, because otherwise we would have a
4.4 Tensorial decompositions 197

decomposition of L with more than n factors. By induction we can infer that all
decompositions of W1 into indecomposable factors have the same number k of
factors. This number, however, must equal n − 2, since
∼ (. . . ((N1 ( , ) N2 ) ( , ) N3 ) . . .) ( , ) Nn−1 .
W 1 ( , ) Mm =

Also by the induction hypothesis, every decomposition of this lattice has pre-
cisely n − 1 factors. 

4.4 Tensorial decompositions

Definition 74 Let T be an index set. The direct product of contexts Kt :=

(Gt , Mt , It ), t ∈ T is defined to be the context

×K
t∈T
t := ( × G , × M , ∇),
t∈T
t
t∈T
t

with
g∇m : ⇔ ∃t∈T gt It mt
for g := (gt )t∈T and m := (mt )t∈T . ♦
We had introduced this definition already in Section 1.4 for the special case of
two factors:
K1 × K2 := (G1 × G2 , M1 × M2 , ∇),
(g1 , g2 )∇(m1 , m2 ) : ⇔ g1 I1 m1 or g2 I2 m2 .
For reasons of simplicity we will use the following abbreviations throughout this
section:

G := ×G ,
t∈T
t M := ×M ,
t∈T
t g := (gt )t∈T and m := (mt )t∈T .

A tiresome complication in the notation stems from the trivial case of the “full
rows” and “full columns”. We use the notation introduced in Section 3.3

:= M ∇ × M ∪ G × G∇ .

However, the reader can assume without great loss of generality that full rows
and columns do not occur, and can set G∇ = M ∇ = = ∅ everywhere.
Every trivial context having only one concept, i.e., every context of the form
(G, M, G×M ), acts like a zero element for the direct product: If one of the factors
is trivial, so is the product. Occasionally, we have to exclude this case.

Proposition 90 For every t ∈ T , the relation

198 4 Decompositions of concept lattices

c d e
a b
3 ×
1 ×
2 × 4
5
×
×
=
K1
K2

a a a b b b a b a b a b
c d e c d e c c d d e e
1 3 × × × × 1 3 × × × ×
1 4 × × × × 2 3 × ×
1
2
5
3
×
×
× ×
×
×
= 1 4 ×
2 4
×
×
× ×
×
2 4 × × 1 5 × × × ×
2 5 × × 2 5 × ×

K1 × K2 K1 × K2

Figure 4.19 The direct product of two small contexts.

e
@
@
@
(a, c) ee(a, d) @
@ e (a, e)
@@ @
@ @ @
@ @ @
(b, c) (b, d) @@ e(b, e)
e @@ e @
@ e
(2,3) @ @ (2,4) (2,5)
@ @
@ @
@e
(1,3)@ e (1,4) @
@ e (1,5)
@
@
@
@@e

B(K1 × K2 )

Figure 4.20 The concept lattices of the direct product of the contexts from Figure 4.19.
4.4 Tensorial decompositions 199

ε1 (1) = ε2 (1)
c
@
@
1 c c c @@c
1
c
@ @ @
@ @ @
c ε2 (x) c @c @ cε2 (y)@
@ c ε2 (z) c cy@ c z
@
w ε1 (w)
@ @ x

c
@ @ @ @
@ @
0
0 c @c c @c
@ @
@
@
@@c
ε1 (0) = ε2 (0)

Figure 4.21 The maps εi : B(Ki ) → B(K1 × K2 ).

∇t := {(g, m) ∈ G × M | gt It mt } ∪

is a closed subrelation. If ∇ 6= G × M , then the corresponding sublattice B(G, M, ∇t )

is isomorphic to B(Kt ), and the map

εt : B(Kt ) → B(G, M, ∇)

with

εt (A, B) := ({g ∈ G | gt ∈ A} ∪ M ∇ , {m ∈ M | mt ∈ B} ∪ G∇ )

is a canonical lattice-embedding.
Proof The fact that ∇t is closed can be proved easily, for instance by means of
Proposition 57: If (g, m) ∈ ∇ \ ∇t , then in particular m∇ 6= G, i.e., we can choose
an object g̃ 6∈ m∇ . Using this object, we define an object h to be

g̃s if s 6= t

hs := .
gs if s = t

We get (h, m) 6∈ ∇ and h∇ = h∇t .

It is a matter of routine to prove that εt has the properties claimed. 

Proposition 91 If g, h are objects of the direct product × t∈T Kt , then

h ∈ M ∇ or

g ∇ ⊆ h∇ ⇔ .
gtIt ⊆ hIt t for all t ∈ T
200 4 Decompositions of concept lattices

Proof “⇒”: We presuppose the negation of the right side. Assume that m 6∈ h∇
and nt ∈ gtIt \ hIt t for some t ∈ T . Consider the attribute m̃, defined by

ms if s 6= t

m̃s := .
nt if s = t

Then m̃ ∈ g ∇ \ h∇ , i.e., g ∇ 6⊆ h∇ . The direction “⇐” is trivial. 

Proposition 92
g . m ⇔ ∀t∈T gt . mt ,
g % m ⇔ ∀t∈T gt % mt .

Proof We prove only the first statement. First of all we notice that, because of
(g, m) 6∈ ∇ ⇔ ∀t (gt , mt ) 6∈ It , we can limit ourselves to non-incident pairs. If,
for some t, gt . mt does not hold, there must be an object ht ∈ Gt with gt0 ⊆ h0t ,
gt0 6= ht0 and (ht , mt ) 6∈ It . The object g̃, defined by

gs if s 6= t

g̃s :=
ht if s = t

satisfies
g ∇ ⊆ g̃ ∇ , g ∇ 6= g̃ ∇ and (g̃, m) 6∈ ∇,
which yields ¬(g . m).
Analogously, we infer the converse direction: If gt . mt holds for all t ∈ T ,
then we certainly have (g, m) 6∈ ∇ and we only have to consider an object h with
g ∇ ⊆ h∇ , g ∇ 6= h∇ . By Proposition 91 we obtain gs0 ⊂ h0s , gs0 6= hs0 for some s ∈ T ,
from which, because of gs . ms , it immediately follows that hs Is m and thus
h∇m. 
Together with Proposition 15 this yields
Corollary 3 An object g of a direct product is irreducible if and only if all gt are irre-
ducible. The corresponding is true for attributes.
The direct product of reduced contexts is reduced, the direct product of doubly founded
contexts is doubly founded. 
Our interest lies in the concept lattice of the direct product. We shall call this
lattice the tensor product of the factor lattices B(Kt ). Thereby we obtain a new
lattice construction and thus a new decomposition principle. However, in order
to do so we have to show that the tensor product is independent (up to iso-
morphism) of the choice of the underlying contexts Kt . This is the result of the
theorem which follows the next definition.
Definition 75 The tensor product of complete lattices Lt , t ∈ T , is defined as

×
O
Lt := B( (Lt , Lt , ≤)),
t∈T t∈T
4.4 Tensorial decompositions 201

i.e., for the special case of two factors as

L1 ⊗ L2 := B(L1 × L2 , L1 × L2 , ∇)

with
(g1 , g2 )∇(m1 , m2 ) : ⇔ g1 ≤ m1 or g2 ≤ m2 . ♦

f
@
f f f f @f
@
@ @ @ @
f ⊗ f f @f f @f @f @ @f
@ ∼
@ @
=
@ @ @
f @f @f f @f
@ @ @
@
@@f

Figure 4.22 B(K1 ) ⊗ B(K2 ) ∼

= B( K 1 × K 2 )

Theorem 31 The concept lattice of a direct product of contexts is isomorphic to the tensor
product of the concept lattices of the factor contexts:

× Kt ) ∼
O
B( = B(Kt ).
t∈T t∈T

For the proof we use the Basic Theorem on Concept Lattices. According to Def-
inition 75, the tensor product t∈T B(Kt ) is the concept lattice of the context
N
(G, M, ∇) with
G=M = B(Kt )×t∈T

and
(At , Bt )t∈T ∇ (Ct , Dt )t∈T ⇔ ∃t∈T At ⊆ Ct .
In order to prove the isomorphy we claimed, we have to give maps

γ̃ : G → B( ×K )
t∈T
t and µ̃ : M → B( ×K )
t∈T
t

which have the properties postulated by the Basic Theorem. We choose

γ̃((At , A0t )t∈T ) := ({g | ∀t gt ∈ At } ∪ M ∇ , {m | ∃t mt ∈ A0t })

202 4 Decompositions of concept lattices

µ̃((Bt0 , Bt )t∈T ) := ({g | ∃t gt ∈ Bt0 }, G∇ ∪ {m | ∀t mt ∈ Bt }).

First, we have to show that these are concepts, i.e., that

{g | ∀t gt ∈ At }∇ = {m | ∃t mt ∈ At0 } and
∇
M ∪ {g | ∀t gt ∈ At } = {m | ∃t mt ∈ At0 }∇ .

The inclusions ⊆ are trivial. Therefore, let m be an attribute with mt 6∈ A0t for
all t ∈ T . Then for every t ∈ T there exists an object gt ∈ At with (gt , mt ) 6∈ It ;
and thus g := (gt )t∈T satisfies (g, m) 6∈ ∇. This proves the inclusion ⊇ in the first
case, the second case as well as the dual proof for µ̃ are analogous.
Next we show that γ̃G is supremum-dense by proving that γ̃G contains all
×
object concepts of B( t∈T Kt ). We have (with g := (gt )t∈T )

γ̃((gt00 , gt0 )t∈T ) = (. . . , {m | ∃t∈T mt ∈ gt }) = (. . . , g ∇ ).

Finally, we have

γ̃((At , Bt )t∈T ) ≤ µ̃((Ct , Dt )t∈T )

⇔ {g ∈ G | ∀t∈T gt ∈ At } ∪ M ∇ ⊆ {g ∈ G | ∃t∈T gt ∈ Ct }
⇔ ∃t∈T At ⊆ Ct ,

since ∀t At 6⊆ Ct ⇔ ∃g∈G ∀t∈T gt ∈ At \ Ct . However, the condition ∃t∈T At ⊆ Ct

is equivalent to (At , Bt )t∈T ∇ (Ct , Dt )t∈T , which remained to be proved. 

Theorem 32 The congruence lattice of a tensor product of finitely many doubly founded
lattices is isomorphic to the tensor product of the congruence lattices:

Lt ) ∼
O O
C( = C(Lt ).
t∈T t∈T

Proof According to Theorem 13, C(Lt ) ∼ = B(Gt , Mt , .

.\ ), where gt . . mt in Lt
if and only if there are objects g1 , . . . , gn and attributes m1 , . . . , mn with

gt = g1 . m1 - g2 . . . . - gn . mn = mt .

If such a sequence of elements of length n exists, then it also exists for every
number which is larger than n, since in this case there is an attribute k ∈ Lt with
. k and consequently gt . k - gt , whereby the sequence can be extended
gt %
arbitrarily. We can apply Proposition 92 and, since T is finite, we obtain

. m in
g. ×K
t∈T
t ⇔ . mt in Kt for all t ∈ T,
gt .

and consequently

g.
. \ mt for some t ∈ T.
\ m ⇔ gt .
.
4.4 Tensorial decompositions 203

Therefore,
(G, M, .
.
\ )=( ×(G , M , ..\)),
t∈T
t t

which together with Theorem 13 proves the assertion. 

The tensor product has been defined as the concept lattice of the context

(G, M, ∇) := ×(L , L , ≤),

t∈T
t t

with G∇ = {m | ∃t mt = 0} and M ∇ = {g | ∃t gt = 0}. The concept lattices of

Nlattices Lt . Therefore, it seems
the factor contexts are naturally isomorphic to the
reasonable to denote the embedding of Ls into t∈T Lt with the same letter as
the
N corresponding embedding in Proposition 90. Hence, we define εs : Ls →
t∈T Lt through

εs (xs ) := ({g ∈ G | gs ≤ xs } ∪ M ∇ , {m ∈ M | xs ≤ ms } ∪ G∇ ),

with (G, M, ∇) := ×t∈T (Lt, Lt, ≤).

Proposition 93 For each object concept y := γ̃(g) and for each attribute concept z :=
µ̃(m) of the tensor product as well as for every subset S ⊆ T , we have
_
y≤ εs (xs ) ⇔ ∃s∈S y ≤ εs (xs )
s∈S
^
z≥ εs (xs ) ⇔ ∃s∈S z ≥ εs (xs ).
s∈S

Proof In other words, the proposition claims that the extent of s∈S εs (xs ) is
W
exactly the union of the extents of the εs (xs ), s ∈ S, and dually. This immediately
results from the explicit descriptions of these sets which were given above. 

Hence, the sublattices εt (Lt ) are mutually distributive: The supremum resp.
infimum of elements from different εt (Lt ) can be obtained by forming the union
of the extents or intents, respectively. This implies a calculation rule which will
be formulated in the following definition:
Definition 76 We call two subsets X and Y of a complete lattice mutually dis-
tributive if the following inequalities hold for every index set S and for every
pair of sequences (xs )s∈S , (ys )s∈S of elements xs ∈ X, ys ∈ Y :
_ ^ _ _
(xs ∧ ys ) ≥ ( xr ∨ ys ),
s∈S R⊆S r∈R s∈S\R

^ _ ^ ^
(xs ∨ ys ) ≤ ( xr ∧ ys ).
s∈S R⊆S r∈R s∈S\R ♦
204 4 Decompositions of concept lattices

The inequalities can be replaced by equations without changing the statement,

since the respective other directions hold in every lattice.
Proposition 94 If Li and Lj , (i 6= j), are factors of a tensor product
O
Lt ,
t∈T

then the sublattices εi (Li ) and εj (Lj ) are mutually distributive.

Proof We only prove the first inequality
_ ^ _ _
(εi (xs ) ∧ εj (ys )) ≥ ( εi (xr ) ∨ εj (ys )).
s∈S R⊆S r∈R s∈S\R

For this purpose, it suffices to prove that every attribute concept z which is ≥
the left side is also ≥ the right side of the inequality. Hence, let z be an attribute
concept and assume that

Rz := {r ∈ S | εi (xr ) ≤ z}.

Then we obviously have r∈Rz εi (xr ) ≤ z and can follow the following chain of
W
inferences:
_
z≥ (εi (xs ) ∧ εj (ys ))
s∈S
⇔ ∀s∈S z ≥ εi (xs ) ∧ εj (ys )
⇔ ∀s∈S z ≥ εi (xs ) or z ≥ εj (ys )
⇔ ∀s∈S\Rz z ≥ εj (ys )
_ _
⇔ z≥ εi (xr ) ∨ εj (ys )
r∈Rz s∈S\Rz
^ _ _
=⇒ z≥ ( εi (xr ) ∨ εj (ys ))
R⊆S r∈R s∈S\R

In the case of the second equivalence we have used Proposition 93. 

The formulation “mutually distributive” suggests the following result:
Theorem 33 The tensor product of completely distributive lattices is completely dis-
tributive.
Proof We leap ahead to the characterization of complete distributivity through a
context condition in Theorem 47 and show that this condition can be transferred
from the factors to a direct product of contexts. In order to improve the readabil-
ity we replace the expression h ∈ k 00 by the (equivalent) statement k 0 ⊆ h0 .
Assume that Kt := (Gt , Mt , It ), t ∈ T are contexts and that (G, M, I) :=
× t∈T Kt . Assume further that g ∈ G and m ∈ M are elements with (g, m) 6∈ I.
Then, we have
4.4 Tensorial decompositions 205

(gt , mt ) 6∈ It
for all t ∈ T . Moreover, if the Kt satisfy the condition from Theorem 47, there
exist elements ht ∈ Gt as well as nt ∈ Mt with

(ht , mt ) 6∈ It , (gt , nt ) 6∈ It and kt ⊆ ht for all k ∈ Gt \ nt0

for every t ∈ T . We set h := (ht )t∈T and n := (nt )t∈T and find (h, m) 6∈ I and
(g, n) 6∈ I. If now k ∈ G \ n0 , i.e., kt ∈ Gt \ n0t for all t ∈ T , then for every t ∈ T it
holds that
kt0 ⊆ ht0
and, according to Proposition 91, consequently

k 0 ⊆ h0 ,

which was to be proved. 

We had generalized the direct product of lattices to the subdirect product in
order to obtain a more versatile decomposition principle. We can proceed simi-
larly in the case of the tensor product. Two possibilities suggest themselves: On
the one hand, we can form a subtensorial product of complete lattices in analogy
to the subdirect product. On the other hand, we can introduce a subdirect product
of contexts. If we do this correctly, both constructions are equivalent.
Definition 77 A subtensorial product of complete lattices Lt , t ∈ T is a factor
lattice
O 
Lt Θ
t∈T

of the tensor product for which the restrictions of the projection mapping
O O 
πΘ : Lt → Lt Θ, x 7→ [x]Θ
t∈T t∈T

onto the sublattices εt (Lt ) are all injective.

A subtensorial decomposition of a complete lattice L then is a sequence (Lt |
t ∈ T ) of complete sublattices  for which there is an isomorphism ψ of L onto
of L
a subtensorial product t∈T Lt Θ with
N

πΘ (εt (Lt )) = ψ(Lt ), t ∈ T. ♦

Subtensorial decompositions can be characterized internally. For reasons of

simplicity, we limit ourselves to the case T = {1, 2}.
Proposition 95 A pair (L1 , L2 ) of complete sublattices is a subtensorial decomposition
of L if and only if L1 and L2 are mutually distributive and their union generates L.
Proof If L1 and L2 are mutually distributive sublattices of L, then by Theo-
rem 42, ⊗3, there is a complete homomorphism
206 4 Decompositions of concept lattices

ϕ : L1 ⊗ L2 → L,

for which it holds that ϕ ◦ εt = idLt for t = 1 and t = 2. If L1 ∪ L2 generates L,

then this morphism must be surjective, i.e., L then is a factor lattice of L1 ⊗ L2 .
If, conversely, (L1 , L2 ) is a subtensorial decomposition, then L1 and L2 are
mutually distributive, since this property is inherited by factor lattices. Their
union generates L, because L1 ⊗ L2 is generated by ε(L1 ) ∪ ε(L2 ). 
In the case of doubly founded concept lattices, the
Nnotation for subtensorial
products can be further simplified. By Theorem 31, t∈T B(Kt ) is isomorphic
to the concept lattice of the direct product

(G, M, ∇) := ×K
t∈T
t

of the respective contexts. The closed relation ∇t always corresponds to the sub-
lattice εt (B(Kt )). In the doubly founded case, we can be sure that a subtensorial
product is always induced by a compatible subcontext (H, N, ∇ ∩ H × N ) of
(G, M, ∇). Such subcontexts are described by the following definition:
Definition 78 A subdirect product of contexts

Kt := (Gt , Mt , It ), t ∈ T,

is a compatible subcontext

(H, N, ∇ ∩ H × N )

of the direct product ×t∈T Kt having the property that for each t ∈ T the sub-
context
(Ht , Nt , It ∩ Ht × Nt )
with
Ht := {ht | h ∈ H} and Nt := {nt | n ∈ N }
is dense in Kt . ♦
Proposition 96 The subdirect products of contexts are precisely the compatible subcon-
texts of subtensorial products.
Proof According to Proposition 44, the condition (here restricted to εt (B(Kt )))
that the map ΠH,N is injective, is equivalent to the fact that the subcontext
(H, N, ∇t ∩ H × N ) is dense in (G, M, ∇t ). This in turn is, according to the same
proposition, equivalent to the fact that

(A ∩ H)∇t ∇t = A and (B ∩ N )∇t ∇t = B

holds for each concept (A, B) of (G, M, ∇t ). If we set At := {gt | g ∈ A}, we rec-
ognize by means of the description of the concepts of (G, M, ∇t ) in Proposition
90 that
4.4 Tensorial decompositions 207

{gt | g ∈ A ∩ H} = At ∩ Ht
and therefore

(A ∩ H)∇t ∇t = {g ∈ G | gt ∈ (At ∩ Ht )It It }.

Consequently,
(A ∩ H)∇t ∇t = A ⇔ (At ∩ Ht )It It = At .
This is again the condition from Proposition 44. Hence, (H, N, ∇t ∩ N × N ) is
dense in (G, M, ∇t ), if and only if (Ht , Nt , It ∩ Ht × Nt ) is dense in Kt . 
The restriction of the closed relations ∇t to such a subcontext then yields the
subrelations Jt := ∇t ∩ H × N with

B(H, N, Jt ) ∼
= B(Kt ).

We now want to find out under which conditions a context is isomorphic to a

subdirect product. For this purpose we define
Definition 79 A subdirect decomposition of a context

K := (G, M, I)

is a family (It )t∈T of subrelations of I with the following properties:

1. I = t∈T It
S
2. There are surjective maps

α : G → H, β:M →N

onto a subdirect product (H, N, ∇ ∩ H × N ) with

g It m ⇔ αg ∇t βm. ♦

If (H, N, ∇∩H ×N ) is a subdirect product, then (∇t ∩H ×N )t∈T is a subdirect

decomposition, as can be easily recognized by choosing the identical map for α
and β, respectively.
The maps α and β do not have to be injective. Nevertheless, from αg = αh it
always follows that g It = hIt for all t ∈ T and in particular g 0 = h0 . Hence, (α, β)
is “up to clarification” an isomorphism of (G, M, I) onto (H, N, ∇ ∩ H × N ) and
even of (G, M, It ) onto (H, N, ∇t ∩H ×N ) for all t ∈ T . In particular, the contexts
Kt := (G, M, It ) and (H, N, ∇t ∩ H × N ) have isomorphic concept lattices. It
suggests itself to use these contexts as the factors of the direct product (note
that such factors are not further specified in the definition). We show that this is
possible, but before that, we clarify the contexts. For this purpose we define for
each of the contexts
Kt := (G, M, It )
equivalence relations Θt on G and Ψt on M through
208 4 Decompositions of concept lattices

(g, h) ∈ Θt : ⇔ g It = hIt
(m, n) ∈ Ψt : ⇔ mIt = nIt .

Hence, Θt = ker γt and Ψt = ker µt . The context

Kt◦ := (G/Θt , M/Ψt , I¯t )

with
([g]Θt , [m]Ψt ) ∈ I¯t : ⇔ (g, m) ∈ It
then is the corresponding clarified context.
Then, we can naturally assign a subcontext of the direct product

×K
t∈T
◦
t =( × G/Θ , × M/Ψ , ∇)
t∈T
t
t∈T
t

of the clarified contexts K◦t to the context K := (G, M, I).

The symbol ∇ is only used for a better distinction. It denotes the incidence of
the direct product, i.e.,

(gt )t∈T ∇s (mt )t∈T ⇔ gs Is ms .

The role of the maps α and β from Definition 79 is taken over by the maps ι and
ι, which are defined as follows:

ι:G→ × G/Θ ,
t∈T
t g 7→ ([g]Θt )t∈T

ι:M → × M/Ψ ,
t∈T
t m 7→ ([m]Ψt )t∈T .

The image context

(ιG, ιM, ∇ ∩ ιG × ιM )
obviously has the property that the projection maps

([g]Θt )t∈T 7→ [g]Θs , ([m]Ψt )t∈T 7→ [m]Ψs

are surjective on the factor contexts whereby their images are certainly dense.
Furthermore, we have

g It m (in K) ⇔ αg ∇t βm (in (ιG, ιM, ∇ ∩ ιG × ιM )).

Hence, if it is compatible, this subcontext is certainly a subdirect product.

Proposition 97 (It )t∈T is a subdirect decomposition, if and only if the subcontext

(ιG, ιM, ∇ ∩ ιG × ιM )

of ×t∈T K◦t is compatible.

4.4 Tensorial decompositions 209

Proof If this subcontext is compatible, it evidently satisfies all conditions of

a subdirect decomposition. The other direction is more laborious. Hence, let
(H, N, ∇ ∩ H × N ) be a compatible subcontext of an arbitrary direct product
× t∈T (G̃t , M̃t , Jt ) and let α : G → H, β : M → N be mappings satisfying
g It m ⇔ αg ∇t βm as in Definition 79. We have to show that under these con-
ditions the subcontext mentioned in the proposition is also compatible. For this
purpose, we use the characterization from Proposition 41. Hence, let h ∈ ιG be
an object of the subcontext, i.e., h = ([g]Θt )t∈T for an object g ∈ G. Furthermore,
let m := ([mt ]Ψt )t∈T be an arbitrary attribute of the direct product of the K◦t with
(h, m) 6∈ ∇. Then, we have to show that there is an attribute n ∈ ιM with

(h, n) 6∈ ∇ and m∇ ⊆ n∇ .

From the preconditions we obtain

([g]Θt , [mt ]Ψt ) 6∈ I¯t

for all t ∈ T , i.e.,

(g, mt ) 6∈ It for all t ∈ T.
Then, we also have
(αg, βmt ) 6∈ ∇t for all t ∈ T.
If
αg =: (gt )t∈T
and
βms =: (mst )t∈T for every s ∈ T,
then we have
(gt , mtt ) 6∈ Jt for all t ∈ T.
If we set
l := (mtt )t∈T ,
then l is an attribute of the direct product with (αg, l) 6∈ ∇. Since (H, N, ∇ ∩ H ×
N ) is compatible, there must be an attribute βn ∈ N with

(αg, βn) 6∈ ∇ and l∇ ⊆ (βn)∇ .

If we set βn =: (nt )t∈T , then with Proposition 91

(mtt )∇t ⊆ (nt )∇t for all t ∈ T,

i.e.,
(βmt )∇t ⊆ (βn)∇t for all t ∈ T,
which yields
mIt t ⊆ nIt t for all t ∈ T.
210 4 Decompositions of concept lattices

Hence, we obtain
¯ ¯
([mt ]Ψt )It ⊆ ([nt ]Ψt )It for all t ∈ T

and with n := ιn
m∇ ⊆ n∇ .
Because of (αg, βn) 6∈ ∇, we have (g, n) 6∈ I and therefore (g, n) 6∈ It for all t ∈ T ,
which yields the statement
(h, n) 6∈ ∇,
which we were still lacking. 
Proposition 97 contains a structural description of subdirect products. It is
particularly easy to make use of this fact when we are dealing with a doubly
founded context. We shall explain this in the following theorem. The notations
used for this purpose are to be understood as follows: A family (It )t∈T of subre-
lations of (G, M, I) is called doubly founded if each of the contexts (G, M, It ) is
doubly founded. The arrow relations .t and %t also refer to these contexts.
Theorem 34 A doubly founded family (It )t∈T of subrelations is a subdirect decomposi-
tion of (G, M, I) if and only if:
1. I = t∈T It ,
S

2. if gt .t m for all t ∈ T , then there exists an object h ∈ G with gtIt = hIt for all
t ∈ T,
3. if g %t mt for all t ∈ T , then there exists an attribute n ∈ M with mIt t = nIt for all
t ∈ T.
Proof From Proposition 97 we infer that (It )t∈T is a subdirect decomposition if
and only if the subcontext

(ιG, ιM, ∇ ∩ ιG × ιM )

is closed. Because of the condition of foundedness, this is equivalent to its be-

ing arrow closed. This is, however, precisely what is postulated in conditions 2)
and 3) of the above theorem. Condition 2), for instance, is a rephrasing of the
condition
([gt ]Θt )t∈T . ιm implies ([gt ]Θt ) ∈ ιG,
since ([gt ]Θt ) forms part of ιG, if there is an object h ∈ G with [gt ]Θt = [h]Θt for
all t ∈ T . 
Together with Proposition 95 this can be extended into a practicable condi-
tion. For this purpose, we call a pair (x, y) of a doubly founded lattice L weakly
distributive if
g ≤ x ∨ y ⇔ g ≤ x or g ≤ y
holds for every -irreducible element g ∈ J(L) and dually
W
4.4 Tensorial decompositions 211

m ≥ x ∧ y ⇔ m ≥ x or m ≥ y

holds for every -irreducible element m ∈ M (L). Hence, two concepts (A1 , B1 )
V
and (A2 , B2 ) certainly form a weakly distributive pair if A1 ∪ A2 is an extent
and B1 ∪ B2 is an intent. We have seen above that this is always the case for
pairs (x, y) of elements of a tensor product with x ∈ εi (Li ), y ∈ εj (Lj ) and
i 6= j. This implied that those sublattices were mutually distributive. In fact, the
following statement can be shown by means of the same proof as as was given
for Proposition 94:
Proposition 98 If all pairs (xt , yt ), t ∈ T are weakly distributive, then
_ ^ _ _
(xt ∧ yt ) = ( xs ∨ yt )
t∈T S⊆T s∈S t∈T \S

holds as well as the dual equation. 

t
f
c @
@
d
@ @
b
@
c @ @ @
@ @ @ @
a
@ @ t. .. .. .. . . . . . . . . . . . .@
.. d @t
c @ cc @ cd .
1 @ 4 @ ..........
.... @
@
@ ...
@ t. . . . . . . . . . . . .. .. .. ..@
@
.. . d
f
@ @
@c @c
2 @ 3 @
@ @
@c @ tf

Figure 4.23 The dotted lines in the diagram on the right link the weakly distributive pairs
of incomparable elements.

Theorem 35 Let L1 and L2 be complete sublattices of L and let L, L1 and L2 be doubly

founded. Then (L1 , L2 ) is a subtensorial decomposition of L if and only if the union
L1 ∪ L2 generates the lattice L and every pair (x1 , x2 ) with x1 ∈ L1 and x2 ∈ L2 is
weakly distributive.

Proof From the Propositions 95 and 98 it immediately follows that the condi-
tions specified are sufficient for a subtensorial decomposition. It remains to be
212 4 Decompositions of concept lattices

shown that weak distributivity is necessary as well. For this purpose we use The-
orem 34 for the standard context

K(L) := (J(L), M (L), ≤)

of L. Consider an element x1 ∈ L1 and an element x2 ∈ L2 as well as a -

W
irreducible element g ∈ J(L) with g 6≤ x1 and g 6≤ x2 . Then there are elements
m1 , m2 ∈ M (L) with x1 ≤ m1 , x2 ≤ m2 and g %1 m1 , g %2 m2 , where %t is
the arrow relation with respect to the closed subrelation It , which belongs to Lt .
According to Theorem 34, there is an attribute m ∈ M (L) with mI1 = mI11 and
mI2 = mI22 . Condition 1 of Theorem 34 forces mI = mI1 ∪ mI2 , from which it
follows that g 6≤ m. Since m = m1 ∨ m2 ≥ x1 ∨ x2 , this implies g 6≤ x1 ∨ x2 , which
is one of the conditions of weak distributivity. The other follows dually. 

Corollary 4 Two doubly founded complete sublattices L1 and L2 whose union generates
a sublattice that is also doubly founded are mutually distributive if and only if every pair
(x1 , x2 ) with x1 ∈ L1 and x2 ∈ L2 is weakly distributive. 

I a b c d I1 a b c d I2 a b c d
1 ×× 1 × × .
% % 1 .
% .
% %
.
2 ××× 2 × × × .
% 2 .
% .
% .
%
3 ××× 3 .
% .
% . × 3 .
% × × ×
4 × 4 .
% .
% . × 4 %
. .
% .
%

I1 a, b c d
I2 a b, c, d
1 ×
1, 2, 4
2 × ×
3 ×
3, 4 ×

Figure 4.24 Context for the lattice from Figure 4.23, together with the closed relations for
the sublattices. Below, the clarified contexts.

Example We examine the lattice L represented in Figure 4.23 to see if it has a

subtensorial decomposition. Apart from pairs of comparable elements (which
are automatically weakly distributive), the lattice only contains three weakly dis-
tributive pairs. They are represented in the diagram on the right side by the dot-
ted lines. We recognize that there is only one non-trivial decomposition into two
sublattices that satisfy the conditions in Theorem 35. Those two sublattices are
marked in the right diagram by the filled or doubled circles, respectively.
Figure 4.24 represents the context K(L) and the closed relations belonging
to the sublattices, including their arrow relations. For those small contexts, it is
easy to verify that the conditions in Theorem 34 are indeed satisfied. Therefore,
4.4 Tensorial decompositions 213

a, b c d
∇
a
b, c, d a b, c, d a b, c, d
1, 2, 4 × × .
% % ←
1
3 × × %
. × % ×
1, 2, 4 × × × × .
% ←
2
3 × × × × .
% ×
1, 2, 4 .
% . × × ←
3,4
3 .
% × . × × × ←
↑ ↑ ↑ ↑

Figure 4.25 The direct product of the two clarified contexts from Figure 4.24. The context
from Proposition 97 is marked by the arrows on the margin.

f
@
({a, b}, {b, c, d})
f @
@
@ ({c},{b, c, d})
({a, b}, {a}) @ ({d}, {b, c, d})
f @f @f
({1}, {1, 2, 4}) @ @ @
@f @f @ f({d}, {a})
@ @ ({3, 4}, {1, 2, 4})
({c}, {a})
f @ f @ f
({2}, {1, 2, 4}) @ ({1}, {3}) ({3, 4}, {3})
@f
({2}, {3}) @
@f

Figure 4.26 The concept lattice for the context from Figure 4.25 is the tensor product of
the sublattices from Figure 4.23.

we can switch over to the clarified contexts to find a concrete representation as

a subtensorial product. Proposition 97 explains how we have to proceed. The
subcontext
(ιG, ιM, ∇ ∩ ιG × ιM )
of× t∈T Kt , which is mentioned there, is marked in Figure 4.25 through arrows
◦

on the margin. From the arrow relations we can see that it is arrow closed and
thus compatible.
Finally, Figure 4.26 shows the concept lattice of the context from Figure 4.25,
i.e., the tensor product of the two sublattices forming the subtensorial decom-
position. The lattice we started with can be recognized as an interval below the
largest element; its elements are indeed separated by the projection map. >
214 4 Decompositions of concept lattices

4.5 Notes, references, and trends

4.1

The introductory lines of this section, approximately up to Proposition 79 are

analogous to known results of General Algebra. The complete lattices, however,
do not form part of the structure classes treated in General Algebra and therefore
require separate proofs. Some of them can be found in Pierce [300]. The concept-
analytic results are based mainly on [390].

4.2

This section follows [394]. The decomposing and gluing technique described in
this section was developed and successfully employed by Herrmann [190], an
updated version can be found in Day and Herrmann [87]. Vogt [376] has em-
ployed the technique of atlas-decomposition when investigating the structure of
subgroup lattices of finite Abelian groups.

4.3

The substitution sum and the substitution product were used by Luksch and
Wille [266] for the concept-analytic evaluation of pair comparison tests. They
were formally introduced in [267] and thoroughly examined by Stephan [351],
[350]. From the dissertation of Stephan we have taken in particular Theorem 30
and the preparatory Proposition 89. However, we have introduced a little change
in the notation compared with the literature we quote and now write U (a, b)V ,
where in earlier publications appeared U (b, a)V .

4.4

Tensor products of complete lattices have been introduced in many articles. Our
presentation (Sections 4.4, 5.4) does not claim to be complete, but is meant to
supply the basic knowledge. The definition of the tensor product discussed in
this section has been taken from [396] and the generalization in [405]. Precursors
can be found among other things in Waterman [383], Mowat [289] and Shmuely
[342]. A description of the extents and intents of direct products of contexts (as
Gκ -ideals) can be found in [405]. The significance of this product for category
theory was discussed by Erné [116]. Other sources are Bandelt [22], Raney [310],
Kalmbach [208], and, more recently, Chornomaz [76].
Subtensorial products are treated in [160].
Chapter 5
Constructions of concept lattices

A construction method by means of which we obtain from two contexts K1 and

K2 a new context, let us say K, can only be a useful construction principle for
concept lattices, if it is invariant under reduction. This means that, if the same con-
struction is applied to contexts whose concept lattices are isomorphic to those of
K1 and K2 , then the concept lattice of the result should be isomorphic to that of
K.
We have already presented some such methods in the first chapter (and
shown in Figure 1.12 that the semiproduct does not fulfill this invariance con-
dition). We will now describe four (further) constructions in detail.
In the case of the subdirect product we consider sublattices of direct products.
In 4.1 we have already examined how we can recognize the corresponding closed
relations of the context sum. Now we are going to show how such relations can
be constructed as a fusion of contexts.
Although the subdirect products are of central significance for General Al-
gebra, they are rarely regarded as means of construction. One reason is their
ambiguity. A subdirect product is not uniquely determined by stating its factors.
This can however be easily remedied by choosing fixed generating systems in
the factors. Thereby we obtain the P -product of algebraic structures and the P -
fusion of contexts. A possible application of this construction consists in jointly
unfolding different data sets which relate to the same situation.
The atlas-gluings introduced in 4.2 will be supplemented by a method in
which the lattices are glued “sideways”. This can be depicted particularly easily
if the the overlap area of the lattices involved is the union of an ideal and a filter.
The third section deals with the technique of doubling convex subsets of a con-
cept lattice. This construction has been used successfully in mathematical lattice
theory, among other things for the examination of free lattices.
Finally we shall return to the tensor product of complete lattices. We shall give
a lattice-theoretic characterization of this product and introduce the tensorial op-
erations, by means of which we can trace back calculation within a tensor product
to calculation within its factors.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 215
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_6
216 5 Constructions of concept lattices

5.1 Subdirect product constructions

We introduced the (direct) sum of contexts in Definition 30, however, we formu-

lated it only for the case of two contexts. More generally, for a family of contexts
Kt := (Gt , Mt , It ), t ∈ T we define the sum by
!
X [˙ ˙
[ ˙
[ [
(Gt , Mt , It ) := Gt , Mt , It ∪ G s × Mt ,
t∈T s6=t

presupposing that the sets Gt , t ∈ T as well as the sets Mt , t ∈ T are pairwise

disjoint. If necessary, this can be enforced by previously replacing every context
Kt := (Gt , Mt , It ) by the isomorphic context K̇t := (Ġt , Ṁt , I˙t ) with Ġt := {t} ×
Gt and Ṁt := {t} × Mt , as in Definition 29 on p. 41. Then we obtain
Theorem 36 The concept lattice of a sum of contexts is isomorphic to the product of its
concept lattices:
Kt ) ∼ ×
X
B( = B(Kt ).
t∈T t∈T

The map
(A, B) 7→ ((A ∩ Gt , B ∩ Mt ) | t ∈ T )
is a natural isomorphism.
The projection map on B(Kt ) combined with this isomorphism is the map

(A, B) 7→ (A ∩ Gt , B ∩ Mt ).

The corresponding compatible subcontext is Kt =: (Gt , Mt , It ).

Proof We only have to show that the concepts of the sum context are precisely
the pairs (A, B) with A ⊆ ˙ Gt , B ⊆ ˙ Mt which have the property that for
S S
every t ∈ T the restriction (A ∩ Gt , B ∩ Mt ) is a concept of Kt . This is easy: By
means of the definition, we realize that for a set At ⊆ Gt the derivation in the
sum context can be determined as follows:
[
A0t = AIt ∪ Mt .
s6=t

For an arbitrary subset A ⊆ G we therefore get (with At := A ∩ Gt )

[ \ [
A0 = ( At ) 0 = A0t = AIt t .
t∈T t∈T t∈T

Dually, the extents of the sum context are precisely the unions of extents of the
summands, which yields the isomorphy we claimed. The statement on the com-
patible subcontexts can be verified by means of Proposition 40. 
5.1 Subdirect product constructions 217

Corollary 5 Let J be a closed relation of a context sum B( t∈T Kt ), and let s ∈ T .

P

Then the projection map from the complete sublattice B ˙ Gt , ˙ Mt , J to B(Ks ) is

S S

surjective if and only if J ∩ (Gs × Ms ) = Is .

Proof This is just a combination of Theorem 36 and Proposition 58 on p. 143,

which says that (Gs , Ms , J ∩ Is ) is a compatible subcontext. According to the
theorem it represents the projection map, which is surjective if and only if every
formal concept of Ks is a formal concept of (Gs , Ms , J ∩ Is ). 

In this section we want to characterize complete subdirect products, i.e., com-

plete sublattices of a direct product for which the projection maps are surjective,
in the language of contexts. We have already taken first steps in this direction,
since to every complete sublattice corresponds a closed relation J in the sum
context and, by Corollary 5, the surjectivity of the projection maps is equiva-
lent to the condition J ∩ (Gs × Ms ) = Is . This is summarized in the following
proposition:
Proposition 99 The subdirect products of concept lattices B(GP t , Mt , It ) are in one-to-
one correspondence to the closed relations J of the sum context t∈T (Gt , Mt , It ) with
J ∩ (Gt × Mt ) = It for all t ∈ T . 

Proposition 100 Assume that Kr := (Gr , Mr , Ir ), Ks := (Gs , Ms , Is ) and Kt :=

(Gt , Mt , It ) are formal contexts and let Jrs and Jst be bonds from Kr to Ks and from
Ks to Kt , respectively. Moreover, let Jrt ⊆ Gr × Mt . Then the following are equivalent:
1. Jrt ⊆ Jrs ◦ Jst ,
2. g Jrt ⊆ g Jrs Is Jst for all g ∈ Gr , and
3. mJrt ⊆ mJst Is Jrs for all m ∈ Mt .

Proof On Page 151, while proving Theorem 18, it was shown that g Jrs ◦Jst =
g Jrs Is Jst holds for all g ∈ Gr . This proves that 1. is equivalent to 2. The other
equivalence is dual. 

Proposition 101 If J ⊆ I is a subrelation in the sum context (G, M, I) := t∈T Kt

P
with the property that
It = Jtt := J ∩ (Gt × Mt )
holds for all t ∈ T , then the following statements are equivalent
1. J is a closed relation.
2. The Jst := J ∩ (Gs × Mt ) are bonds and Jrt ⊆ Jrs ◦ Jst holds for all r, s, t ∈ T .

Proof If J is closed, then, for every object g ∈ G, (g JJ , g J ) is a concept of Kt .

P
We learn from Theorem 36 that the subcontexts Ks are all compatible, so that
(g JJ ∩ Gs , g J ∩ Ms ) always is a concept of Ks .
In particular is g J ∩ Ms an intent of Ks , and if g ∈ Gr , then g J ∩ Ms = g Jrs .
This shows, together with the dual argument for attributes, that all Jrs must be
bonds.
218 5 Constructions of concept lattices

Moreover, since (g JJ ∩ Gs , g Jrs ) is a concept of Ks , the concept extent must be

the derivation of the intent, i.e.,

g JJ ∩ Gs = g Jrs Jss .

From g JJ ∩ Gs ⊆ g JJ we get (g JJ ∩ Gs )J ⊇ g JJJ = g J , which, when restricted

to Kst , gives
(g JJ ∩ Gs )J ∩ Mt ⊇ g J ∩ Mt ,
which abbreviates to
(g JJ ∩ Gs )Jst ⊇ g Jrt .
Combined with g JJ ∩ Gs = g Jrs Jss this yields

g Jrs Jss Jst ⊇ g Jrt .

Proposition 100 shows that this proves (2).

If, on the other hand, we presuppose (2), then we can use Proposition 57 on
p. 142 to prove that J is a closed relation. Hence, let (g, m) ∈ I \ J, which because
of It = Jtt implies that g ∈ Gr and m ∈ Ms hold for suitable r 6= s. Since Jrs is
a bond from Kr to Ks , g Jrs is an intent of Ks . Therefore, there must be an object
h ∈ g Jrs Is with (h, m) 6∈ I. According to Proposition 100 we have

g Jrt ⊆ g Jrs Is Jst ⊆ hJst

for every t ∈ T , and consequently g J ⊆ hJ . This, together with the dual argu-
ment, shows that the condition (C) of Proposition 57 is satisfied: J is closed. 
The Propositions 99 and 101 can be summarized as follows:
Theorem 37 For a subrelation J ⊆ I in the sum context (G, M, I) := Kt , the
P
t∈T
following statements are equivalent:
1. J is a closed relation and corresponds to a subdirect product of the B(Kt ), t ∈ T .
2. J is a closed relation and It = J ∩ (Gt × Mt ) holds for all t ∈ T .
3. The Jst := J ∩ (Gs × Mt ) are bonds from Ks to Kt with Jtt = It and Jrt ⊆ Jrs ◦Jst
for all r, s, t ∈ T . 
Note that the last point of the theorem does not contain the requirement that
J is closed. The latter follows (as in Proposition 101) as a consequence if J is
made up of bonds, as specified above. If the contexts Kt are all reduced, then
(G, M, J) is reduced as well. This follows from Proposition 58.
In order to be able to use the subdirect product as a construction method, we
introduce the following notion:
Definition 80 If P is a set, L is a complete lattice and α : P → L is a map, then
we call (L, α) a (complete) P -lattice if L is generated by {αp | p ∈ P }.
When P := {1, 2, . . . , n} we also speak of a (complete) n-lattice. If (P, ≤) is
an ordered set, then we call (L, α) a (complete) (P, ≤)-lattice if α is furthermore
order-preserving.
5.1 Subdirect product constructions 219

To a family (Lt , αt ), t ∈ T , of complete P -lattices we can naturally assign a

complete sublattice of the direct product ×
t∈T Lt , namely the sublattice which
is generated by the elements

αp := (αt p | t ∈ T ), p ∈ P.

We call this lattice the P -product of the lattices (Lt , αt ). As a symbol for P -
P n
products of two lattices we use × or ×, respectively. ♦

Example In Figure 5.1 a 4-product of three small chains is represented as a sub-

lattice of the direct product. The elements of the set P = {1, 2, 3, 4} are written
below those lattice elements on which they are mapped by α. In the diagram on
the right, only the elements represented by small circles belong to the 4-product;
the additional lines have been drawn to indicate the situation of the lattice within
the direct product.

d
@
d @d
1 d @ d @ @ d3
@ @ @
d1 d3
@ @ @ @ @
d1,3 d4 d2 @d @d @d @ @
4 4
× × ∼
=
@ @ @ @ @ @
d d3 d1 @ @d @d @d @ @
2,4 @ @ @ @ @
d2 d4 @d @ @d @ @d
4 @ @ @ 2
@d @ @d
@
@d

Figure 5.1 A simple example of a 4-product.

If the lattices involved are concept lattices, we use the following obvious
terms: (K, α) is called P -context, if (B(K), α) is a P -lattice. In this case, α maps
the elements of P onto concepts of K; for those images we most often write
(Ap , B p ) := αp. Then we call (B(K), α) “the concept lattice of the P -context
(K, α)”, for short.
Evidently, the P -product is a complete subdirect product, since the canonical
projection πt maps the generating system {αp | p ∈ P } of the P -product onto
{αt p | p ∈ P }, i.e., onto a generating system of Lt . Therefore, πt must be sur-
jective. Hence, according to Theorem 37, to the P -product of P -concept lattices
there corresponds in a natural way a closed relation J in the sum context. This is
220 5 Constructions of concept lattices

described more precisely in the following theorem, in which we shall again use
the abbreviation Jst := J ∩ (Gs × Mt ):
Theorem 38 The closed relation J of the sum context t∈T Kt which belongs to a P -
P
product of P -concept lattices (B(Kt ), αt ) is characterized by the following properties:
1. for all t ∈ T , Jtt = It ,
2. for all s, t ∈ T, s 6= t, Jst is the smallest bond from Ks to Kt which contains the sets

Aps × Btp , p ∈ P,

Aps and Btp being defined to be αs p =: (Aps , Bsp ) and αt p =: (Apt , Btp ).
Proof From the definition of a bond (Definition 61 on p. 148) it immediately fol-
lows that the intersection of bonds is a bond. Therefore, there is always a smallest
bond from Ks to Kt , s 6= t which entirely contains the sets

Aps × Btp , p ∈ P.

If we denote this bond by Jst and set Jtt := It , then obviously

[
J := Jst
s,t∈T

is precisely the relation which is characterized by the two conditions of the the-
orem. We shall show first that J satisfies condition 3) of Theorem 37.
For this purpose we consider, for fixed r, s, t ∈ T , p ∈ P , an attribute m ∈ Btp .
Because of Aps × Btp ⊆ Jst we certainly have mJst ⊇ Aps . Likewise, because of
Apr × Bsp ⊆ Jrs , we can infer for each object g ∈ Apr that g Jrs ⊇ Bsp and, since
(Aps , Bsp ) is a concept of Ks , even g Jrs Is ⊆ Aps . Hence, we have g Jrs Is ⊆ Aps ⊆ mJst
and obtain, together with Definition 62 on p. 150,

(g, m) ∈ Apr × Btp ⇒ g Jrs Is ⊆ mJst

⇒ (g, m) ∈ Jrs ◦ Jst .

Since this is correct for all g ∈ Apr and all m ∈ Btp , we have

Aps × Btp ⊆ Jrs ◦ Jst .

Theorem 18 on p. 151 furthermore states that Jrs ◦ Jst is a bond. Since we have
assumed that Jrt is the smallest bond containing all those sets, it follows that

Jrt ⊆ Jrs ◦ Jst .

Thus, the third condition of Theorem 37 is satisfied and consequently J is the

closed relation of a subdirect product of the B(Kt ).
It remains to be shown that J is the right closed relation, i.e., that it really
corresponds to the P -product specified. This is generated by the elements {αp |
p ∈ P }; the corresponding concepts in the sum context are
5.1 Subdirect product constructions 221
[ [
(Ap , B p ), Ap := Apt , B p := Btp , t ∈ T.

J contains all sets Ap × B p , and every closed relation containing those sets be-
longs to a subdirect product and therefore has to satisfy the third condition of
Theorem 37. J is the smallest closed relation for which this is true, and is there-
fore, according to Proposition 55, the closed relation belonging to the sublattice
generated by (Ap , B p ). 

The context construction described in Theorem 38, which corresponds to the P -

product, is here given a name:
Definition 81 The P -fusion of a family (Kt , αt ), t ∈ T of P -contexts is the P -
context
((G, M, J), α),
in the case that J ⊆ I is the subrelation in the sum context (G, M, I) := t∈T Kt
P
which is characterized by the conditions of Theorem 38 and in the case that α is
the map defined as follows: If for any t ∈ T αt p =: (Atp , Btp ), then

˙
[ ˙
[
αp := ( Atp , Btp ).
t∈T t∈T

In the case of two P -context we use

P
(K1 , α1 ) + (K2 , α2 )

as the symbol for the P -fusion. ♦

Then, of course we have:
Corollary 6 The concept lattice of a P -fusion of contexts is isomorphic to the P -product
of its concept lattices. 
Example We calculate the 4-product of the two 4-lattices
b
b
b 2 b @ b4
@b
1,2 b @ b3,4 and .
b
@b
b 1 b @ b3
@b

In this case, the method used in the preceding example would lead to compli-
cated intermediate steps. Therefore, we determine the corresponding 4-standard
contexts and obtain:
222 5 Constructions of concept lattices

bg
4
b
m1 m1 m2 m3 m4
m2 b @ bm3 g1 × × × α1 (1) = α1 (2) = ({g1 , g2 }, {m1 , m2 })
g2 g3 g2 × ×
@b α1 (3) = α1 (4) = ({g1 , g3 }, {m1 , m3 })
g1 g3 × ×
m4 b g4
b
n1 n2 n3 n4 n5 α2 (1) = ({h1 }, {n1 , n2 , n3 , n5 })
n2 b @ b n1
h4 h5 h1 × × × × α2 (2) = ({h1 , h2 , h3 , h4 }, {n2 })
@b h2 × × × ×
h3
n3 b h3 × × α2 (3) = ({h2 }, {n1 , n2 , n3 , n4 })
n5 b @ b n4 h4 ×
h1 h2 α2 (4) = ({h1 , h2 , h3 , h5 }, {n1 })
h5 ×
@b

Now we form the 4-fusion of these two contexts, as described in Theorem 38,
i.e., we form the disjoint union of the two contexts and add the sets {g1 , g2 } ×
{n1 , n2 , n3 , n5 }, {g1 , g2 } × {n2 }, {g1 , g3 } × {n1 , n2 , n3 , n4 }, {g1 , g3 } × {n1 } for J1,2
as well as the sets {h1 } × {m1 , m2 }, {h1 , h2 , h3 , h4 } × {m1 , m2 }, {h2 } × {m1 , m3 },
{h1 , h2 , h3 , h5 } × {m1 , m3 } for J2,1 to the incidence. In the present case this has
already resulted in bonds. In general, the incidence must be extended until bonds
are obtained. As a result we obtain the following 4-context:
m 1 m 2 m 3 m 4 n1 n2 n3 n4 n5
g1 × × × × × × × ×
g2 × × × × × × α(1) = ({g1 , g2 , h1 }, {m1 , m2 , n1 , n2 , n3 , n5 })
g3 × × × × × × α(2) = ({g1 , g2 , h1 , h2 , h3 , h4 }, {m1 , m2 , n2 })
g4
h1 × × × × × × × α(3) = ({g1 , g3 , h2 }, {m1 , m3 , n1 , n2 , n3 , n4 })
h2 × × × × × × ×
α(4) = ({g1 , g3 , h1 , h2 , h3 , h5 }, {m1 , m3 , n1 })
h3 × × × × ×
h4 × × ×
h5 × × ×

The corresponding concept lattice is isomorphic to the 4-product we have been

looking for. It is presented in Figure 5.2.
For the special case that the contexts Kt have the same objects and attributes,
there is a natural choice for the set P . In the following definition we presuppose
merely for reasons of convenience that G and M are disjoint.
Definition 82 Contexts Kt := (G, M, It ) with a fixed object set G and a fixed
attribute set M can be interpreted as P -contexts (Kt , αt ) with P := G ∪˙ M and

αt g := (g tt , g t ) ∈ B(Kt ) for g ∈ G,
t tt
αt m := (m , m ) ∈ B(Kt ) for m ∈ M.

If ((G, M, I), α) is the P -fusion of such a family of P -contexts, then (G, M, I) is

called the fusion of the contexts Kt . ♦
5.2 Gluings 223

d
d
@
d d d @d
@ @ @
d 2 d
@ d4 2 d
@d @ d4
@ @ @ @
1,2 d @ d3,4 4 @d ∼ @d d @d
× =
@ @ @
@d d d @d @d
@ @ @
d 1 d
@ d3 1 d
@d @ d3
@ @ @
@d @d @d
@
@d

Figure 5.2 The 4-lattice on the right is the concept lattice of the 4-fusion calculated in
Example 5.1 and is consequently isomorphic to the 4-product of the factors on the left.

5.2 Gluings

The geometric nature of the lattice diagrams suggests a simple construction,

namely that of putting together lattices to form larger lattices by gluing them
together along common substructures. Such a possibility has already been intro-
duced in Section 4.2, but as a decomposition principle.
Such methods do indeed play a role in the construction, but they turn out to be
complicated in the details and are not always easy to manage. The same is true
for the corresponding context operation, the union. Under suitable additional
conditions, however, we obtain a smooth and practicable theory.
Proposition 102 Let (G0 , G00 ) and (M00 , M0 ) be concepts of (G, M, I). Then

(M00 , M0 ) ≤ (X, Y ) or (X, Y ) ≤ (G0 , G00 ),

holds for every (X, Y ) of (G, M, I) if and only if

I ⊆ G × M0 ∪ G0 × M.
M0
Proof “⇒”: If (g, m) ∈ I, then (M00 , M0 ) ≤ (g 00 , g 0 ) (i.e., m ∈ G0
M0 ) or (g 00 , g 0 ) ≤ (G0 , G00 ) (i.e., g ∈ G0 ). “⇐”: If I ⊆ G ×
M0 ∪ G0 × M and (X, Y ) is a concept with X 6⊆ G0 , then ∅
X 0 ⊆ M0 and therefore (M00 , M0 ) ≤ (X, Y ). 
224 5 Constructions of concept lattices

If we are confronted with the situation described in the proposition, the con-
cept lattice is made up in a simple way of two lattices, namely of the ideal
((G0 , G00 )] and the filter [(M00 , M0 )), which overlap in the (possibly empty) in-
terval [(M00 , M0 ), (G0 , G00 )]. We speak of the Hall-Dilworth gluing (cf. Defini-
tion 70 on page 175), in the special case (G0 , G00 ) = (M00 , M0 ) of the vertical sum,
which has already been mentioned on page 42. There we also introduced the
horizontal sum, where two lattices are “glued together sideways” by identify-
ing the two largest and the two smallest elements. In Section 4.3 we showed that
the substitution product generalizes both constructions.
Whereas the Hall-Dilworth gluing is the simplest case of an atlas construction
in the sense of Section 4.2, another way of generalizing the horizontal sum sug-
gests itself, namely the horizontal gluing, where we allow the lattices involved to
overlap in more than the two border elements. The general situation, namely the
situation that a concept lattice is the union of sublattices, is treated in the next
(rather trivial) proposition:
Proposition 103 For relations Jt ⊆ G × M , t ∈ T , the following statements are equiv-
alent:
1. B(G, M, t∈T Jt ) = t∈T B(G, M,
S S
SJt ).
2. The Jt are closed relations of (G, M, t∈T Jt ) and
[
A×B ⊆ Jt ⇒ ∃s∈T A × B ⊆ Js .
t∈T


The proof is simple, but the result is not very rewarding. In order to obtain a con-
dition which is easier to manage, we limit ourselves to two lattices and assume
that the overlapping is the union of an ideal and a filter, i.e., that it has the form
described in Proposition 102.
Definition 83 A complete lattice L is an ideal-filter gluing of two sublattices U1
and U2 if:
1. L = U1 ∪ U2
2. x ≤ y in L implies {x, y} ⊆ U1 or {x, y} ⊆ U2 .
3. U1 ∩ U2 = (a] ∪ [b) for suitable elements a, b ∈ L. ♦
An ideal-filter gluing can be recognized by the context:
Theorem 39 The following conditions are equivalent:
1. B(G, M, I) is an ideal-filter gluing of complete sublattices U1 and U2 . J1 := C(U1 )
and J2 := C(U2 ) are the corresponding closed relations.
2. J1 and J2 are closed relations of (G, M, I) and
a. g I = g J1 or g I = g J2 holds for every object g and, dually, mI = mJ1 or mI =
mJ2 holds for every attribute m,
b. there is an extent G0 and an intent M0 of (G, M, I) with J1 ∩ J2 = (G0 × M ∪
G × M0 ) ∩ I.
5.2 Gluings 225

M0

G0

J1 ∅

∅ J2

Figure 5.3 Context of an ideal-filter-gluing. In the hatched area J1 and J2 coincide with
I ; this is at the same time the closed relation belonging to the overlapping of the lattices.

Proof 1) ⇒ 2a): Every object concept γg belongs to one of the sublattices, but
γg ∈ Ui is equivalent to g I = g Ji .
1) ⇒ 2b): By assumption U1 ∩ U2 = (a] ∪ [b) for suitable concepts a =: (G0 , G00 )
and b =: (M00 , M0 ). It is evident that J1 ∩J2 ⊇ (G0 ×M ∪G×M0 )∩I, what remains
to be shown is the other inclusion. Assume that (g, m) ∈ J1 ∩ J2 and m 6∈ M0 .
Then there are concepts (X1 , Y1 ) ∈ U1 and (X2 , Y2 ) ∈ U2 with (g, m) ∈ Xi × Yi .
The infimum of these concepts is comparable with both of them, i.e., one of the
three concepts (X1 , Y1 ), (X2 , Y2 ) and (X1 , Y1 ) ∧ (X2 , Y2 ) is not in U1 ∩ U2 . The
intent of this concept contains m, i.e., it cannot be a subset of M0 . Hence, the
concept is contained in the ideal ((G0 , G00 )], and we obtain X1 ∩ X2 ⊆ G0 , which
implies g ∈ G0 .
2) ⇒ 1): We have to show that all concepts and all comparabilities between
concepts of (G, M, I) originate from one of the sublattices Ui . Let us assume
that there is a concept (X, Y ) ∈ B(G, M, I) that belongs neither to U1 nor to
U2 , so that neither X × Y ⊆ J1 nor X × Y ⊆ J2 . Then there must be pairs
(g, m), (h, n) ∈ X × Y with (g, m) ∈ J1 \ J2 and (h, n) ∈ J2 \ J1 , and from the pre-
suppositions it follows that g, h 6∈ G0 and m, n 6∈ M0 . Hence, (g, n) cannot belong
to J1 ∩ J2 , but certainly (g, n) ∈ I, since (g, n) ∈ X × Y . From (g, m) ∈ J1 \ J2 we
infer g I = g J1 , i.e., (g, n) ∈ J1 , and from (h, n) ∈ J2 \ J1 we dually infer nI = nJ2 ,
i.e., (g, n) ∈ J2 , which is a contradiction.
If (X, Y ) ∈ B(G, M, J1 ) and (U, V ) ∈ B(G, M, J2 ) and (U, V ) ≤ (X, Y ), then
U ⊆ X and V ⊆ Y , i.e., U × Y ⊆ J1 ∩ J2 . This implies U ⊆ G0 (i.e.,
(U, V ) ∈ B(G, M, J1 )) or Y ⊆ M0 (i.e., (X, Y ) ∈ B(G, M, J2 )), in any case
{(U, V ), (X, Y )} ⊆ Ui for i = 1 or i = 2. 
The characterization in Theorem 39 leads the way to the corresponding con-
text construction. The following definition is illustrated by Figure 5.4:
Definition 84 The union of two formal contexts K1 := (G1 , M1 , I1 ) and K2 :=
(G2 , M2 , I2 ) is the context
226 5 Constructions of concept lattices

K1 ∪ K2 := (G1 ∪ G2 , M1 ∪ M2 , I1 ∪ I2 ).

We call K1 ∪ K2 a gluing of the contexts K1 and K2 if the following conditions

are satisfied:
1. G0 := G1 ∩ G2 is an extent of K1 ∪ K2 .
2. M0 := M1 ∩ M2 is an intent of K1 ∪ K2 .
3. I0 := I1 ∩ I2 = I1 ∩ G0 × M0 = I2 ∩ G0 × M0 . ♦

M0

∅
K1
G0

K2
∅

Figure 5.4 With reference to Definition 84

The context gluing is not the exact counterpart of the lattice gluing. The pre-
conditions are weaker. The condition that K1 and K2 coincide on G0 × M0 , does
not at all enforce that the extents contained in G0 and the intents contained in
M0 are also the same in both contexts. This is however necessarily true in the
case of an ideal-filter gluing. Therefore, it is rather surprising that the following
theorem holds true. There is a snag in the theorem, which we have to point out.
It says that the concept lattice of the gluing of two contexts K1 and K2 is the
ideal-filter gluing of two sublattices, but it does not say that those sublattices are
isomorphic to B(K1 ) and B(K2 ). In fact, this is generally not the case.

Theorem 40 B(K) is the ideal-filter gluing of two complete sublattices U1 and U2 with

U1 ∩ U2 = ((G0 , G00 )] ∪ [(M00 , M0 ))

if and only if K is the gluing of two subcontexts K1 := (G1 , M1 , I1 ) and K2 :=

(G2 , M2 , I2 ) with

G0 = G1 ∩ G2 and M0 = M1 ∩ M2 .

Proof One direction immediately follows from Theorem 39: If K := (G, M, I)

satisfies condition 2 of Theorem 39, then with

G1 := {g ∈ G | g I = g J1 },
5.2 Gluings 227

M1 := {m ∈ M | mI = mJ1 },
G2 := G0 ∪ (G \ G1 ),
M2 := M0 ∪ (M \ M1 )
and I1 := I ∩ G1 × M1 ,
I2 := I ∩ G2 × M2

we evidently obtain contexts K1 and K2 wit K = K1 ∪ K2 and G0 = G1 ∩ G2 ,

M0 = M1 ∩ M2 and I0 = I1 ∩ I2 = I1 ∩ G0 × M0 = I2 ∩ G0 × M0 . Since G0 is an
extent and M0 is an intent of K, K is a gluing of K1 and K2 .
For the opposite direction we have to show that K satisfies the second condi-
tion of Theorem 39. For this purpose we set

J1 := I1 ∪ I ∩ (G0 × M ∪ G × M0 )
J2 := I2 ∪ I ∩ (G0 × M ∪ G × M0 ).

Then, we have to prove that J1 and J2 are closed relations. We show this for
J := J1 , with the help of Proposition 56: Let X ⊆ G be arbitrary. If X 6⊆ G0 , then
X J ⊆ M1 and because of

J ∩ G × M1 = I ∩ G × M1

it follows that X JJ = X II . If X ⊆ G0 , then X J = X I and therefore X JJ ⊆

X JI = X II ⊆ G0 , since G0 is an extent of K. Because of J ∩G0 ×M = I0 ∩G0 ×M
and X II ⊆ G0 we have X IJ = X II , i.e., X JJ = X II = X JI . 
The closed relations corresponding to the sublattices have been specified in
the proof. They coincide with I1 resp. I2 up to a modification “below G0 ” and
“above M0 ”. This modification does not apply if I1 is dense in J1 and I2 is dense
in J2 . This is the case if G0 is an extent of K1 as well as of K2 and if furthermore
every subset T ⊆ G0 satisfies

T is extent of K1 ⇔ T is extent of K2

and the corresponding is true for M0 . Those conditions have the effect that the
ideals generated by the concept with the extent G0 are isomorphic in B(K1 ) and
B(K1 ) and that the same is true for the filters of the concepts with the intents
⊆ M0 . The fact that K1 and K2 coincide in I0 implies that those isomorphisms
can be generated by a single map.
In particular, we have: An ideal-filter gluing of two concept lattices is isomor-
phic to the concept lattice of the gluing of the contexts involved.
In practice, the task that usually crops up is the slightly generalized one of
having to glue two lattices together which do not have elements in common, but
in which an isomorphism of the ideal-filter pair of one lattice onto a correspond-
ing pair of the other lattice is given. In order to implement this construction for
concept lattices, one first modifies the respective contexts K1 and K2 in such a
228 5 Constructions of concept lattices

way that both the object concepts in the two ideals and the attribute concepts in
the two filters coincide. This can be achieved through mutual enrichment, and
in the case of doubly founded contexts even through reduction. The objects and
attributes of these concepts are given the same name, if they are mapped onto
each other by the isomorphism. For the remainder one makes the two contexts
disjoint. The concept lattice of the gluing of the contexts modified in this way is
then the ideal-filter gluing of suitable isomorphic copies of their concept lattices,
as desired.

e.......... e e
H H
 @ HH  @ HH
b e ec a
e.......... a e @c e H ed b e e c ea @ ec H ed
   
HH  HH  HH  HH 
@ H
@ @ @ −→ @ H@ @ 
 @
e e e .......... e 3 e H@ e4 e e @ e@ He 3H
@e
H
2 1H
2 H 1
 4
 @ H@ @  @
0
@ 0 0
HH  HH
 @ 
H@ e.......... e
 HH @ e

Figure 5.5 Ideal-filter-gluing of two cubes

a b c d e
a b c a d e 0 × × × ×
0 × × 0 × × 1 × ×
1 × × ∪ 3 × × = 2 × ×
2 × × 4 × × 3 × ×
4 × ×
K1 K2
K1 ∪ K2

Figure 5.6 The context gluing belonging to Figure 5.5.

As an example, Figure 5.5 shows an ideal-filter gluing of two Boolean lattices.

The corresponding context gluing is presented in Figure 5.6. We recognize that
K1 ∪K2 is furthermore a context sum and that consequently the lattice generated
by the gluing is a direct product. In general, for context gluings we have:

(K0 + K1 ) ∪ (K0 + K2 ) = K0 + (K1 ∪ K2 ).

5.3 Local doubling 229

5.3 Local doubling

A further construction principle consists in suitably doubling a part of a lattice,

for example an interval. We first describe the context construction and then de-
rive the corresponding lattice construction.
A context manipulation which has no influence at all on the structure of the
concept lattice is the “inverse reduction”, i.e., the addition of reducible attributes
or objects: To a context (G, M, I) we add for example a set N of new attributes and
supplement the relation I in such a manner that for every n ∈ N , the attribute
extent n0 is a extent of (G, M, I). In this case, (G, M, I) is a dense subcontext of
the new context and, consequently, the concept lattices are isomorphic. The same
is true, if instead we add a set H of new objects to (G, M, I) and make sure that
every such object h ∈ H is reducible with respect to G, i.e., that h0 is an intent of
(G, M, I).
However, if we carry out both extensions simultaneously, the concept lattice
changes considerably. A first clue in this connection is contained in the next few
propositions.
Proposition 104 Let (G ∪ H, M ∪ N, J) be a context with G ∩ H = M ∩ N = ∅ and
J ∩ H × N = ∅. The subcontext (G, M, I) with I := J ∩ G × H satisfies the following
conditions:
1. for every object h ∈ H, h0 is an intent of (G, M, I),
2. for every attribute n ∈ N , n0 is an extent of (G, M, I)
if and only if (G, M, I) is compatible.
This can be proved without effort by means of conditions a1) and a2) of Proposi-
tion 41. 

Hence, under the conditions of the proposition, the map ΠG,M with

(A, B) 7→ (A ∩ G, B ∩ M )

is a surjective complete homomorphism (Proposition 40, p. 120), which, as the

next proposition shows, has small pre-image sets:
Proposition 105 Let (C, D) be a concept of (G, M, I). Then there is at least one and
at most two concepts (A, B) of (G ∪ H, M ∪ N, J) with (C, D) = (A ∩ G, B ∩ M ),
namely
(C, C J ) or (DJ , D).
(C, C J ) is a concept of (G ∪ H, M ∪ N, J) if and only if
– there is an attribute n ∈ N with C ⊆ nJ or
– there is no object h ∈ H with D ⊆ hJ .
Proof Because of J ∩ H × N = ∅, one of the possibilities A ⊆ G or B ⊆ M holds
for every concept (A, B) of (G ∪ H, M ∪ N, J), which, under the condition that
(C, D) = (A ∩ G, B ∩ M ), implies
230 5 Constructions of concept lattices

A = C (and thus B = C J ) or B = D (and thus A = DJ ).

(C, C J ) is not a concept, if C J ⊆ M (i.e., C J = C I = D), but DJ 6⊆ G (i.e.,
DJ 6= C). For D we argue correspondingly. 
The proposition states the possible pre-images for a formal concept (C, D) ∈
B(G, M, I) in a somewhat tricky formulation. Therefore, we repeat the descrip-
tion of the different cases in the form of a table:
C ⊆ nJ for D ⊆ hJ for (C, C J ) is (DJ , D) is equal ?
some n ∈ N some h ∈ H a concept a concept
yes yes yes yes (C, C J ) < (D J , D)
yes no yes no
no yes no yes
no no yes yes (C, C J ) = (D J , D)

Thus it is only in the first case that (C, D) has two different pre-images with
respect to ΠG,M . Additionally, we note down:
Proposition 106 If (G ∪ H, M ∪ N, J) is a context with the properties specified in the
preceding proposition, then:
(A, B) is a concept of (G ∪ H, M ∪ N, J) if and only if (A ∩ G, B ∩ M ) is a concept of
(G, M, I) and we are dealing with one of the following three cases
1. A ⊆ G, B ⊆ M , A = B J , B = AJ
2. A ⊆ G, B = AJ 6⊆ M
3. B ⊆ M, A = B J 6⊆ G.
Proof According to Proposition 104, for every concept (A, B) of (G ∪ H, M ∪
N, J), the restriction (A ∩ G, B ∩ M ) is a concept of (G, M, I), and, since we have
presupposed that J ∩ H × N = ∅, it follows that A × B ∩ H × N = ∅, i.e., A ⊆ G
or B ⊆ N , and thus one of the cases 1)–3) must hold.
If, conversely, A is an extent of (G, M, I) and AJ 6⊆ M , then (A, AJ ) by the
preceding proposition is a concept of (G∪H, M ∪N, J). If AJ = AI , then B = AJ
is an intent of (G, M, I), and, under the condition that B J 6= B I , we can argue
dually. What remains is the trivial case AI = AJ and B I = B J . 

In the case of a doubly founded context, it is particularly easy M N

to check whether the conditions of Proposition 104 are sat- G %
\
isfied. We can apply Proposition 42 and obtain a condition
H .
\ ∅
which is easy to manage algorithmically.

Proposition 107 A doubly founded context (G ∪ H, M ∪ N, J) with G ∩ H = ∅ and

M ∩ N = ∅ has the properties specified in Proposition 104 if and only if the following
conditions are satisfied:
1. J ∩ H × N = ∅,
2. h . m, h ∈ H together imply m ∈ N ,
5.3 Local doubling 231

3. g % n, n ∈ N together imply g ∈ H.


We have made no restrictions concerning the choice of the sets H and N . How-
ever, it turns out that we can make a very special choice without loss of generality.
Definition 85 Assume that C ⊆ B(G, M, I) is a convex set of concepts and
w.l.o.g. that C ∩ (G ∪ M ) = ∅. Then

K[C] := (G ∪ C, M ∪ C, IC ), ♦

IC being defined as follows:

IC ∩ G × M := I, IC ∩ C × C := ∅

and, for (C, D) ∈ C, g ∈ G and m ∈ M ,

g IC (C, D) : ⇔ g ∈ C,
(C, D) IC m : ⇔ m ∈ D.

Evidently, the context defined in this way satisfies the conditions from Propo-
sition 104, C assuming the role of H as well as of N .

M N M C

G I G I ∈

H ∅ C
∈ ∅

(G ∪ H, M ∪ N, J) K [C ]

Figure 5.7 In the case of the doubling construction H and N can be replaced by the same
convex set C.

Proposition 108 If we define C ⊆ B(G, M, I) by specifying for (C, D) ∈ B(G, M, I)

that
(C, D) ∈ C : ⇔ ∃h∈H ∃n∈N C ⊆ nJ and D ⊆ hJ ,
then C is convex and
B(G ∪ H, M ∪ N, J) ∼
= B(K[C]).
Proof Let H0 := {h ∈ H | hJIJ ∩N 6= ∅}. N0 is defined dually. If h ∈ H \H0 , then
hJ = (hJI )J , i.e., h is reducible. (G ∪ H0 , M ∪ N0 , J) is dense in (G ∪ H, M ∪ N, J)
232 5 Constructions of concept lattices

and therefore has the same concept lattice up to an isomorphism. Therefore, we

may presuppose H = H0 , N = N0 . This simplifies the argumentation, since we
have for every h ∈ H some n ∈ N with hJI ⊆ nJ , from which we may infer that
the concept γI (h) := (hJI , hJ ) of (G, M, I) belongs to C. Likewise, we may assign
the concept µI (n) := (nJ , nJI ) ∈ C to every attribute n ∈ N . In the context K[C]
those concepts are objects and attributes. For the object γI (h) we have

γI (h)IC = hJ ,

and for the attribute µI (n) we have

µI (n)IC = nJ .

If (A, B) is a concept of (G, M, I) and h ∈ H, then we have

B ⊆ hJ ⇔ B ⊆ γI (h)IC

and dually. Since furthermore for every concept (C, D) ∈ C by the definition of C
there exists some h ∈ H and some n ∈ N with µI (n) ≤ (C, D) ≤ γI (h), we have

∃h∈H B ⊆ hJ ⇔ ∃c∈C B ⊆ cIC ,

and the dual statement, which implies

AJ ∩ H 6= ∅ ⇔ AIC ∩ C 6= ∅ and B J ∩ N 6= ∅ ⇔ B IC ∩ C 6= ∅.

Now we define for (A, B) ∈ B(G ∪ H, M ∪ N, J)

(A, AIC ) if A ⊆ G

ϕ(A, B) := ,
(B IC , B) if B ⊆ M

and claim that we have thereby defined an isomorphism

ϕ : B(G ∪ H, M ∪ N, J) → B(K[C]).

First of all, we note that ϕ(A, B) by Proposition 106 is defined for every concept
(A, B) ∈ B(G ∪ H, M ∪ N, J). By means of the equivalence proved above and
again by means of Proposition 106 we conclude that ϕ(A, B) is in fact always
a concept of B(K[C]) and even that every such concept occurs. Hence, ϕ is a
bijection. The fact that ϕ is also an order isomorphism is elementary because of
the simple shape of the concepts involved. 
Thanks to Proposition 107 we may concentrate on the context construction
K 7→ K[C], because it covers the general case. The content of Proposition 105,
specialized to the context K[C], reads as follows:
Proposition 109 For every concept (C, D) of K there is at least one and at most two
concepts (A, B) of K[C] with (A ∩ G, B ∩ M ) = (C, D), namely
5.3 Local doubling 233

(C, C IC ) or (DIC , D).

(C, C IC ) and (DIC , D) are both concepts of K[C] and distinct from it if and only if
(C, D) ∈ C.
Proof What remains to be proved is only the last sentence. By Proposition 105
there are two concepts (A, B) with (A ∩ G, B ∩ M ) = (C, D) if and only if there
are elements h, n ∈ C with C ⊆ hIC , D ⊆ nIC , i.e., h ≤ (C, D) ≤ n. Since C is
convex, this is equivalent to (C, D) ∈ C. 
Definition 86 For a convex subset C of a complete lattice L we define the com-
plete lattice L[C] := (L[C], ≤) to be

L[C] := (L \ C) ∪ (C × {0, 1})

and

 x, y ∈ L \ C and x ≤ y in L

or x ∈ L \ C, y = (y0 , i), y0 ∈ C, x ≤ y0 in L

x ≤ y :⇔
 or y ∈ L \ C, x = (x0 , i), x0 ∈ C, x0 ≤ y in L

or x = (x0 , i), y = (y0 , j) ∈ C × {0, 1}, i ≤ j and x0 ≤ y0 . ♦

The assertion that a complete lattice is defined in this way requires a proof. It
follows from the next theorem.

Theorem 41 If C ⊆ B(K) is convex, then

B(K)[C] ∼
= B(K[C]).

Proof We show that the rule

if (A ∩ G, B ∩ M ) 6∈ C,

 (A ∩ G, B ∩ M ),
ϕ(A, B) := ((A ∩ G, B ∩ M ), 0), if (A, B ∩ M ) ∈ C,
((A ∩ G, B ∩ M ), 1), if (A ∩ G, B) ∈ C,


defines an isomorphism

ϕ : B(K[C]) → B(K)[C].

By Proposition 106, for every concept (A, B) of K[C], at least one of (A ∩ G, B)

and (A, B ∩M ) is always a concept of K and, by Proposition 109, a concept (C, D)
of K has two pre-images under ΠG,M if and only if (C, D) ∈ C.
Therefore, ϕ is a bijective map. It remains to be shown that ϕ is also an order
isomorphism. We have

(A1 , B1 ) < (A2 , B2 ) ⇔ A1 ⊂ A2

⇔ A1 ∩ G ⊂ A2 ∩ G or
A1 ∩ G = A2 ∩ G and A1 ⊂ A2
234 5 Constructions of concept lattices

⇔ A1 ∩ G ⊂ A2 ∩ G or (A1 ∩ G, B1 ∩ M ) ∈ C
⇔ ϕ(A1 , B1 ) < ϕ(A2 , B2 ).

In practice, we would if possible reduce the context K[C]. A look at the arrow
relations shows us how: If c ∈ C is an object of K[C], then by Proposition 107
c . d can only hold for one attribute d ∈ C. We discover quickly that precisely
the minimal resp. maximal elements of C are irreducible. If we define

Cmin := {c ∈ C | c is minimal in C}

and Cmax := {c ∈ C | c is maximal in C},

then for c, d ∈ C we have

c . d ⇔ c ∈ Cmin and c ≤ d,

c % d ⇔ d ∈ Cmax and c ≤ d.
If, therefore, we assume that C has enough minimal and maximal elements, i.e.,
that [
C = {[c, d] | c ∈ Cmin , d ∈ Cmax , c ≤ d},

then K[C] is doubly founded (provided that K is doubly founded), and the con-
text K[C] has (up to isomorphism) the same concept lattice as

K[C]r := (G ∪ Cmin , M ∪ Cmax , IC ∩ (G ∪ Cmin ) × (M ∪ Cmax )).

Particularly simple is K[C]r in the case of the interval doubling, that is in the
case that C is an interval C = [(B 0 , B), (C, C 0 )] of B(K), because in this case Cmin
and Cmax are both one-element (sets) and we have

K[C]r = (G ∪ {(B 0 , B)}, M ∪ {(C, C 0 )}, J)

with
J ∩ G × M = I, (B 0 , B)J := B and (C, C 0 )J := C.
We note this down as a proposition:
Proposition 110 A doubly founded context is of the form K[C] for an interval C ⊆
B(K), if and only if there is an object h and an attribute n with

h%
. n, g . n ⇒ g = h, h % m ⇒ m = n.

Because in this case with H := {h} and N := {n} the conditions of Proposition
107 are evidently satisfied. 
Example We consider the possible bracketings of a product x0 x1 · · · xn of n+1
variables x0 , . . . , xn . Since the names of the variables are of no consequence, we
replace them by dots. Thus, (..)((..).) stands for (x0 x1 )((x2 x3 )x4 ), etc. We can
5.3 Local doubling 235

order these bracketings by agreeing that a term becomes larger if subterms are
replaced according to the rule

A(BC) −→ (AB)C.

Tamari [367] observed that this induces an order which turns the set of all brack-
etings of n + 1 symbols into a lattice; this lattice is therefore called the Tamari
lattice Tn . Bennett and Birkhoff [43] have determined the irreducibles of these
lattices. This makes it possible to state a (reduced) context for the Tamari lattice
Tn . With S := {1, 2, . . . , n} and P2 (S) := {{i, j} | i, j ∈ S, i 6= j} this is the
context
(P2 (S), P2 (S), I),
the incidence I for i < j and p < q being defined by

{i, j} I {p, q} : ⇔ not (p ≤ i < q ≤ j).

Geyer [171] has stated a recursion rule for these contexts, which can be recog-
nized by means of the example n = 5 in Figure 5.8.

1 2 1 3 2 1 4 3 2 1
2 3 3 4 4 4 5 5 5 5
1 2 %
. × × × × × × × × ×
2 3 × .
% . × × × × × × ×
1 3 % × .
% × × × × × × ×
3 4 × × × .
% . . × × × ×
2 4 × % × .
% . × × × ×
1 4 % × % × × .
% × × × ×
4 5 × × × × × × .
% . . .
3 5 × × × % × .
% . .
2 5 × % × % × × .
% .
1 5 % × % × × % × × × .
%

Figure 5.8 The reduced context belonging to the Tamari lattice T5 .

The context shows a salient structure of the arrow relations: The (square)
cross table can be arranged in such a way that all double arrows are on the main
diagonal, all upward arrows are above and all downward arrows are below the
main diagonal. In this particular case the “lowest” object h and the correspond-
ing attribute with regard to %
. n evidently satisfy the conditions 110. This means
that the context is generated by interval doubling from the subcontext obtained
by omitting h and n.
236 5 Constructions of concept lattices

However, this subcontext has again the same structure of the arrow relations.
Hence, the procedure can be repeated until there remains nothing. This means
that the Tamari lattice can be generated by iterated interval doubling from the
one-element lattice. Figure 5.9 shows the Tamari lattice T4 including its “gene-

(((..).).).
e
4
@
@
((.(..)).). e
@
@
5 @ e((..)(..)). @
@ @
(.((..).)). e
@ @
@ @
4
6
@ @
e @ @ e((..).)(..)
.(((..).).) e @ @
@ e (..)((..).)
@ @ 4
(.(.(..))). @ @@
2
@ e
(.(..))(..)
@ @
4
6 @ @
.((.(..)).) e @ @
@ @
3 5
1
@ @
@
@ @ e(..)(.(..))
@
.(.((..).)) e @ e.((..)(..))
@
@
@ 1 4
@ 2
@
@
@
@
@
@e
.(.(.(..)))

Figure 5.9 The Tamari lattice T4 . The numbers at the edges indicate the recursive struc-
ture of the lattice, which was generated by interval doubling.

sis”: At the edges, we have noted at which stage of the iterated interval doubling
they have been generated. In descending order, congruences arise, which grad-
ually factorize the lattice until a one-element lattice is reached. >
5.4 Tensorial constructions 237

5.4 Tensorial constructions

By means of the law of mutual distributivity, introduced in 4.4, a lattice-theoretic

characterization of the tensor product can be formulated that does not fall back
upon the notion of a formal context. We only deal with the case of tensor prod-
ucts with two factors L1 := (L1 , ≤) and L2 := (L2 , ≤). The general case is not
significantly different, but takes some getting used to.
Theorem 42 The tensor product L1 ⊗ L2 has the following properties:
⊗1) L1 ⊗ L2 is a complete lattice, and

ε1 : L1 ,→ L1 ⊗ L2 , ε2 : L2 ,→ L1 ⊗ L2

are complete lattice embeddings.

⊗2) The complete sublattices ε1 (L1 ) and ε2 (L2 ) are mutually distributive.
⊗3) If L is a complete lattice satisfying ⊗1) and ⊗2), i.e., if there are embeddings

α1 : L1 ,→ L and α2 : L2 ,→ L

such that the complete sublattices α1 (L1 ) and α2 (L2 ) are mutually distributive,
then there is a complete homomorphism

ϕ : L1 ⊗ L2 → L

with
α 1 = ϕ ◦ ε1 , and α 2 = ϕ ◦ ε2 .
⊗4) The union ε1 (L1 ) ∪ ε2 (L2 ) of the two sublattices generates L1 ⊗ L2 .
By these properties the tensor product is characterized up to isomorphism.
Proof The properties ⊗1), ⊗2) and ⊗4) have already been proved. We can easily
see that the properties ⊗1)–⊗4) are characteristic, since for every lattice with
these properties, from ⊗3) we immediately obtain an isomorphism to the tensor
product.
What remains to be shown is ⊗3). Hence, let L be a lattice with the properties
specified in ⊗3). First, we work out the following sub-claim:
For every subset X ⊆ L1 × L2 we have
_ ^
(α1 (x1 ) ∧ α2 (x2 )) = (α1 (y1 ) ∨ α2 (y2 )).
(x1 ,x2 )∈X (y1 ,y2 )∈X ∇

For this purpose, we make use of the condition that the two image sets are mu-
tually distributive and obtain:
_ ^ _ _
(α1 (x1 ) ∧ α2 (x2 )) = ( α1 (x1 ) ∨ α2 (x2 ))
(x1 ,x2 )∈X R⊆X x1 ∈R x2 ∈X\R
238 5 Constructions of concept lattices
^ _ _
= (α1 ( x1 ) ∨ α2 ( x2 )).
R⊆X x1 ∈R x2 ∈X\R

By means of the notations y1R := x1 ∈R x1 and y2R := x2 ∈X\R x2 we simplify

W W
this to
_ ^
(α1 (x1 ) ∧ α2 (x2 )) = (α1 (y1R ) ∨ α2 (y2R )).
(x1 ,x2 )∈X R⊆X

Every element of X belongs to R or X \ R, therefore, either its first com-

ponent must be ≤ y1R or its second component ≤ y2R , in any case we have
(x1 , x2 )∇(y1R , y2R ) for all (x1 , x2 ) ∈ X, and consequently (y1R , y2R ) ∈ X ∇ , inde-
pendent of R. This proves
^ ^
(α1 (y1R ) ∨ α2 (y2R )) ≥ (α1 (y1 ) ∨ α2 (y2 )).
R⊆X (y1 ,y2 )∈X ∇

If, on the other hand, for (y1 , y2 ) ∈ X ∇ we specifically choose R := {(x1 , x2 ) ∈

X | x1 ≤ y1 }, then X \ R ⊆ {(x1 , x2 ) ∈ X | x2 ≤ y2 } and therefore α1 (y1R ) ∨
α2 (y2R ) ≤ α1 (y1 ) ∨ α2 (y2 ), from which it follows that
^ ^
(α1 (y1R ) ∨ α2 (y2R )) ≤ (α1 (y1 ) ∨ α2 (y2 ))
R⊆X (y1 ,y2 )∈X ∇

and thus the sub-claim. Now we define a map ϕ : L1 ⊗ L2 → L by

_ ^
ϕ(A, B) := α1 (x1 ) ∧ α2 (x2 ) = α1 (y1 ) ∨ α2 (y2 ).
(x1 ,x2 )∈A (y1 ,y2 )∈B 

We have to show that ϕ is a complete homomorphism. W Because of the symmetry

of the definition it suffices to prove the property “ -preserving”. For this pur-
pose we use the sub-claim and obtain for an arbitrary subset {(At , Bt ) | t ∈ T } ⊆
L1 ⊗ L2
_ _ _
ϕ(At , Bt ) = (α1 (x1 ) ∧ α2 (x2 ))
t∈T t∈T (x1 ,x2 )∈At
_
= (α1 (x1 ) ∧ α2 (x2 ))
S
(x1 ,x2 )∈ At
^
= (α1 (x1 ) ∨ α2 (x2 ))
At ) ∇
S
(y1 ,y2 )∈(
^
= (α1 (x1 ) ∨ α2 (x2 ))
T
(y1 ,y2 )∈ Bt
_
= ϕ( (At , Bt )),
t∈T
5.4 Tensorial constructions 239

as desired.
Finally, we have to examine the connection between the maps αi and εi . For
this purpose we recall the definition of the εi (in particular of ε1 ), from which it
follows that, for an arbitrary x ∈ L1 , the extent of ε1 (x) is given by {(x1 , x2 ) ∈
L1 × L2 | x1 ≤ x or x2 = 0}. Thereby we obtain
_ _
ϕ(ε1 (x)) = (α1 (x1 ) ∧ α2 (x2 )) ∨ (α1 (x1 ) ∧ α2 (x2 ))
x1 ≤x x1 ≤1
x2 ≤1 x2 ≤0

= α1 (x) ∧ α2 (1) = α1 (x).


The maps γ and µ mapping onto the object and attribute concepts, respec-
tively, are related to the εt . We have γ(x1 , x2 ) =

({(g1 , g2 ) | g1 ≤ x1 and g2 ≤ x2 } ∪ M ∇ , {(m1 , m2 ) | x1 ≤ m1 or x2 ≤ m2 }),

and µ(x1 , x2 ) =

({(g1 , g2 ) | g1 ≤ x1 or g2 ≤ x2 }, G∇ ∪ {(m1 , m2 ) | x1 ≤ m1 and x2 ≤ m2 }),

and therefore

γ(x1 , x2 ) = ε1 (x1 ) ∧ ε2 (x2 ) and µ(x1 , x2 ) = ε1 (x1 ) ∨ ε2 (x2 ).

It has proved worthwhile to introduce special symbols for these maps.

Definition 87 If L1 and L2 are complete lattices, then the tensorial operations

∧ : L1 × L2 → L1 ⊗ L2 and ∨ : L1 × L2 → L1 ⊗ L2

are defined by
x1 ∧ x2 := γ(x1 , x2 ) = ε1 (x1 ) ∧ ε2 (x2 ),
x1 ∨ x2 := µ(x1 , x2 ) = ε1 (x1 ) ∨ ε2 (x2 ). ♦

Proposition 111 The tensorial operations satisfy the following arithmetic rules:

x1 ∧ 1 = ε1 (x1 ) = x1 ∨ 0, 1 ∧ x2 = ε2 (x2 ) = 0 ∨ x2 ,

x1 ∧ x2 = x1 ∨ 0 ∧ 0 ∨ x2 , x 1 ∨ x2 = x1 ∧ 1 ∨ 1 ∧ x2 ,
^ ^ ^ _ _ _
x1s ∧ xs2 = ( xs1 ) ∧ ( x2s ), x1s ∨ xs2 = ( xs1 ) ∨ ( xs2 ),
s∈S s∈S s∈S s∈S s∈S s∈S
_ _ ^ ^
x1 ∧ x2s = x1 ∧ ( xs2 ), x1 ∨ xs2 = x1 ∨ ( xs2 ),
s∈S s∈S s∈S s∈S
_ _ ^ ^
x1s ∧ x2 = ( x1s ) ∧ x2 , xs1 ∨ x2 = ( xs1 ) ∨ x2 ,
s∈S s∈S s∈S s∈S
240 5 Constructions of concept lattices

1∨ 0=1∧ 1=0∨ 1 (= 1 ∨ x = 1 ∨ y = 1 ∨ z
e = w ∨ 1 = 1 ∨ 1)
@
@
@
w∨x w y
e ∨ @ @e ∨ w z
e
@
@ @
@
@ @
@ @ @
0∨x
e @ e0 ∨ y @
@ e w ∨ 0@
w ∧ 1@ @ e0 ∨ z
1∧x @ @ 1∧y 1∧z
@ @
@ @
@@e @e
ew ∧ y @
w∧x @ w∧z
@
@
@e
@
= w ∧ 0 = 0 ∧ 0)
0∧ 1=0∨ 0=1∧ 0 (= 0 ∧ x = 0 ∧ y = 0 ∧ z

1
1 c c
Figure 5.10 w c ⊗ x c c@
y cz
@ with the tensorial operations.
0 c
@@c
0

^ _ ^ ^
(xs1 ∨ xs2 ) = (( xr1 ∧ 1) ∧ ( 1 ∧ xs2 )),
s∈S R⊆S r∈R s∈S\R
_ ^ _ _
(xs1 ∧ xs2 ) = (( xr1 ∨ 0) ∨ ( 0 ∨ xs2 )).
s∈S R⊆S r∈R s∈S\R

Proof All these rules result immediately from the definitions, apart from the last
two, for which we have to consult Proposition 94: Because of x1 ∧ x2 = ε1 (x1 ) ∧
ε2 (x2 ), x1 ∨ x2 = ε1 (x1 ) ∨ ε2 (x2 ) and the rules mentioned in the first line, the
equations are precisely the translation of the circumstance that the sublattices
ε1 (L1 ) and ε2 (L2 ) are mutually distributive. 
Thiele [369] has impressively demonstrated how to obtain readable diagrams
of tensor products of small lattices. First, the idea of the P -product developed in
Section 5.1 is transferred to the tensor product and it is agreed that:
Definition 88 For a P -lattice (L1 , α1 ) and a Q-lattice (L2 , α2 ) with P ∩ Q = ∅,
5.4 Tensorial constructions 241

(L1 , α1 ) ⊗ (L2 , α2 ) := (L, α)

is the P ∪˙ Q-sublattice of L1 ⊗ L2 for which the map

α : P ∪˙ Q → L1 ⊗ L2

is defined as follows:
if r ∈ P ,

ε1 α1 (r)
α(r) :=
ε2 α2 (r) if r ∈ Q.

♦
By means of Theorem 42 we quickly convince ourselves of the fact that this
indeed defines a P ∪˙ Q-lattice.
If L1 and L2 are concept lattices, we can introduce the corresponding context
operation:
Definition 89 For a P -context (K1 , α1 ) and a Q-context (K2 , α2 ) with P ∩ Q = ∅
we define
(K1 , α1 ) × (K2 , α2 ) := (K1 × K2 , α)
to be the P ∪˙ Q-context for which the map

α : P ∪˙ Q → B(K1 × K2 )

is explained as in Definition 88. ♦

From these two definitions it immediately follows that

B((K1 , α1 ) × (K2 , α2 )) ∼
= (B(K1 ), α1 ) ⊗ (B(K2 ), α2 ).

Thiele has shown that the product defined in this way is distributive over the
P -fusion, i.e., that it is possible to transfer Proposition 18 to this case. This is the
content of the following theorem.
Theorem 43 If (K1 , α1 ) and (K2 , α2 ) are both P -contexts and if (K3 , α3 ) is a Q-context
with P ∩ Q = ∅, then
P
((K1 , α1 ) + (K2 , α2 )) × (K3 , α3 )
˙
P ∪Q
= ((K1 , α1 ) × (K3 , α3 )) + ((K2 , α2 ) × (K3 , α3 )).
Proof Both sides of the equation claimed describe closed relations of

K := (K1 + K2 ) × K3 = K1 × K3 + K2 × K3

(cf. Proposition 18). If we are able to show that the map α is also the same in both
cases, nothing remains to be proved, since in this case the sublattices generated
by
242 5 Constructions of concept lattices

{αx | x ∈ P ∪˙ Q}
and thus the corresponding closed relations must also be the same. This can be
checked easily; the main problem is that of a transparent notation. We again use
the abbreviations

(Axt , Btx ) := αt x for t ∈ {1, 2, 3} and x ∈ P ∪˙ Q.

P
Furthermore, we write (K12 , α12 ) for the P -context (K1 , α1 ) + (K2 , α2 ) and agree
on the abbreviation

(A1 , B1 ) + (A2 , B2 ) := (A1 ∪ A2 , B1 ∪ B2 ).

Then we have
α12 p = α1 p + α2 p.
For the embedding maps we use the symbols ε1 , ε2 , ε12 and ε3 in the obvious
way, in the case of ε3 , however, we have to differentiate: We write εi3 , if we are
working in the product Ki × K3 . For reasons of readability we presuppose that
none of the contexts contains any full columns or full rows. Thus, the trivials
terms M ∇ and G∇ disappear when we evaluate the maps εi by means of the
formula stated in Proposition 90.
For the left-hand side we obtain, if p ∈ P ,

αp = ε12 α12 p
= ε12 (A1p ∪ Ap2 , B1p ∪ B2p )
= ((Ap1 ∪ A2p ) × G3 , (B1p ∪ B2p ) × M3 )

and for q ∈ Q

αq = ε12
3 α3 q
= ε312 (A3q , B3q )
= ((G1 ∪ G2 ) × A3q , (M1 ∪ M2 ) × B3q ).

On the right-hand side we calculate for p ∈ P

αp = ε1 α1 p + ε2 α2 p
= (A1p × G3 , B1p × M3 ) + (Ap2 × G3 , B2p × M3 )
= ((A1p ∪ Ap2 ) × G3 , (B1p ∪ B2p ) × M3 )

and for q ∈ Q

αq = ε13 α3 q + ε23 α3 q
= (G1 × Aq3 , M1 × B3q ) + (G2 × A3q , M2 × B3q )
= ((G1 ∪ G2 ) × Aq3 , (M1 ∪ M2 ) × B3q ).
5.4 Tensorial constructions 243


Because of the nice applications offered by this theorem, we substantiate it by

several examples. The first one simply demonstrates the situation reflected by
the theorem.
Example Consider the two 3-contexts (K1 , α1 ) and (K2 , α2 ), given as follows

M1 M2 c
c  @ c3
× × ×
c1 c3@ c2
@ c
1
G1 × G2 × c2
\
× @@c × \c

as well as the {4}-context (K3 , α3 ) with the following illustration:

M3 c
c4
G3 × c
With P := {1, 2, 3} and Q := {4} we recognize in Figure 5.11 on the left the
context
P
((K1 , α1 ) + (K2 , α2 )) × (K3 , α3 )
and on the right
˙
P ∪Q
((K1 , α1 ) × (K3 , α3 )) + ((K2 , α2 ) × (K3 , α3 )).

We also recognize that both contexts are equal. >

M1 M2 M1 M2 M1 M1 M2 M2
× ×××× ××× × × ××××××
G1 × ×× × ×× G1 × × ×× ××
××× ××× × ××× ××
××××× ××××× × ×××××××××
G2 ×××× ×××× G1 × ××××× ×××
× ×× × = ×××××× ×××
× ××××××××× ×××××××× ××
G1 × ××××××××× G2 ××××××× ×
××× ×××××× × × × ×
××××× ×××××× ×××××××× ×××
G2 ×××× ×××××× G2 ××××××× ×××
× ××××××× × ××× ××××

Figure 5.11 According to Theorem 43 both contexts are equal.

Corollary 7 For two any P -lattices (L1 , α1 ), (L2 , α2 ) and a Q-lattice (L3 , α3 ) (with
P ∩ Q = ∅) we have
244 5 Constructions of concept lattices

P
((L1 , α1 ) × (L2 , α2 )) ⊗ (L3 , α3 )
P ∪Q
∼ ((L1 , α1 ) ⊗ (L3 , α3 )) × (L2 , α2 ) ⊗ (L3 , α3 )).
=

This corollary also goes back to Thiele. He has used it skillfully in order to
draw diagrams of tensor products. As a first application we calculate the tensor
product of two four-element chains.
Example In order to calculate the tensor product of two four-element chains,
we make use of the fact that such a chain can be written as a 2-product of a two-
element and a three-element chain:
c c
c
c2 2 c @ 2 c2
= c 1,2 × c .
@
∼
c1 = @c
1
1 c
c @c
@

If we decompose both factors in this way, we obtain

c c  c   c 
c2 c4 {1,2} c2 {3,4} c4
∼  c 1,2  ⊗  c 3,4
c1 ⊗ c3 = × c1 × c3 .

c c
c c

Using Corollary 7 this can be multiplied out. The convention “tensor product
first, then the P -product” saves brackets. The above expression yields
c c c c
c 4 c2 c2 c4
c 1,2 ⊗ 4 ×
4 4
∼ c 1,2 ⊗ c 3,4 × c
= c3 c 1 ⊗ 3,4 × c1 ⊗ c3 .
c c c c

As can be easily seen by means of the contexts, the tensor product of two three-
element chains has six elements. With respect to the tensor product, two-element
chains behave like neutral elements. Using these observations, we can convert the
expression into a mere 4-product:
c
c c4 c2
c 2,4
∼ 1,2 c @ c3,4 ×
4
c 1,2 ×
4
c 3,4 ×
4
=
@c c 1,3 .
c3 c1
c
c c2
c c4
4 c @
c @ c3,4 4 c
4 @
∼
= 1,2 × 1,2 × @c
@c c3 3
@c
c
@
1
5.4 Tensorial constructions 245

c c2
c c4
c4
1,2 c @ c3,4 ×
4 4
∼
= c 1,2 ×
@c c3
c3
c c1
c
c
c 4 c @ c2
@c @
1,2 c @ c3,4 ×
4
∼
=
@
@c @c @
@ @
c @@c @c
1 3
@c
b
c 4 b
@ b2
c @b
∼
= 1,2 c @ c3,4 4
× b
@c b @ b3
1
c @b
This product has already been calculated in Example 5.1. The result is presented
in Figure 5.2. >
Example A particularly nice application of this method is Thiele’s represen-
tation of a free distributive lattice with four generators as a subdirect product.
The nested line diagram obtained thereby is presented in Figure 1.15.
In general, it is true that FCD(n) is isomorphic to the n-th tensor power of the
three-element lattice [395]. This means that FCD(4) can be obtained as the tensor
product of four three-element chains.
We make use of the fact that a tensor product of two three-element chains can
be rewritten as a 2-product:
c
c
c c c @
c1 2 c2
@
c1 ⊗ c2 ∼ 1 c @c ∼
= 2 = c2 × c1
c c @c
c
@c
@

Thereby we obtain
 c c   c c 
FCD(4) ∼
=  c1 ⊗ c2  ⊗  c3 ⊗ c4 
c c c c
 c   c 
 c1 {1,2} c2   c3 {3,4} c4
∼

= c2 × c1
⊗
  c4 × c3


c c
246 5 Constructions of concept lattices
c c c c
c1 c3 4 c1 c4 4 c2 c3 4 c2 c4
∼
= c2 ⊗ c4 × c2 ⊗ c3 × c1 ⊗ c4 × c1 ⊗ c3
c c c c
c1 c3 c c
c 1,3 4 c 4 4 c2 4 c 2 c4
∼
= × × × ⊗ .
c 2,4 c3 c1 c1 c3
c2 c4 c c
The tensor product of two four-element chains has already been calculated in
Example 16; the 4-product of the remaining factors was treated in Example 5.1,
see Figure 5.1. Hence the 4-product presented in Figure 5.12 is isomorphic to the
tensor product of four three-element chains, and thus also to the free completely
distributive lattice with four generators. This is how the diagram in Figure 1.15
has been obtained.

d
d d
@
d @d
@
d @d
1 d
@ @ 3
@d @d
@ @
2 d
@d @ d4
@ @
@d d @d
@ @
@d d @d
4
@ @ ×
d @d @d
@ @
d @d @d
@ @
d @d @d
@ @
1 d
@d @ d3
4 @ @ 2
@d @d
@ @
@d @d
@
@d
@
@d

d
Figure 5.12 The free distributive lattice FCD(4) as a 4-product.
5.5 Notes, references, and trends 247

5.5 Notes, references, and trends

5.1

Section 5.1 follows [399] and the predecessor of this article, [393].
P -products of lattices are a long-standing subject of one of the authors of this
book, see [387] and [388]. Bartenschlager [25] and Thiele [369] in their work
make ample use of the P -product as a mathematical construction tool. [142] uses
P -products to describe rough sets in the language of Formal Concept Analysis.

5.2

The results of this section are based on the doctoral thesis of S. Gürgens [180], we
have, however, changed the notations. It also contains further-reaching results.
Gürgens states an algorithm which determines whether K is the gluing of two
contexts. Furthermore, she studies the simultaneous gluing of several lattices
or contexts, respectively. Her model were the gluings of Boolean lattices in the
theory of orthomodular lattices, cf. Greechie [176].

5.3

Local doubling was introduced by Day [85], first for intervals and then more
generally. It played an important role in the framework of the examination of free
lattices, see also Day [86], Nation [291] and Day, Nation & Tschantz [88]. Day
even stated a concept-analytic version of the construction of interval doubling.
Our representation mainly follows Geyer [170].
The fact that the bracketings form a lattice was first published by Tamari [367].
Later, Huang and Tamari [195] gave a simpler proof. The concept-analytic inves-
tigation goes back to Geyer [171].

5.4

The direct product of contexts in particular has proved to be a natural product

for conceptual scales. Products of the elementary scales have been examined and
illustrated by many diagrams in Thiele [369]. This article, which has already been
cited several times, also contains the result that the direct product of two closed
relations is again closed. Hence, tensor products of sublattices lead to sublattices
of the tensor product.
Strahringer [354] describes products of convex-ordinal scales. [398] uses the
direct product for the general modeling of dependencies between many-valued
attributes. Stumme [361] uses it for distributive concept exploration.
248 5 Constructions of concept lattices

The set of all order-preserving maps from an ordered set P into a complete
lattice L also forms a complete lattice, when ordered point-wise. This lattice is
denoted by LP . Occasionally, a formula is used which establishes a connection
between this lattice and the lattice 2P of all order-preserving maps of P into the
two-element lattice 2 (for which a context construction was given on Page 48):

LP ∼
= 2P ⊗ L.
Chapter 6
Properties of concept lattices

It is not uncommon to encounter the claim that it is “a fundamental problem of

Formal Concept Analysis that the size of the concept lattices is exponential in the
size of the formal contexts”. In fact, there are such examples. Typical are the con-
tranominal scales, which have 2n formal concepts for n objects and n attributes
(p. 46). In fact, the contranominal scale is always the reason, because formal con-
texts that do not contain large contranominals do not have large concept lattices.
This needs to be specified and proved, and this is done in the first section.
But the size of the lattices is not the most important problem. Even small con-
cept lattices can be confusingly complicated. On the other hand, some lattices are
so well structured that they are easy to understand despite their size. Structural
properties, as typically studied in algebraic lattice theory, can therefore also be
important for concept lattices. This applies in particular to the distributive law,
which we will discuss in more detail in the second section. Generalizations of
this law are discussed there and in the third section.
The fourth section is devoted to the notion of dimension. The order dimension
of the concept lattice is reflected in the Ferrers dimension of the formal context,
and this translation can even be broken down to specializations such as the k-
dimension. The special case k = 2 is particularly interesting because it has cross-
connections to the current topic of Boolean matrix factorization.
At the beginning of the fifth section, we mention some of the many approaches
to achieve simplification by not representing all concepts of a formal context, but
only particularly important ones.

6.1 Size and breadth

In this section, we focus on the size of concept lattices. A first result is that
this is not easy to predict: Kuznetsov [245] has shown that it is difficult (“#P -
complete”) to compute the number of formal concepts for given contexts. A. Al-
bano and B. Chornomaz have jointly made an important cause of the size of

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 249
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_7
250 6 Properties of concept lattices

concept lattices understandable by combining Formal Concept Analysis with

extremal set theory. Our presentation follows Chapter 5 of Albano’s disserta-
tion [1].
Recall that Nkc denotes the contranominal scale

Nck := ({1, . . . , k}, {1, . . . , k}, 6=)

with k objects and attributes, defined on Page 46. Its concept lattice is isomor-
phic to the powerset lattice of a k-element set and therefore has 2k elements. In
mathematical literature it is known as the “Boolean lattice with k atoms”. We
abbreviate it to B(k) := B(Nck ).
Definition 90 A formal context K is called Nck -free if it does not contain a sub-
context isomorphic to Nck . A complete lattice L is B(k)-free whenever B(k) does
not order-embed into L. ♦
Figure 6.1 shows a lattice without a B(4)-sublattice, which is not B(4)-free.

Figure 6.1 A lattice containing B(4) as a suborder, but not as a sublattice. From [1].

Proposition 112 A formal context K is Nck -free if and only if B(K) is B(k)-free.

Proof It follows from Proposition 38 on p. 119 that B(k) embeds into B(K)
whenever Nck ≤ K. To prove the converse, let (A1 , A01 ), . . . , (Ak , A0k ) be the atoms
of an B(k) embedded in B(K) and let (D10 , D1 ), . . . , (Dk0 , Dk ) be the coatoms of
the same B(k) in such a way that (Ai , A0i ) ≤ (Dj0 , Dj ) ⇔ i 6= j. For each i we can
choose an object-attribute pair (gi , mi ) with gi ∈ Ai , mi ∈ Di and (gi , mi ) ∈ / I
because of (Ai , A0i ) 6≤ (Di0 , Di ). Since (Ai , A0i ) ≤ (Dj0 , Dj ) holds true for all i 6= j,
it follows that gi I mj for i 6= j. It also follows that the gi are pairwise differ-
ent, as are the mj . That is, the objects gi and attributes mj form a contranominal
subcontext of K. 
6.1 Size and breadth 251

The following theorem is due to Albano & Chornomaz [2].

Theorem 44 Let K := (G, M, I) be a formal context and A ⊆ G. There exists a contra-
nominal subcontext of K having A as its object set if and only if A is a minimal extent
generator1 .
Proof Suppose A is a minimal extent generator of (G, M, I) and g ∈ A. From
/ I and
(A \ {g})0 ) A0 we infer that there exists an attribute mg with (g, mg ) ∈
h I mg for every h ∈ A \ {g}. Clearly mg 6= mh if g 6= h, and thus N := {mg |
g ∈ A} gives a contranominal subcontext (A, N, I ∩ A × N ). For the converse
let (A, B, I ∩ A × B) be a contranominal subcontext of K and denote for each
g ∈ A by mg the unique attribute in B which is not incident to g. Then mg ∈ / A0 ,
but mg ∈ (A \ {g}) , which implies (A \ {g}) ) A for each g ∈ A. Thus A is a
0 0 0

minimal extent generator. 

Corollary 8 K is Nkc -free if and only if |A| < k for every minimal extent generator A.

The number of subsets of an n-element set which have less than k elements is
X n 
f (n, k) := .
i
i<k

According to Corollary 8 an Nck -free formal context (G, M, I) with |G| =: n < ∞
cannot have more than f (n, k) minimal extent generators. And since every finite
extent has at least one minimal generator, (G, M, I) cannot have more that f (n, k)
extents.
Birkhoff [49] defines2 the breadth of a finite lattice to be the smallest integer
b such that any join x1 ∨ · · · ∨ xn (n > b) is the join of a subset of b of the xi .
W X of the lattice can be called an irredundant join representation if
A subset
x 6≤ (X \ {x}) holds for all x ∈ X. The breadth of a finite lattice is then the
maximum size |X| of such an irredundant representation.
These notions fit the topic perfectly:
Corollary 9 The breadth of a finite concept lattice is the largest size of a minimal extent
generator and also the largest size of a minimal intent generator.
It is moreover the largest number b for which B(b), the Boolean lattice with b atoms,
can be order-embedded.
Proof Let X be an irredundant join-representation. Then for each xW ∈ X there
mustWbe some join-irreducible element γgx ≤ x for which γgx 6≤ (X \ {x}),
and {γgx | x W ∈ X} also is irredundant. {gx | x ∈ X}00 is the extent of the
formal concept {γgx | x ∈ X} and {gx | x ∈ X} is a minimal generator of
that extent. This means that a minimal generator of the same size is obtained
1 I.e., a minimal generator wrt. the closure system of concept extents of K in the sense of Defi-
nition 16 on p. 14.
2 In fact, Birkhoff gives the dual definition, but mentions an observation by I. Rose that it is

self-dual.
252 6 Properties of concept lattices

from each irredundant join representation. TheWreverse is also true, because if

n
{g1 , . . . , gn } is a minimal extent generator, then i:=1 γgi is an irredundant join
representation.
Therefore, the breadth, which equals the maximum size of an irredundant
join representation, is also the maximum size of an extent generator. By The-
orem 44 this is the largest number of objects of a contranominal subcontext,
which, according to Proposition 112, is the number of atoms in the largest order-
embeddable Boolean lattice. 
Because Boolean lattices are self-dual, the breadth of a lattice is equal to the
breadth of its dual.
Theorem 44 and its corollaries make it clear what makes concept lattices large:
the contranominal subcontexts. If a formal context contains the contranominal
scale Nkc as a subcontext, then the concept lattice contains a Boolean lattice B(k)
as a suborder and therefore has at least 2k elements. If the context does not con-
tain Nck , then the concept lattice can have at most f (n, k) elements, where n is the
number of objects of the context. Without proof we cite two further results from
Albano & Chornomaz [2]:
Theorem 45 The size of a finite lattice (L, ≤) of breadth b is bounded by

2b ≤ |L| ≤ f (n, b + 1),

where n is the number of join-irreducibles. The bounds are sharp. 

To prove that the bounds are sharp, the authors construct an extremal lattice for
each pair of values n, k, i.e. an example of a Nkc -free context with n objects whose
concept lattice has f (n, k) elements. Using an estimate of the function f (n, k)
they also obtain:
|G|
Corollary 10 A finite concept lattice B(G, M, I) of breadth b < 2 satisfies

b+1
2b ≤ |B(G, M, I)| ≤ · |G|b .
b!

Albano and Chornomaz also show that for a formal context with n objects
f (n, k) not only places an upper bound on the size of the concept lattice, but also
on the number of pseudo-extents and the number of proper premises (of object
implications). The reasons for this are that proper premises are always minimal
generators and that the canonical basis according to Proposition 28 on p. 98 is
not greater than the number of proper premises.

6.2 Distributivity, modularity and semimodularity

Already in Definition 14 we had introduced variants of the distributive law: A

complete lattice L is called distributive if the following two (mutually equiva-
6.2 Distributivity, modularity and semimodularity 253

lent) laws

(D∧ ) x ∧ (y ∨ z) = (x ∧ y) ∨ (x ∧ z)
(D∨ ) x ∨ (y ∧ z) = (x ∨ y) ∧ (x ∨ z)

hold, and it is called completely distributive if the following generalization to

arbitrary infima and suprema is satisfied for all index sets S, T 6= ∅:
^ _ _ ^
(DW V ) xs,t = xs,ϕ(s) .
s∈S t∈T ϕ:S→T s∈S

This law is also equivalent to its dual (DV W ). One direction of the law (DW V ),
namely the inequality
^ _ _ ^
xs,t ≥ xs,ϕ(s) ,
s∈S t∈T ϕ:S→T s∈S

holds in every complete lattice, since, for fixed ϕ, s∈S xs,ϕ(s) is always less than
V
or equal to the left-hand side.
We frequently use a stricter version of the law of complete distributivity, which
can however be derived from the one mentioned above. We allow the set T to
vary with s ∈ S; for this purpose we replace T by a family of sets

{Ts | s ∈ S}.

The place of the maps ϕ : S → T is taken by the elements

ϕ∈ ×T
s∈S
s

of the direct product of these sets (which we abbreviate as × Ts). This version
of the law (DV W ) then reads
^ _ _ ^
xs,t = xs,ϕ(s) .
s∈S t∈Ts
ϕ∈ ×T s
s∈S

The inequality “≥” again holds in every complete lattice.

Proofs for the equivalences we have claimed can be found in the books cited on
lattice theory, in particular Balbes & Dwinger [17]. The following useful charac-
terization of distributive lattices has been taken from Birkhoff’s “Lattice Theory”:

Proposition 113 A lattice is distributive if and only if a ∧ x = a ∧ y and a ∨ x = a ∨ y

always imply x = y. 

Examples of completely distributive complete lattices are the powerset lat-

tices, and more generally lattices of the order ideals of ordered sets, as stated by
the following well known theorem:
254 6 Properties of concept lattices

Theorem 46 (TheoremW of Birkhoff) If D is a completely distributive complete lattice

in which the set J(D) of -irreducible elements is supremum-dense, then

x 7→ (x] ∩ J(D)

describes an isomorphism of D onto the closure system of all order ideals of (J(D), ≤).
Conversely, for every ordered set (P, ≤) the closure system of all order ideals is a com-
pletely distributive lattice D, in which

J(D) = {(x] | x ∈ P }

is supremum-dense.

Proof For x ∈ D, (x] ∩ J(D) is obviously an order ideal of (J(D), ≤). If A is

an order ideal of (J(D), ≤) and a := A, then A ⊆ (a] ∩ J(D) and even A =
W
(a] ∩ J(D), as the following consideration shows: For x ∈ D we have
_ _
x ∈ (a] ⇔ x ≤ a ⇔ x ≤ A ⇔ x=x∧ A.

By means of the distributive law we obtain

_ _
x ∈ (a] ⇔ x = x ∧ A = {x ∧ y | y ∈ A}.

If additionally x is -irreducible, then x = {x ∧ y | y ∈ A} can only occur if

W W
x = x ∧ y holds for some y ∈ A, i.e., if x ≤ y holds for some y ∈ A. Since A is an
order ideal in J(D), this yields x ∈ A.
Hence if J(D) is supremum-dense in D,

x 7→ (x] ∩ J(D)

describes a bijection, which, because of

x ≤ y ⇔ (x] ∩ J(D) ⊆ (y] ∩ J(D),

is even a lattice isomorphism.

The intersection and the union of an arbitrary number of order ideals of an
ordered set (P, ≤) are again order ideals. Therefore, the lattice of all order ideals
is a complete sublattice of the powerset lattice of P and thus is completely dis-
tributive. Every order ideal A is the union and hence the supremum of principal
ideals: [
A= (a],
a∈A

and, because (a]∗ = (a] \ {a}, every principal ideal is -irreducible.

W

The following theorem gives a context characterization without assuming dou-
bly foundedness.
6.2 Distributivity, modularity and semimodularity 255

Theorem 47 A concept lattice B(G, M, I) is completely distributive if and only if for

every non-incident object-attribute pair

(g, m) 6∈ I

there exist an object h ∈ G and an attribute n ∈ M with (g, n) 6∈ I, (h, m) 6∈ I and

h ∈ k 00 for all k ∈ G \ {n}0 .

Proof The following statement holds for every concept lattice:

^ _ _ ^
(As,t , Bs,t ) ≥ (As,ϕ(s) , Bs,ϕ(s) ).
s∈S t∈Ts
ϕ∈ ×T s
s∈S

Assume that the left-hand side is strictly greater than the right-hand side. Then
there exist
\ \ \ \
g∈ ( Bs,t )0 and m ∈ ( As,ϕ(s) )0
s∈S t∈T
ϕ∈ ×T s
s∈S

with (g, m) 6∈ I. Now if there are

h ∈ G and n ∈ M with (g, n) 6∈ I, (h, m) 6∈ I and h ∈ k 00 for all k ∈ G \ n0 ,

then n ∈ t∈T Bs,t cannot hold for s ∈ S, because of g ∈ ( t∈T Bs,t )0 . Therefore,
T T

there is a ϕ̂ ∈× s∈S Ts with n 6∈ Bs,ϕ̂(s) for all s ∈ S. From h ∈Tk and k ∈ G

00
\ n0
it follows that h ∈ As,ϕ̂(s) for all s ∈ S. Since, however, m ∈ ( s∈S As,ϕ̂(s) ) , this
0

results in a contradiction with (h, m) 6∈ I. Hence, the equation follows from the
conditions specified. In order to be able to use complete distributivity to prove
the inverse direction, we first argue that
^ _
(g 00 , g 0 ) = (ϕ(n)00 , ϕ(n)0 )
ϕ n∈M \g 0

holds for every object g ∈ G, provided the maps ϕ under the -operator are
V
chosen as follows:
ϕ∈ ×
(G \ n0 ).
n∈M \g 0

Thus ϕ runs over all maps that assign to each attribute n which is not incident
with g an object ϕ(n) which is not incident with n. Hence, a possible choice is
ϕ(n) := g for all n, whereby we obtain the direction “≥” of the statement. For
the other direction we note that
W n 6∈ ϕ(n) . Hence,
0
n can still less be contained in
the intent of the supremum n∈M \g0 (ϕ(n)00 , ϕ(n)0 ), i.e., this intent is a subset of
g0 .
We have to show that – provided that the concept lattice is completely dis-
tributive – it is possible, for arbitrary g ∈ G and m ∈ M with (g, m) 6∈ I, to
256 6 Properties of concept lattices

find some h ∈ G and some n ∈ M which satisfy the conditions specified in the
theorem.
With the help of the preliminary considerations and by applying the distribu-
tive law we obtain
^ _ _ ^
(g 00 , g 0 ) = (ϕ(n)00 , ϕ(n)0 ) = (k 00 , k 0 ).
ϕ n∈M \g 0 n∈M \g 0 k∈G\n0

Thus, there is some n ∈ M \ g 0 with m 6∈ ( k∈G\n0 k 00 )0 and consequently some

T

h ∈ k∈G\n0 k 00 with (h, m) 6∈ I. The elements elements h and n we obtain satisfy

T
the conditions specified. 
It is easy to give complete lattices which are distributive but not completely
distributive. However, these examples cannot be doubly founded, as the follow-
ing theorem shows:
Theorem 48 For a doubly founded concept lattice L := B(G, M, I), the following con-
ditions are equivalent:
1. L is distributive.
2. L is completely distributive.
3. From g %. m and g % n it follows that µm = µn.
4. From g %. m and h . m it follows that γg = γh.
5. L is isomorphic to a complete subdirect product of two-element lattices.
6. L is isomorphic to a complete sublattice of a powerset lattice.
7. L is isomorphic to the complete lattice of all order filters of an ordered set.
If (G, M, I) is reduced, the following conditions are equivalent to those stated so far:
8. Every proper premise is a singleton set.
9. g % m implies g %
. m, g . m implies g % . m,
g%. m and g %. n imply m = n and g % . m and h %
. m imply g = h.
Proof We show that 1 ⇒ 3 ⇔ 4 ⇒ 9 ⇒ 5 ⇒ 6 ⇒ 2 ⇒W8 ⇒ 7 ⇒ 1.
1 ⇒ 3: From g % . m and g % n we infer that γg is -irreducible and that γg ∧
µm ≤ γg∗ and γg ∧ µn ≤ γg∗ . Hence, µm ∨ µn 6≥ γg, because by the distributive
law we obtain γg ∧ (µm ∨ µn) = (γg ∧ µm) ∨ (γg ∧ µn) ≤ γg∗ ∨ γg∗ = γg∗ < γg.
Since we had presupposed µm and µn to be maximal 6≥ γg, we obtain µm =
µm ∨ µn = µn, q.e.d.
3 ⇒ 4: Assume that g % . m, h . m and γg 6= γh. Then h0 6⊆ g 0 , i.e., there exists
an attribute n with h I n, (g, n) ∈/ I. Then there is an attribute ñ with g % ñ and
n0 ⊆ ñ0 , however, because of 3, γm = γ ñ, from which we infer that n0 ⊆ m0 . This
is however contradictory to h ∈ n0 , h 6∈ m0 .
Correspondingly we show 4 ⇒ 3.
3, 4 ⇒ 9: If g % n and g is irreducible, then there exists an m with g % . m,
i.e., by 3, m = n and thus g %. n. This shows, together with the dual argument,
that a reduced context satisfying 3 and 4 allows only double arrows. The further
assertion, namely that the double arrow relation represents a bijection between
G and M , now follows immediately from 3 and 4, respectively.
6.2 Distributivity, modularity and semimodularity 257

9 ⇒ 5: According to Proposition 81, the subdirectly irreducible factors of L

exactly correspond to the concept lattices of the one-generated subcontexts of
(G, M, I) (we may presuppose (G, M, I) to be reduced). By 9, however, all one-
generated subcontexts are of the same, trivial form: they are composed of an
object and an attribute which are not related to each other by I. Hence, every
subdirect factor of B(G, M, I) is a two-element lattice.
5 ⇒ 6: The mapping (xt )t∈T 7→ {t ∈ T | xt = 1} is an isomorphism of the
T -fold power of the two-element lattice onto the complete lattice of all subsets of
T.
6 ⇒ 2: The fact that the powerset lattices are completely distributive is known
from elementary set theory (and is moreover easy to prove); hence, we have 2.
2 ⇒ 8: If (G, M, I) is a reduced context with a completely distributive concept
lattice, then [
B 00 = b00
b∈B

holds for every set B ⊆ M , since

^
m ∈ B 00 ⇔ µm ≥ {µn|n ∈ B}
^
⇔ µm = µm ∨ {µn|n ∈ B}
^
⇔ µm = {µm ∨ µn|n ∈ B}
⇔ µm = µm ∨ µn for some n ∈ B
(since µm is -irreducible)
V

⇔ m ∈ n00 for some n ∈ B.

This is obviously equivalent to condition 8.

8 ⇒ 7: The set M of (irreducible) attributes is ordered by m ≤ n :⇔ n0 ⊆ m0 .
From 8 it follows that the intents are precisely the order filters of M with respect
to this order.
7 ⇒ 1: Union and intersection of order filters again yield order filters, i.e., the
distributivity follows from that of the set operations. 
Every subdirect product3 of (completely) distributive lattices is (completely)
distributive.

Semimodularity and modularity

The lattice of the subspaces of a vector space, or more generally the lattice of the
submodules of a module, has a particular structural property: it satisfies the mod-
ular law. The lattices of normal subgroups of groups are also modular. A weaker

3 compare the footnote on Page 164.

258 6 Properties of concept lattices

property, semimodularity can be defined in different ways. Some of these defini-

tions make use of the neighborhood relation x ≺ y (compare Definition 3) and
therefore refer meaningfully only to lattices with certain finiteness requirements.
Definition 91 A lattice L is called semimodular if

x∧y ≺y ⇒x≺x∨y

holds for each two elements x, y. It is modular if it satisfies the following law for
all x, y and z:
x ≤ z ⇒ x ∨ (y ∧ z) = (x ∨ y) ∧ z,
and graded if there is a rank function r(x) assigning a natural number to each
element of L with

r(0) = 0 and x ≺ y ⇒ r(y) = r(x) + 1.

We say that L satisfies the weak condition of semimodularity if

x ∧ y ≺ x, y ⇒ x, y ≺ x ∨ y,

and the strong condition of semimodularity if the following is true:

If x, y and z are elements with x < z, y ∨ x = y ∨ z and y ∧ x = y ∧ z, then
there exists an element d ≤ y with y ∧ x < d and (x ∨ d) ∧ z = x. ♦
There is a characterization of modular lattices which is quite analogous to
that of the distributive lattices in Proposition 113. The proof can again be found
in Birkhoff’s book.
Proposition 114 A lattice is modular if and only if x ≤ y, a∧x = a∧y and a∨x = a∨y
always imply x = y.
A modular lattice which is not distributive contains elements x, y, z with x ∨ y =
x ∨ z = y ∨ z and x ∧ y = x ∧ z = y ∧ z, but x 6= y. 
The hierarchy of these lattice properties has been thoroughly examined and
is described in detail elsewhere. Here we note only the simplest statements:
Proposition 115 Every distributive lattice is modular. The modular law implies the
strong condition of semimodularity (and its dual), and from this follows semimodu-
larity.
Proof The first statement results from a comparison of Propositions 113 and 114.
The strong condition of semimodularity holds in every modular lattice already
because its premise is never satisfied: If x, y, z are elements with x < y and y ∨ x
= y ∨ z as well as y ∧ x = y ∧ z, the modular law yields

x ∨ (y ∧ z) = (x ∨ y) ∧ z.

However, the left-hand side of this equation equals x, the right-hand side equals
z, which is contradictory to x < z.
6.2 Distributivity, modularity and semimodularity 259

The fact that the strong condition of semimodularity implies semimodularity

can be seen as follows: Let y, x be elements with y ∧ x ≺ y, but x 6≺ y ∨ x. Then
there must be an element z with x < z < y ∨ x, and y ∧ x ≺ y forces y ∧ z = y ∧ x.
Condition 2 now yields an element d with y ∧ x < d ≤ y, from which, because of
y∧x ≺ y, we can immediately infer d = y. This implies, however, that x∨d = y∨x,
i.e., d cannot satisfy the required condition (x ∨ d) ∧ z = x. 
In a reduced context g . m is equivalent to (γg)∗ = γg ∧ µm. If the concept
lattice is semimodular, this implies γg ∨ µm = (µm)∗ , i.e., g % m. Hence, such
a context cannot contain “proper downward arrows”. In the modular case all
arrows must even be double arrows. However, these conditions are by no means
sufficient for semimodularity, to say nothing of modularity.
Theorem 48 states, among other things, that a doubly founded complete lat-
tice is distributive if and only if its standard context contains only double ar-
rows and exactly one in each row and column. This implies that the number of
-irreducible elements equals that of the -irreducibles, from which it follows
W V
that the standard context has the same number of objects as attributes, i.e. is
square in shape. A well-known theorem by R. P. Dilworth [94]W generalizes this
to (finite) modular lattices. They also have the same number of -irreducible as
-irreducible elements and thus a square-shaped standard context. And it was
V
already said above that they too have double arrows only.
One therefore could hope that the description by the arrow relations can be
generalized, e.g. to the effect that doubly founded modular lattices can be char-
acterized by a maximum of two double arrows per row and column. But this is
false, as the example of the two-dimensional dichotomous scale (see Fig. 8.2 on
p. 319) shows, the concept lattice of which is not modular.
A theorem by S. Doerfel shows that a generalization is possible, but it does not
go very far. For modular, doubly founded concept lattices, the 2-arrow condition
characterizes those in the smallest non-distributive equational class. We cite the
result without giving his elaborate proof.
Theorem 49 (Doerfel [99]) For a clarified formal context (G, M, I) of a modular,
doubly founded concept lattice the following conditions are equivalent:
1. B(G, M, I) belongs to the equational class generated by the five-element modular
non-distributive lattice,
2. ∀g∈G |{m ∈ M | g % . m}| ≤ 2,
3. ∀m∈M |{g ∈ G | g % . m}| ≤ 2.

In addition, Doerfel gives an example which shows that this theorem cannot be
generalized in an obvious way.
The next theorem provides a characterization of the abovementioned semi-
modularity properties in the language of formal contexts. Preparatory to it, we
need an abbreviation: If g is an object in a context (G, M, I), let

g• := {x ∈ G | γx < γg}.
260 6 Properties of concept lattices

If γg is -irreducible, then g• is precisely the extent of (γg)∗ .

W

Theorem 50 For a doubly founded concept lattice L := B(G, M, I), the following con-
ditions are equivalent:
1. L is semimodular.
2. L satisfies the strong condition of semimodularity.
3. The following exchange condition holds in (G, M, I):

g• ⊆ A, h ∈ (A ∪ {g})00 and h 6∈ A00 ⇒ g ∈ (A ∪ {h})00 .

4. From g . m, g . n, h I m and (h, n) ∈/ I it follows that there is an attribute p with

/ I, g I p and m0 ∩ n0 ⊆ p0 .
(h, p) ∈
If L is finite, the following conditions are equivalent to those stated so far:
5. L satisfies the weak condition of semimodularity.
6. L is graded and has a rank function with

r(x) + r(y) ≥ r(x ∧ y) + r(x ∨ y).

Note: Contrary to the formulation of the theorem, the proof only uses one of
the conditions of foundedness. If we add the other one, we can slightly improve
the result. For instance, in condition (4) (h, p) ∈
/ I can in this case be replaced
by h % p. This makes it possible to show that factor lattices of doubly founded
semimodular lattices are again semimodular.
Proof 1 ⇒ 2: Let y, x and z be elements with x < z and y ∨ x = y ∨ z and
y ∧ x = y ∧ z. First we show that there is an element d with y ∧ x ≺ d ≤ y. Because
of the foundedness thereWexists an element s minimal with respect to s ≤ y,
s 6≤ y ∧ x. s is necessarily -irreducible, and s∗ ≤ y ∧ x, i.e., s ∧ (y ∧ x) = s∗ ≺ s.
If now we set d := s ∨ (y ∧ x), we obtain y ∧ x ≺ d from the first condition of the
proposition. A repeated application of the condition yields x ≺ x ∨ d. x ∨ d ≤ z
would imply z ≥ x ∨ d ≥ d ≥ s and, because of y ≥ s, also y ∧ z = y ∧ x ≥ s,
which would be contradictory to the definition of s. Hence z ∧ (x ∨ d) = x. 2 ⇒
1 has already been proved in Proposition 115.
3 ⇒ 1: Let x, y be concepts with x ∧ y ≺ y and x = (A, B), and let furthermore
z be a concept between x and x ∨ y: x < z < x ∨ y. The foundedness W yields an
element s which is minimal with respect to s ≤ y, s 6≤ x, s then is -irreducible
and thus an object concept, i.e., s = γh for some h ∈ G. Because of s∗ ≤ x we
have h• ⊆ A, furthermore we have x ∨ y = x ∨ s. If we now choose an object g
which is contained in the extent of the concept z but which does not form part
of A, we obtain
g 6∈ (A ∪ h• )00 , g ∈ (A ∪ {h})00 ,
from which by (3) it follows that h ∈ (A∪{g})00 . This extent is however contained
in that of z, which results in a contradiction with γh 6≤ z.
1 ⇒ 3: Trivially, condition (3) is satisfied if g is reducible. Therefore, we can
restrict ourselves to the case (γg)∗ ≺ γg. The preconditions of (3) describe three
6.2 Distributivity, modularity and semimodularity 261

concepts, namely (A00 , A0 ), ((A ∪ {h})00 , (A ∪ {h})0 ) and ((A ∪ {g})00 , (A ∪ {g})0 )
with A00 ⊂ (A ∪ {h})00 ⊆ (A ∪ {g})00 . Since g• ⊆ A is true, γg ∧ (A00 , A0 ) = γg ∗ ,
and thus, because of semimodularity,

((A ∪ {g})00 , (A ∪ {g})0 )

is an upper neighbor of (A00 , A0 ), which forces

((A ∪ {g})00 , (A ∪ {g})0 ) = ((A ∪ {h})00 , (A ∪ {h})0 ).

1 ⇒ 4: Let g, h, m, n be as specified. If we choose x minimal with respect to

x ≤ γh, x 6≤ µn, then x is an object concept, satisfying the same preconditions as
γh. Furthermore, for every attribute p with x 6≤ µp we have (h, p) ∈ / I. Hence, we
may assume without loss of generality that x = γh, i.e., in particular that γh is
-irreducible and that γh∗ ≤ µm∧µn. Because of (1), γh∨(µm∧µn) is an upper
W
neighbor of µm ∧ µn which is ≤ µm. Also because of (1) y := γg ∨ (µm ∧ µn) is
an upper neighbor of µm ∨ µn, but because of γg 6≤ µm it is different from γh ∨
(µm ∧ µn). Hence, there must be a distinguishing attribute p which is contained
in the intent of the concept y but which h does not have. This attribute satisfies
the conditions specified.
4 ⇒ 1: First, we convince ourselves that it suffices to prove condition (1) for
object concepts y: If y is arbitrary and x∧y ≺ y is a lower neighbor of y, we find an
object concept γg ≤ y which is minimal with respect to γg 6≤ x ∧ y, which implies
that γg ∗ = γg ∧ (x ∧ y) = γg ∧ x is a lower neighbor of γg. If we now may apply
(1), we obtain x ≺ x ∨ γg = x ∨ γg ∨ (x ∧ y) = x ∨ y, since x ∧ y < γg ∨ (x ∧ y) ≤ y,
i.e., γg ∨ (x ∧ y) = y.
Now we assume that there are an object g and an x such that x ∧ γg = γg ∗
is a lower neighbor of γg but x ∨ γg is not an upper neighbor of x, i.e., that x <
z < x ∨ γg for some z. Then there exist an object h in the extent of z which is not
contained in the extent of x, an attribute n which is contained in the intent of x
but not in the intent of z and an attribute m which is contained in the intent of z
but does not apply to g. Thereby the prerequisites of (4) are all satisfied and we
may conclude that there has to be an attribute p which has m0 ∩ n0 in its extent,
i.e., which satisfies x ≤ µp, with g I p and (h, p) ∈ / I. However, from g I p it follows
that µp ≥ x ∨ γg ≥ z and from (h, p) ∈ / I it follows that µp 6≥ z, a contradiction!
1 ⇒ 5 is trivial.
5 ⇒ 6: Let n be the length of a maximal chain in L. We define a function r on
L by r(1) := n and, for x 6= 1,

r(x) := max{r(y) | x ≺ y} − 1.

If r were not a rank function, then there would have to be elements x ≺ y with
r(x) + 1 6= r(y), and, of all such examples, we could choose one with maximal
y, furthermore, by the definition of r, there would be a further upper neighbor
z of x with r(z) = r(x) + 1. By the condition of semimodularity y ∨ z would be
262 6 Properties of concept lattices

an upper neighbor of y and of z, by the condition of maximality; then we would

have r(y) = r(y ∨ z) − 1 = r(z), a contradiction!
In order to prove the rank inequality claimed, we again derive a contradiction
from the assumption of a counter-example: Let a, b, x, y be elements with a =
x ∧ y and b = x ∨ y which do not satisfy the inequality. Let these elements be
chosen such that a is maximal among all counter-examples and that b is minimal
among all counter-examples with the smallest element a. If now y is an upper
neighbor of x ∧ y, then we choose an upper neighbor x of x ∧ y lying below x.
Since y cannot lie below x (otherwise the inequality would be satisfied), x 6= y,
and y := y ∨ x is an upper neighbor of y with x ∧ y = x and x ∨ y = x ∨ y. For
the elements x and y the rank inequality holds by assumption and thus also for
x and y. If y is not an upper neighbor of x, we can find an upper neighbor y of
x ∧ y which lies below y. The application of the rank inequality to the elements
x and y and to y and x ∨ y yields the statement.
6 ⇒ 1 is again trivial. 

In the atomistic case we have g• = ∅ for all g ∈ G. Thus, the exchange condition
simplifies to yield the known form

h ∈ (A ∪ {g})00 , h 6∈ A00 ⇒ g ∈ (A ∪ {h})00 .

6.3 Semidistributivity and local distributivity

Definition 92 A complete lattice L is called

• semidistributive, if the following laws hold for all x, y, z ∈ L:

(SD∨ ) x ∨ y = x ∨ z ⇒ x ∨ y = x ∨ (y ∧ z) = x ∨ z
(SD∧ ) x ∧ y = x ∧ z ⇒ x ∧ y = x ∧ (y ∨ z) = x ∧ z.

If L satisfies (SD∨ ), L is called join-semidistributive, dually, a complete lat-

tice satisfying (SD∧ ) is called meet-semidistributive.
• locally distributive or join-distributive, if L is semimodular and every mod-
ular sublattice is distributive. Lattices satisfying the dual condition are called
meet-distributive. ♦
A representation W of a lattice element a as a supremum a = X is called irre-
W
dundant if a 6= (X \ {x}) holds for every x ∈ X. Obviously, the elements
of X Wmust be pairwise incomparable. Here, we will mainly deal with irredun-
dant -representations in chain-finite lattices. This constitutesW a simplification
in many respects, because in this kind of lattice from each -representation we
can choose a finite and then also anWirredundant -representation.
W
Special attention is given to the -representations through -irreducible ele-
W
ments. If a has exactly one such representation, we say that a has a unique irre-
6.3 Semidistributivity and local distributivity 263

dundant -representation. The addition “through irreducibles” is often omit-

W
ted; it is however implied.
An irredundant -representation W a = X is called canonical if, for every
W W
irredundant representation a = Y and for every x ∈ X, there exists some
y ∈ Y with x ≤ y. If an element ofW finite lattice has a canonical representation,
a
the latter necessarily consists of -irreducible elements. Note, however, that the
refinement property is required with respect to all irredundant representations.
AsW is well known, in a doubly founded lattice every element a is the supremum
of -irreducible elements: a = {x ∈ J(L) | x ≤ a}. An element e ∈ J(L) is an
W
extremal point of a if e is indispensable, i.e., if
_
a 6= {x ∈ J(L) \ {e} | x ≤ a}.

The extremal points are part W of every -representation of a through irre-

W
ducibles. A base point of a is a -irreducible element b ≤ a with
_
b 6≤ {x ∈ J(L) | x ≤ a, b 6≤ x}.

In a concept lattice the extremal points of a concept (A, A0 ) are precisely the
object concepts γg with g ∈ A but

g 6∈ (A \ {h | g 0 = h0 })00 .

Therefore, such an object is called an extremal point of the extent A. Corre-

spondingly, g is a base point of the extent A if g ∈ A but

g 6∈ (A \ {h | g 0 ⊇ h0 })00 .

Theorem 51 For a doubly founded concept lattice L = B(G, M, I), the following state-
ments are equivalent:
1. L satisfies (SD∨ ).
2. For all g, h ∈ G and all m ∈ M we have:

. m and h %
g% . m imply g 0 = h0 .

If L is finite, or, more generally, if L satisfies the additional condition

(*) if x < a, then there exists a lower neighbor u of a with x ≤ u ≺ a,
then the following conditions are equivalent to the above-stated ones:
3. Every element of L has a canonical -representation.
W
4. Every extent is the closure of its base points.
The condition (SD∧ ) can be characterized dually.
Proof 1 ⇔ 2: We assume without loss of generality that (G, M, I) is the context
of a concept lattice satisfying (SD∨ ). From g %
. m and h %
. m we infer that
264 6 Properties of concept lattices

µm ∨ γg = µm∗ = µm ∨ γh

and then with the help of semidistributivity

µm∗ = µm ∨ (γg ∧ γh).

Now γg ∗ as well as γh∗ is less than or equal to µm. The above equation can only
be true if neither γg ∧ γh ≤ γg ∗ nor ≤ γh∗ . This forces γg = γh.
Now we conversely assume that (G, M, I) satisfies the above-stated condition
for the arrow relations and show that B(G, M, I) is join-semidistributive. Let
furthermore x, y and z be elements of B(G, M, I) with

x ∨ y = x ∨ z > x ∨ (y ∧ z).

Then there exists an element t maximal with respect to

t ≥ x ∨ (y ∧ z), t 6≥ x ∨ y.

t is -irreducible, i.e., t = µm for some m ∈ M , and µm∗ ≥ x ∨ y. Now, y 6≤ µm,

V
since
y ∨ µm ≥ y ∨ x ∨ (y ∧ z) = x ∨ y,
and we obtain
y ∨ µm = µm∗ . 
Hence there is a concept s which is minimal with respect to s ≤ y, s 6≤ µm. s is
-irreducible; consequently it is an object concept s = γg for some g ∈ G and
W
γg ∗ ≤ µm, i.e., g %. m.
Likewise, we can find some h ∈ G with h % . m and γh ≤ z. The presuppo-
sition forces γg = γh, from which in turn it follows that γg ≤ z, i.e., γg ≤ y ∧ z
≤ x ∨ (y ∧ z) ≤ µm. Contradiction!
1 ⇒ 4 for lattices satisfying the additional condition: Let a be a concept. For
every lower neighbor u of a there exists, because of the foundedness, an element
ū ≤ a which is minimal with respect to ū 6≤ u. This element is uniquely deter-
mined by u, since if x is an arbitrary element satisfying x ≤ a, x 6≤ u, we have

u ∨ ū = a = u ∨ x

and by application of (SD∨ ) we obtain

u ∨ (ū ∧ x) = a,

which, because of the minimality of ū, immediately yields ū ≤ x. Hence for x ≤ a

we have
x 6≤ u ⇔ ū ≤ x.
ū is -irreducible, and consequently there exists an object g with ū = γg. We
W
show that g is a base point of a: If A is the extent of a and h ∈ A is arbitrary then
6.3 Semidistributivity and local distributivity 265

we have
h0 ⊆ g 0 ⇔ γh ≥ γg ⇔ γh ≥ ū ⇔ γh 6≤ u.
Hence, in this case, A \ {h | g 0 ⊇ h0 } is entirely contained in the extent of u, from
which follows the desired base point property:

g 6∈ (A \ {h | g 0 ⊇ h0 })00 .

Now assume that _

x := {ū | u ≺ a}.
We claim that x = a. If this were not the case, because of the additional condition
there would exist a lower neighbor u of a with x ≤ u ≺ a, and ū 6≤ x would yield
a contradiction.
4 ⇒ 3: If every extent is the closure of its base points, i.e., for every lattice
element a we have that
_
a = {b | b is base point of a},

W this representation is canonical. It is furthermore irredundant and, if a =

then
Y is anW arbitrary irredundant representation and b is a base point of a, then
from b ≤ Y and the fact that J(L) is -dense it follows immediately that there
W
has to be some y ∈ Y with b ≤ y.
3 ⇒ 1: We consider an element v and a canonical -representation v = X.
W W
From
v = v0 ∨ y = v0 ∨ z
it follows that y ≥ x for all x ∈ X with x 6≤ v0 and z ≥ x for all x ∈ X with x 6≤ v0 .
Together this forces
v = v0 ∨ (y ∧ z).
Examples of lattices satisfying the conditions of Theorem 39 can easily be ob-
tained by means of the technique of local doubling (5.3). If L is a doubly founded
semidistributive lattice and C ⊆ L is a convex subset with a smallest element,
then L[C] is also semidistributive, as can be seen by the arrow relations. The
Tamari lattice (cf. Figures 5.8, 5.9, p. 236) satisfies (SD∨ ) and (SD∧ ).
Theorem 52 For a doubly founded concept lattice L := B(G, M, I) the following state-
ments are equivalent:
1. If g and h are irreducible objects then
V g % m and h % m imply g = h .
0 0

2. L has a neighborhood-preserving -embedding into a powerset lattice.

3. Every extent is the closure of its extremal points.
4. For every concept extent A and for all g, h ∈ G \ A with g 0 6= h0 we have:

g ∈ (A ∪ {h})00 implies h 6∈ (A ∪ {g})00 (Anti-exchange Axiom).

If L is chain-finite, then the following condition is equivalent to those we have mentioned

so far:
266 6 Properties of concept lattices

5. L is meet-distributive.
6. Every element has a unique irredundant -representation.
W

Proof 1 ⇒ 4: If A is an extent and g 6∈ A, we find an attribute m with A ⊆ m0 and

g % m. From h I m it would follow that g 6∈ (A ∪ {h})00 ; hence (h, m) ∈ / I would
hold and we would find an attribute n with h % n and m0 ⊆ n0 . m0 = n0 would
imply g 0 = h0 already because of (1), hence g ∈ n0 ; this would mean, however,
that A ∪ {g} ⊆ n0 and consequently h ∈ / (A ∪ {g})00 .
4 ⇒ 2: We pass on to the concept lattice of the clarified context and show
that the closure system of the extents is embedded in the powerset of G in such a
way that it preserves neighborhoods. Let A1 , A2 be extents with A1 ⊂ A2 , and let
(A1 , A01 ) be a lower neighbor of (A2 , A02 ). If now g, h ∈ A2 \ A1 , then, because of
the neighborhood (A1 ∪ {g})00 = A2 = (A1 ∪ {h})00 , which, by the Anti-exchange
Axiom, yields g = h.
2 ⇒ 3: We begin this part of the proof with three preliminary considerations:
Every lattice having a neighborhood-preserving ∧-embedding

ϕ : L → P(X)

into a powerset lattice is dually semimodular, since from a ≺ a ∨ b it follows that

ϕ(a) and ϕ(a ∨ b) only differ in one element. This transfers to ϕ(a ∧ b) and ϕ(b),
which means that they also must be neighbors, which implies a ∧ b ≺ b.
Secondly we show that a doubly founded lattice satisfying (2) also satisfies
the additional condition (*)Vin Theorem 51. If a < b, then because L is doubly
founded, there exists some -irreducible element t ∈ L with

a ≤ t, b 6≤ t, b ≤ t∗ ,

and with t ≺ t∗ because of the dual semimodularity it follows that

a ≤ b ∧ t ≺ b.

Hence every element of L is either -irreducible or is the supremum of its lower

W
neighbors.
For the third preliminary consideration we take an arbitrary element x ∈ X.
Among all elements of L whose image contains x, there is a smallest one, namely
^
vx := {v ∈ L | x ∈ ϕ(v)}.

vx is -irreducible (or the smallest element of L), since there can only be one
W
lower neighbor u of v because the image

ϕ(u) = ϕ(vx ) \ {x}

of the latter is uniquely determined. If w is an arbitrary -irreducible element

W
with x ∈ ϕ(w), x 6∈ ϕ(w∗ ), then w = w∗ ∨ vx , i.e., w = vx .
6.3 Semidistributivity and local distributivity 267

Now we consider an arbitrary element a of L. If the extent a were not the

closure of its extremal points, then there would be a lower neighbor u ≺ a, the
extent of which would contain all extremalWpoints of a. ϕ(a) and ϕ(u) only differ
by one element, let us say x of X. If w is a -irreducible element with w ∨ u = a,
then u ∧ w is a lower neighbor of w, i.e., w∗ , and we have x = ϕ(w) \ ϕ(w∗ ), from
which it follows that w = vx . Hence vx is an extremal point of a that does not lie
below u. Contradiction!
3 ⇒ 1: If m is an irreducible attribute and A is the extent of µm∗ , then by
(3) there is an extremal point h of A with h 6∈ m0 . A− := A \ {x | h0 = x0 }
then is an extent containing m0 . Now consider an arbitrary object g ∈ A \ m0 . We
have (m0 ∪ {g})00 = A, and therefore A− cannot contain m0 ∪ {g}. Consequently,
g ∈ {x | h0 = x0 }, i.e., g 0 = h0 . Hence µm and µm∗ differ only by one single object
concept.
3⇒W 6: If an extent is the closure of its extremal points, this is its only irredun-
dant -representation.
3 ⇐ 6 for chain-finite lattices: In a chain-finite lattice every -representation
W
contains an irredundant one, i.e., the uniqueness implies that every extent A has
a smallest subset E with E 00 = A. This must however consist of extremal points,
since otherwise there would be some e ∈ E with e ∈ (A \ {e})00 , and A \ {e}
would be a generating set of A not containing E.
2 ⇒ 5: We have already shown above that from (2) follow dual semimodular-
ity and the additional condition from Theorem 51. It remains to be shown that ev-
ery modular sublattice is distributive. If this were not the case, by Proposition 114
there would be elements x, y, z with x ∨ y = x ∨ y = y ∨ z, x ∧ y = x ∧ z = y ∧ z,
but x 6= y. Choose a lower neighbor u of x ∨ y with z ≤ u. ϕ(x ∨ y) \ ϕ(u) con-
sists of exactly one element, let us say g. Since ϕ(x) 6⊆ ϕ(u) and ϕ(y) 6⊆ ϕ(u),
g ∈ ϕ(x) ∩ ϕ(y) = ϕ(x ∧ y), from which, because of x ∧ y ≤ z follows the contra-
diction g ∈ ϕ(z) ⊆ ϕ(u).
5 ⇒ 1 for chain-finite lattices: Because of dual semimodularity g % m implies
g . m. Hence we can presuppose g % . m, h %. m and we have to show γg = γh.
The element γg ∨ γh is ≤ µm∗ , but not ≤ µm. Therefore if γg 6= γh we can find
three different lower neighbors a0 , b0 , c0 of the element γg ∨ γh with a0 ≥ γg,
b0 ≥ γh, b0 6≥ γg and c0 ≤ µm, and furthermore a descending chain a0  a1 
. . .  an = γg of neighboring elements. Because of meet-distributivity the meets
b1 := a0 ∧b0 and c1 := a0 ∧c0 are distinct from each other. Furthermore, they both
cannot be equal to a1 , since otherwise it would follow that γg ≤ µm or γg ≤ b0 . If
we continue this argument, we get that in each case the elements ai , bi and ci are
different lower neighbors of ai−1 for i ∈ {1, . . . , n} and that therefore an cannot
be -irreducible; this however is a contradiction because an = γg.
W

The first condition of Theorem 52 obviously implies the second condition of
Theorem 51. Hence finite meet-distributive lattices also satisfy (SD∨ ).
In a context in which there are no “proper upward arrows” (which means that
g % m always implies g % . m), the converse is also true. This holds in particular
for atomistic contexts, in which from (g, m) 6∈ I it already follows that g . m.
268 6 Properties of concept lattices

e
@
@
@
g % m, g . n ⇒ g .
%m semi convex @ g . m, h % m ⇒ g .
%m
e e e
@ @ @
dually
@ SD∧ @ semimodular
@ @ @
semimodular SD∨ @
e e @ e @e @e
HH @ join-

 H
meet-
@  @ @
 H @ distributive
distributive
@ @ @ HH
 @ @ @ H@
e @e @e @e @e

b
H H
b @
b @b @
@
@ @ b b@ b
b @b
@
b@ b b @
b @
@ @ @
b
B b@e b
@b e
@ @ @
b
A@b @ modular @
b b b
@

A@
b A@ @ b @ @@ b
b
b b
AA@ b@ b
@
A @
b
A@b @ b
@A b @
A @ @b @ b A@ @ b@ A
@
@b b
@ b b

@ b
@A b
A
@
@ A


b b@ A b @ @
@ b @A b

AA b
@
b @ @
@@
b b @ A
A
@b @b @b
@
@b @ A
@A b b @b
distributive@A e @b @ @b
@ @ @b
@@b
Figure 6.2 Generalizations of the Distributive Law.

Figure 6.2 shows the implications between the above mentioned lattice prop-
erties (for doubly founded complete lattices). Two of the attributes still lack an
explanation: According to A. Day, a lattice is semiconvex if it satisfies the follow-
ing condition:

(C) x ∧ y = x ∧ z, x∨y =x∨z ⇒ x ≤ z.

B stands for the property of being a bounded homomorphic image of a free lat-
tice. This property has a simple characterization in the language of contexts: it
is equivalent to the fact that the objects and attributes of the (reduced) context
can be ordered in such a way that every % . is on the diagonal, every % below the
diagonal and every . above the diagonal.
The convex-ordinal scales C(P,≤) (see Section 1.4), at least if (P, ≤) is finite,
are doubly founded. The extents are the convex subsets of (P, ≤), the extremal
points are the maximal and minimal elements of such subsets. Hence Condition
6.3 Semidistributivity and local distributivity 269

(3) of Theorem 52 is satisfied and therefore finite convex-ordinal scales are meet-
distributive.

d
1
d
##
#d
d#
# cc
#
#d
d d## c d
cc
#
#J##Jc c ##J
# d#J d2Jcdc #d
cc Jd
d #J
# c # cc d #J #
#
# # c J# # #
dc # # Jd
c
#J Jc
# #
d# J d# # d# c J d## 3
c c c
d
c
#c# c## #
c
J c
# c # c c J ###
J d#
c # J #
cJ d#
4 cc d#
c ccJ d## 6
7 5 #
cJ #
ccJ d#

Figure 6.3 The additively saturated subsets of {1, . . . , 7}.

The Anti-exchange Axiom holds in the closure system of the convex sets of an
arbitrary metric space. But it also holds in other connections, for example, if we
can assign a weight wt(g) to every object g, such that wt(g) = wt(h) only if g 0 = h0
and such that
g 6∈ A00 , g ∈ (A ∪ {h})00 ⇒ wt(g) ≥ wt(h).
The last condition can be interpreted as saying that the weight of g must be at
least as big as that of h if g is generated by means of h. A simple example in
this connection is presented in Figure 6.3. If G ⊆ N is an arbitrary set, then a
subset T ⊆ G will be called additively saturated if from a, b ∈ T, a + b ∈ G
it already follows that a + b ∈ T . The additively saturated subsets of G form a
closure system which with wt(g) := g obviously satisfies the above-mentioned
condition. Hence the corresponding lattice is meet-distributive. Figure 6.3 shows
an example with G := {1, 2, . . . , 7}.
270 6 Properties of concept lattices

6.4 Dimension

Order dimension

Definition 93 An ordered set (P, ≤) has order dimension

dim(P, ≤) = n

if and only if it can be embedded in a direct product of n chains4 and n is the

smallest number for which this is possible. ♦
We will show that the order dimension can also be well described in the lan-
guage of Formal Concept Analysis. For this purpose we start with some simple
observations, on which we will base our statements and which at the same time
are formulated in such a way that they can also be used for possible variations
of the notion of order dimension. An example of such a variation is that of the k-
dimension: The k-dimension dimk (P, ≤) of an ordered set (P, ≤) is the smallest
number of chains of cardinality k in whose product it can be order-embedded.
What is usually examined are embeddings of arbitrary ordered sets. We shall
concentrate on embeddings of concept lattices. This is not a serious restriction,
however: The Dedekind Completion Theorem (Theorem 4, p. 46) shows that
(P, ≤) can be embedded in a complete lattice if and only if B(P, P, ≤) can also
be embedded in this lattice. Therefore, we have the following theorem:
Theorem 53

dim(P, ≤) = dim B(P, P, ≤)

dimk (P, ≤) = dimk B(P, P, ≤)


As a direct corollary of Proposition 39 we obtain:
Proposition 116 There is an order embedding of B(G, M, I) in a product

×(L , ≤)
t∈T
t

if and only if there are pairs of maps (αt , βt ), t ∈ T , with the following properties:
1. αt : G → (Lt , ≤), βt : M → (Lt , ≤),
2. (g, m) ∈ I ⇒ αt g ≤ βt m for all t ∈ T ,
3. (g, m) ∈
6 I ⇒ αt g 6≤ βt m for some t ∈ T .

We can express this fact differently by replacing the maps αt and βt by rela-
tions Jt with (g, m) ∈ Jt : ⇔ αt g ≤ βt m:
4 Meaning an order embedding according to Definition 6.
6.4 Dimension 271

Proposition 117 There is an order-embedding of B(G, M, I) in a product

×(L , ≤)
t∈T
t

if and only if there exist contexts (G, M, Jt ) and order-embeddings of B(G, M, Jt ) in

(Lt , ≤) for t ∈ T such that \
I= Jt .
t∈T


In the case of a doubly founded context,

T Proposition 117 can be further im-
proved by weakening the condition I = t∈T Jt : The statement remains correct
if instead we only presuppose that
\ \
I⊆ Jt and {(g, m) | g . m or g % m} ∩ Jt = ∅.
t∈T t∈T

This follows from Proposition 59.

Ferrers dimension

Definition 94 A Ferrers relation is a relation F ⊆ G × M with

(g, m) ∈ F, (h, n) ∈ F, (g, n) 6∈ F ⇒ (h, m) ∈ F.

The Ferrers dimension fdim(G, M, I) of a context (G,TM, I) is the smallest num-

ber of Ferrers relations Ft ⊆ G × M, t ∈ T , with I = t∈T Ft . ♦
If we imagine (G, M, F ) as a cross table, it is easy to visualize the Ferrers con-
dition: The definition excludes the subcontext × × , which does not occur if and
only if the table can be brought into a stair-shaped form by rearranging the rows
and columns. This is also the basis of the following proposition:
Proposition 118 F ⊆ G × M is a Ferrers relation if and only if B(G, M, F ) is a chain.
Proof Let F be a Ferrers relation and let (A1 , B1 ), (A2 , B2 ) be two concepts of
B(G, M, F ). If (A1 , B1 ) 6≤ (A2 , B2 ), then there are an object g ∈ A1 and an at-
tribute n ∈ B2 with (g, n) 6∈ F . For every m ∈ B1 we have (g, m) ∈ F and for
every h ∈ A2 we have (h, n) ∈ F . Hence from the Ferrers condition it follows
that (h, m) ∈ F for all h ∈ A2 , m ∈ B1 and thus (A2 , B2 ) ≤ (A1 , B1 ). This means
that any two concepts of B(G, M, F ) are comparable and B(G, M, F ) thus is a
chain. The reverse direction is even easier. 
The determination of the Ferrers dimension is a task which is generally dif-
ficult to solve (since for k ≥ 3 it is N P-complete to decide if fdim(K) ≤ k).
272 6 Properties of concept lattices

Nevertheless, in the case of small contexts it can be carried out by hand. In this
connection it is convenient to make use of the fact that the complement of a Fer-
rers relation is again a Ferrers relation. Hence the Ferrers dimension of (G, M, I)
is also equal to the smallest number of Ferrers S relations Ft covering the empty
cells of the cross table, i.e., with G × M \ I = t∈T Ft . It is, however, not always
possible to choose this covering to be disjoint.
Theorem 54 The Ferrers dimension of (G, M, I) is equal to the order dimension of the
concept lattice B(G, M, I):

dim B(G, M, I) = fdim(G, M, I).

The order dimension of an ordered set (P, ≤) is equal to the Ferrers dimension of
(P, P, ≤):
dim(P, ≤) = fdim(P, P, ≤).
Proof This follows immediately from the Propositions 117 and 118, because, ob-
viously, a complete lattice can be embedded in a chain if and only if it is a chain
itself. 

× × 1 × 1 1 1 (3, 0, 0, 0) (3, 0, 3, 3)
2 × × 4 × 1 1 (2, 1, 0, 1) (3, 3, 0, 3)
2 3 × × 2 × 1 (1, 2, 1, 0) (2, 2, 1, 3)
2 3 3 × × 3 × (0, 2, 3, 0) (3, 2, 3, 0)
× 3 3 4 × × 1 (1, 0, 2, 1) (2, 1, 3, 2)
2 × 2 2 2 × × (0, 3, 0, 0) (1, 3, 2, 2)
× 3 × 4 4 4 × (0, 0, 1, 3) (0, 3, 3, 3)

Figure 6.4 Point-line context of the projective plane PG(2,2), with Ferrers relations. The
Ferrers dimension, and thus the order dimension of the plane, is 4; a covering of the 28
empty boxes with less than four Ferrers relations is impossible, since each Ferrers relation
in this example can have eight elements at most. On the right is an embedding of PG(2,2)
into a product of four chains; the first column gives the images of the points, the second
those of the lines.

k-dimension

Theorem 54 can be strengthened in several respects. We can include the lengths

of the chains involved and we can examine whether it would not suffice to cover
the arrow relations. Both are possible and the two possibilities can even be com-
bined. For this purpose, we first define the length of a Ferrers relation:
6.4 Dimension 273

Definition 95 The length of a Ferrers relation F ⊆ G × M is the length of the

concept lattice B(G, M, F ). By fdimk (G, M, I) let us denote theTsmallest number
of Ferrers relations Ft ⊆ G × M , t ∈ T , of length ≤ k with I = t∈T Ft . A Ferrers
relation is k-step, if k = |{g F | g ∈ G}|. ♦

If we imagine a k-step Ferrers relation represented as a table and arranged

in the form of stairs, then these stairs really have k steps. The number of steps
is equal to the length if full rows and full columns are counted separately. The
following trivial observation is quite helpful:

Proposition 119 The complement G × M \ F of a k-step Ferrers relation F ⊆ G × M

is a Ferrers relation of length k.
fdimk (G, M, I) is equal to the smallest number of at most (k − 1)-step Ferrers rela-
tions whose union is G × M \ I.

Now Theorem 54 can be transferred without a new proof:

Theorem 55

dimk B(G, M, I) = fdimk (G, M, I)

dimk (P, ≤) = fdimk (P, P, ≤).



× × 1 × 1 1 1
2 × × 2 × 1 1
3 3 × × 3 × 3
3 3 4 × × 5 ×
× 4 4 4 × × 4
2 × 2 2 2 × ×
× 5 × 5 5 5 ×

Figure 6.5 The 3-dimension of PG(2,2) is 5. In arithmetic terms a covering with four Fer-
rers relations of the length 3 would be conceivable, however, such a covering does not
exist. The 2-dimension is 7.

Of particular interest is the 2-dimension, i.e., the smallest number of rectan-

gles by which the complement of I can be covered. The direct products of two-
element chains are precisely the powerset lattices. Hence the 2-dimension of a
complete lattice is also the smallest possible size of a set representation in the
sense of Definition 39 on p. 88. We show a connection with another form of rep-
resentation:
274 6 Properties of concept lattices

Set representations

Definition 96 A set representation of (G, M, I) on a set T is a pair of maps α :

G → P(T ), β : M → P(T ) assigning a subset of T to each object and to each
attribute, in such a way that:

g I m ⇔ αg ∩ βm 6= ∅.

We speak of a complementary set representation if the condition

g I m ⇔ αg ∩ βm = ∅

is satisfied. ♦
In Section 1.4 we had defined the contexts for free distributive lattices by
means of set representations.

Proposition 120 The following statements are equivalent:

1. (G, M, I) has a complementary set representation in T .
2. B(G, M, I) has a set representation in T .
3. dim2 B(G, M, I) ≤ |T |.

Proof 1 ⇒ 3: If (α, β) is a complementary set representation in T , then the re-

lations

Ft := {g | t ∈ αg} × {m | t ∈ βm}
= {(g, m) | t ∈ αg ∩ βm}
⊆ G×M \I

are 1-step Ferrers relations (or empty) with t∈T Ft = G × M \ I, and by

S
Theorem 55 it follows that dim2 B(G, M, I) ≤ |T |.
3 ⇒ 2: The products of 2-element chains are precisely the powerset lattices
and, therefore, dim2 B(G, M, I) ≤ n is equivalent to the fact that B(G, M, I)
can be order-embedded in the powerset lattice of an n-element set.
2 ⇒ 1: If ϕ is an order embedding in a powerset lattice P(T ), then by αg :=
ϕγg, βm := T \ ϕµm we obtain a complementary set representation of
(G, M, I), since

g I m ⇔ γg ≤ µm ⇔ ϕγg ⊆ ϕµm ⇔ ϕγg ∩ (T \ ϕµm) = ∅. 

We know already from the observations following Proposition 117 that in or-
der to determine the Ferrers dimension of a doubly founded context B(G, M, I),
it suffices to cover those pairs (g, m) of the complement of I for which g . m or
g % m holds. In fact, we can even restrict ourselves to the double arrows, since
they play the role for the Ferrers dimension which the critical pairs known from
order theory play for the order dimension.
6.4 Dimension 275

Theorem 56 A doubly founded context (G, M, I) has Ferrers dimension ≤ n if and

only if there are n Ferrers relations Ft ⊆ G × M \ I whose union contains all pairs
(g, m) with g % . m.
If, in this connection, all Ft are at most (k − 1)-step, then

fdimk (G, M, I) ≤ n

also holds.
Proof First we describe the possibility of suitably extending a given Ferrers re-
lation F . The basic idea is that the Ferrers condition is not affected if we double
some row or column of the context. Formally, this can be described as follows:
For a Ferrers relation F ⊆ G × M and objects g, h ∈ G,

F ∪ {(g, m) | (h, m) ∈ F }

is also a Ferrers relation. Moreover, if F ∩ I = ∅, then from g 0 ⊆ h0 it follows

that F ∪ {(g, m) | (h, m) ∈ F } is also disjoint to I. The number of steps does
not increase. The corresponding is true when we copy column n into column m
(m, n ∈ M ), i.e., when we pass from F to

F ∪ {(g, m) | (g, n) ∈ F }.

If, therefore, we define for a Ferrers relation F ⊆ G × M \ I

[
Fe := F ∪ {(g, m) | (h, m) ∈ F }
g 0 ⊆h0
[
and thus F := Fe ∪ {(g, m) | (g, n) ∈ F },
m0 ⊆n0

we obtain again a Ferrers relation F ⊆ G × M \ I. If F is at most k-step, then F

is at most k-step.
Now, assume that for t ∈ T Ferrers relations Ft ⊆ G × M \ I are given which
together cover all double arrows. We claim that the relations F t , t ∈ T , then
completely cover G × M \ I:
If (g, m) 6∈ I, then because of doubly foundedness, there is an attribute n with
g % n and m0 ⊆ n0 , and furthermore an object h with h % . n and g 0 ⊆ h0 . If
(h, n) ∈ Ft , then (g, n) ∈ Ft and (g, m) ∈ F t .
e 
From Definition 96 it follows immediately that a set representation of (G, M, I)
is a complementary set representation of the complementary context (G, M, G ×
M \ I), and vice versa. The determination of the set dimension of K, i.e., the
smallest possible cardinality of a set representation, is therefore equivalent to
the determination of the 2-dimension of Kc (and difficult, since it is also N P-
complete). According to the above-stated results, the task consists in filling up
the relation I with as few as possible 1-step Ferrers relations, i.e., with as few as
possible “rectangular” subrelations. For such rectangleses (A, B) with A×B ⊆ I
276 6 Properties of concept lattices

we had earlier introduced the term preconcept of the context (G, M, I). Hence the
set dimension is equal to the smallest number of preconcepts (more precisely:
sets A × B, where (A, B) is a preconcept) whose union fills up I. Since, however,
every preconcept can be extended to a concept, the set dimension is also equal
to the smallest number of concepts whose union is I. We give an example:

Example We want to find out whether the digits of the seven-segment display

can be represented as the unions of less than seven parts.

a b c d e f g
× ×⊗×××
⊗×
a ××× ⊗×

d f ××⊗ ××
b ⊗ × ××
e g ×××× ×
c ××××× ×
⊗ ××
×××××××
×××× ××

Figure 6.6 The seven-segment display and the context of the digits. The crosses sur-
rounded by circles mark a blocking set.

For this purpose we consider the context in Figure 6.6, whose objects are those
digits and whose attributes are the segments of the display; the incidence is ex-
plained in the obvious way. The intents naturally correspond to partial figures of
the display; a set of concepts fills out I if and only if every object intent can be rep-
resented as the union of (concept) intents from this set. Thus, the task of getting
along with as few partial figures as possible is equivalent to the determination
of the set dimension. In fact, six intents are sufficient, namely

The fact that a smaller number cannot suffice is evidenced by the “blocking
set” of incidences marked in Figure 6.6: No two of these crosses can belong to
a common preconcept. >
6.4 Dimension 277

Boolean matrix factorization

The procedure demonstrated in the example above can be elegantly formu-

lated with the Boolean matrix product introduced in Definition 33 on p. 44. The
sources for the results presented here are given in Section 6.5.
In (exact) Boolean factor analysis, the task is to decompose a given formal
context (G, M, I) into a Boolean matrix product

(G, M, I) = (G, F, IGF ) · (F, M, IF M ),

where the set F of “factors” is initially unknown. The aim is usually to get by
with as few factors as possible.
A look at Definition 33 shows that for each factor f ∈ F the set of pairs

f IGF × f IF M

must be a subset of the incidence relation I, i.e. it describes a preconcept of

(G, M, I). In fact, it is even true that these preconcepts together cover the whole
of I, so that [
I= f IGF × f IF M
f ∈F

must hold. Conversely, any covering of I with preconcepts results in a Boolean

factorization.
The number of factors required for a Boolean factorization is therefore exactly
the number of preconcepts required to cover the incidence relation I. And be-
cause each preconcept is contained in a formal concept, it is also the number of
concepts required to cover I. This gives us the following theorem.
Theorem 57 The smallest number of factors in a Boolean matrix factorization of K
equals the 2-dimension of Kc . 
There is another approach to Boolean factors which is perhaps more intuitive.
For a given formal context one may ask if its attributes can be interpreted as
disjunctions of attributes of another, hopefully simpler context. More formally,
let us say that a disjunctive attribute representation of (G, M, I) over (G, N, J)
is a mapping δ : M → P(N ) such that

g I m ⇔ ∃n∈δ(m) g J n.

The existence of such an attribute representation leads to a factorisation

(G, M, I) = (G, N, J) · (N, M, K) with n K m : ⇔ n ∈ δ(m).

Conversely any such factorisation leads to a disjunctive attribute representation

via δ(m) := mK for all m ∈ M .
278 6 Properties of concept lattices

a b c d e f g
× × × × × × × × × × × a b c d e f g
× × ×
× ×
× × × × × × × × ×
× × × × × × ×
× × × ×
× ×
× × × × = × × × ·
× × × × × × × × ×
× × × × × × × × × × × ×
× × × × × × ×
× × × × × × × × × × × × ×
× × × × × × × × × × ×

Figure 6.7 The formal context from Figure 6.6 allows a Boolean matrix factorization with
six factor attributes.

W
-dimension

Embeddings in direct productsWof chains are also of interest for W the drawing of lat-
tice diagrams. In this context, -embeddings, i.e., injective, -preserving maps,
have proved helpful, and we are faced with the question of the existence of such
embeddings. The theoretical background of this question is elaborated in Chap-
ter 7 in a different context. We give the following result without proof, because
it can be obtained as a special case of the Propositions 132 and 134.
Theorem 58 There exists a -embedding of a finite lattice L in the direct product of
W
n chains of the lengths l1 , . . . , ln if and only if M (L) can be covered by n chains of
cardinality m1 , . . . , mn , where mi ≤ li holds for i ∈ {1, . . . , n}. 
As an immediate consequence, we can W determine two parameters for this
problem of embedding: If we define the -dimension to W be the smallest number
of chains in whose product the lattice can be embedded -preservingly and the
-rank to be the least length of such a product, the theorem gives us satisfactory
W
information on those numbers, at least for finite lattices. For this purpose we use
a well known result of Dilworth ([81], p. 3), which says that the smallest num-
ber of chains by which a finite ordered set can be covered is equal to the width
of this ordered set.
Corollary 11 InW the case of a finite lattice L, the -dimension is equal to the width of
W
M (L) and the -rank is equal to the cardinality of M (L). 
6.5 Notes, references, and trends 279

6.5 Notes, references, and trends

Various authors have suggested highlighting particularly important formal con-

cepts and made proposals as to what makes such concepts important. Kuznetsov
& Makhalova [249] have compared “interestingness measures” for formal con-
cepts. Stumme et al. [362] introduce the iceberg of a concept lattice. This con-
sists of the concepts with a particularly large concept extent. These correspond
to the (closed) “frequent itemsets” that play a role for association rules. The au-
thors develop a special algorithm, Titanic, that finds the iceberg. Ventos and
Soldano [374] pursue a related, but more flexible approach.
It was already mentioned in Section 0.5 that Kuznetsov [247] has introduced
a definition of stability of closed sets. Formal concepts can be assigned both an
extent stability and an intent stability. Roughly speaking, a concept extent is par-
ticularly stable if it has many generators relative to its size.
Belohlavek and Trnecka [36, 37, 39] propose a formalization of a phenomenon
well known from psychology, with which humans apparently simplify the nam-
ing of their concepts, namely that of basic level concepts.
Hanika and Hirth [183] examine dense parts of the formal context. They in-
troduce the p, q-kernel of a context as that (uniquely determined) maximal sub-
context in which each object has at least p incidences and each attribute has at
least q incidences, and investigate the interplay with the concept lattice of the full
context both theoretically and experimentally.
The lattice properties treated in this chapter, namely distributivity, its gener-
alizations, and modularity, as well as questions of dimension, are basic topics
of the mathematical theories of lattices and order. We did not intend to offer a
concept-analytic representation of the general state of the research; rather, we
want to show that these subjects can be included without difficulty in our lan-
guage and can thus be used. Naturally, our summary lays no claim to complete-
ness. We refer to the textbooks on order and lattice theory which were mentioned
in the beginning and to current discussion in the respective scientific journals,
e.g. Order. A detailed comparison of many properties of finite lattices has been
elaborated by Reeg and Weiß [311]. Algorithms and complexity estimations can
be found in the doctoral thesis [344] of Skorsky.

6.1

This section is based on A. Albano’s dissertation [1].

6.2

A detailed examination of the distributivity of concept lattices is attributable to

Erné. His study [117] contains many results which go beyond the contents of this
chapter. Furthermore, he was the first to work out that the different variants of
280 6 Properties of concept lattices

the distributive law coincide in the doubly founded case. A connection is estab-
lished between Theorem 47 and topological statements. This theorem was taken
from [396] and goes back to ideas of Raney [310]. Finite distributive concept lat-
tices are described in [395]. That distributivity can be described by means of the
arrow relations has been known for quite a long time in the version of Theorem
48(9). The elegant concise version in Theorem 48(3,4) also goes back to Erné.
There is a book by Stern [352] on semimodular lattices. The characterization of
semimodularity by means of the arrow relations was derived by Skorsky from
a result of Faigle and Herrmann [121]. For the arrow relations in contexts of
modular lattices, see Doerfel [99].

6.3

The literature on the generalizations of the distributive laws is so extensive that

we have to refer to the relevant textbooks, in particular to Crawley & Dilworth
[81]. A substantial amount of information and links can also be found in the
dissertation by H. Reppe [313], where, among other things, n-distributivity is
examined, which is shown to be related to the maximum size of proper premises.
See also Reppe [312].
Semidistributivity is of great importance in connection with the examination
of free lattices, as Jónsson and Kiefer [203] have demonstrated; see also Nation
[291]. Day [86] already uses the characterization by means of the double arrow
relations. There are generalizations of semidistributivity which can also be de-
scribed by means of the arrow relations; see Day, Nation & Tschantz [88] as well
as Geyer [168].
Locally distributive lattices have been introduced by Dilworth [95] and appear
in different connections as natural structures. A concentrated overview over the
development of this term is given by Monjardet [286]. The different parts of The-
orem 52 have been compiled from publications by Green and Markowski [177],
Jamison [200] and Edelman [113]. Closure systems with the anti-exchange prop-
erty have been intensively investigated. They are also referred to as convex ge-
ometries. Antimatroids [200] and, with a different interpretation, learning spaces
[122] are finite -closed set systems consisting of the complements of the closed
S
sets of a closure system with anti-exchange property.
The comparison of lattice properties in Figure 6.2 again goes back to Skors-
ky; see also R. Schmidt [338]. Grötzsch [178] translates Day’s characterization
by proving that a finite lattice L is a (lower) bounded homomorphic image of
a free lattice if and only if J(L) admits a linear order  such that g % m and
h . m together always imply g  h. This, combined with its dual, yields prop-
erty B. Skorsky also has worked on neighborhood-preserving embeddings into
distributive lattices [343], since this is interesting for the automatic generation of
diagrams. Wild [385] had solved the problem for powerset lattices.
6.5 Notes, references, and trends 281

6.4

A standard reference for the dimension theory of ordered sets is the book by
Trotter [370]. The fact that the order dimension of (P, ≤) is equal to that of the
Dedekind-MacNeille-completion has been known for quite a long time, compare
[20], [337]. Baldy & Mitas [19] have generalized this.
The terms Ferrers relation and -dimension go back to Riguet [320] and Cogis
[78], [79], The works of the latter already contain parts of the content of Theo-
rem 54. Bouchet [65] proves the theorem for ordered sets, the generalized version
can be found in [403]. With regard to the Ferrers dimension see also Doignon,
Ducamp & Falmagne [100] and Koppen [226]. Applications can be found in
Reuter [315], [316], Ganter, Nevermann, Reuter and Stahl [153]. Closely related
b b relations are the interval orders. These are ordered sets (P, ≤) in
to the Ferrers
which b b cannot be embedded. Formally, the condition reads:

u < v, x < y, u 6< y ⇒ x < v.

By means of Definition 94 we recognize that this is precisely equivalent to the

fact that < is a Ferrers relation. Hence examinations of the “interval dimension”
are closely related to those of the Ferrers dimension. It is possible to generalize
this by considering the order of the intervals of an arbitrary ordered set. This was
done by Mitas [285].
The example of the seven-segment display goes back to Stahl and Wille [348].
With respect to set representations see also Markowsky [276].
Numerous publications deal with Boolean matrix factorization, and there are
powerful algorithms for it. However, these are usually approximate factoriza-
tions for large data sets, and not exact factorizations as discussed here. To our
knowledge, the initial idea of combining Boolean factorizations with Formal
Concept Analysis comes from Keprt [215], see also Keprt & Snášel [216] and
was then perfected by Belohlavek and his co-authors [40, 38].
Glodeanu [172], Belohlavek and Vychodil [35] extended the methodology
to the triadic case. In [147] an ordinal factor analysis is presented, where the 2-
dimension is replaced by the order dimension and the covering concepts are re-
placed by covering chains of concepts.
In [148] it is shown how the “Competence based knowledge space theory” designed
by Korossy [227] can be expressed with the help of Boolean factorizations of
formal contexts.
Chapter 7
Context comparison and conceptual
measurement

The mappings between concept lattices which are important for structure com-
parison are primarily the complete homomorphisms. In Section 3.2 we have
worked out the connection between compatible subcontexts and complete con-
gruences, i.e., the kernels of complete homomorphisms. Another approach is to
connect morphisms between lattices with morphisms between formal contexts.
But what are suitable morphisms between formal contexts? For isomorphisms
and thus especially for automorphisms the answer is obvious, even if mapping
pairs are required. We show in the first section how to compress concept lattice
diagrams if there are non-trivial automorphisms. In a folded diagram, the nodes
do not represent the individual concepts, but entire orbits of the automorphism
group.
In the second section we then ask for suitable general morphism terms for
formal contexts. We start with structure-preserving or -reflecting mappings and
come across the extensionally continuous mappings, which become important in
the third section under the name scale measures and even fill the fourth section.
Mapping pairs are also considered again. However, the most natural morphisms
for formal contexts are probably the bonds that have already been considered in
section 3.4.
Scale measures are needed for conceptual measurement, i.e. for the process of
mapping an observation (noted as a formal context) into a scale (which is also
a formal context). The third section summarizes the most important properties,
the fourth contains first approaches to a conceptual measurement theory.

7.1 Automorphisms and foldings

Definition 97 An isomorphism between contexts K1 := (G, M, I) and K2 :=

(H, N, J) is a pair (α, β) of bijective maps α : G → H, β : M → N with

g I m ⇔ α(g) J β(m).

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 283
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_8
284 7 Context comparison and conceptual measurement

In the case K1 = K2 we call this an automorphism; the group of automorphisms

of a context K is denoted by Aut(K). ♦
Isomorphic contexts have isomorphic concept lattices, since every context iso-
morphism (α, β) through

(A, B) 7→ (α(A), β(B)) for (A, B) ∈ B(K1 )

induces a lattice isomorphism of B(K1 ) onto B(K2 ). If both contexts are reduced,
then every lattice isomorphism is induced by one (and only one) context iso-
morphism. More generally: If the contexts are clarified, a lattice isomorphism ϕ
is induced by a context isomorphism if and only if ϕ surjectively maps object
concepts onto object concepts and attribute concepts onto attribute concepts.
An observation by W. Xia shows that in turn we can interpret the isomor-
phisms themselves as concepts of a suitable context.
Definition 98 For contexts K1 := (G, M, I) and K2 := (H, N, J) we define
∼
K1 × K2 := (G × H, M × N, ∼)

with
(g, h) ∼ (m, n) :⇔ (g I m ⇔ h J n). ♦

Theorem 59 If K2 is clarified and if α ⊆ G×H, β ⊆ M ×N are bijective maps between

G and H or M and N , respectively, then the following conditions are equivalent:
1. (α, β) is a context isomorphism of K1 onto K2
∼
2. (α, β) ∈ B(K1 × K2 ).
Proof For A ⊆ G × H, B ⊆ M × N it holds that

A∼ = {(m, n) | g I m ⇔ h J n for all (g, h) ∈ A}

as well as
B ∼ = {(g, h) | g I m ⇔ h J n for all (m, n) ∈ B}.
Hence, a pair (α, β) of bijective maps is an isomorphism of K1 onto K2 if and only
if α ⊆ β ∼ or, equivalently, if β ⊆ α∼ . This, however, also implies that α = β ∼
and β = α∼ , since, if (g, h) ∈ β ∼ , then

hJβm ⇔ g I m ⇔ αg J βm

for all m ∈ M , from which we infer that hJ = (αg)J . If K2 is clarified, this implies
h = αg and we obtain α = β ∼ . 
Furthermore, we learn from the theorem that, in the case of clarified contexts,
an isomorphism (α, β) is determined already by each of its two components.
A further application of Theorem 6 on p. 82 allows us to calculate a concept
lattice “modulo automorphisms”. This means: If Γ ≤ Aut(G, M, I) is a group
7.1 Automorphisms and foldings 285

of context automorphisms, then B(G, M, I) divides into orbits under Γ . We se-

lect exactly one concept from each such orbit, namely the one with the lectically
largest extent. Such an extent is called orbit-maximal. We want to avoid calculat-
ing first all extents and then determining the orbit-maximal ones among them.
The following theorem shows that the Next Extent algorithm (p. 80) can be suit-
ably modified such that it only generates the desired extents.
Theorem 60 The smallest orbit-maximal concept extent lectically greater than a set A ⊆
G is
A+ := A ⊕ i,
i being the largest element of G for which A <i A ⊕ i and, at the same time, α(A ⊕ i) ≤
A ⊕ i for all (α, β) ∈ Γ .
Proof What we have to show is that the system of orbit-maximal extents satis-
fies the prerequisites of Theorem 6. We prove that, more generally, the following
holds true: If B is an extent and A = (B ∩ {1, . . . , i − 1})00 , then

A < α(A) ⇒ B < α(B)

holds for every automorphism (α, β) of (G, M, I).

A < α(A) means A <j α(A) for some j ∈ G. If i ≤ j, this would mean that

A ∩ {1, . . . , i − 1} = α(A) ∩ {1, . . . , i − 1},

which, because of
A = (A ∩ {1, . . . , i − 1})00
would imply that A ⊆ α(A) and thus A = α(A), contradictory to A < α(A).
Hence, j < i must hold. But then we have

B ∩ {1, . . . , j − 1} = A ∩ {1, . . . , j − 1} = α(A) ∩ {1, . . . , j − 1} ⊆ α(B)

and j ∈ α(B) \ B, from which it follows that B < α(B). 

Lattice foldings

Example Figure 7.1 shows the lattice of the subgroups of the alternating group
A5 . However, this is not a usual diagram of the subgroup lattice, because firstly,
only nine elements are shown, although the group A5 has 59 subgroups. Sec-
ondly, it is not a lattice line diagram as introduced in Section 0.1, because the
elements labeled Z2 and Z3 have no supremum. And finally, one of the edges
(far left) bears an unusual annotation “(24)”. Further explanation is obviously
needed to make the diagram comprehensible.
Represented are the nine orbits of the automorphism group Γ of A5 , which is
isomorphic to S5 . Each circle of the diagram represents several subgroups that
286 7 Context comparison and conceptual measurement

eA5
  HH
 HH
  HH
 H
D3 eH eA4
 HH e
  D5
HH  
(24)
HH eV4 
 H 
 HH 
Z3 e H e eZ5
HH Z2 
HH 
H 
HH  
HH e
S1
Figure 7.1 The lattice of subgroups of the alternating group A5 “modulo automor-
phisms”. The representatives of the subgroup orbits are: S1 := {id}, Z2 := {id, (01)(23)},
Z3 := {id, (012), (021)}, V4 := {id, (01)(23), (03)(12), (02)(13)}, Z5 := h(01234)i,
D3 := h(01)(23), (014)i, D5 := h(01)(23), (01234)i, A4 := h(01)(23), (012)i, A5 :=
h(012), (01234)i.

can be mapped to each other by the automorphism group. The diagram shows
a factor structure that is completely different from the homomorphic images dis-
cussed in section 3.1. A representative for each subgroup orbit is given in the
figure caption.
The order relation is defined as follows: an orbit is less or equal to another
one, if one of its elements is a subgroup of some element of the second orbit. It
is easy to see that this implies that every element of the first orbit is contained in
some element of the second. And it also implies, that every representative of the
second orbit contains some element of the first. The order relation is independent
of the choice of representatives.
The additional annotation is needed to express which elements of one orbit
are contained in which elements of the second. It is surprising that in Figure 7.1 a
single edge label is sufficient. This is only possible because the labeling has been
strictly simplified, and does depend on the choice of representatives.
In Figure 7.1, the representatives were chosen in such a way that they almost
always actually reflect the order relation, i.e. one orbit is contained in another if
the chosen representative is a subgroup of the second representative. The excep-
tion is Z3 , because this group is not a subgroup of the group D3 , which represents
the orbit above it. But D3 does contain the group (24)Z3 (24), which is obtained
from Z3 by conjugation with the permutation (24). This is the reason why the
element (24) is entered at the corresponding edge of the line diagram. >
In the case of lattices with many automorphisms a diagram of the orbits is
often easier to read than a diagram showing all elements of the lattice. Such di-
7.1 Automorphisms and foldings 287

agrams “modulo automorphisms” are used in group theory, but can be applied
more generally.
Definition 99 Let (P, ≤) be a finite ordered set and let Γ ≤ Aut(P, ≤) be a sub-
group of the automorphism group of (P, ≤). A folding of (P, ≤) under Γ , also
called a Γ -folding of (P, ≤), is

(T, ≤f , λ, (G, ◦)),

where
– T is some transversal of the orbits of Γ on P ,
– a ≤f b : ⇔ ∃γ∈Γ a ≤ γ(b) (for a, b ∈ T ),
– (G, ◦) with G := {gγ | γ ∈ Γ } is a group for which γ 7→ gγ is an isomorphism
from Γ to (G, ◦), and
– λ : T × T → P(G) is defined by

λ(a, b) := {gγ ∈ G | a ≤ γ(b)}.

If (P, ≤) is a lattice, we speak of a lattice folding. λ is called the (full) annotation

of the folding. ♦
Finiteness is assumed here (unlike in [59]) to ensure that all orbits are an-
tichains.
Proposition 121 The annotation λ of an order folding satisfies
1. λ(a, b) 6= ∅ is and only if a ≤f b,
2. each set λ(a, a), a ∈ T , is a subgroup Ga of (G, ◦), and
3. λ(a, b) ◦ λ(b, c) ⊆ λ(a, c) for all a ≤f b ≤f c in T .
Proof (1) is immediate from the definition, (2) follows from the fact that a ≤
γ(a) is possible only for a = γ(a), and (3) can be concluded since a ≤ γ1 (b),
b ≤ γ2 (c) together imply a ≤ γ1 γ2 (c). 
At this point we need a standard notion from group theory: When S1 and S2 are
subgroups of a group (G, ◦) and g ∈ G is an element, then the set

S1 ◦ g ◦ S2 = {s1 ◦ g ◦ s2 | s1 ∈ S1 , s2 ∈ S2 }

is called a double coset of S1 and S2 . Any two double cosets S1 ◦ g ◦ S2 and

S1 ◦h◦S2 are either equal or disjoint, i.e., the double cosets of S1 and S2 partition
G. Note that g ∈ S1 ◦ g ◦ S2 always holds.
Proposition 122 Each set λ(a, b) is a union of double cosets of Ga and Gb .
Proof It follows from Condition (3) of the definition that

λ(a, a) ◦ λ(a, b) ◦ λ(b, b) = λ(a, b).

Thus g ∈ λ(a, b) implies Ga ◦ g ◦ Gb ⊆ λ(a, b). 

288 7 Context comparison and conceptual measurement

The fact that the double cosets partition the group allows for a much more read-
able annotation.
Definition 100 A simplified annotation λ• corresponding to λ specifies
– for each element a ∈ P the group Ga = λ(a, a), usually by annotating a struc-
ture whose automorphism group is Ga , and
– for every pair a < b a transversal of the double cosets in
[
λ(a, b) \ λ(a, c) ◦ λ(c, b).
a<c<b

♦
Note that λ• (a, b) may be empty. The convention is that such pairs are not
listed at all. Likewise, if only the neutral element occurs as a representative for
neighboring pairs a ≺ b, nothing is written. It also happens that λ• (a, b) 6= ∅ even
if a 6= b and a is not a lower neighbor of b. It may then be necessary to add an
edge to the folding diagram that carries this annotation. Figure 7.3 shows such
a situation.

(243)

(243)

(23)

2 1

3 4

Figure 7.2 An S4 -folding of the lattice of all graphs on four points, with simplified anno-
tation.

Example Figure 7.2 shows the powerset lattice of a six-element set, not as a usual
lattice diagram, but as a lattice folding. Undirected graphs on the four-element
7.1 Automorphisms and foldings 289

set {1, 2, 3, 4} are shown, i.e. subgraphs of the complete graph K4 . This graph has
six edges, and each of the 64 subsets of this edge set determines a graph on the
same nodes. However, the folding diagram does not show all 64 graphs, but only
the 11 isomorphism classes, which are the orbits of the automorphism group S4
of K4 , acting on its subgraphs.
At this point, we will explain in detail which groups we are talking about.
The lattice which was folded is the lattice of all 64 graphs on {1, 2, 3, 4}. The
automorphism group of this lattice has a subgroup Γ , which is induced by the
permutations of the node set and which maps graphs to isomorphic graphs. For
the folded structure, the information on how Γ operates on the original lattice is
no longer given, but only that this group (G, ◦) is isomorphic to S4 .
From each graph isomorphism class, a representative was chosen and entered
with its graph diagram, thereby specifying the groups λ(a, a) as being the auto-
morphism groups of these representatives.1
We explain the edge labeling using two examples. The circle labeled is
lower neighbor of the circle labeled , and the edge between these two circles
carries no annotation, meaning that (in the unsimplified annotation λ) this edge
is annotated by the double coset containing the neutral element, i.e. by the prod-
uct of the two automorphism groups. Indeed, is a subgraph of , and thus
the identity is one of the exactly eight embeddings. is not a subgraph of its
upper neighbor , but can be embedded. One such embedding is the annotated
mapping (23), the others fill the double coset of which (23) is a representative.>
Figure 7.3 shows a further simple example: on the left, a lattice whose automor-
phism group is apparently isomorphic to the cyclic group Z3 and on the right,
the orbit diagram with the necessary information.
Now that we are able to fold ordered sets we also would like to unfold them
in a way that reconstructs the original order. Before doing so, we reconstruct the
full annotation.
Proposition 123
[
λ(a, b) = λ(a, a) ◦ λ• (a, c) ◦ λ(c, c) ◦ λ• (c, b) ◦ λ(b, b).
a≤c≤b

Proof The inclusion ⊇ is obviously true. Suppose there is a counterexample for

the other inclusion, and choose this conterexample such that the longest chain
from a to b is of minimal possible length. For a ≺ b, λ• (a, b) is a transversal of
the double cosets in λ(a, b), which verifies the claim for length one. For greater
length, an element could only be removed from λ(a, b) for simplification when
a double coset of the form λ(a, c) ◦ λ(c, b) was subtracted for some a < c < b.

1 To avoid the term “isomorphism class”, some also speak of the unlabeled case. In this sense,

Figure 7.2 shows a diagram of the unlabeled graphs on four vertices. However, this may be
confusing, as it is a diagram in which each node is labeled with a labeled graph.
290 7 Context comparison and conceptual measurement

c cZ3
@
c c @c c{id}
@

c
c
c (012) c{id}






c
c
c c{id}
@
@c cZ3
@

Figure 7.3 The diagram of orbits (on the right) has an additional edge.

But since the chain length for both a < c and c < b are shorter, these sets are
contained in the RHS of the proposition. 
Definition 101 The unfolding of an order folding (T, ≤f , λ, (G, ◦)) is

˙
[
where g ◦ Ga ≤r h ◦ Gb : ⇔ g −1 ◦ h ∈ λ(a, b).


G/Ga , ≤r ,
a∈T

♦
Remark: For any a ∈ T , G/Ga equals the set {g ◦ Ga | g ∈ G} of all cosets of
the “stabilizer” subgroup Ga . The intuition behind this is that a represents one
isomorphism type and the cosets g ◦ Ga stand for all possible isomorphic copies
of this type. However, Ga and Gb can be the same, even if a and b are different.
For example, in Figure 7.2 the representatives and have the same auto-
morphisms. This is why a disjoint union is required in the above definition. But
the somewhat vague definition of a disjoint union can lead to difficulties in no-
tation, which can be overcome by using the “dot notation” as in Definition 29 on
p. 41, replacing ˙ p∈P G/Gp by p∈P ({p} × G/Gp ). Having said this, we will
S S
avoid this exact but more cumbersome notation and write as if cosets g ◦ Ga and
h ◦ Gb are always distinct when a 6= b.
Theorem 61 The unfolding of any folding of a finite ordered set (P, ≤) is isomorphic to
(P, ≤).
Proof Let Γ ≤ Aut(P, ≤) be a subgroup of the automorphism group, Y be a
transversal of the Γ -orbits and (T, ≤f , λ, (G, ◦)) be a folding of (P, ≤) under Γ ,
as in Definition 99. Recall that γ 7→ gγ denotes a group isomorphism from Γ to
(G, ◦). We have to show that

˙
[

G/Ga , ≤r
a∈T
7.1 Automorphisms and foldings 291

××
× ×
××

××
× ×
×
×

××
×
××
×
× ×
×
×
(bc)

×× ×× ×× ×
× × × × ×
× × ×

(bc)
××
×
×× ×
× × @ ×
(ac)@
××
×

×× ×
× ××

×× ×

a b c

Figure 7.4 An S3 -folding of the lattice of the 61 closure systems on the set {a, b, c}. Each
closure system is the system of intents of an object-reduced formal context, which is
unique up to permutations of the attribute set {a, b, c}. The context at the least element
has an empty object set. From [178], modified. An unfolded diagram can be found at
Higuchi [191].

is an ordered set isomorphic to (P, ≤) (assuming disjointness, as discussed in

the remark after Definition 101). Proving that ≤r is an order is easy: Clearly ≤r
is reflexive, since the neutral element of Ga is in λ(a, a). Antisymmetry follows
from Proposition 121, which proved that for a 6= b at least one of the sets λ(a, b)
must be empty, and transitivity follows from the multiplicativity condition for
annotation maps. What remains to show is isomorphism of the two ordered sets.
We claim that [˙
Ψ: G/Ga → P, gγ ◦ Ga 7→ γ(a),
a∈T
292 7 Context comparison and conceptual measurement

is an order isomorphism. It is

gγ ◦ Ga ≤r gδ ◦ Gb ⇔ gγ−1 ◦ gδ ∈ λ(a, b)
⇔ gγ −1 δ ∈ λ(a, b)
⇔ a ≤ γ −1 δ(b)
⇔ γ(a) ≤ δ(b),

which completes the proof. 

7.2 Morphisms

Definition 102 If K1 := (G, M, I) and K2 := (H, N, J) are formal contexts, we

call a map α : G → H
- extensionally continuous, if for every extent U of K2 the pre-image α−1 (U )
is an extent of K1 .
- extensionally closed, if the image α(U ) of an extent U of K1 is always an extent
of K2 .
The extensionally continuous maps are also called scale measures; they will be
examined in detail in the next section. Dually, we define in which cases a map-
ping β : M → N is intensionally continuous or intensionally closed.
A pair of maps (α, β) : K1 → K2 is called
- incidence-preserving if

g I m ⇒ α(g) J β(m)

for all g ∈ G, m ∈ M ,
- incidence-reflecting if
g I m ⇐ α(g) J β(m)
for all g ∈ G, m ∈ M ,
- continuous if α is extensionally continuous and β is intensionally continuous,
- concept-preserving if, for every concept (A, B) ∈ B(K1 ), the pair

(β(B)0 , α(A)0 )

is a concept of K2 , and
- concept faithful if it is both concept-preserving and continuous.
♦
Even if (α, β) is incidence-preserving and -reflecting, the two maps need not
be injective. However, in this case from α(g) = α(h) follows that g 0 = h0 and
7.2 Morphisms 293

dually; i.e., this kind of map is “injective up to clarification”. If both maps are in-
jective, incidence-preserving and -reflecting, then (α, β) may be called a context-
embedding. It has already been explained in Section 3.1 that this results in
order-embeddings, and that order-preserving maps correspond to incidence-
preserving maps. The following proposition follows directly from Proposition 38
on p. 119 and the text that follows it.
Proposition 124 If (α, β) : K1 → K2 is incidence-preserving, then the two maps
_ ^
ϕ(A, B) := γαg and ψ(A, B) := µβm
g∈A m∈B

both are order-preserving. If (α, β) is also incidence-reflecting, then these mappings are
order-embeddings. 
We cannot formulate the opposite direction so smoothly. It is obviously not
the case that every order embedding also entailes a context-embedding, as even
the simplest examples show.2 But one can also specify the reason for this: Order-
preserving mappings (and thus also order embeddings) depend only on the
isomorphism type of the embedding order, not on its representation as a con-
cept lattice. If you allow the second concept lattice, i.e. B(K2 ), to be replaced by
an isomorphic concept lattice, then you can also represent any order-preserving
mapping by an incidence-preserving one. The formal context replacing K2 can
always be chosen as the largest possible one, which is the ordinal scale of B(K2 ).
This is the statement of Proposition 38 and the preceding text on page 119, which
gives the following:
Proposition 125 If ϕ : B(K1 ) → B(K2 ) is order-preserving. then α := ϕ ◦ γ and
β := ϕ ◦ µ are maps for which

(α, β) : K1 → (B(K2 ), B(K2 ), ≤)

is incidence-preserving. If ϕ is an order-embedding, then (α, β) is also incidence-reflecting.


Concept-preserving maps are necessarily incidence-preserving, since with
(g 00 , g 0 ) also (β(g 0 )0 , α(g 00 )0 ) must be a concept, and from αg ∈ α(g 00 ) one obtains
(αg)0 ⊇ α(g 00 )0 and thus (αg)00 ⊆ β(g 0 )0 . With m ∈ g 0 this implies α(g) ∈ β(m)0 ,
i.e., α(g) J β(m).
The map stated in the definition, (A, B) 7→ (β(B)0 , α(A)0 ), is order-preserving.
These maps, however, need not be lattice homomorphisms. It is different in the
case of the concept faithful maps:
Theorem 62 If (α, β) : K1 → K2 is a concept faithful map, then

(A, B) 7→ (β(A)0 , α(A)0 )

2An exceptional case occurs when K1 is the contranominal scale, see Proposition 112 on p. 250.
Whether this can be generalized is unknown to us.
294 7 Context comparison and conceptual measurement

is a complete homomorphism from B(K1 ) to B(K2 ).

A complete homomorphism is induced by a concept faithful map if and only if it maps
object concepts onto object concepts and attribute concepts onto attribute concepts. 
Erné has proved this theorem with the aim of representing lattice homomor-
phisms through concept faithful maps; the restriction made in the theorem dis-
appears if we consider the lattices only up to an isomorphism: According to
the Basic Theorem, every lattice L is isomorphic to B(L, L, ≤) and for this con-
text every concept is an object and an attribute concept, and consequently, every
complete homomorphism between such concept lattices is induced by a concept
faithful map. Erné furthermore gives a simple characterization of the concept
faithful maps, which we add:
Proposition 126 An incidence-preserving map

(α, β) : (G, M, I) → (H, N, J)

is concept faithful if and only if the following condition is satisfied for all h ∈ H, n ∈ N :

(h, n) 6∈ J ⇒ ∃(g,m)6∈I α−1 (n0 ) ⊆ m0 and β −1 (h0 ) ⊆ g 0 .


One can learn from Theorem 62 the somewhat sobering lesson that mapping
pairs (α, β) are not really suitable to represent the complete homomorphisms be-
tween concept lattices elegantly and comprehensively. An alternative is to use re-
lations instead of mappings. For this, we recall Proposition 66 on p. 149, in which
it was established that the bonds from K1 to K2 are in 1-1-correspondence to the
adjunctions between the concept lattices,Wwhich, according to Proposition 11 on
p. 19, are in 1-1-correspondence to the -preserving maps. W SinceV the complete
lattice homomorphisms are those maps which are both - and -preserving, it
seems promising to describe them using pairs of bonds. This is done in Theo-
rem 63, which is illustrated by Figure 7.5.

Theorem 63 For every complete homomorphism

ϕ : B(G, M, I) → B(H, N, J)

R := Rϕ = {(g, n) | ϕγg ≤ µn} ⊆ G × N

defines a bond from (G, M, I) to (H, N, J) and

S := Rϕ = {(h, m) | γh ≤ ϕµm} ⊆ H × M

defines a bond from (H, N, J) to (G, M, I), and

ARJ = B S , B SJ = AR

for all (A, B) ∈ B(G, M, I).

If, conversely, R ⊆ G × N and S ⊆ H × M are bonds satisfying this condition, then
7.2 Morphisms 295

ϕ(A, B) := (B S , AR )

defines a complete homomorphism from B(G, M, I) to B(H, N, J). 

M N

G I R := Rϕ

H S := Rϕ J

Figure 7.5 With reference to Theorem 63

Hence, Theorem 63 shows that complete homomorphisms between concept

lattices can be satisfactorily described by means of suitable pairs of bonds be-
tween the contexts. Now we have different possibilities of turning these bonds
into set-valued maps. For example, we can assign maps

α : G → P(N ), β : M → P(H)

to each homomorphism

ϕ : B(G, M, I) → B(H, N, J)

through

αg := g R = {n | ϕγg ≤ µn}
βm := mS = {h | γh ≤ ϕµm}.

The homomorphism ϕ can be reconstructed from (α, β) through

\ \
ϕ(A, B) = ( βm, αg).
m∈B g∈A

It is not difficult to characterize the pairs of maps which result in this way from
complete homomorphisms by means of the conditions stated in Theorem 63.
The symmetrical situation in Theorem 63 permits further variations. We can
also describe the bonds through maps in the other direction, i.e., α : H → P(M ),
β : N → P(G), and so on. We shall give only one further example of this kind:
296 7 Context comparison and conceptual measurement

Proposition 127 For every complete homomorphism

ϕ : B(G, M, I) → B(H, N, J)

the definition

g ∈ αh : ⇔ γh ≤ ϕγg, m ∈ βn : ⇔ ϕµm ≤ µn

yields a pair of maps

α : H → P(G), β : N → P(M )
with
ϕ(A, B) = ({h ∈ H | αh ⊆ A}, {n ∈ N | βn ⊆ B}).
If, conversely, α : H → P(G) and β : N → P(M ) are maps with the property that, for
every concept (A, B) ∈ B(G, M, I),

({h ∈ H | αh ⊆ A}, {n ∈ N | βn ⊆ B})

is a concept of (H, N, J), then the map

(A, B) 7→ ({h ∈ H | αh ⊆ A}, {n ∈ N | βn ⊆ B})

is a complete homomorphism of B(G, M, I) to B(H, N, J). 

Infomorphisms

These morphisms were introduced by Barwise and Seligman [28]. They are the
same as the Chu maps, see H. Mori [288]. Replacing the second context by its
dual gives the context Galois connections of W. Xia [425].
Definition 103 An infomorphism from a formal context K1 := (G, M, I) to
K2 := (H, N, J) is a pair (~f, f~) of mappings ~f : G → H, f~ : N → M, such
that
g I f~n ⇔ ~fg J n
holds for all g ∈ G, n ∈ N . ♦

Proposition 128 Every infomorphism (~f, f~) from K1 to K2 is also an infomorphism

from Kc1 to Kc2 , and (f~, ~f) is an infomorphism from Kd2 to Kd1 . ~f is extensionally contin-
uous, f~ is intensionally continuous.

Proof The defining condition in Definition 103 retains its truth value when I and
J are replaced by r
I and J
r. For an object n of Kd and an attribute g of Kd we find
2 1
that
n J d ~fg ⇔ f~n I d g
7.2 Morphisms 297

is the same as in the definition. To see that ~f is extensionally continuous, consider

some g ∈ G and n ∈ N . Then

g ∈ ~f −1 (nJ ) ⇔ ~f(g) J n ⇔ g I f~(n) ⇔ g ∈ f~(n)I ,

which proves the claim since ~f −1 preserves set intersections. 

Relational morphisms

Morphisms do not have to be mappings. The examples we discussed above, e.g.,

included pairs of mappings. Another format has already proven itself for formal
contexts because it provides a natural concept of morphism: relations!
We mention this here because it belongs in this section. However, we have
long since introduced the core results under the name “bonds”. According to
Definition 61 on p. 148, a bond from (G, M, I) to (H, N, J) is a relation R ⊆ G×N
with the property that for each object g ∈ G the set g R is an intent of (H, N, J)
and at the same time for each attribute n ∈ N the set nR is an extent of (G, M, I).
Proposition 66 on p. 149 then describes the meaning of such bonds for the
concept lattices: To each bond belongs an adjunction W between the concept lat-
tices, and vice versa. Such an adjunction consists of a -preserving mapping ϕ
from B(G, M, I) to B(H, N, J) and a residual -morphism ψ from B(H, N, J)
V
to B(G, M, I).
The conditions for a residual mapping are specified in Proposition 11:

ϕ(x) ≤ y ⇔ x ≤ ψ(y).

It is also shown in this proposition that the two mappings ϕ and ψ determine
each other and that every -morphism is part of an adjunction, as is every -
W V
morphism.
We have also discussed other relations between formal contexts and resulting
mappings between concept lattices, such as dual bonds and relational Galois
connections. From the point of view of category theory, however, adjunctions
are particularly suitable as morphisms, because they can be concatenated, as can
bonds.
This results in two equivalent categories, namely the category of formal con-
texts with the bonds as morphisms and the category of complete lattices with
adjunctions. A careful presentation of the equivalence can be found in Borg-
wardt [64] together with details about special properties of these categories.
298 7 Context comparison and conceptual measurement

7.3 Scale measures

Definition 104 Let K := (G, M, I) be a context and let S := (GS , MS , IS ) be a

scale. A map σ : G → GS is called an S-measure if the preimage σ −1 (U ) of every
extent of S is an extent of K. An S-measure σ is called full if every extent of K is
the preimage of an extent of S. ♦
In order to visualize this definition, we imagine a new context Kσ whose ob-
jects are the objects of K and whose attributes are the attributes of S. Into the
g-row of this context, we enter the σ(g)-row of the scale. This means that for-
mally we define Kσ := (G, MS , Iσ ) by

g Iσ m :⇔ σ(g) IS m.

Now, the definition says that σ is a S-measure if and only if every extent of Kσ is
also an extent of K; σ is full if and only if K and Kσ have the same extents. Since
the context Kσ is defined on the same object set as K, we can imagine the two
contexts joined together to form the apposition

K Kσ ,

whose extents are the same as those of K, provided that σ is a measure. In this
way a S-measure is understood as the possibility of extending the given context
by attributes from the scale S without changing the extents. σ is full if the new
attributes render the old ones dispensable.

Proposition 129 For a map σ : G → GS the following conditions are equivalent:

1. σ is a S-measure.
2. For all subsets A ⊆ G it holds that σ(A00 ) ⊆ σ(A)00 .
3. For all subsets A, B ⊆ G it holds that A → B ⇒ σ(A) → σ(B)
(In accordance with Section 2.3, A → B is used as an abbreviation of B ⊆ A00 ).
σ is full if and only if the inverse implication also holds in (3).

Proof (1) ⇒ (2): Every scale measure satisfies condition (2), since, for every
A ⊆ G, σ −1 (σ(A)00 ) is an extent containing A and thus also A00 .
(2) ⇒ (3): A → B ⇔ B ⊆ A00 ⇒ σ(B) ⊆ σ(A00 ) ⊆ σ(A)00 ⇒ σ(A) → σ(B).
(3) ⇒ (1): If U is an extent of S and g is an arbitrary object from (σ −1 (U )), i.e.,
with σ −1 (U ) → g, this yields U → σg, i.e. σg ∈ U and consequently g ∈ σ −1 (U ),
hence this set must be an extent of K.
σ is full if and only if for A ⊆ G it always holds that A00 = σ −1 (σ(A)00 ); this is
however equivalent to
g ∈ A00 ⇔ σ(g) ∈ σ(A)00 ,
i.e., to
A → g ⇔ σ(A) → σg. 
7.3 Scale measures 299

We do not really have to check the definition of the S-measure for all extents. It
suffices that the preimages of the attribute extents of S are extents of K (since the
preimage of an intersection of sets is the intersection of the preimages). Likewise,
an S-measure is already full if every column extent of K is the preimage of an
extent of S. If we call a subset T of the attribute set of a context K dense in the
case that the set {µm | m ∈ T } is infimum-dense in B(K), then we can continue
as follows: A scale measure is full if and only if the set

{m ∈ M | m0 is the preimage of an extent}

is dense.
A surjective S-measure is not automatically full. Indeed, every scale measure
σ can be replaced by a surjective one if we switch to a subscale (by a subscale of
a scale S we understand a subcontext (T, MS , IS ∩ (T × MS )) with T ⊆ GS ). This
is the content of the following proposition:

Proposition 130 For every subscale (T, MS , IS ∩ (T × MS )) of S, the identical map is

a full S-measure. For a context (G, M, I), σ : G → GS is a S-measure if and only if σ is
a (σ(G), MS , IS ∩ (σ(G) × MS ))-measure.

Proof The extents of the subscale are precisely the sets of the form U ∩ σ(G),
with U being an extent of S. 
From a scale measure σ we obtain two maps between the concept lattices of
the corresponding contexts K and S, i.e., one in each direction. This is described
by the following two propositions. From the circumstance that the set mapping
σ −1 is intersection-preserving we immediately infer:

Proposition 131 If σ is a S-measure of K,

(A, A0 ) 7→ (σ −1 (A), σ −1 (A)0 )

defines a -preserving map of B(S) to B(K). If σ is surjective, this map is injective.

V

(We denote this map also by σ −1 .)

In other words: If σ is a S-measure, we can rediscover a copy of the subscale
(σ(G), MS , IS ∩ (σ(G) × MS )) in the system of the extents of K.
We shall now show that it is possible to assign to every S-measure of K a -
W
preserving map of B(K) to B(S) in a unique manner.

Proposition 132 For every S-measure σ of K,

σ̃(A, A0 ) := (σ(A)00 , σ(A)0 )

defines a -morphism
W
σ̃ : B(K) → B(S).
σ̃ maps the object concepts of K onto object concepts of S. If S is a scale
W in which g 6= h
always implies g 0 6= h0 (for all g, h ∈ GS ), then, conversely, every -preserving map of
300 7 Context comparison and conceptual measurement

B(K) to B(S) with this property results from an S-measure in the manner specified. σ
and σ̃ uniquely determine each other. σ is full if and only if σ̃ is injective.

Proof Let ψ be the map residual to σ −1 and let X be an extent of K. We have

^
ψ(X, X 0 ) = {(Y 00 , Y 0 ) | X ⊆ σ −1 (Y 00 )}
^
= {(Y 00 , Y 0 ) | σ(X) ⊆ Y 00 }
\ \ 
= {Y 00 | σ(X) ⊆ Y 00 }, ( {. . .})0
= (σ(X)00 , σ(X)0 ) = σ̃(X, X 0 ).

Hence, σ̃ is residual to σ −1 and thus -preserving.

W
Next we show that σ̃ maps object concepts onto object concepts: Let x ∈ G
be an object and let g := σ(x) be the image of x in GS . σ −1 (g 00 ) is an extent of
K containing x and thus also x00 , hence, σ(x00 ) ⊆ g 00 , which implies σ(x00 )00 ⊆
g 00 . However, since g ∈ σ(x00 ), equality must hold, i.e., σ(x00 )00 = g 00 , and thus
σ̃(x00 , x0 ) = (g 00 , g 0 ).
It remains to be shown that every -preserving map that maps object concepts
W
to object concepts goes back to a measure. Hence, let ψ be such a map. We define
σ : G → GS through

σ(x) = g :⇔ ψ(x00 , x0 ) = (g 00 , g 0 ).

(Here the assumption g 0 = h0 ⇒ g = h is used.) Then, we have

_
ψ(X, X 0 ) = ψ( {(x00 , x0 ) | (x ∈ X})
_
= {ψ(x00 , x0 ) | x ∈ X}
_
= {(σ(x)00 , σ(x)0 ) | x ∈ X}
[ 
= ( {σ(x)00 | x ∈ X})00 , (. . .)0
= (σ(X)00 , σ(X)0 ).

Hence, every map of this kind is induced by a measure. The remaining state-
ments follow from Proposition 11 on p. 19. 
Definition 105 If S1 and S2 are scales with the same scale values, i.e., with GS1 =
GS2 , we call S1 finer than S2 if every extent of S2 is also an extent of S1 . S2 then is
coarser than S1 . S1 and S2 are called (scale-)equivalent if there is a full bijective
S2 -measure of S1 .3 ♦
If σ is a bijective full measure, so is σ −1 . Hence, the equivalence of scales
is symmetrical. Since the concatenation of (full) scale measures again yields a

3 See Section 7.5 for a more general approach.

7.3 Scale measures 301

(full) scale measure, the equivalence of scales is in fact an equivalence relation.

If S1 is simultaneously finer and coarser than S2 , the two scales are equivalent.
The possibility of scaling variably fine is very useful for data analysis. The fact
that S1 is a finer scale than S2 means that S1 (up to equivalence) can be written as
an apposition of S2 with another context. For example, S1 is equivalent to S2 S1 .
In the case of plain scaling, this is inherited by the derived one-valued context:
finer scaling yields a finer derived context. If we use finer scales, the derived
context simply has “some more columns”, i.e., it can be written as an apposition
of two contexts, one of which is the derived context with respect to the coarser
scaling. From Section 2.2 we know that the concept lattice of an apposition can
be adequately represented by a nested line diagram, the concept lattice of the
coarser scale then represents a rough structure which is further differentiated by
the attributes which are added through the finer scale.
An example: Questionnaires are often formulated in such a way that they
present opinions to which the participants can express approval or rejection,
offering an alternative with intermediate values, more or less in the following
form:
agree 2 2 2 2 2 2 2 do not agree

In Section 1.4 we have suggested the interordinal scale for the ≤2≥6
scaling of such attributes, however, it is usual first to scale
1 ×
coarser, for example with the threshold scale (displayed on the
2 ×
right), which only uses two elements of the interordinal scale.
3
In this way we obtain an approximate impression of the results.
However, this coarsening is quite correct: According to Propo- 4
sition 132,Wthe concept lattice obtained in this way is an image 5
(under a -morphism) of the concept lattice which is the result 6 ×
of interordinal scaling. 7 ×
The scales Nn , On and Ma,n−a are also (equivalent to) coarsenings of the in-
terordinal scale In . The finest scale with n values is the contranominal scale Ncn .
In order to describe the role of the scale measures in the case of plain scaling,
we first convince ourselves that the semiproduct
 

×
` [
a
Sj :=  Gj , Ṁj , ∇
j∈J j∈J j∈J

with
(gj )j∈J ∇ (k, m) : ⇔ gk Ik m
of scales introduced in Definition 31 on p. 43 is also a product in the sense of
category theory, namely in the category of the scales with the scale measures as
morphisms. For this purpose we check that the projections

πk : ×G
j∈J
j → Gk
302 7 Context comparison and conceptual measurement

with πk ((gj )j∈J ) := gk

are surjective Sk -measures of j∈J Sj . Furthermore, we have to show that the
`
a

product map is a scale. The property claimed follows from the fact that the prod-
uct in the category of sets is the Cartesian product.
Definition 106 If K is a context and if for every j ∈ J the map σj is an Sj -measure
of K, then the product measure
`
σ : K →aj∈J Sj

is defined by
σ(g) := (σj (g))j∈J . ♦

Proposition 133 The product measure σ is a aj∈J Sj -measure with πk ◦ σ = σk .

`

Proof The extents of the semiproduct are precisely the products ×

j∈J Uj of
extents of the individual scales. The preimage of such an extent with respect to
σ is given by
×
\
σ −1 ( Uj ) = σj−1 (Uj ).
j∈J j∈J

Hence, the preimages under σ are precisely such subsets that are intersections of
preimages under the measures σj . Each set of this kind is an extent of K, hence
σ is a scale measure. 
We can slightly refine the argument in this proof: Every extent is the inter-
section of attribute extents. Hence, the preimages of extents of the product scale
under σ are precisely the intersections of preimages of the attribute extents of
Sj , j ∈ J. This leads to the following observation:
Proposition 134 The product measure is full if and only if the set of concepts of the form

(σj−1 (mIj ), (σj−1 (mIj )0 ), j ∈ J, m ∈ Mj

is infimum-dense in B(K). 
If all attributes of K are irreducible, the following holds true: The product
measure is full if and only if every attribute extent of K is the preimage of an
attribute extent under one of the scale measures σj , i.e., if for every attribute m
I
of K there exists some j ∈ J and some attribute mj ∈ Mj with m0 = σj−1 (mjj ).
Finally, we can use the notion of the product measure to give an alternative
definition of the derived context with respect to plain scaling:
Proposition 135 Let (G, M, W, I) be a complete many-valued context and let Sm , m ∈
M , be scales for the attributes of M . Furthermore, let K be the derived context with respect
to plain scaling. Then, for every many-valued attribute m ∈ M , the map

g 7→ m(g)
7.4 Measurability theorems 303

is an Sm -measure of K, and K is isomorphic to the subscale of the semi-product of the

Sm which is the image of the product measure of those scales.

The proof results immediately from the definitions. 

7.4 Measurability theorems

Proposition 135 has shown that full scale measures into semiproducts of scales
can be understood as a kind of inversion of plain scaling. Now, we can try to rec-
ognize derived contexts, i.e., to decide in the case of a given one-valued context
whether it could have been derived from a many-valued context through scaling
with given scales. Hence, the question is which contexts can be fully measured
in a semiproduct of nominal scales, ordinal scales etc. and, if so, which size the
necessary semiproduct must have. Proposition 133 is very useful in this context,
since it can be used to split up the problem. Therefore, we first examine how we
can recognize whether a given context allows a measure in one of the standard
scales. If this is the case, some of the attributes of the context can be combined to
form a many-valued attribute (with given scaling) and a repeated implementa-
tion of this procedure according to Proposition 133 finally yields a full measure
into the semiproduct.
Since every scale measure is surjective onto a subscale, it is useful to know the
subscales of the standard scales. In many cases they belong to the same family
of scales.

Proposition 136 The following families of scales have the property that every subscale
of a scale belonging to the family is equivalent to a scale of the same family:

a) nominal scales, e) contranominal scales,

b) one-dimensional ordinal scales, f) contraordinal scales,
c) one-dimensional interordinal scales, g) convex-ordinal scales.
d) multiordinal scales,
We shall omit the proof. 
Theorem 64 The context K := (G, M, I) allows a surjective S-measure for
a) S = Nn , if and only if there is a partition of the object set G into n extents.
b) S = On , if and only if there is a chain U1 ⊂ U2 ⊂ . . . ⊂ Un of n non-empty
extents.
c) S = In , if and only if there is a chain of n non-empty extents of K whose comple-
ments are also extents.
d) S = Mn1 ,...,nk , if and only if there are k chains, each made up of ni non-empty
extents, whose largest elements form a partition of G.
e) S = Nnc , if and only if there is a partition of G into n extents whose unions are also
extents.
304 7 Context comparison and conceptual measurement

f) cd
S = OP , if and only if there is a set P of extents with the following properties:
• The set P, ordered by set inclusion ⊆, is isomorphic to P.
• Every union of extents from P is an extent.
• For every object g ∈ G there is a largest extent Ug ∈ P which does not contain
g.
• P = {Ug | g ∈ G}.
g) S = CP , if and only if there is a set P of extents which satisfy the conditions under
f) and for which additionally the following statement is true:
• The complements of extents from P and the unions of such complements are
also extents.
Proof From Proposition 131 we know that K allows a surjective S-measure if
and only if there is a family US of extents of K and a map σ : G −→ GS such
that the (attribute) extents of S are precisely the images of US under σ. In other
words: In the system of the extents of K, those attribute extents of a scale must
occur which are isomorphic to S after clarification.
This makes a), b), c) and d) obvious. In order to show e), we first convince our-
selves of the fact that the complementary nominal scale has the property spec-
ified: For every scale value g ∈ GS , {g} is an extent, but also GS \ {g}. Hence,
the preimage sets of the scale values under a surjective S-measure form a par-
tition with the property specified in the Proposition. However, the converse is
also true: Given such a partition, a map of G onto GS mapping the classes of the
partition onto the values of the scale is an S-measure.
For f) we argue similarly: According to the definition, the extents of a contraor-
dinal scale are precisely the order ideals of P, the attribute extents are precisely
the complements of principal filters. Hence, the system of attribute extents of
S := Ocd P satisfies the conditions specified in f), and so do the preimages under
an S-measure. If, conversely, a system of such extents of a context is given, and
if ϕ is the order isomorphism of this system onto P, then we obtain a S-measure
through σ(g) := ϕ(Ug ) for all g ∈ G. Under these premises, the preimage of the
attribute extent {x ∈ P | x 6≥ p} is equal to

σ −1 (p0 ) = {g | ϕ(Ug ) 6≥ p} = {g | Ug 6⊇ Up̄ },

for any scale attribute p ∈ P, provided that p̄ is an object of K with ϕ(Up̄ ) = p.

Since Ug is the largest among the selected extents which does not contain g, Ug 6⊇
Up̄ is equivalent to g ∈ Up̄ , i.e., we get

σ −1 (p0 ) = {g | Ug 6⊇ Up̄ } = Up̄ .

The preimage of every column extent is an extent: σ is a measure.

g): The convex-ordinal scale is the apposition of two contraordinal scales,
therefore a CP -measure is in particular a OP
cd
-measure and has to satisfy the con-
ditions under f). The convex-ordinal scale even satisfies the condition addition-
ally required under g), which in this particular case demands that the unions of
7.4 Measurability theorems 305

principal filters are convex sets. Hence this condition is necessary. It remains to
be shown that it is also sufficient.
Hence, let P be a system of extents of K which satisfies the conditions under
f), and let σ be the OP cd
-measure constructed in the proof of f). We shall prove
that under the additional condition the same map σ is also an OcP -measure (from
which the statement follows).
For this purpose we define a set system Q := {Vg | g ∈ G} through
[
Vg := {G \ U | U ∈ P, g ∈ U },

and show that Q satisfies the conditions specified under f), namely for the order
dual to P. The additional condition in g) guarantees that every Vg and all unions
of such sets are extents. According to the definition, we have furthermore

Vg ⊆ Vh ⇔ {U ∈ P | g ∈ U } ⊆ {U ∈ P | h ∈ U }
⇔ {U ∈ P | g 6∈ U } ⊇ {U ∈ P | h 6∈ U }
⇔ {U ∈ P | U ⊆ Ug } ⊇ {U ∈ P | U ⊆ Uh }
⇔ Ug ⊇ Uh .

Hence, Q is order-isomorphic to Pd by means of the isomorphism ψVg := ϕUg .

In the proof of f) we have shown that in this case the map g 7→ ψVg is an Ocd
Pd -
measure from which, because of σg = ϕUg = ψVg and Ocd P d = O c
P , everything
else follows. 
Now we can turn our attention to the question we asked in the beginning: Let
S be a family of scales, for example the family of nominal scales or that of or-
dinal scales. We want to characterize those derived contexts which result from
many-valued contexts through plain scaling with scales from S. According to
Proposition 134 these are precisely the contexts which are equivalent to a sub-
scale of the semiproduct of scales from S. We coin a shorter name for this:
Definition 107 Let K be a context and let S be a family of scales. We say that K
is fully S-measurable if K can be fully measured into a semiproduct of scales
from S.
If, in particular, S is the family of the nominal scales, we say fully nominally
measurable instead of “fully S-measurable”. A fully {Nn }-measurable context
is called fully n-valued nominally measurable, in the special case that n = 2 it
is also called fully dichotomially measurable. ♦

Proposition 137 For every family S of scales, one of the following alternatives holds:
1. Every context is fully S-measurable.
2. Every fully S-measurable context is fully nominally measurable.

Proof First, we show that every context K is fully ordinally measurable, even
fully {O2 }-measurable. This follows immediately from Proposition 133: If we de-
fine for every attribute m of K an O2 -measure σm through
306 7 Context comparison and conceptual measurement

1 if g I m

σm (g) := ,
2 if (g, m) ∈
/I

then, because of σm −1
(1) = m0 , the product measure is full.
In this argument we have only made use of the fact that in O2 there are two
objects g, h with g 0 ⊂ h0 and g 0 6= h0 , i.e., that the context is not atomistic in the
sense of the definition on Page 45. Hence, the first alternative only does not occur
in the case that all scales in S are atomistic. Therefore, it only remains to be shown
that every atomistic scale is fully nominally measurable itself. This follows from
Proposition 138 below. 

Concept lattices of atomistic contexts are atomistic, and every atomistic com-
plete lattice is isomorphic to the concept lattice of an atomistic context. The con-
text property “atomistic” means precisely that the extents of the object concepts
form a partition of the object set. It is inherited by semiproducts and by the
preimages under full scale measures: A context which can be fully measured
into a semiproduct of atomistic scales has to be atomistic itself. A reduced con-
text is atomistic if and only if from (g, m) ∈
/ I, g . m always follows.
Definition 108 An extent U of a context K = (G, M, I) is called n-valent if G \ U
is the disjoint union of n−1, but not of fewer extents. If K is atomistic, every extent
has a valence. In this case we define the valence of a set of extents as the supre-
mum of the valences involved. The attribute-valence VM (K) of an atomistic con-
text is the valence of the set of attribute extents, provided that K is reduced. In the
general case we say that an atomistic context K has attribute-valence VM (K) ≤ n
if and only if there is an infimum-dense set of concepts of K whose extents all
have a valence ≤ n. ♦

Proposition 138 A context is fully n-valued nominally measurable if and only if it has
attribute-valence ≤ n. Every atomistic context is fully nominally measurable.

Proof Every Nn -measure of K induces a partition of the object set into no more
than n extents, i.e., all those extents have a valence ≤ n. Hence, according to
Proposition 133 we find, for every full measure into a semiproduct of Nn -scales,
an infimum-dense set of concepts whose extents all have a valence ≤ n. If, on
the other hand, K has such a set of concepts, it is possible, by means of the same
proposition, to construct a product measure with the desired properties: An Nn -
measure can be assigned to every partition of the object set in at most ≤ n classes
by mapping the classes of the extent partition onto different objects of the nom-
inal scale.
An atomistic context (G, M, I) has an attribute-valence ≤ |G|, i.e., it is certainly
fully N|G| -measurable. 
There has been little investigation as to which “measurability classes” there
are within the class of the atomistic scales. A first clue can be obtained if we also
define a valence for objects g ∈ G: g has valence n, if there are n − 1 objects
g1 , g2 , . . . , gn−1 (but no more) with the property that g, g1 , . . . , gn−1 generate the
7.4 Measurability theorems 307

same concept pairwise: {g, g1 }00 = {g, g2 }00 . . . = {g1 , g2 }00 . . . = {gn−2 , gn−1 }00
6= g 00 . VG (K) denotes the supremum of the valences of objects of K.
Proposition 139 If S consists of atomistic scales and if n is a natural number, the fol-
lowing statement is true: Nn is fully S-measurable if and only if S contains a scale S of
the object-valence VG (S) ≥ n.
Proof Every -preserving map of B(Nn ), n > 2 into a lattice which is not injec-
W
tive maps two atoms onto comparable elements. Therefore, by Proposition 132,
every non-trivial measure of Nn , n > 2 into an atomistic context is injective and
thus full. If N2 is measured fully into a semiproduct of atomistic scales, then at
least one of the factors must separate the two objects, and we have: If Nn is fully S-
measurable, then there is a scale S ∈ S, suchW that Nn is fully S-measurable. Again
by Proposition 132, then there must be a -embedding of B(Nn ) into B(S), in
the case of which object concepts are mapped onto object concepts, i.e., S has an
object-valence ≥ n.
Conversely, in a scale which has an object-valence ≥ n, we also find an object
of the valence n and thus a measure mapping Nn injectively and fully onto S. 
From the last two propositions we draw a simple conclusion for a special case:
Proposition 140 If S consists of atomistic scales with an attribute-valence ≤ n and if S
contains a scale of the object-valence n (n ∈ N), then the following holds true: A context
K is fully S-measurable if it is fully n-valued nominally measurable. 
This already suffices to provide us with an overview over the measurability
classes with respect to the atomistic standard scales. We have for all n ≥ 2

VG (Nn ) = VM (Nn ) = n,

VG (In ) = VM (In ) = 2,
VG (Ncn ) = VM (Ncn ) = 2.
From this follows:
Proposition 141 For a context K the following statements are equivalent:
1. K is fully dichotomially measurable.
2. K is fully interordinally measurable.
3. K is fully contranominally measurable.

The proposition also makes it possible to characterize the concept lattices of
many-valued contexts which are scaled plainly by means of elementary scales.
For this purpose, we need another definition: We say that an element x of an
atomistic complete lattice L has valence ≤ n if there is an n-element subset T
of L which contains x and which has the property that each atom of L is less
than or equal to precisely one element of T . If, in the special case n = 2, the set
T = {x, y} has the property specified, then we call y a pseudo-complement of
x.
308 7 Context comparison and conceptual measurement

Theorem 65 Every complete lattice is isomorphic to the concept lattice of an ordinally

scaled many-valued context.
A complete lattice is isomorphic to the concept lattice of a nominally scaled complete
many-valued context, if and only if it is atomistic.
A complete lattice is isomorphic to the concept lattice of a nominally scaled complete n-
valued context if and only if it is atomistic and contains an infimum-dense set of elements
of a valence ≤ n.
For the special case n = 2, we obtain:
A complete lattice is isomorphic to the concept lattice of a nominally scaled complete 2-
valued context if and only if it is atomistic and contains an infimum-dense set of elements
with a pseudo-complement. This at the same time characterizes the concept lattices of in-
terordinally scaled many-valued contexts as well as the concept lattices of complementary
nominally scaled many-valued contexts. 

7.5 Notes, references, and trends

7.1

Not only isomorphisms but also other classes of maps can be represented as con-
cepts, this has been worked out by W. Xia [425]. Theorem 60 has been taken from
[135], compare also Ganter and Reuter [155].
With regard to the group-theoretical background of Figure 7.1 extensive in-
formation can be found in Kerber ([217], in particular Chapter 3). M. Zickwolff
(from whose publications [428], [429] the example has been taken), has worked
out which information has to be entered into the diagram of the orbits of a (finite)
lattice (or, more generally, a relational structure) L with automorphism group
Γ , so that L can be reconstructed from it. Zickwolff’s approach focused on rule
exploration, about which more can be found in Section 6.4 of [154]. Later, the topic
was taken up, deepened and expanded by D. Borchmann [59]. The presentation
in this book is based on [62], where more detailed examples and further results
can be found.
Borchmann [59] also introduces context foldings and has GAP-programs for
computing concept lattice foldings directly from such context foldings. T. Schlem-
mer [334] suggests to further simplify the annotation of lattice foldings by using
group generators.

7.2

Definition 102 follows, with slight modifications, the article [119] by Erné, from
which we have also taken Theorem 62 and Proposition 126 and which contains
a lot of additional information concerning this subject. The question to which
7.5 Notes, references, and trends 309

extent lattice morphisms can be represented by context maps so that a duality is

created has also been examined by G. Hartung [187].
J. van Benthem [44] has characterized the logical formulas (flow formulas) that
are preserved by such morphisms. His results were strengthened by S. Fefer-
man [124].
Atif et al. [6] connect the morphisms of Formal Concept Analysis to the no-
tions used in Mathematical Morphology.
Scale measures, i.e. extensionally continuous mappings, also lead to adjunc-
tions, as the Propositions 131 and 132 in Section 7.3 show. However, these have
the special property that object concepts are mapped onto object concepts. An in-
depth investigation of bonds and dual bonds with such properties can be found
in the M.Sc. thesis by M. Krötzsch [236] and the accompanying publication by
Krötzsch et al. [237].

7.3

Conceptual scales and conceptual measuring have first been discussed in [156].
Many results of this section can be found in different formulations in books such
as that of Blyth and Janowitz [52]. Otherwise, this section and the following one
make use of results taken from [159].
Hanika & Hirth [184] generalize Definition 105 by dropping the condition of
equal scale values. They introduce a finer-coarser relation on the set of all scale
measures (which start from a fixed formal context K) by calling a scale measure
σ1 finer than σ2 if every extent that is the preimage of an extent under σ2 is also
the preimage of an extent under σ1 . When this quasi-order is factorized by the
induced equivalence, the result is the complete lattices of the conceptual scalings of
K, which is isomorphic to the lattice of closure systems contained in the closure
system of all extents of K. This opens up the possibility of navigating between
possible scalings without having to commit to a conceptual scale beforehand.

7.4

We again refer to [159]. Parts of Theorem 65 are contained in [389].

Chapter 8
Contextual concept logic

The idea of concepts as the building blocks of human thought has been the object
of philosophical theories for centuries, as has logic as the doctrine of the forms of
thought. Formal Concept Analysis provides mathematical tools for the concep-
tual analysis of concrete facts and thereby supports human thinking and com-
munication. Of course, the rules of formal logic apply here as well, but additional
tasks arise. Therefore R. Wille and his co-workers have worked out a “Contextual
Logic” which provides a formal framework for Immanuel Kant’s “three essential
main functions of thinking”,
concepts – judgments – conclusions.
Here we discuss only the simplest parts of Contextual Logic in detail, begin-
ning with contextual attribute logic in Section 8.1. It mainly provides language
tools that allow attributes to be combined using logical connectives. A mathemat-
ically minded reader might object that this is merely reformulated Propositional
Logic, and in fact, we connect here to G. Boole, who gave the impulse in the mid-
dle of the 19th century. However, the language of contemporary mathematical
logic is noticeably shaped by its applications to mathematics itself. Boole spoke
of signs, classes, and of a universe of discourse, which in our understanding corre-
sponds to attributes, concept extents, and to the notion of a formal context. This does
not mean that the remarkable results of mathematical logic could be ignored. We
will use them whenever appropriate, but in doing so we will maintain a way of
speaking that is close to the logic of concepts. Logical Concept Analysis by Ferre
and Ridoux [128] offers an approach that is closer to the language of mathemat-
ical logic.
Concepts are not true or false, and formal concepts have no truth values. It
is therefore not obvious how a negation for (formal) concepts could be formu-
lated. Concept lattices are usually not complemented and therefore do not allow
a negation operation with the usual properties. Nevertheless, negation is dis-
cussed in the theory of concepts, and also the case where concepts are in oppo-
sition to each other. Wille [412, 413] has worked out how this can be expressed
for formal concepts. To map negation and opposition of concepts into Formal

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 311
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2_9
312 8 Contextual concept logic

Concept Analysis, he investigates two approaches: one uses generalizations of

formal concepts, that is, attenuations of the conditions that define a formal con-
cept. This is discussed, along with some results for the resulting mathematical
structures, in Section 8.2.
A second possibility is to lower the requirements for the properties of negation
and opposition to the point where a mathematical rendering becomes possible.
This leads in Section 8.3 to the concept algebras, whose mathematical theory is not
yet fully understood.
In order to give Contextual Logic sufficient expressiveness, relations must be
included in the theory. Several groups of authors have worked on this and have
arrived at differently formulated approaches. In Section 8.4, after an introduc-
tory example, we only give a few language elements and appetizing results. The
difficult problems that then arise can be solved using methods and algorithms
of Description Logics. However, these cannot be explained here.
A list of some of the different approaches has been moved to Section 8.5.

8.1 Contextual attribute logic

In addition to the attributes in the set M , one can consider compound attributes
formed by iteratively combining the attributes in M using the symbols ¬, , and
W
. The interpretation of when an object has a compound attribute follows the
V
standard definitions of propositional logic1 , with
¬m representing the negation of an attribute m ∈ M , with the extent being
V (¬m) := G \ m ,
0 0

T 0 the conjunction of the attributes in A ⊆ M , with extent

AV representing
( A) 0
:= {m | m ∈ A}, and
representing the disjunction of the attributes m ∈ B, with extent ( B)0 :=
W W
B
S 0
{m | m ∈ B}.
Iteration of the above compositions leads to further compound attributes, the
extents of which are determined in the obvious manner.
So an object g has the compound V attribute ¬m if and only if g does not haveW m, g
has the compound attribute A iff g has every attribute m ∈ A, and g has B iff
g has at least one of the attributes m ∈ B. Note that, unlike in propositional logic,
the sets A and B need not be finite. The compound attributes include, e.g., the
implications as introduced in the previous section,
V but also the trivial attributes
⊥ := ∅ (which now object has) and > := ∅ (which all objects have). The
W
same symbols are used for the bottom and the top element of a lattice, but we
see no danger of confusion.
Compound attributes can be added to the attribute set of a formal context: If
K := (G, M, I) and C is a set of compound attributes, then we call
1 One difference, which we skip in the following definition, is that in mathematical logic con-
junctions and disjunctions are formulated by default only for finite sets A, B .
8.1 Contextual attribute logic 313

(G, M ∪ C, I ∪ IC )

a Boolean enrichment of K, provided that IC realizes the above definitions of

the extents. The enriched context is the apposition

(G, M, I) | (G, C, IC )

and therefore has all concept extents of K and usually more, depending on the
choice of C. Context enrichment thus increases expressiveness, but also the num-
ber of formal concepts. It is usually not a good idea to add many or even all pos-
sible compound attributes because this can easily lead to a drastic increase in the
size of the concept lattice.
We must be careful when stating that two compound attributes “mean the
same”: we call such attributes
• extensionally equivalent over (G, M, I) if they have the same extent in this
formal context (G, M, I) and
• globally equivalent if they have the same extent in every context with at-
tribute set M .

Proposition 142 Two compound attributes of an attribute set M are globally equivalent
if and only if they are extensionally equivalent in the formal context

(P(M ), M, 3).

For this reason (P(M ), M, 3) is called the test context over M .

Clause logics of formal contexts

Since we have permitted iterations of the composition process, compound at-

tributes may be arbitrarily complex. Up to equivalence, however, they can be
represented by rather simple ones, the sequents (called clauses in propositional
logic):
Definition 109 A sequent (A, S) over M is a compound attribute, represented
by two subsets A, S ⊆ M , which is globally equivalent to the compound attribute
_
(S ∪ {¬m | m ∈ A}).

The extent of a sequent (A, S) consists of all objects g which have at least one of
the attributes in S or do not have at least one of the attributes in A. This is why
we sometimes find the notation
^ _
A→ S
314 8 Contextual concept logic

more intuitive, especially when examining the interaction of such sequents with
implications.
We will only consider disjoint sequents with A ∩ S = ∅, because every non-
disjoint sequent is globally equivalent to >. Sequents are naturally ordered by

(A1 , S1 ) ≤ (A2 , S2 ) : ⇔ A1 ⊆ A2 and S1 ⊆ S2 .

A sequent (A,S) is full, if S = M \ A. A full sequent has a simple extent:

Proposition 143 The extent of a full sequent (A, M \ A) in the test context is P(M ) \
{A}. 
It is immediate from this proposition that the extent of a conjuction of full
sequents (At , M \ At ), t ∈ T , is just P(M ) \ {At | t ∈ T }. This gives the following
well known result:
Theorem 66 (Conjunctive normal form) Every compound attribute is globally
equivalent to a unique conjunction of disjoint full sequents. 
A compound attribute is all-extensional in (G, M, I) if its extent is the set G
of all objects. The set of all all-extensional compound attributes may be called
the Boolean attribute logic of (G, M, I).
One can see from this way of speaking that we understand compound at-
tributes not only as “generalized attributes” but also as logical rules that may
or may not hold in a formal context. For example, the sequent (A, S) may be
interpreted as the implication
^ _
A→ S,

and in fact does the compound attribute A imply the compound attribute S
V W
if and only if (A, S) is all-extensional.
The clause logic of a formal context (G, M, I) is the set of all sequents which
are all-extensional in (G, M, I). A clause set, i.e., a set C of disjoint sequents over
M , is called regular if it satisfies the condition

(A, S) ∈ C ⇔ (A ∪ X, M \ (A ∪ X)) ∈ C for all X ⊆ M \ (A ∪ S) .

In words: (A, S) belongs to C iff every full sequent which contains (A, S) belongs
to C. An immediate consequence is that two regular clause sets containing the
same full sequents must be equal.
Theorem 67 A clause set is regular iff it is the clause logic of some formal context. Up
to clarification and isomorphism, this context is unique.
Proof It is easy to see that the clause logic of any formal context satisfies the
regularity condition. To construct a formal context for a given regular clause set
C, let {At | t ∈ T } denote the set of all first components of full sequents in C, and
let
G := P(M ) \ {At | t ∈ T }.
8.1 Contextual attribute logic 315

G is the extent of the conjunction of these full sequents in the test context. There-
fore these sequents are exactly the ones which are all-extensional in (G, M, 6 3),
and the clause logic of this formal context must be equal to C. For the unique-
ness, see below. 

Attribute logic can, of course, not distinguish between objects having the same
object intent, and also does not encode information about the object names. But
up to object clarification, any two formal contexts having the same clause logic
must be isomorphic, as is easily proven. Thus the clause logic of (G, M, I) com-
pletely describes the system of object intents, in the same way as the implication
logic describes the system of concept intents.
Note that the intersection of regular clause sets is regular, and that the regular
clause sets thus form a closure system. The corresponding context is, up to clari-
fication, just the subposition of the individual ones. A full sequent (X, M \ X) is
in the regular closure of (A, S) iff (A, S) ≤ (X, M \ X).

Implications and inconsistencies

It is sometimes helpful to replace the clumsy formulation “the extent of a com-

pound attribute c (or of the conjunction of a clause set C) in the test context” by
something shorter. We shall use the term free extent of c (or of C) instead.
Definition 110 Let A, S ⊆ M be nonempty sets of attributes and let m ∈ M be
some attribute. We call a sequent of the form
(A, ∅) an inconsistency (because the sets inconsistent with this clause are
precisely the sets containing A),
(∅, S) a disjunction (in accordance with the definition above),
(A, {m}) an implication with singleton conclusion and
({m}, S) a surmise relationship (cf. [101]).
♦
We say that S is the free extent of inconsistencies if S is the free extent of some
clause set consisting of inconsistencies only. For the other cases, the wording is
similar.
Proposition 144 A set S of subsets of M is
• an order ideal in (P(M ), ⊆) iff it is the free extent of inconsistencies,
• an order filter in (P(M ), ⊆) iff it is the free extent of disjunctions,
• a closure system on M iff it is the free extent of implications, and
• a kernel system on M iff it is the free extent of surmise relationships.

316 8 Contextual concept logic

The free extent of the union C1 ∪ C2 of two clause sets C1 and C2 is just the in-
tersection of the free extents of the two. This allows to combine the results of
the proposition to give further characterizations, e.g., that the convex subsets of
P(M ) are precisely the free extents of inconsistencies and disjunctions, etc.

Cumulated clauses

How do you find the most compact clause set possible for a formal context?
This question is well studied in mathematical logic (topic: “prime implicants”).
However, one should not expect such an elegant solution as for the implications.
When calculating with clauses, one quickly encounters intractable problems, i.e.
problems wich are NP-complete. On the other hand, computer science has de-
veloped algorithms –precisely for such calculations– (e.g., “SAT-solver”) that are
breathtakingly fast in practice. For formal contexts of moderate size it is therefore
quite possible to work with their clause logic.
Take a look at Figure 1.4 (page 30), the concept lattice to “Living beings and
water”. The largest element is attribute concept to the attribute “needs water to
live”, but it is not an object concept. So the intent of this concept is not an object
intent, and it does not occur as a “context row”. Instead, we see that every object
has one of the attributes a: needs water to live, g: can move around, b: lives on land, c:
lives in water, or d: needs chlorophyll. Using the abbreviations we can translate that
to saying that the sequent
({a}, {b, c, d, g}),
which can also be written as

a → b ∨ c ∨ d ∨ g,

is all-extensional in the formal context. Repeating the same for the attribute con-
cept for the attribute b: lives on land, we find that every such living being (of the
formal context) lives in water, needs chlorophyll, or can move around and has limbs.
What we get is the rule
c → b ∨ d ∨ (g ∧ h),
which is not a clause, but a combination of two clauses.
Such cumulated clauses, as we call expressions of the form
^ _^
A→ At ,
t∈T

show some surprising similarities to implications and are sometimes more intu-
itive than clauses alone. Since they include clauses, they have the same expres-
sivity, which at the same time means that they are algorithmically more complex
to handle than implications. We refer to Chapter 5 of [154] for details.
8.2 Pre-, proto-, and semiconcepts 317

8.2 Pre-, proto-, and semiconcepts

By weakening the definition of a formal concept, further interesting structures

can be assigned to a formal context. Intuitively, the following simple definitions
are particularly easy to understand if one remembers the way of thinking of
“rectangles” in a formal context (G, M, I). Here, a rectangle is to be understood
as a set A × B with A ⊆ G and B ⊆ M . A formal concept corresponds to a
maximum full rectangle, where the word full means that A × B ⊆ I holds, so in
crosstable language, every cell in A × B contains a cross. By omitting the con-
dition of being maximal, one gets preconcepts, so they simply correspond to the
full subrectangles. Requiring that one of the sets A and B is maximal, but not
necessarily the other, leads to the semiconcepts. And a full rectangle contained in
only one maximal full rectangle is a protoconcept. The following definitions spec-
ify this.
Definition 111 A pair (A, B) with A ⊆ G and B ⊆ M is called a preconcept
of the context (G, M, I) if A × B ⊆ I (which is equivalent to both A ⊆ B 0 and
B ⊆ A0 ). A protoconcept fulfills A0 = B 00 and B 0 = A00 . If A0 = B or B 0 = A,
this is called a semiconcept, see Figure 8.1, where a further distinction is made
between u-semiconcepts for A0 = B and t-semiconcepts for A = B 0 . ♦

A⊆G B⊆M

preconcept A ⊆ B 0 ( ⇔ B ⊆ A0 ) “rectangle”
protoconcept A0 = B 00 ( ⇔ B 0 = A00 ) A ⊆ B 0 B ⊆ A0
t-semiconcept A = B0
preconcept
u-semiconcept B = A0
concept A = B 0 and B = A0 A0 = B 00 B 0 = A00
protoconcept

A = B0 B = A0
t-semiconcept u-semiconcept

formal concept

Figure 8.1 Pre-, proto-, and semiconcepts.

While preconcepts are very general structural elements (just some objects
together with some attributes incident to them), protoconcepts (and thus also
semiconcepts) are quite natural for human understanding. It is not unusual to
describe a concept by some of its attributes and objects, if that is enough to de-
fine it unambiguously. It is not necessary to know the entire concept extent and
concept intents for this purpose.
318 8 Contextual concept logic

The set of all preconcepts of a formal context K is denoted by V(K),2 the set
of all protoconcepts by P(K). There are two natural ways to order preconcepts.
One ordering relation, noted as ⊆2 , is simply the containment of rectangles:

(A1 , B1 ) ⊆2 (A2 , B2 ) :⇔ A1 ⊆ A2 and B1 ⊆ B2 .

The maximal elements with respect to this order are exactly the formal concepts,
and the protoconcepts are precisely those less or equal to exactly one maximal
element.
The second definition generalizes the subconcept–superconcept order to pre-
concepts:
(A1 , B1 ) v (A2 , B2 ) :⇔ A1 ⊆ A2 and B2 ⊆ B1 .

The lattice of preconcepts

It is surprisingly easy to prove that the set of all preconcepts, when ordered by
v, is also a complete lattice, which is even distributive. The following proposi-
tion also shows that the preconcepts can efficiently be generated using the Next
closure algorithm.
Proposition 145 (from [417]) For a formal context K := (G, M, I), the ordered set
(V(K), v) of all preconcepts is a completely distributive complete lattice, which is iso-
morphic to the concept lattice of the formal context

V(K) := (G ∪˙ M, G ∪˙ M, I ∪ (6= \ (G × M ))).

Remark: To make the somewhat arcane definition I ∪ (6= \ (G × M )) of the in-

cidence relation better understandable, perhaps a look at the G M
graphic to the right will help, which symbolizes V(K). G and
M are assumed to be disjoint, the object and the attribute set G 6= I
of the new formal context are both equal to G ∪˙ M , and the
M × 6=
incidences are as follows: An object from G, understood as an
object of the new context, is incident to all elements from G except to itself, and to
the attributes from M to which it is incident in K. An element of M , understood
as an object, is incident to all elements of G and of M , except to itself.
Proof of Proposition 145 For (At , Bt ) ∈ V(K), t ∈ T , we obviously have
\ [
inf (At , Bt ) = ( At , Bt )
t∈T
t∈T t∈T

and

2 The letter V (Fraktur V ) resembles the German syllable „vor“, which translates to “pre”.
8.2 Pre-, proto-, and semiconcepts 319
[ \
sup(At , Bt ) = ( At , Bt ).
t∈T
t∈T t∈T

Hence (V(K), v) is a complete sublattice of the direct product of the two pow-
erset lattices (P(G), ⊆) and (P(M ), ⊇) and therefore completely distributive.
This proves the first assertion of the proposition. For the second one we consider
the assignment
ι
(A, B) 7→ (A ∪ (M \ B), (G \ A) ∪ B).
It can easily be checked that (A∪(M \B), (G\A)∪B) is a formal concept of V(K)
whenever (A, B) is a preconcept of K, and that the mapping ι is one-to-one. To
see that it is also onto, consider an arbitrary formal concept (C, D) of V(K), and
let A := C ∩ G and B := D ∩ M . From the definition of the incidence relation of
V(K) one concludes that A × B ⊆ I, and thus that (A, B) is a preconcept of K.
Moreover, C = A ∪ (M \ D) and D = B ∪ (G \ C), but since M \ D = M \ B and
ι
G \ C = G \ A, this shows that (A, B) 7→ (C, D). It remains to show that ι is not
only a bijection, but also an order-isomorphism, and this is easy:

(A1 , B1 ) v (A2 , B2 ) ⇔ A1 ⊆ A2 and B1 ⊇ B2

⇔ A1 ∪ (M \ B1 ) ⊆ A2 ∪ (M \ B2 ) and
(G \ A1 ) ∪ B1 ⊇ (G \ A2 ) ∪ B2
⇔ ι(A1 , B1 ) ≤ ι(A2 , B2 ).

male female old young

father × ×
mother × ×
son × ×
daughter × ×

Figure 8.2 A formal context of family members

Example Naturally, the number of preconcepts is usually large. We therefore

discuss a rather small example, namely the context in Figure 8.2 (its concept
lattice is isomorphic to the upper lattice in Figure 1.12 on p. 43). There are 47
preconcepts, ten of which are formal concepts, as shown in Figure 8.3. >

Example For the contranominal scale Ncn , which we abbreviate with the symbol
6= , the context for the preconcepts takes a striking special form,

6= 6=
.
× 6=
320 8 Contextual concept logic

old young

male female

father daughter

son mother

Figure 8.3 Preconcept lattice for the formal context of family members in Figure 8.2. The
black nodes represent the ten formal concepts. The labelling is not the standard one de-
fined in Section 1.1. Not all irreducible elements are labelled. But it allows, with the usual
reading rule, to read off the sets A and B for each preconcept (A, B). Most of the nodes
in this diagram correspond to preconcepts (A, B) for which A or B are empty (31 of 47).
These form the two 16-element Boolean sublattices, one containing the smallest and the
other containing the largest element. Eight of the remaining elements are formal concepts.
The other eight correspond to preconcepts of size 1 × 1. The semicircles mark the proper
semiconcepts. There are no other protoconcepts in this example.

We already know (see the text following Definition 32 on p. 44) that its concept
lattice is isomorphic to the order relation of the contranominal scale, and more-
over we can infer from Proposition 18 on p. 45, that this context is isomorphic to
the context sum of n copies of the standard context of the three element chain.
Indeed, the order relation of the powerset of an n-element set consists of 3n pairs
and is isomorphic to the n-th power of a three-element chain. >
8.2 Pre-, proto-, and semiconcepts 321

Proposition 145 does not precisely charaterize the preconcept lattices, but this
is not an open problem, because Burgmann and Wille give a Basic Theorem on
preconcept lattices. We omit the technical details and instead quote their charac-
terization without proof.
Theorem 68 (from [68]) The preconcept lattices are (up to isomorphism) the com-
pletely distributive complete lattices in which the supremum of all atoms is equal or
greater than the infimum of all coatoms. 
The lattice of preconcepts is useful because it provides an approach in the the-
ory of Formal Concept Analysis. Well-known algorithms such as Next closure
can thus be applied to preconcepts. However, this lattice does not represent a
generalization of the concept lattice to preconcepts in that the arithmetic opera-
tions do not coincide. The concept lattice is indeed order-embedded in the lattice
of preconcepts, but not as a sublattice. In the preconcept lattice, join and meet of
formal concepts are not necessarily formal concepts again. Therefore, another al-
gebraic structure is introduced next, whose elements are again the preconcepts,
but with other algebraic operations.

The algebra of preconcepts

Definition 112 The preconcept algebra V(K) of a formal context K is the set
¬
V(K) endowed with operations u, t, ¬, , ⊥, and >, defined as follows:

(A1 , B1 ) u (A2 , B2 ) := (A1 ∩ A2 , (A1 ∩ A2 )0 ) meet

(A1 , B1 ) t (A2 , B2 ) := ((B1 ∩ B2 )0 , B1 ∩ B2 ) join
¬(A, B) := (G \ A, (G \ A)0 ) negation
¬
(A, B) := ((M \ B)0 , M \ B) opposition
⊥ := (∅, M ) nothing
> := (G, ∅) all.

♦
Two things stand out immediately: First, each result of these operations is a semi-
concept, so the semiconcepts form a subalgebra of each preconcept algebra. And
second, for formal concepts meet and join coincide with the concept lattice op-
erations. Generally for preconcepts, however, these operations are not lattice op-
erations and, except in special cases, do not have the usual properties. Meet and
join are not idempotent, double negation is not identity, etc. For details we refer
to the literature, in particular to Wille [417], where the equational theory of such
algebras is studied and finally a complete characterization is proved in terms of
a Basic Theorem on preconcept algebras (and for protoconcept algebras as well).
322 8 Contextual concept logic

Double Boolean algebras

Even if we restrict ourselves to the subalgebra of semiconcepts, the operations

introduced in Definition 112 still have some surprising properties. Join and meet
are still not idempotent, which is quite obvious because the u operation always
results in a u-semiconcept. For a proper t-semiconcept (B 0 , B) (i.e., for one
which is not a formal concept), (B 0 , B) u (B 0 , B) cannot be equal to (B 0 , B).
Nevertheless, the algebra H(K) of a formal context K semiconcepts is an in-
teresting algebraic structure called a double Boolean algebra. This name is sugges-
tive for the following reason: The set H(K) of all semiconcepts of K splits into
the set Hu (K) of all u-semiconcepts and the set Ht (K) of all t-semiconcepts.
u-semiconcepts are exactly the pairs of the form (A, A0 ), where A ⊆ G, and
(A1 , A01 ) v (A2 , A02 ) holds exactly if A1 ⊆ A2 (since that implies A01 ⊇ A02 ).
The ordered set (Hu (K), v) is therefore naturally isomorphic to the powerset
lattice (P(G), ⊆). And by analogous argument is (Ht (K), v) isomorphic to the
Boolean lattice (P(M ), ⊇). Thus, the ordered set (H(K), v) of all semiconcepts
is the union of two Boolean lattices. These are not disjoint, because their inter-
section is exactly the set of all formal concepts of K. The semiconcept algebra
is called a pure double Boolean algebra because the two Boolean parts cover all
elements. The algebra of protoconcepts is also a double Boolean algebra, but not
pure in general.
Definition 113 A double Boolean algebra is a universal algebra
¬
(D; u, t, ¬, , ⊥, >)

of type (2, 2, 1, 1, 0, 0) that satisfies all equations which hold in all protoconcept
algebras.3 A double Boolean algebra is called pure, if

x u x = x or x t x = x

holds for each x. ♦

For protoconcept algebras, it is clear that the condition x u x = x is satisfied
only by u-semiconcepts and that x t x = x holds only for t-semiconcepts. How-
ever, it can be shown that this holds for all double Boolean algebras:
¬
Theorem 69 (see [381]) If (D; u, t, ¬, , ⊥, >) is a double Boolean algebra and if
Du := {x ∈ D | x u x = x}, then

(Du ; u, t, ¬, ⊥, >)

is a Boolean algebra for x t y := ¬(¬x u ¬y) and > := ¬⊥, and

¬
(Dt ; u, t, , ⊥, >)

3 An explicit equational base is given in [412].

8.3 Concept algebras 323
¬¬ ¬
is a Boolean algebra for Dt := {x ∈ D | x t x = x}, x u y := ( x t y), and
¬
⊥ := >. 
Note that in a protoconcept algebra, ⊥ is just the smallest formal concept and >
is the largest one.
Double Boolean algebras are more general than protoconcept algebras. Not
every double Boolean algebra is isomorphic to some protoconcept algebra. A
complete characterization has been given by Vormbrock and Wille [381] in the
form of Basic Theorems for proto- and for semiconcept algebras.

8.3 Concept algebras

In the previous section, a mathematization of “negation” and of “opposition”

for formal concepts was established in Definition 112. When these operations
are applied to a formal concept, the result is usually not a formal concept but a
semiconcept. Each semiconcept is a protoconcept and therefore can be extended
to formal concept in an unambiguous way. It seems natural to combine these
two steps, i.e. to first apply the negation or opposition operator and then to com-
plete the resulting semiconcept to a formal concept. This leads to two new unary
operations on arbitrary concept lattices B(G, M, I).
Definition 114 For every formal concept (A, B) its weak negation is defined by

(A, B)M := ((G \ A)00 , (G \ A)0 ),

and its weak opposition by

(A, B)O := ((M \ B)0 , (M \ B)00 ).

The concept algebra of a formal context K then is

A(K) := (B(K); ∧, ∨, M , O , 0, 1),

where ∧ and ∨ denote meet and join in the concept lattice, and 0 is the least and
1 the largest concept of K. ♦
Note that weak negation and weak opposition can be expressed in terms of
the operations in Definition 112: For each formal concept x := (A, B) one has
¬ ¬
xM = (¬x) t (¬x) and xO = ( x) u ( x).

Concept algebras are concept lattices with two additional complement opera-
tions. Therefore we speak of dicomplemented lattices. However, an algebraic defi-
nition is usually done by equations or at least quasi-equations, but these are not
yet clear. Therefore, for the time being, we define as follows:
324 8 Contextual concept logic

Definition 115 (from4 [252]) A dicomplemented lattice is a bounded lattice

with two additional unary operations which satisfies all equations and quasi-
equations that hold in all concept algebras. ♦

To understand the algebraic properties of concept algebras, a Basic Theorem

on concept algebras is to be aimed at, but this goal has not been achieved so far.
Such a theorem would have to describe exactly which dicomplemented lattices
are isomorphic to concept algebras.
One can name steps towards this goal. First, the class of dicomplemented lat-
tices is to be described algebraically, preferably by means of equations, if neces-
sary also by quasi-equations. In a second step, conditions must then be found
which ensure that a given dicomplemented lattice is isomorphic to a concept al-
gebra. Since concept algebras arise from concept lattices, the condition of being
complete is certainly one of them5 . A first open problem we encounter on the
outlined path is the following:

Can the class of dicomplemented lattices be defined by equations, and if so, what is
a defining set of equations?
Some simple laws of concept algebras are compiled in the next proposition,
which is easily derived from the definitions of weak negation and opposition:

Proposition 146 The following hold in every concept algebra for all elements x and y:

(1) xMM ≤ x (10 ) xOO ≥ x

(2) x ≤ y ⇒ xM ≥ y M (20 ) x ≤ y ⇒ xO ≥ y O
(3) (x ∧ y) ∨ (x ∧ y M ) = x (30 ) (x ∨ y) ∧ (x ∨ y O ) = x.


All six conditions can be formulated as equations (for example, (2) is equiva-
lent to xM ∨(x∨y)M = xM ). These six conditions therefore might give an equational
base for the theory of dicomplemented lattices, but it is unknown if they do. As
a precaution, therefore, lattices that meet these conditions are given a different
name.
Definition 116 A weakly dicomplemented lattice is a bounded lattice equipped
with two unary operations M and O, satisfying the conditions (1)–(3) and (1’)–
(3’), listed in Proposition 146. ♦

A second problem, also unsolved, is

Is every weakly dicomplemented lattice dicomplemented?
From the harmless looking rules of Proposition 146, numerous other ones can be
derived. Some are compiled in the proposition below.

4 This definition is slightly different from and supersedes the one in [412].
5 but does not suffice, as was shown by Kwuida & Machida [255].
8.3 Concept algebras 325

Proposition 147 Each weakly dicomplemented lattice satisfies the following equations:

(4) (x ∧ y)M = xM ∨ y M (40 ) (x ∨ y)O = xO ∧ y O

(5) x ∨ xM = 1 (50 ) x ∧ xO = 0
(6) (x ∧ xM )M = 1 (60 ) (x ∨ xO )O = 0
(7) 0M = 1 = 0O (70 ) 1O = 0 = 1 M
(8) xOOO = xO ≤ xM = xMMM
(9) xMO ≤ xMM ≤ x ≤ xOO ≤ xOM .

Proof From x ∧ y ≤ x and x ∧ y ≤ y we get with (2) that (x ∧ y)M ≥ xM ∨ y M .

If u is an upper bound of xM and of y M , then uM ≤ xMM ∧ y MM ≤ x ∧ y by (1).
Thus (x ∧ y)M ≤ uMM ≤ u, and (4) is proved. The dual argument proves (4’). (5)
is obtained from (3) when 1 is substituted for x. To obtain (6), replace y by xM
in (4) and then apply(5). The first identity of (7) follows from (5), the second
by deriving 0O = 1OO from (7’) and then using (1’). The first equality of (8)
follows from the combination of (1’) and (2’), and replacing x by y O in (3) gives
y O ≤ y M , using (5’). With y := xM this gives the first inequality in (9), while the
second is just (1). 
The equations of Proposition 147 can be found in [252], together with many
others. It follows, among other things, that the mapping x 7→ xMM is an interior
operator and the mapping x 7→ xOO is a closure operator. However, other obvious
laws do not apply in general, as the following result exemplifies:
Theorem 70 (Kwuida [253]) For a doubly founded weakly dicomplemented lattice
(L; ∧, ∨, M, O, 0, 1) the following conditions are equivalent:
• xM = 1 ⇒ x = 0 and xO = 0 ⇒ x = 1 for all x ∈ L, and
• M = O.
In this case, (L; ∧, ∨, M, 0, 1) is a Boolean algebra. 
Kwuida’s result also makes understandable why concept algebras, as general-
izations of Boolean algebras, require dicomplementation and cannot get by with
only one unary complementation. See also [254].

Partial results

At the time of the second edition of this book, the representation issues for
dicomplemented lattices are open. Remarkable advances by Kwuida, however,
provide clues.
The question whether the quasi-equational class of dicomplemented lattices
is definable by equations alone refers to an important result of universal algebra,
namely to a theorem of Birkhoff that a class of algebras (of fixed signature) is
equationally definable if and only if it is closed under forming subalgebras, direct
products and homomorphic images. We had already mentioned that concept
326 8 Contextual concept logic

algebras cannot form an equational class because the lattice property of being
complete cannot be expressed by equations. But at least parts of the conditions
are fulfilled:
Theorem 71 (Kwuida [252])
• Direct products of concept algebras are isomorphic to concept algebras.
• Complete subalgebras of concept algebras are concept algebras.
• Homomorphic images of doubly founded concept algebras are isomorphic to concept
algebras.
For a proof we refer to the cited literature. 
The weakly dicomplemented lattices form an equational class. We do not
know whether this is identical to the class of dicomplemented lattices. If this
can be shown, then the first open problem is also answered.
Note that a given lattice may have many weak dicomplementations. Kwuida
studied this for the case of the direct product of a two-element chain with an
n-element chain and showed that this lattice carries n2 weak dicomplementa-
tions, all of which are indeed dicomplementations. In other words, for each of
the n2 weak dicomplementations of this 2n-element lattice L, there exists a for-
mal context K whose concept algebra A(K) is isomorphic to L, equipped with
this dicomplementation.
Kwuida shows that this extends to arbitrary finite distributive lattices:
Theorem 72 (Kwuida [252]) Every finite distributive weakly dicomplemented lattice
is isomorphic to a concept algebra. 

8.4 Contextual logic with relations

The example chosen to introduce the topic of this section comes (with some mod-
ifications) from the dissertation by S. Rudolph [323]. There it was used to sug-
gest the possibility of process exploration, i.e., of applying attribute exploration to
processes.
Example (Media player) A simple media player is considered. As visible at-
tributes, symbols are used to indicate which of six states the player is in:

“ready to start”, “playing”, “ready to continue”

“playback has ended”, “rewinding”, and “forwarding”.

Figure 8.4 shows a formal context with some obvious attributes for these states,
along with its simple concept lattice.
However, it is possible to influence the states. The media player provides four
commands for this purpose, which are also represented by symbols:
8.4 Contextual logic with relations 327

not playing

forwarding
rewinding
playing

idle
not playing

× ×
idle rewinding forwarding playing
×

× ×

× ×

× ×

× ×

Figure 8.4 States of a simple media player

“play”, “stop”, “forward”, “rewind”.

We add a hourglass symbol, , to indicate that no command is given for an
unspecified period of time. This is used as if it were a command:
“no action”.
The effects of these commands on the states are listed in Figure 8.5.
How can these commands be treated conceptually? Obviously, they are not
attributes in the usual sense, so they cannot be simply inserted into the formal
context. But they can be used for describing sets of objects. The first column of
the table in Figure 8.5 indicates that the set of states for which the command
causes the device to play is just
 
, , ,

which is not an extent of the formal context in Figure 8.4. We have thus intro-
duced a new attribute that can be called “ playing after ”, and can be symbol-
ized by
.playing
and which is defined to have the extension

0
.playing = {g ∈ G | changes g to h and (h, playing) ∈ I }.

This is, however, too special and only works, because is functional, meaning
that h is uniquely determined by g. The command is different, since the last
328 8 Contextual concept logic

Command given

Initial state State after command

or

or

or

Figure 8.5 The effect of the possible commands on the state of the media player.

column in Figure 8.5 in some cases has more than one entry per row. We then
must be more precise and specify, which of the possible outcomes is meant. For
a simpler notation, let us write
g h
to express that g, h ∈ G are objects such that h is one of the possible outcomes
when the command is given while the player is in state g. We introduce two
possibilities, ∃ .m and ∀ .m, with the extensions

0
∃ .m := {g ∈ G | ∃h∈G (g h and (h, m) ∈ I)}

0
∀ .m := {g ∈ G | ∀h∈G (g h implies (h, m) ∈ I)}.

0
In the above example, ∃ .idle consists of all states, as the media player al-
ways enters an idle state after a sufficiently long wait. After a short wait, however,
it may be that the previous state of playing, rewinding or forwarding has not yet
been completed. Thus

0 
∀ .idle = , ,

in this example. Adding such attributes for all possible combinations can enlarge
the concept lattice considerably, even in this tiny example, see Figure 8.6. >
8.4 Contextual logic with relations 329

not playing
∀ .rewinding
∀ .f orwarding
∀ .idle ∀ .playing
idle

∀ .idle

rewinding playing

Figure 8.6 The concept lattice of the formal context from Figure 8.4, extended by relational
compound attributes. Only one name is given to each irreducible attribute.

In the example, we talked about “states” and “commands” that lead from one
state to another. However, the approach of relational concept analysis is more
general and not only suitable for processes, but only requires that additional re-
lations are given on the objects of a formal context. Recall the very first definition
in this book (Definition 1 on p. 1). There we had explained what a binary relation
on a set G is, namely a subset of G × G, i.e., a set of pairs of elements of G. This
can be generalized to k-ary relations on G, which are subsets R ⊆ Gk , i.e., sets
of k-tuples with entries from G.
Definition 117 A formal context K := (G, M, I) together with6 a set R of rela-
tions on G is a relational formal context. ♦
Contextual logic handles relations of all arities, but binary relations, i.e., relations
of arity 2, are the most frequently used ones. In this section we therefore only
deal with the case of binary relations and call them roles, which follows the way
of speaking that is common in Description Logics. Thus, a role is the same as a
binary relation on the object set.
For a relational context, further compound attributes can be defined in addi-
tion to those discussed in Section 8.1:

6 Admittedly, the definition remains somewhat vague at this point, since we do not specify a
notation for relational contexts. However, this is intentional, as there are different notations in
the literature, see Section 8.5.
330 8 Contextual concept logic

Definition 118 Let K := (G, M, I) be a formal context and let R ⊆ G × G be a

role, i.e., a binary relation on G. For an attribute m ∈ M and a set B ⊆ M of
attributes we define the terminological7 attributes ∀R.m and8 ∃R.B to have the
extents

(∀R.m)0 := {g ∈ G | ∀h∈G (g, h) ∈ R implies h I m }

(∃R.B)0 := {g ∈ G | ∃h∈G (g, h) ∈ R and h ∈ B 0 }.

For B = {m} this gives

(∃R.m)0 := {g ∈ G | ∃h∈G (g, h) ∈ R and h I m }.

To make these terminological attributes more intuitive, a simple pizza example

may help: for a pizza to be vegetarian, all the toppings must be vegetarian, so the
pizza must have the attribute

∀has-topping.vegetarian.

To make a pizza non-vegetarian, a single topping is sufficient, giving it the at-

tribute
∃has-topping.non-vegetarian.
Note that it is not necessary to introduce ∀R.B, since that would have the same
extent as {∀R.m | m ∈ B}.
It is common to understand the definition of terminological attributes as re-
cursive. This means that nested attribute definitions such as

∀R2 .(∃R1 .m)

are allowed (with the parenthesis usually omitted). Staying with the pizza ex-
ample: If one of the toppings contains bacon, i.e., if the pizza has the attribute

∃has-topping.∃contains.bacon

then it is already non-vegetarian.

The recursive definition allows the construction of infinitely many terminolog-
ical attributes, and one can easily give examples of contexts where there are also
infinitely many distinct extents of terminological attributes. However, this is only
possible for infinite sets of objects, since a finite set has only a finite number of
subsets. But even in the finite it is often necessary to limit complexity. The next
definition serves this purpose.
Definition 119 The role depth of an attribute m is defined as follows:
• If m contains no quantifier symbols, then its role depth is zero.
7 Description Logics were previously also called Terminological Logics.
8 ∃R.B may be regarded as an abbreviation of ∃R.∧B .
8.4 Contextual logic with relations 331

• If m is a compound attribute formed by a Boolean combination of attributes

mt , t ∈ T , then its role depth is the supremum of the role depths of the mt .
• The role depth of ∃R.m and the role depth of ∀R.m both equal the role depth
of m, plus 1. ♦

Description Logics are a family of logics with different expressiveness. Im-

pressively fast algorithms, based on solid theory, have been developed for these
logics, and the translation presented in this section gives hope that Formal Con-
cept Analysis can benefit from these. This has indeed been demonstrated by the
publications cited below. Naturally, the fastest algorithms exist for the least ex-
pressive logics. Each increase also increases the complexity, and one soon reaches
undecidable logics, i.e. logics for which it is impossible to answer certain ques-
tions algorithmically.
Each individual description logic has a sequence of symbols as its name, often
written in calligraphic characters. We will mention three of them, ALC, EL, and
FLE.
Definition 120 From a relational formal context with attribute set M and role
set R one obtains
ALC attributes by adding > and ⊥ to M and recursively applying negation,
finite conjunctions and disjunctions, and the formation of ∃ R.m, ∀ R.m for
previously obtained attributes and roles R ∈ R.
EL attributes by adding > to M and recursively applying finite conjunctions
and the formation of ∃R.m for previously obtained attributes and roles R ∈ R.
FLE attributes by adding > to M and recursively applying finite conjunctions
and the formation of ∃R.m, ∀R.m for previously obtained attributes and roles
R ∈ R. ♦
An ALC-enriched, EL-enriched, or FLE-enriched context is obtained from a
relational context by adding ALC attributes, EL attributes, or FLE attributes,
respectively.

Terminological attribute logic, as designed by Prediger [306], has provided the

language for handling terminological attributes and is based on ALC. See also
[154] for an introduction. Rudolph [323] and later Sertkaya [340] have carefully
worked out the possibility of extending attribute exploration to terminological at-
tributes. This poses a very fundamental problem: the ability to concatenate roles
means that an infinite number of terminological attributes can be created even
from a finite set M . Therefore, it is not clear from the outset that an attribute ex-
ploration will terminate at all, that it will come to an end after a finite number of
steps. Rudolph proves this for FLE attributes of bounded role depth, and pro-
vides the necessary algorithms. Distel [97] does not restrict the role depth, but
uses the less expressive description logic EL, which is however well suited for
conceptual data.
We are not aware of mathematical results on the contextual logic of termi-
nological attributes that correspond Proposition 142 in Section 8.1. We have no
332 8 Contextual concept logic

“test context” for terminological attributes, and thus no obvious way to decide
whether two such attributes are globally equivalent, an indispensable prerequi-
site for attribute exploration.
At this point, however, Formal Concept Analysis can benefit from the results
obtained for description logics. There, the satisfiability problem for the respective
logic is studied both theoretically and algorithmically. Many decision questions
that arise for terminological attributes can be reformulated as satisfiability prob-
lems, e.g. the question whether such an attribute follows (globally) from other
attributes. Typical results characterize the computational complexity of satisfia-
bility and the development of powerful reasoners, which can solve the problem
effectively. The cited authors have therefore pursued the goal of establishing con-
nections from Formal Concept Analysis to Description Logics that make these
reasoners accessible.
Sebastian Rudolph has pointed out (in personal communication) that it is per-
fectly possible to construct relational test contexts using description logics. How-
ever, these are infinite and inferior to reasoners for testing global equivalence.

8.5 Notes, references, and trends

8.1

This section is based on [162]. The combination of implications (with one-

element premises) and inconsistencies leads to “truncated distributive lattices”
[416], which often occur as conceptual scales. For incomplete formal contexts,
Obiedkov [294] suggests using modal logic formulas.

8.2

The notion of preconcepts occurs in Stahl & Wille [348], that of semiconcepts in
Luksch & Wille [268], and protoconcepts are discussed in [412]. Proofs of the
“basic theorems” for semi- and for protoconcept lattices can be found in Vorm-
brock & Wille [381], for preconcept lattices in Burgmann & Wille [68]. Figure 8.3
is from [417].

8.3

Concept algebras were introduced in [412], where also a series of results can be
found. Later results are largely contained in Kwuida’s dissertation [252] and the
publications based on it.
8.5 Notes, references, and trends 333

8.4

We try to summarize important developments and results related to this section

in the “Research developments” report below.

Research developments

Contextual logic and Description Logics

In 1996, the first version of this book (in German) made it visible that by then an
extensive mathematical basis of Formal Concept Analysis had been worked out.
Rudolf Wille therefore turned more to the question of how this theory could
be embedded in a larger framework and began to develop a contextual logic.
Here, “logic” is to be understood as the doctrine of the forms of human thought,
and the addition of “contextual” refers to the goal of uncovering the regulari-
ties of the data at hand in each case, of the “context”. Wille was referring to the
philosopher Immanuel Kant, who taught elementary logic as “the theory of the
three main essential functions of thinking – concepts, judgments, and conclusions”,
with concepts as the basic units of thought, judgments as combinations of con-
cepts and conclusions as entailments between judgments, see [414] and Predi-
ger [305]. Was it possible to complement this philosophical approach, developed
over centuries, with a theory of formal concepts, formal judgments, and formal
conclusions? To mathematize the formal judgments, concept graphs were intro-
duced, a mathematically adapted variant of J. Sowa’s conceptual graphs [346].
A first systematic presentation of the approach is Prediger’s dissertation [304].
However, a comprehensive account of contextual logic does not yet exist, except
for the condensed “summary” in [414].
Description Logics (DL) are a family of logics that restrict the full expressive
power of predicate logic in favor of practicality for knowledge representation
and knowledge processing. There is far advanced theory and extremely effective
algorithms for this. They are also the basis of the web ontology language OWL.
And they offer a natural language for a relational extension of Formal Concept
Analysis. There was early interest in combining description logics and Formal
Concept Analysis. F. Baader, editor of the Description Logic Handbook [8], used
FCA methods as early as 1995 in his research [7]. Together with his students
and co-authors, he has consistently continued these efforts. This has resulted in
numerous publications with substantial results. Of particular note are the disser-
tations by S. Rudolph [323] B. Sertkaya [340], F. Distel [97], D. Borchmann [60],
and F. Kriegel [232, 233], but also Baader & Molitor [10], Baader & Sertkaya [11],
Baader et al. [9]. Other contributions come e.g. from Stumme [360].
Conversely, there are numerous efforts to enrich Formal Concept Analysis
with elements of DL languages. Prediger & Stumme [307] use DL for introduc-
ing logical scaling, and Prediger [306] works out a terminological attribute logic.
334 8 Contextual concept logic

Rouane-Hacene et al. [322, 196] suggest to combine methods from the two fields.
An early paper on formal contexts with a relational structure is due to Priss [308].
For a contextual logic, as introduced by R. Wille, the standard version of For-
mal Concept Analysis mathematizes the part of Contextual Concept Logic. But its
expressive power of the Formal Concept Analysis is not sufficient for Contex-
tual Judgment Logic of Conclusion Logic. To capture the logical content of concept
graphs, Wille [415] introduced power context families, which encode relations
of arbitrary arity.

Graph contexts and RCA

Other authors chose different approaches. Ferre & al. [130] use predicate logic,
Ferre & Cellier introduce Graph-contexts for which the object set is replaced by
words of objects to describe knowledge graphs (both formalisms allow for re-
lations of arbitrary arity). Several authors restrict to binary relations. Rudolph
uses binary power context families, Huchard and her coauthors, in development
of a relational concept analysis (see [196] for an early and [290] for a more re-
cent contribution), use relational context families, where each relation is given by
its incidence matrix and has a domain and a range. Their data type is similar to
what Dörflein & Wille [104] call a coherence network and for which they have even
proved a “Basic Theorem”.
For our presentation in Section 8.4, the choice of the data type is of secondary
importance.
References

[1] Albano, A.: Polynomial growth of concept lattices, canonical bases and
generators: extremal set theory in Formal Concept Analysis. Ph.D. thesis,
TU Dresden (2017). Qucosa Dresden → 113, 250, 279
[2] Albano, A., Chornomaz, B.: Why concept lattices are large: extremal the-
ory for generators, concepts, and VC-dimension. International Journal of
General Systems 46(5), 440–457 (2017). → 63, 251, 252
[3] Almazaydeh, A., Behrisch, M., Vargas-García, E., Wachtel, A.: Arrow re-
lations in lattices of integer partitions. Preprint arXiv:2403.07217 (2024).
→ 71
[4] Almeida, J., Bordalo, G., Dwinger, P.: Lattices, Semigroups, and Universal
Algebra. Plenum Press, New York–London (1990). doi:10.1007/978-1-
4899-2608-1. → 348
[5] Armstrong, W.W.: Dependency structures of data base relationships. In:
IFIP congress, vol. 74, pp. 580–583. Geneva, Switzerland (1974). → 94
[6] Atif, J., Bloch, I., Distel, F., Hudelot, C.: Mathematical morphology oper-
ators over concept lattices. In [73] pp. 28–43 (2013). doi:10.1007/978-3-
642-38317-5_2. → 309
[7] Baader, F.: Computing a minimal representation of the subsumption lat-
tice of all conjunctions of concepts defined in a terminology. In: Proc. Intl.
KRUSE Symposium, August, pp. 168–178. Santa Cruz, USA (1995). → 333
[8] Baader, F., Calvanese, D., McGuinness, D.L., Nardi, D., Patel-Schneider,
P.F.: The Description Logic handbook: Theory, implementation and
applications. Cambridge University Press (2nd edition, 2007).
doi:10.1017/CBO9780511711787. → 333
[9] Baader, F., Ganter, B., Sattler, U., Sertkaya, B.: Completing description logic
knowledge bases using formal concept analysis. In: IJCAI, vol. 7, pp. 230–
235 (2007). doi:10.25368/2022.155. → 333
[10] Baader, F., Molitor, R.: Building and structuring description logic knowl-
edge bases using least common subsumers and concept analysis. In [152]
pp. 292–305 (2000). doi:10.1007/10722280_20. → 333

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 335
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2
336 References

[11] Baader, F., Sertkaya, B.: Applying formal concept analysis to description
logics. In [114] pp. 261–286 (2004). doi:10.1007/978-3-540-24651-0_24.
→ 333
[12] Babin, M.A., Kuznetsov, S.O.: Recognizing pseudo-intents is coNP-
complete. In [242] pp. 294–301 (2010). → 115
[13] Baixeries, J.: Lattice characterization of Armstrong and symmetric depen-
dencies. Ph.D. thesis, Universitat Politècnica de Catalunya, Spain (2007).
→ 115
[14] Baixeries, J.: Database dependencies and formal concept analysis. Preprint
arXiv:2403.13914 (2024). → 115
[15] Baixeries, J., Codocedo, V., Kaytoue, M., Napoli, A.: Three views on de-
pendency covers from an FCA perspective. In [111] pp. 78–94 (2023).
doi:10.1007/978-3-031-35949-1_6. → 115
[16] Baker, K.A., Wille, R.: Lattice Theory and Its Applications: In Celebration
of Garrett Birkhoff’s 80th Birthday. Heldermann Verlag (1995). → 343,
357
[17] Balbes, R., Dwinger, P.: Distributive lattices. Journal of Symbolic Logic
42(4) (1977). doi:10.2307/2271885. → 253
[18] Balcázar, J.L., Baixeries, J.: Characterizations of multivalued dependencies
and related expressions. In: Discovery Science: 7th International Confer-
ence, DS 2004, Padova, Italy, October 2-5, 2004. Proceedings 7, pp. 306–313.
Springer (2004). doi:10.1007/978-3-540-30214-8_25. → 115
[19] Baldy, P., Mitas, J.: Generalized dimension of an ordered set and its Mac-
Neille completion. Order 11(2), 135–148 (1994). doi:10.1007/bf01108598.
→ 281
[20] Banaschewski, B.: Hüllensysteme und Erweiterung von Quasi-
Ordnungen. Mathematical Logic Quarterly 2(8-9), 117–130 (1956).
doi:10.1002/malq.19560020803. → 68, 281
[21] Banaschewski, B., Bruns, G.: Categorical characterization of the MacNeille
completion. Arch. Math. 18, 369–377 (1967). doi:10.1007/bf01898828.
→ 71
[22] Bandelt, H.J.: Tolerance relations on lattices. Bulletin of
the Australian Mathematical Society 23(3), 367–381 (1981).
doi:10.1017/s0004972700007255. → 160, 214
[23] Barbut, M.: Note sur l’algèbre des techniques d’analyse hiérarchique. B.
Matalon (1965). doi:10.1515/9783111540924-010. → 68
[24] Barbut, M., Monjardet, B.: Ordre et classification, vols. 1 and 2. Hachette,
Paris, France (1970). → 68
[25] Bartenschlager, G.: Free bounded distributive lattices generated by finite
ordered sets. Ph.D. thesis, Technische Hochschule Darmstadt (1994).
→ 71, 247
[26] Bartenschlager, G.: Free bounded distributive lattices over finite ordered
sets and their skeletons. Acta Math. Univ. Comenianae 64(1), 1–23 (1995).
→ 71
References 337

[27] Bartl, E., Krupka, M.: Residuated lattices of block relations: size reduction
of concept lattices. International Journal of General Systems 45(7-8), 773–
789 (2016). doi:10.1080/03081079.2016.1194021. → 160
[28] Barwise, J., Seligman, J.: Information flow: the logic of distributed sys-
tems. Cambridge University Press (1997). doi:10.1017/cbo9780511895968.
→ 75, 296
[29] Bazin, A.: A triadic generalisation of the Boolean concept lattice. In [111]
pp. 95–105 (2023). → 73
[30] Bazin, A., Beaudou, L., Kahn, G., Khoshkhah, K.: Bounding the num-
ber of minimal transversals in tripartite 3-uniform hypergraphs. Discrete
Mathematics & Theoretical Computer Science 23 (special issues) (2023).
doi:10.46298/dmtcs.7129. → 73
[31] Bazin, A., Kahn, G.: Reduction and introducers in d-contexts. In [82] pp.
73–88 (2019). doi:10.1007/978-3-031-35949-1_7. → 73
[32] Behrisch, M., Chavarri Villarello, A., Vargas-García, E.: Representing par-
tition lattices through FCA. In [66] (2021). doi:10.1007/978-3-030-77867-
5_1. → 71
[33] Belohlavek, R.: Fuzzy relational systems: foundations and principles,
vol. 20. Kluwer Academic/Plenum Publishers, New York (2002).
doi:10.1108/k.2003.06732iae.005. → 73
[34] Bĕlohlávek, R.: Concept lattices and order in fuzzy logic. An-
nals of pure and applied logic 128(1-3), 277–298 (2004).
doi:10.1016/j.apal.2003.01.001. → 73
[35] Belohlavek, R., Glodeanu, C., Vychodil, V.: Optimal factorization of three-
way binary data using triadic concepts. Order 30, 437–454 (2013).
doi:10.1007/s11083-012-9254-4. → 281
[36] Belohlavek, R., Trnecka, M.: Basic level of concepts in formal concept ana-
lysis. In [102] pp. 28–44 (2012). doi:10.1007/978-3-642-29892-9_9. → 279
[37] Belohlavek, R., Trnecka, M.: Basic level in formal concept analysis: Inter-
esting concepts and psychological ramifications. In: Twenty-Third Inter-
national Joint Conference on Artificial Intelligence (2013). → 279
[38] Belohlavek, R., Trnecka, M.: From-below approximations in Boolean ma-
trix factorization: Geometry and new algorithm. Journal of Computer and
System Sciences 81(8), 1678–1697 (2015). doi:10.1016/j.jcss.2015.06.002.
→ 281
[39] Belohlavek, R., Trnecka, M.: Basic level of concepts in formal concept ana-
lysis 1: formalization and utilization. International Journal of General Sys-
tems 49(7), 689–706 (2020). doi:10.1080/03081079.2020.1837794. → 279
[40] Belohlavek, R., Vychodil, V.: Discovery of optimal factors in binary data
via a novel method of matrix decomposition. Journal of Computer and
System Sciences 76(1), 3–20 (2010). doi:10.1016/j.jcss.2009.05.002. → 281
[41] Belohlavek, R., Vychodil, V.: Factorizing three-way binary data with triadic
formal concepts. In: International Conference on Knowledge-Based and
Intelligent Information and Engineering Systems, pp. 471–480. Springer
(2010). doi:10.1007/978-3-642-15387-7_51. → 73
338 References

[42] Belohlavek, R., Vychodil, V.: Formal concept analysis and linguistic
hedges. International Journal of General Systems 41(5), 503–532 (2012).
doi:10.1080/03081079.2012.685936. → 73
[43] Bennett, M.K., Birkhoff, G.: Two families of Newman lattices. Algebra Uni-
versalis 32, 115–144 (1994). doi:10.1007/bf01190819. → 235
[44] Benthem, J.: Information transfer across Chu spaces. Logic Journal of the
IGPL 8(6), 719–731 (2000). doi:10.1093/jigpal/8.6.719. → 309
[45] Berry, A., Sigayret, A.: Representing a concept lattice by a
graph. Discrete Applied Mathematics 144(1-2), 27–42 (2004).
doi:10.1016/j.dam.2004.02.016. → 71
[46] Berry, A., Sigayret, A.: Dismantlable lattices in the mirror. In [73] pp. 44–59
(2013). doi:10.1007/978-3-642-38317-5_3. → 70, 161
[47] Bertet, K., Borchmann, D., Cellier, P., Ferré, S. (eds.): Proceedings of the
14th ICFCA, vol. LNAI 10308. Springer (2017). doi:10.1007/978-3-319-
59271-8. → 344, 348
[48] Biedermann, K.: A foundation of the theory of trilattices. Ph.D. thesis, TU
Darmstadt (1998). → 73
[49] Birkhoff, G.: Lattice theory, vol. 25. American Mathematical Soc. (1940).
doi:10.1090/coll/025. → 20, 68, 251
[50] Birkhoff, G.: Lattice Theory 3rd. ed. American Mathematical Soc. (1967).
→ 20
[51] Birkhoff, G., Bennett, M.K.: The convexity lattice of a poset. Order 2, 223–
242 (1985). doi:10.1007/bf00333128. → 71
[52] Blyth, T.S., Janowitz, M.F.: Residuation theory. Pergamon Press (1972).
doi:10.1016/b978-0-08-016408-3.50004-8. → 20, 309
[53] Bock, H.H.: Classification and related methods of data analysis. Dis-
tributors for the USA and Canada, Elsevier Science Pub. Co. (1988).
doi:10.5771/0943-7444-1989-2-115. → 351, 358
[54] Bock, H.H., Ihm, P.: Classification, Data Analysis, and Knowledge Organi-
zation: Models and Methods with Applications. Springer Science & Busi-
ness Media (2012). doi:10.1007/978-3-642-76307-6. → 351, 356, 357
[55] Bock, H.H., Lenski, W., Richter, M.M.: Information Systems and Data Ana-
lysis: Prospects-Foundations-Applications. Springer Science & Business
Media (2013). doi:10.1007/978-3-642-46808-7. → 349, 357
[56] Bock, H.H., Polasek, W. (eds.): Data Analysis and Information Systems:
Statistical and Conceptual Approaches. Springer Science & Business Me-
dia, University of Basel (1996). doi:10.1007/978-3-642-80098-6. → 354,
356
[57] Bokowski, J., Kollewe, W.: On representing contexts in line arrangements.
Order 8, 393–403 (1991). doi:10.1007/bf00571189. → 113
[58] Boley, M., Grosskreutz, H., Gärtner, T.: Formal concept sampling for
counting and threshold–free local pattern mining. In: Proc. of
the SIAM Int. Conf. on Data Mining (SDM 2010). SIAM (2010).
doi:10.1137/1.9781611972801.16. → 69
References 339

[59] Borchmann, D.: Context orbifolds. Diplomarbeit, TU Dresden (2009).

→ 287, 308
[60] Borchmann, D.: Learning terminological knowledge with high confidence
from erroneous data. Ph.D. thesis, TU Dresden (2014). Qucosa Dresden
→ 333
[61] Borchmann, D.: Exploring faulty data. In: Formal Concept Analysis: 13th
International Conference, ICFCA 2015, Nerja, Spain, June 23-26, 2015, Pro-
ceedings 13, pp. 219–235. Springer (2015). doi:10.1007/978-3-319-19545-
2_14. → 74
[62] Borchmann, D., Ganter, B.: Concept lattice orbifolds–first steps. In: Formal
Concept Analysis: 7th International Conference, ICFCA 2009 Darmstadt,
Germany, May 21-24, 2009 Proceedings 7, pp. 22–37. Springer (2009).
doi:10.1007/978-3-642-01815-2_2. → 20, 308
[63] Borchmann, D., Hanika, T.: Some experimental results on randomly gen-
erating formal contexts. In: CLA’16, pp. 57–69. CEUR-WS 1624 (2016).
→ 69
[64] Borgwardt, S.: Bindungen. Diplomarbeit, TU Dresden (2010). → 20, 150,
152, 153, 297
[65] Bouchet, A.: Etude combinatoire des ordonnés finis. Ph.D. thesis, Univer-
sité Joseph Fourier Grenoble I (1971). → 281
[66] Braud, A., Buzmakov, A., Hanika, T., Le Ber, F. (eds.): Proceedings of the
16th ICFCA, vol. LNAI 12733. Springer (2021). doi:10.1007/978-3-030-
77867-5. → 337, 343
[67] Brylawski, T.: The lattice of integer partitions. Discrete Mathematics 6(3),
201–219 (1973). doi:10.1016/0012-365x(73)90094-0. → 54
[68] Burgmann, C., Wille, R.: The basic theorem on preconcept lattices. In [284]
pp. 80–88 (2006). doi:10.1007/11671404_5. → 321, 332
[69] Burmeister, P.: Merkmalimplikationen bei unvollständigem Wissen. In: In
[263], 15–46 (1991) 113
[70] Burmeister, P., Holzer, R.: Treating incomplete knowledge in formal con-
cept analysis. In [157] pp. 114–126 (2005). doi:10.1007/11528784_6. → 74,
113
[71] Burosch, G., Demetrovics, J., Katona, G.O.H., Kleitman, D.J., Sapozhenko,
A.A.: On the number of databases and closure operations. Theoretical
computer science 78(2), 377–381 (1991). doi:10.1016/0304-3975(91)90359-
a. → 21
[72] Caspard, N., Monjardet, B.: The lattices of closure systems, closure opera-
tors, and implicational systems on a finite set: a survey. Discrete Applied
Mathematics 127(2), 241–269 (2003). doi:10.1016/s0166-218x(02)00209-3.
→ 71
[73] Cellier, P., Distel, F., Ganter, B. (eds.): Proceedings of the 11th ICFCA, vol.
LNAI 7880. Springer (2013). doi:10.1007/978-3-642-38317-5. → 335, 338,
342
340 References

[74] Cellier, P., Ducassé, M., Ferré, S., Ridoux, O.: Formal concept analysis
enhances fault localization in software. In [279] pp. 273–288 (2008).
doi:10.1007/978-3-540-78137-0_20. → 74
[75] Chaudron, L., Maille, N.: Generalized formal concept analysis. In [152]
pp. 357–370 (2000). doi:10.1007/10722280_25. → 75
[76] Chornomaz, B.: Algebraicity and the tensor product of concept lattices.
In [173] pp. 54–66 (2014). doi:10.1007/978-3-319-07248-7_5. → 214
[77] Codd, E.F.: A relational model of data for large shared data
banks. Communications of the ACM 13(6), 377–387 (1970).
doi:10.7551/mitpress/12274.003.0034. → 71
[78] Cogis, O.: Le dimension Ferrers des graphs orientés. Ph.D. thesis, Univer-
sité Curie, Paris (1980). → 281
[79] Cogis, O.: On the Ferrers dimension of a digraph. Discrete Mathematics
38(1), 47–52 (1982). doi:10.1016/0012-365x(82)90167-4. → 281
[80] Colomb, P., Irlande, A., Raynaud, O.: Counting of Moore families for n=
7. In [256] pp. 72–87 (2010). doi:10.1007/978-3-642-11928-6_6. → 13, 20
[81] Crawley, P., Dilworth, R.P.: Algebraic theory of lattices. Prentice Hall
(1973). → 278, 280
[82] Cristea, D., Le Ber, F., Sertkaya, B. (eds.): Proceedings of the 15th ICFCA.
Springer (2019). doi:10.1007/978-3-030-21462-3. → 337, 344, 349, 352
[83] Czédli, G.: Factor lattices by tolerances. Acta Sci. Math.(Szeged) 44(1-2),
35–42 (1982). → 160
[84] Davey, B.A., Priestley, H.A.: Introduction to lattices and order. Cambridge
University Press (2002). doi:10.1017/cbo9780511809088. → 20
[85] Day, A.: A simple solution to the word problem for lattices. Canadian
Mathematical Bulletin 13(2), 253–254 (1970). doi:10.4153/cmb-1970-051-
0. → 247
[86] Day, A.: Characterizations of finite lattices that are bounded-homomor-
phic images or sublattices of free lattices. Canadian Journal of Mathemat-
ics 31(1), 69–78 (1979). doi:10.4153/cjm-1979-008-x. → 70, 247, 280
[87] Day, A., Herrmann, C.: Gluings of modular lattices. Order 5, 85–101
(1988). doi:10.1007/bf00143900. → 214
[88] Day, A., Nation, J., Tschantz, S.: Doubling convex sets in lattices and
a generalized semidistributivity condition. Order 6, 175–180 (1989).
doi:10.1007/bf02034334. → 247, 280
[89] Degens, P.O., Hermes, H.J., Opitz, O. (eds.): Die Klassifikation und ihr
Umfeld. Indeks–Verlag, Frankfurt (1986). doi:10.5771/0943-7444-1987-2-
104. → 344
[90] Deiters, K., Erné, M.: Negations and contrapositions of complete lat-
tices. Discrete Mathematics 181(1-3), 91–111 (1998). doi:10.1016/s0012-
365x(97)00047-2. → 70
[91] Deiters, K., Erné, M.: Sums, products and negations of contexts
and complete lattices. Algebra Universalis 60(4), 469–496 (2009).
doi:10.1007/s00012-009-2141-1. → 70
References 341

[92] Delugach, H.S., Stumme, G. (eds.): Proceedings of the 9th ICCS, vol.
LNAI 2120. Springer Science & Business Media (2001). doi:10.1007/3-540-
44583-8. → 344, 359
[93] Diday, E.: Introduction à l’approche symbolique en analyse des
données. RAIRO-Operations Research 23(2), 193–236 (1989).
doi:10.1051/ro/1989230201931. → 75
[94] Dilworth, R.: Proof of a conjecture on finite modular lattices. The Dilworth
Theorems: Selected Papers of Robert P. Dilworth pp. 219–224 (1990).
doi:10.1007/978-1-4899-3558-8_21. → 259
[95] Dilworth, R.P.: Lattices with unique irreducible decompositions. In: The
Dilworth Theorems: Selected Papers of Robert P. Dilworth, pp. 93–99.
Springer (1990). doi:10.1007/978-1-4899-3558-8_10. → 280
[96] Distel, F.: Hardness of enumerating pseudo-intents in the lectic order.
In [256] pp. 124–137 (2010). doi:10.1007/978-3-642-11928-6_9. → 115
[97] Distel, F.: Learning description logic knowledge bases from data using
methods from formal concept analysis. Ph.D. thesis, TU Dresden (2011).
Qucosa Dresden, → 331, 333
[98] Distel, F., Sertkaya, B.: On the complexity of enumerating
pseudo-intents. Discrete Applied Mathematics 159(6) (2011).
doi:10.1016/j.dam.2010.12.004. → 115
[99] Doerfel, S.: A context-based description of the doubly founded concept
lattices in the variety generated by M3 . In [372] pp. 93–106 (2011).
doi:10.1007/978-3-642-20514-9_9. → 70, 259, 280
[100] Doignon, J.P., Ducamp, A., Falmagne, J.C.: On realizable biorders and the
biorder dimension of a relation. Journal of Mathematical Psychology
28(1), 73–109 (1984). doi:10.1016/0022-2496(84)90020-8. → 281
[101] Doignon, J.P., Falmagne, J.C.: Knowledge spaces. Springer Science & Busi-
ness Media (1999). doi:10.1007/978-3-642-58625-5. → 75, 315
[102] Domenach, F., Ignatov, D., Poelmans, J. (eds.): Proceedings of the 10th
ICFCA, vol. LNAI 7278. Springer (2012). doi:10.1007/978-3-642-29892-9.
→ 337, 344, 349, 354
[103] Domenach, F., Leclerc, B.: Biclosed binary relations and Galois connec-
tions. Order 18, 89–104 (2001). doi:10.1023/a:1010662327346. → 161
[104] Dörflein, S.K., Wille, R.: Coherence networks of concept lattices: the basic
theorem. In [149] pp. 344–359 (2005). doi:10.1007/978-3-540-32262-7_24.
→ 334
[105] Dorn, G., Frank, R.D., Ganter, B., Kipke, U., Poguntke, W., Wille, R.:
Forschung und Mathematisierung — Suche nach Wegen aus dem Elfen-
beinturm. Wechselwirkung 15, 20–23 (1982). → 68
[106] Dorninger, D., Eigenthaler, G., Kaiser, H.K., Müller, W.B. (eds.): Contri-
butions to general algebra, vol. 7. Hölder–Pichler–Tempsky, Wien (1991).
→ 355, 359, 360
[107] Dubois, D., Prade, H.: Possibility theory and formal concept analysis in
information systems. In: International Fuzzy Systems Association World
Congress and Conference of the European Society for Fuzzy Logic and
342 References

Technology (IFSA-EUSFLAT 2009), pp. 1021–1026. European Society for

Fuzzy Logic and Technology (2009). doi:10.1016/j.fss.2011.02.008. → 74
[108] Duffus, D., Rival, I.: Crowns in dismantlable partially ordered sets. Col-
loquia Mathematica Societatis János Bolyai 18 (1976), 271-292 (1976).
→ 161
[109] Duffus, D., Rival, I.: A structure theory for ordered sets. Discrete Mathe-
matics 35(1-3), 53–118 (1981). doi:10.1016/0012-365x(81)90201-6. → 20
[110] Duquenne, V.: Contextual implications between attributes and some rep-
resentation properties for finite lattices (orig. 1987 in [165], reprinted).
In [73] pp. 1–27 (2013). doi:10.1007/978-3-642-38317-5_1. → 113
[111] Dürrschnabel, D., Rodríguez, D.L. (eds.): Proceedings of the 17th ICFCA,
vol. LNAI 13934. Springer Nature (2023). doi:10.1007/978-3-031-35949-1.
→ 336, 337
[112] Dürrschnabel, D., Hanika, T., Stumme, G.: Drawing order diagrams
through two-dimension extension. J. Graph Algorithms Appl. 27(9), 783–
802 (2023). doi:10.7155/JGAA.00645. → 113
[113] Edelman, P.H.: Meet-distributive lattices and the anti-exchange closure.
Algebra Universalis 10, 290–299 (1980). doi:10.1007/bf02482912. → 280
[114] Eklund, P. (ed.): Proceedings of the 2nd ICFCA, vol. LNAI 2961. Springer
(2004). → 336, 343, 350, 359
[115] Emilion, R., Lévy, G.: Size of random Galois lattices and number of closed
frequent itemsets. Discrete Applied Mathematics 157(13), 2945–2957
(2009). doi:10.1016/j.dam.2009.02.025. → 69
[116] Erné, M.: The Dedekind-MacNeille completion as a reflector. Order 8,
159–173 (1991). doi:10.1007/bf00383401. → 214
[117] Erné, M.: Distributive laws for concept lattices. Algebra Universalis 30,
538–580 (1993). doi:10.1007/bf01195382. → 279
[118] Erné, M.: Tensor products of contexts and complete lattices. Algebra Uni-
versalis 31, 36–65 (1994). doi:10.1007/bf01188179. → 70
[119] Erné, M.: Categories of contexts. Preprint arXiv:1407.0512 (2014). → 308
[120] Everett, C.: Closure operators and Galois theory in lattices. Trans-
actions of the American Mathematical Society 55, 514–525 (1944).
doi:10.1090/s0002-9947-1944-0010556-9. → 20
[121] Faigle, U., Herrmann, C.: Projective geometry on partially ordered sets.
Transactions of the American Mathematical Society 266(1), 319–332
(1981). doi:10.1090/s0002-9947-1981-0613799-9. → 280
[122] Falmagne, J.C., Doignon, J.P.: Learning spaces: Interdisciplinary applied
mathematics. Springer Science & Business Media (2010). doi:10.1007/978-
3-642-01039-2. → 75, 280
[123] Farley, J.D.: The uniqueness of the core. Order 10, 129–131 (1993).
doi:10.1007/bf01111296. → 161
[124] Feferman, S.: Ah, Chu! JFAK, Essays Dedicated to Johan Van Benthem on
the Occasion of his 50th Birthday (1999) 309
[125] Felde, M., Hanika, T., Stumme, G.: Null models for formal contexts. Infor-
mation 11(3) (2020). doi:10.3390/info11030135. → 69
References 343

[126] Felde, M., Koyda, M.: Interval-dismantling for lattices. Inter-

national Journal of Approximate Reasoning 159, 108931 (2023).
doi:10.1016/j.ijar.2023.108931. → 161
[127] Felde, M., Stumme, G.: Triadic exploration and exploration with multiple
experts. In [66] (2021). doi:10.1007/978-3-030-77867-5_11. → 73
[128] Ferré, S., Ridoux, O.: A logical generalization of formal concept analysis.
In [152] pp. 371–384 (2000). doi:10.1007/10722280_26. → 74, 311
[129] Ferré, S., Ridoux, O.: Introduction to logical information systems.
Information Processing & Management 40(3), 383–419 (2004).
doi:10.1016/s0306-4573(03)00018-9. → 74
[130] Ferré, S., Ridoux, O., Sigonneau, B.: Arbitrary relations in formal concept
analysis and logical information systems. In: Conceptual Structures: Com-
mon Semantics for Sharing Knowledge. Proceedings of the 13th ICCS., pp.
166–180. Springer (2005). doi:10.1007/11524564_11. → 334
[131] Flath, S.: The order dimension of multinomial lattices. Order 10, 201–219
(1993). doi:10.1007/bf01110542. → 71
[132] Flath, S.: Multinomial lattices. an approach to the structure of multiper-
mutations via formal concept analysis. Ph.D. thesis, TH Darmstadt (1994).
→ 71
[133] Ganter, B.: Algorithmen zur formalen Begriffsanalyse. In [165] pp. 196–
212 (1987). → 112, 114
[134] Ganter, B.: Composition and decomposition of data in formal concept ana-
lysis. In: 1. Conference of the International Federation of Classification
Societies, pp. 561–566 (1987). → 70
[135] Ganter, B.: Finding closed sets under symmetry. FB4 preprint 1307, TH
Darmstadt (1990). → 112, 308
[136] Ganter, B.: Lattice theory and formal concept analysis: a subjective intro-
duction. In [16] (1995). → 142, 161
[137] Ganter, B.: Attribute exploration with background knowledge. Theo-
retical Computer Science 217(2), 215–233 (1999). doi:10.1016/s0304-
3975(98)00271-0. → 114
[138] Ganter, B.: Begriffe und Implikationen. In: [364] pp. 1–24 (2000).
doi:10.1007/978-3-642-57217-3_1. → 108, 114
[139] Ganter, B.: Conflict avoidance in additive order diagrams. J. Univers. Com-
put. Sci. 10(8), 955–966 (2004). → 113
[140] Ganter, B.: Congruences of finite distributive concept algebras. In [114]
pp. 128–141 (2004). doi:10.1007/s11083-006-9045-x. → 161
[141] Ganter, B.: Relational Galois connections. In [251] pp. 1–17 (2007).
doi:10.1007/978-3-540-70901-5_1. → 161, 162
[142] Ganter, B.: Lattices of rough set abstractions as P -products. In [279] pp.
199–216 (2008). doi:10.1007/978-3-540-78137-0_15. → 247
[143] Ganter, B.: Two basic algorithms in concept analysis (1984), reprinted. In
[256] pp. 312–340 (2010). doi:10.1007/978-3-642-11928-6_22. → 112, 114
[144] Ganter, B.: Random extents and random closure systems. In: CLA’11, vol.
CEUR-WS 959, pp. 309–318 (2011). → 69
344 References

[145] Ganter, B.: “Properties of Finite Lattices” by S. Reeg and W. Weiß, re-
visited. In memoriam Peter Burmeister (1941–2019). In [82] pp. 99–109
(2019). doi:10.1007/978-3-030-21462-3_8. → 98, 353
[146] Ganter, B.: Notes on integer partitions. International Journal of Approxi-
mate Reasoning 142, 31–40 (2022). doi:10.1016/j.ijar.2021.11.004. → 71
[147] Ganter, B., Glodeanu, C.V.: Ordinal factor analysis. In [102] pp. 128–139
(2012). doi:10.1007/978-3-642-29892-9_15. → 281
[148] Ganter, B., Glodeanu, C.V.: Factors and skills. In [173] pp. 173–187 (2014).
doi:10.1007/978-3-319-07248-7_13. → 281
[149] Ganter, B., Godin, R. (eds.): Proceedings of the 3rd ICFCA 2005, vol.
LNAI 3403. Springer (2005). doi:10.1007/b105806. → 341, 355, 357
[150] Ganter, B., Krauße, R.: Pseudo-models and propositional Horn in-
ference. Discrete Applied Mathematics 147(1), 43–55 (2005).
doi:10.1016/j.dam.2004.06.019. → 107, 114
[151] Ganter, B., Kuznetsov, S.O.: Pattern structures and their projections.
In [92] pp. 129–142 (2001). doi:10.1007/3-540-44583-8_10. → 74
[152] Ganter, B., Mineau, G. (eds.): Proceedings of the 8th ICCS 2000, vol.
LNAI 1867. Springer (2000). doi:10.1007/10722280. → 335, 340, 343,
359
[153] Ganter, B., Nevermann, P., Reuter, K., Stahl, J.: How small can a lattice of
order-dimension n be? Order 3, 345–353 (1987). doi:10.1007/bf00340776.
→ 281
[154] Ganter, B., Obiedkov, S.: Conceptual exploration. Springer (2016).
doi:10.1007/978-3-662-49291-8. → 74, 77, 82, 94, 98, 99, 106, 112, 113, 114,
115, 308, 316, 331
[155] Ganter, B., Reuter, K.: Finding all closed sets: A general approach. Order
8, 283–290 (1991). doi:10.1007/bf00383449. → 112, 308
[156] Ganter, B., Stahl, J., Wille, R.: Conceptual measurement and many valued
contexts. In [166] (1986). → 309
[157] Ganter, B., Stumme, G., Wille, R.: Formal concept analysis: foundations
and applications, vol. 3626. Springer (2005). doi:10.1007/978-3-540-31881-
1. → 339, 349, 357
[158] Ganter, B., Wille, R.: Implikationen und Abhängigkeiten zwischen Merk-
malen. In [89] pp. 171–185 (1986). English version in [47]. → 113, 114
[159] Ganter, B., Wille, R.: Conceptual scaling. In: F.S. Roberts (ed.) Appli-
cations of combinatorics and graph theory to the biological and social
sciences, pp. 139–167. Springer (1989). doi:10.1007/978-1-4684-6381-1_6.
→ 71, 114, 309
[160] Ganter, B., Wille, R.: Subtensorial decompositions of concept lattices.
Tech. rep., Bericht MATH-AL-1-1994, TU Dresden FB Mathematik (1994).
→ 214
[161] Ganter, B., Wille, R.: Formale Begriffsanalyse – Mathematische Grundla-
gen. Springer-Verlag (1996). doi:10.1007/978-3-642-61450-7. → vii, 345
References 345

[162] Ganter, B., Wille, R.: Contextual attribute logic. In: International
Conference on Conceptual Structures, pp. 377–388. Springer (1999).
doi:10.1007/3-540-48659-3_23. → 332
[163] Ganter, B., Wille, R.: Formal Concept Analysis – Mathematical Founda-
tions. Springer-Verlag (1999). doi:10.1007/978-3-642-59830-2. Translated
from [161]. → vii, 345
[164] Ganter, B., Wille, R.: Xingshi Gainian Fenxi. Science Press (2005). Trans-
lated from [163]. → vii
[165] Ganter, B., Wille, R., Wolff, K.E. (eds.): Beiträge zur Begriffsanalyse. B.I.-
Wissenschaftsverlag Mannheim (1987). → 342, 343, 350, 358
[166] Gaul, W., Schader, M. (eds.): Proceedings of the 9th Annual Meeting of
the Classification Society, Classification as a Tool of Research, vol. 9. North
Holland, University of Karlsruhe, Germany (1986). → 344, 351, 355
[167] Gepperth, U.: Automatisches Zeichnen begriffsanalytischer Liniendi-
agramme unter Verwendung flußalgorithmischer Berechnungen von
Merkmalsketten. Diplomarbeit, FH Darmstadt (1990). → 113
[168] Geyer, W.: Generalizing semidistributivity. Order 10, 77–92 (1993).
doi:10.1007/bf01108710. → 280
[169] Geyer, W.: On context patterns associated with concept lattices. Order
10(4), 363–373 (1993). doi:10.1007/bf01108830. → 70
[170] Geyer, W.: The generalized doubling construction and formal concept ana-
lysis. Algebra Universalis 32, 341–367 (1994). doi:10.1007/bf01235175.
→ 247
[171] Geyer, W.: On Tamari lattices. Discrete Mathematics 133(1-3), 99–122
(1994). doi:10.1016/0012-365x(94)90019-1. → 235, 247
[172] Glodeanu, C.V.: Triadic factor analysis. In [242] pp. 127–138 (2010). → 73,
281
[173] Glodeanu, C.V., Kaytoue, M., Sacarea, C. (eds.): Proceedings of the 12th
ICFCA, vol. LNAI 8478. Springer (2014). doi:10.1007/978-3-319-07248-7.
→ 340, 344, 354
[174] Grätzer, G.: On the complete congruence lattice of a complete lattice with
an application to universal algebra. CR Math. Rep. Acad. Sci. Canada 11,
105–108 (1989). → 160
[175] Grätzer, G.: General lattice theory. Springer Science & Business Media
(2002). doi:10.1007/978-3-0348-9326-8. → 20, 70
[176] Greechie, R.: On generating pathological orthomodular structures. In:
Technical Report no. 13. Kansas State University (1970). → 247
[177] Greene, C., Markowsky, G.: A combinatorial test for local distributivity. In:
Notices of the American Mathematical Society, vol. 22 (2), pp. A299–A299.
AMS (1975). → 280
[178] Grötzsch, A.: Begriffsanalytische Darstellung verallgemeinerter Hüllen-
systeme. Diplomarbeit, TU Dresden (2005). → 71, 280, 291
[179] Guigues, J.L., Duquenne, V.: Familles minimales d’implications informa-
tives résultant d’un tableau de données binaires. Mathématiques et Sci-
ences humaines 95, 5–18 (1986). → 96, 97, 113
346 References

[180] Gürgens, S.: Ideal-Filter–Verklebung von Verbänden. Ph.D. thesis, TH

Darmstadt (1992). → 247
[181] Habib, M., Nourine, L.: The number of Moore families on n= 6. Discrete
Mathematics 294(3), 291–296 (2005). doi:10.1016/j.disc.2004.11.010. → 20
[182] Hanika, T.: Discovering knowledge in bipartite graphs with formal con-
cept analysis. Ph.D. thesis, Universität Kassel (2019). → 71
[183] Hanika, T., Hirth, J.: Knowledge cores in large formal contexts. An-
nals of Mathematics and Artificial Intelligence 90(6), 537–567 (2022).
doi:10.1007/s10472-022-09790-6. → 279
[184] Hanika, T., Hirth, J.: On the lattice of conceptual measurements. Informa-
tion Sciences 613, 453–468 (2022). doi:10.1016/j.ins.2022.09.005. → 309
[185] Hartung, G.: Darstellung beschränkter Verbände als Verbände offener Be-
griffe. Diplomarbeit, TH Darmstadt (1989). → 75
[186] Hartung, G.: A topological representation of lattices. Algebra Universalis
29, 273–299 (1992). doi:10.1007/bf01190610. → 75
[187] Hartung, G.: An extended duality for lattices. In: K. Denecke, H.J. Vo-
gel (eds.) General algebra and applications. Heldermann–Verlag, Berlin
(1993). → 309
[188] Hartung, G.: Sublattices of topologically represented lattices. Acta Math.
Univ. Comenianae 64(1), 25–41 (1995). → 75
[189] Hermann, M., Sertkaya, B.: On the complexity of computing generators
of closed sets. In [279] pp. 158–168 (2008). doi:10.1007/978-3-540-78137-
0_12. → 21
[190] Herrmann, C.: S-verklebte Summen von Verbänden. Mathematische
Zeitschrift 130, 255–274 (1973). doi:10.1007/bf01246623. → 214
[191] Higuchi, A.: Lattices of closure operators. Discrete Mathematics 179(1-3),
267–272 (1998). doi:10.1016/s0012-365x(97)00099-x. → 291
[192] Hillman, C.: What is a concept? Technical paper review, University of
Washington (2001). → 57
[193] Hitzler, P., Krötzsch, M., Zhang, G.Q.: A categorical view on algebraic lat-
tices in formal concept analysis. Fundamenta Informaticae 74(2-3), 301–
328 (2006) 70
[194] Hoch, S.: Begriffsverbände mit Anti-Endomorphismus. Ph.D. thesis, TH
Darmstadt (1989). → 75
[195] Huang, S., Tamari, D.: Problems of associativity: A simple proof for the
lattice property of systems ordered by a semi-associative law. Journal
of Combinatorial Theory, Series A 13(1), 7–13 (1972). doi:10.1016/0097-
3165(72)90003-9. → 247
[196] Huchard, M., Rouane-Hacene, M., Roume, C., Valtchev, P.: Relational con-
cept discovery in structured datasets. Annals of Mathematics and Artifi-
cial Intelligence 49, 39–76 (2007). doi:10.1007/s10472-007-9056-3. → 334
[197] Ignatov, D.I.: On the cryptomorphism between Davis’ subset lattices,
atomic lattices, and closure systems under T1 separation axiom. arXiv
preprint arXiv:2209.12256 (2022). doi:10.48550/arXiv.2209.12256. → 69
References 347

[198] Ignatov, D.I.: On the number of maximal antichains in Boolean lat-

tices for n up to 7. Lobachevskii J. Math. 44, 137–146 (2023).
doi:10.1134/s1995080223010158. → 71
[199] Ignatov, D.I., Gnatyshak, D.V., Kuznetsov, S.O., Mirkin, B.G.: Triadic for-
mal concept analysis and triclustering: searching for optimal patterns. Ma-
chine Learning 101, 271–302 (2015). doi:10.1007/s10994-015-5487-y. → 73
[200] Jamison, R.E.: Copoints in antimatroids. Congr. Numer 29, 535–544
(1980). → 280
[201] Janowitz, M.F., Wille, R.: On the classification of monotone equivariant
cluster methods. In: I. Cox, P. Hansen, B. Jolesz (eds.) Partitioning data
sets, DIMACS Series in Discrete Mathematics and Theoretical Computer Sci-
ence, vol. 19. Am. Math. Soc. (1995). doi:10.1090/dimacs/019/08. → 72
[202] Jäschke, R., Hotho, A., Schmitz, C., Ganter, B., Stumme, G.: Trias–
an algorithm for mining iceberg tri-lattices. In: Sixth International
Conference on Data Mining (ICDM’06), pp. 907–911. IEEE (2006).
doi:10.1109/icdm.2006.162. → 73
[203] Jónsson, B., Kiefer, J.E.: Finite sublattices of a free lattice. Canadian Journal
of Mathematics 14, 487–497 (1962). doi:10.4153/cjm-1962-040-1. → 280
[204] Juandeaburre, A.B., Fuentes-Gonzalez, R.: The study of the l-fuzzy con-
cept lattice. Mathware & Soft Computing 3, 209–218 (1994) 73
[205] Jürgensen, H., Loewer, J.: Drawing Hasse diagrams of partially ordered
sets. In [208] pp. 331–345 (1979). → 179
[206] Kaarli, K., Kuchmei, V., Schmidt, S.: Sublattices of the direct product. Al-
gebra Universalis 59(1-2), 85–95 (2008). doi:10.1007/s00012-008-2067-z.
→ 161
[207] Kaarli, K., Radeleczki, S.: Ordered relations, concept lattices and self-
bonds. Journal of Multiple-Valued Logic & Soft Computing 36(4/5), 321
– 336 (2021) 161
[208] Kalmbach, G.: Orthomodular lattices. Academic Press (1983). → 214,
347
[209] Kark, M.: Interaktives Zeichnen geometrischer Darstellungen von Be-
griffsverbänden. Master’s thesis, FH Darmstadt (1989). → 113
[210] Kästner, F.: Retrakte und Retraktionen von Begriffsverbänden. Master’s
thesis, TU Dresden (2010). → 161
[211] Kauer, M., Krupka, M.: Removing an incidence from a formal context. In:
CLA’14, vol. CEUR-WS 1252, pp. 195–206 (2014). → 161
[212] Kauer, M., Krupka, M.: Generating complete sublattices by methods of
formal concept analysis. International Journal of General Systems 46(5),
475–489 (2017). doi:10.1080/03081079.2017.1354964. → 161
[213] Kaytoue, M., Kuznetsov, S.O., Napoli, A.: Revisiting numerical pattern
mining with formal concept analysis. In: T. Walsh (ed.) IJCAI 2011, Pro-
ceedings of the 22nd International Joint Conference on Artificial Intel-
ligence, Barcelona, Catalonia, Spain, July 16-22, 2011, pp. 1342–1347. IJ-
CAI/AAAI (2011). doi:10.5591/978-1-57735-516-8/IJCAI11-227. → 74
348 References

[214] Kent, R.E.: Rough concept analysis. In: Rough Sets, Fuzzy Sets and Knowl-
edge Discovery: Proceedings of the International Workshop on Rough Sets
and Knowledge Discovery (RSKD’93), Banff, Alberta, Canada, 12–15 Oc-
tober 1993, pp. 248–255. Springer (1994). doi:10.1007/978-1-4471-3238-
7_30. → 74
[215] Keprt, A.: Using blind search and formal concepts for binary factor analy-
sis. In: DATESO, vol. 98, pp. 128–140 (2004). → 281
[216] Keprt, A., Snášel, V.: Binary factor analysis with help of formal concepts.
In: CLA’04, vol. CEUR-WS 110, pp. 90–101 (2004). → 281
[217] Kerber, A.: Algebraic combinatorics via finite group actions. BI-
Wissenschaftsverlag, Mannheim (1991). → 308
[218] Kerber, A., Lex, W.: Kontexte und ihre Begriffe (1998). Lecture notes. →
69
[219] Kerkhoff, S., Pöschel, R., Schneider, F.M.: A short introduction to clones.
Electronic Notes in Theoretical Computer Science 303, 107–120 (2014).
doi:10.1016/j.entcs.2014.02.006. → 58
[220] Kipke, U., Wille, R.: Formale Begriffsanalyse erläutert an einem Wortfeld.
LDV-Forum 5, 31–36 (1987). → 113
[221] Kleitman, D.J.: Extremal properties of collections of subsets containing no
two sets and their union. Journal of Combinatorial Theory, Series A 20(3),
390–392 (1976). doi:10.1016/0097-3165(76)90037-6. → 69
[222] Klotz, U., Mann, A.: Begriffexploration. Diplomarbeit, TH Darmstadt
(1988). → 114
[223] Knecht, S., Wille, R.: Congruence lattices of finite lattices as concept lat-
tices. In [4] pp. 323–325 (1990). doi:10.1007/978-1-4899-2608-1_32. → 160
[224] Kollewe, W.: Representation of data by pseudoline arrangements. In [295]
pp. 113–122 (1993). → 113
[225] Kollewe, W., Skorsky, M., Vogt, F., Wille, R.: Toscana – ein Werkzeug zur
begrifflichen Analyse und Erkundung von Daten. In [419] pp. 267–288
(1994). → 113
[226] Koppen, M.: On finding the bidimension of a relation. Journal of
Mathematical Psychology 31(2), 155–178 (1987). doi:10.1016/0022-
2496(87)90013-7. → 281
[227] Korossy, K., et al.: Modeling knowledge as competence and perfor-
mance. In: D. Albert, J. Lukas (eds.) Knowledge spaces: Theories, em-
pirical research, and applications, pp. 103–132. Psychology Press (1999).
doi:10.4324/9781410602077-14. → 281
[228] Krajci, S.: Cluster based efficient generation of fuzzy concepts. Neural Net-
work World 13(5), 521–530 (2003). → 74
[229] Krantz, D., Luce, D., Suppes, P., Tversky, A.: Foundations of measurement,
Vol. I, II, 3. Academic Press, New York (1971–1990). → 71
[230] Kriegel, F.: Implications over probabilistic attributes. In [47] pp. 168–183
(2017). doi:10.1007/978-3-319-59271-8_11. → 74
References 349

[231] Kriegel, F.: Joining implications in formal contexts and inductive learning
in a Horn description logic. In [82] pp. 110–129 (2019). doi:10.1007/978-
3-030-21462-3_9. → 115
[232] Kriegel, F.: Constructing and Extending Description Logic Ontologies us-
ing Methods of Formal Concept Analysis: A Dissertation Summary. KI -
Künstliche Intelligenz 34, 399–403 (2020). doi:10.1007/s13218-020-00673-
8. → 333
[233] Kriegel, F.: Efficient axiomatization of OWL 2 EL ontologies from data by
means of Formal Concept Analysis. To appear (2024). → 333
[234] Kriegel, F., Borchmann, D.: Nextclosures: Parallel computation of the
canonical base with background knowledge. International Journal of Gen-
eral Systems 46(5), 490–510 (2017). doi:10.1080/03081079.2017.1349570.
→ 115
[235] Krolak-Schwerdt, S., Orlik, P., Ganter, B.: Tripat: a model for analyzing
three-mode binary data. In [55] pp. 298–307 (1994). doi:10.1007/978-
3-642-46808-7_27. → 73, 112
[236] Krötzsch, M.: Morphisms in Logic, Topology, and Formal Concept Analy-
sis. Master’s thesis, TU Darmstadt (2005). → 70, 309
[237] Krötzsch, M., Hitzler, P., Zhang, G.Q.: Morphisms in context. In: Concep-
tual Structures: Common Semantics for Sharing Knowledge. Proceedings
of the 13th ICCS., pp. 223–237. Springer (2005). doi:10.1007/11524564_15.
→ 157, 309
[238] Krötzsch, M., Malik, G.: The tensor product as a lattice of regular Galois
connections. In [284] pp. 89–104 (2006). doi:10.1007/11671404_6. → 162
[239] Krupka, M.: On factorization of concept lattices by incompatible toler-
ances. In [372] pp. 167–182 (2011). doi:10.1007/978-3-642-20514-9_14.
→ 160
[240] Krupka, M.: Basic theorem of fuzzy concept lattices revisited. Fuzzy Sets
and Systems 333, 54–70 (2018). doi:10.1016/j.fss.2017.04.007. → 73
[241] Krupka, M., Lastovicka, J.: Concept lattices of incomplete data. In [102]
pp. 180–194 (2012). doi:10.1007/978-3-642-29892-9_19. → 74
[242] Kryszkiewicz, M., Obiedkov, S. (eds.): Proceedings of the 7th CLA, vol.
CEUR 672 (2010). → 336, 345
[243] Kühn, R., Ries, M.: Methoden der Formalen Begriffsanalyse beim Aufbau
kooperativer Information. Diplomarbeit, TH Darmstadt (1988). → 113
[244] Kuznetsov, S.: Stability as an estimate of degree of substantiation of
hypotheses derived on the basis of operational similarity. Nauchno-
Tekhnicheskaya Informatsiya Ser. 2 24(12), 21–29 (1990). → 14
[245] Kuznetsov, S.O.: On computing the size of a lattice and related decision
problems. Order 18, 313–321 (2001). → 113, 249
[246] Kuznetsov, S.O.: Galois connections in data analysis: Contributions from
the soviet era and modern russian research. In [157] pp. 196–225 (2005).
doi:10.1007/11528784_11. → 20
350 References

[247] Kuznetsov, S.O.: On stability of a formal concept. Annals of Mathematics

and Artificial Intelligence 49(1-4), 101–115 (2007). doi:10.1007/s10472-
007-9053-6. → 21, 279
[248] Kuznetsov, S.O.: Pattern structures for analyzing complex data. In:
International Workshop on Rough Sets, Fuzzy Sets, Data Mining, and
Granular-Soft Computing, pp. 33–44. Springer (2009). doi:10.1007/978-
3-642-10646-0_4. → 74
[249] Kuznetsov, S.O., Makhalova, T.: On interestingness measures
of formal concepts. Information Sciences 442, 202–219 (2018).
doi:10.1016/j.ins.2018.02.032. → 279
[250] Kuznetsov, S.O., Obiedkov, S.: Counting pseudo-intents and #P-
completeness. In [284] pp. 306–308 (2006). doi:10.1007/11671404_21.
→ 115
[251] Kuznetsov, S.O., Schmidt, S. (eds.): Proceedings of the 5th ICFCA, vol.
LNAI 4390. Springer (2007). doi:10.1007/978-3-540-70901-5. → 343, 354
[252] Kwuida, L.: Dicomplemented lattices. A contextual generalization of
Boolean algebras. Ph.D. thesis, TU Dresden, Germany (2004). Qucosa
Dresden, → 324, 325, 326, 332
[253] Kwuida, L.: When is a concept algebra Boolean? In [114] pp. 142–155
(2004). doi:10.1007/978-3-540-24651-0_14. → 325
[254] Kwuida, L.: A single axiom for Boolean algebras. Discrete Applied Math-
ematics 249, 85–90 (2018). doi:10.1016/j.dam.2018.08.002. → 325
[255] Kwuida, L., Machida, H.: On the isomorphism problem of concept alge-
bras. Annals of Mathematics and Artificial Intelligence 59, 223–239 (2010).
doi:10.1007/s10472-010-9194-x. → 324
[256] Kwuida, L., Sertkaya, B. (eds.): Proceedings of the 8th ICFCA, vol.
LNAI 5986. Springer Science & Business Media (2010). doi:10.1007/978-
3-642-11928-6. → 340, 341, 343
[257] Lehmann, F., Wille, R.: A triadic approach to formal concept analysis. In:
Conceptual Structures: Applications, Implementation and Theory: Third
International Conference on Conceptual Structures, ICCS’95 Santa Cruz,
CA, USA, August 14–18, 1995 Proceedings 3, pp. 32–43. Springer (1995).
doi:10.1007/3-540-60161-9_27. → 73
[258] Lehmann, K.: Der vollständige Verband minimaler Separatoren. Diplo-
marbeit, TU Dresden (2009). → 71
[259] Lengnink, K.: Formale Begriffsanalyse von Ähnlichkeitsdaten. Staatsexa-
mensarbeit, TH Darmstadt (1991). → 113
[260] Lengnink, K.: Diagrams of similarity lattices. In [295] pp. 99–107 (1993).
doi:10.1007/978-3-642-50974-2_10. → 113
[261] Leonhard, I., Winterberg, M.: Begriffliches Clustern mit konvex-ordinalen
Skalen. Diplomarbeit, TH Darmstadt (1994). → 72
[262] Lex, W.: Eine Darstellung zur maschinellen Behandlung von Begriffen. In
[165] (1986). → 70
[263] Lex, W. (ed.): Arbeitstagung Begriffsanalyse und Künstliche Intelligenz,
Informatik-Bericht, vol. 89/3 (1991). → 339
References 351

[264] Luksch, P.: Distributive lattices freely generated by an ordered set of width
two. Discrete Mathematics 88(2-3), 249–258 (1991). doi:10.1016/0012-
365x(91)90013-r. → 71
[265] Luksch, P., Skorsky, M., Wille, R.: On drawing concept lattices with a com-
puter. In [166] (2001). → 113
[266] Luksch, P., Wille, R.: Formal concept analysis of paired comparisons. In
[53] pp. 567–576 (1987). → 214
[267] Luksch, P., Wille, R.: Substitution decomposition of concept lattices. In:
G. Eigenthaler, H.K. Kaiser, W.B. Müller, W. Nöbauer (eds.) Contributions
to general algebra, vol. 5, pp. 213–220. Hölder–Pichler–Tempsky, Wien
(1987). → 214
[268] Luksch, P., Wille, R.: A mathematical model for conceptual knowledge
systems. In [54] pp. 156–162 (1991). doi:10.1007/978-3-642-76307-6_21.
→ 332
[269] Luxenburger, M.: Partielle Implikationen und partielle Abhängigkeiten
zwischen Merkmalen. Diplomarbeit, TH Darmstadt (1988). → 113
[270] Luxenburger, M.: Implications partielles dans un contexte. Mathéma-
tiques informatique et sciences humaines 113, 35–55 (1991). → 113
[271] Luxenburger, M.: Implikationen, Abhängigkeiten und Galois Abbildun-
gen: Beiträge zur Formalen Begriffsanalyse. Ph.D. thesis, TH Darmstadt
(1993). → 113
[272] MacNeille, H.M.: Partially ordered sets. Transactions of the American
Mathematical Society 42(3), 416–460 (1937). doi:10.2307/1989739. → 71
[273] Maier, D.: The theory of relational databases, vol. 11. Computer science
press Rockville (1983). → 113
[274] Markowsky, G.: Some combinatorial aspects of lattice theory. Contract
14(67-A), 0298–0015 (1973). → 71
[275] Markowsky, G.: The factorization and representation of lattices. Trans-
actions of the American Mathematical Society 203, 185–200 (1975).
doi:10.2307/1997078. → 68
[276] Markowsky, G.: The representation of posets and lattices by sets. Algebra
Universalis 11(1), 173–192 (1980). → 281
[277] Marty, R.: Foliated semantic networks: concepts, facts, qualities. Com-
puters & mathematics with applications 23(6-9), 679–696 (1992).
doi:10.1016/0898-1221(92)90129-6. → 75
[278] Medina, J., Ojeda-Aciego, M., Ruiz-Calvino, J.: Formal concept analysis via
multi-adjoint concept lattices. Fuzzy sets and systems 160(2), 130–144
(2009). doi:10.1016/j.fss.2008.05.004. → 74
[279] Medina, R., Obiedkov, S. (eds.): Proceedings of the 6th ICFCA, vol.
LNAI 4933. Springer (2008). doi:10.1007/978-3-540-78137-0. → 340, 343,
346
[280] Mehrtens, H.: Die Entstehung der Verbandstheorie, vol. 6. Gerstenberg
(1979). → 20
[281] Meschke, C.: Concept approximations – approximative notions for con-
cept lattices. Ph.D. thesis, TU Dresden (2012). Qucosa Dresden, → 74
352 References

[282] Meschke, C.: Lattices of orders. In [82] pp. 130–151 (2019).

doi:10.1007/978-3-030-21462-3_10. → 71
[283] Missaoui, R., Kwuida, L.: Mining triadic association rules from ternary
relations. In [372] pp. 204–218 (2011). doi:10.1007/978-3-642-20514-9_16.
→ 73
[284] Missaoui, R., Schmid, J. (eds.): Proceedings of the 4th ICFCA 2006, vol.
LNAI 3874. Springer (2006). doi:10.1007/11671404. → 339, 349, 350
[285] Mitas, J.: Interval orders based on arbitrary ordered sets. Discrete Mathe-
matics 144(1-3), 75–95 (1995). doi:10.1016/0012-365x(94)00288-t. → 281
[286] Monjardet, B.: A use for frequently rediscovering a concept. Order 1(4),
415–417 (1985). doi:10.1007/bf00582748. → 280
[287] Monjardet, B., Wille, R.: On finite lattices generated by their doubly
irreducible elements. Discrete Mathematics 73(1-2), 163–164 (1989).
doi:10.1016/0012-365x(88)90144-6. → 161
[288] Mori, H.: Functorial properties of formal concept analysis. In [309] pp.
505–508 (2007). doi:10.1007/978-3-540-73681-3_45. → 20, 150, 296
[289] Mowat, D.G.: A Galois problem for mappings. Ph.D. thesis, University of
Waterloo (1968). doi:10.1090/s0002-9904-1968-12060-9. → 214, 357
[290] Muller, E., Huchard, M., Martin, P., Poncelet, P., Sallaberry, A.: Rcaviz: Vi-
sualizing and exploring relational conceptual structures. In: CLA’22, vol.
CEUR-WS 3308 (2022). → 334
[291] Nation, J.: Finite sublattices of a free lattice. Transactions of the American
Mathematical Society 269(1), 311–337 (1982). doi:10.1090/s0002-9947-
1982-0637041-9. → 247, 280
[292] Neouchi, R., Tawfik, A.Y., Frost, R.A.: Towards a temporal extension of
formal concept analysis. In: E. Stroulia, S. Matwin (eds.) Advances in
Artificial Intelligence, pp. 335–344. Springer Berlin Heidelberg (2001).
doi:10.1007/3-540-45153-6_33. → 75
[293] Nourine, L., Raynaud, O.: A fast algorithm for building lattices. Infor-
mation processing letters 71(5-6), 199–204 (1999). doi:10.1016/s0020-
0190(99)00108-8. → 114
[294] Obiedkov, S.: Modal logic for evaluating formulas in incomplete con-
texts. In: Proceedings ICCS 2002 Borovets, pp. 314–325. Springer (2002).
doi:10.1007/3-540-45483-7_24. → 332
[295] Opitz, O., Lausen, B., Klar, R. (eds.): Proceedings of the 16th Annual
Conference of the “Gesellschaft für Klassifikation eV” University of Dort-
mund, April 1–3, 1992. Information and Classification. Concepts, Meth-
ods and Applications. Springer (1993). doi:10.1007/978-3-642-50974-2.
→ 348, 350, 354, 356
[296] Ore, O.: Galois connexions. Transactions of the American Mathematical
Society 55, 493–513 (1944). doi:10.2307/1990305. → 20
[297] Pawlak, Z.: Rough sets. International journal of computer & information
sciences 11, 341–356 (1982). doi:10.1007/978-94-011-3534-4. → 73
[298] Pawlak, Z.: Rough concept analysis. Bull. Pol. Acad. Sci. Math 33, 9–10
(1985). → 71
References 353

[299] Pensa, R.G., Boulicaut, J.F.: Towards fault-tolerant formal concept analysis.
In: Congress of the Italian Association for Artificial Intelligence, pp. 212–
223. Springer (2005). doi:10.1007/11558590_22. → 74
[300] Pierce, R.S.: Introduction to the theory of abstract algebras. Courier Cor-
poration (2014). doi:10.2307/2317361. → 214
[301] Poelmans, J., Ignatov, D.I., Kuznetsov, S.O., Dedene, G.: Fuzzy and rough
formal concept analysis: a survey. International Journal of General Sys-
tems 43(2), 105–134 (2014). doi:10.1080/03081079.2013.862377. → 74
[302] Pollandt, S.: Fuzzy-Begriffe: Formale Begriffsanalyse unscharfer Daten.
Springer-Verlag (2013). doi:10.1007/978-3-642-60460-7. → 357
[303] Pratt, V.: Chu spaces: Automata with quantum aspects. School on
Category Theory and Applications (Coimbra, 1999) 21, 39–100 (1999).
doi:10.1109/phycmp.1994.363682. → 71
[304] Prediger, S.: Kontextuelle Urteilslogik mit Begriffsgraphen: ein Beitrag zur
Restrukturierung der mathematischen Logik. Shaker (1998). → 333
[305] Prediger, S.: Mathematische Logik in der Wissensverar-
beitung. Mathematische Semesterberichte 47(2), 165–192 (2000).
doi:10.1007/s005910070002. → 333
[306] Prediger, S.: Terminologische Merkmalslogik in der Formalen Begriffs-
analyse. In [364] (2000). doi:10.1007/978-3-642-57217-3_5. → 331, 333
[307] Prediger, S., Stumme, G.: Theory-driven logical scaling. In: International
Workshop on Description Logics, vol. 22 (1999). → 333
[308] Priss, U.: The formalization of WordNet by methods of relational concept
analysis. In: WordNet: An Electronic Lexical Database and Some of its Ap-
plications. MIT press (1998). doi:10.7551/mitpress/7287.003.0013. → 334
[309] Priss, U., Polovina, S., Hill, R. (eds.): Proceedings of the 15th ICCS, vol.
LNAI 4604. Springer (2007). doi:10.1007/978-3-540-73681-3. → 352, 353
[310] Raney, G.N.: A subdirect-union representation for completely distribu-
tive complete lattices. Proceedings of the American Mathematical Society
4(4), 518–522 (1953). doi:10.2307/2032514. → 214, 280
[311] Reeg, S., Weiß, W.: Properties of finite lattices. Diplomarbeit, TH Darm-
stadt (1990). See also [145]. → 114, 279
[312] Reppe, H.: An FCA perspective on n-distributivity. In [309] pp. 255–268
(2007). doi:10.1007/978-3-540-73681-3_19. → 71, 280
[313] Reppe, H.: Three generalisations of lattice distributivity: An FCA perspec-
tive. Ph.D. thesis, TU Dresden (2010). Shaker, ISBN 978-3-8440-0037-5
→ 280
[314] Reuter, K.: Counting formulas for glued lattices. Order 1, 265–276 (1985).
doi:10.1007/bf00383603. → 160
[315] Reuter, K.: On the dimension of the cartesian product of relations and or-
ders. Order 6, 277–293 (1989). doi:10.1007/bf00563528. → 70, 281
[316] Reuter, K.: On the order dimension of convex polytopes. European Journal
of Combinatorics 11(1), 57–63 (1990). doi:10.1016/s0195-6698(13)80056-
x. → 281
354 References

[317] Reuter, K.: The jump number and the lattice of maximal antichains.
Discrete Mathematics 88(2-3), 289–307 (1991). doi:10.1016/0012-
365x(91)90016-u. → 71
[318] Reuter, K., Wille, R.: Complete congruence relations of concept lattices.
Acta Sci. Math. 51, 319–327 (1987). → 160
[319] Riguet, J.: Relations binaires, fermetures, correspondances de Galois.
Bulletin de la Société Mathématique de France 76, 114–155 (1948).
doi:10.24033/bsmf.1401. URL http://www.numdam.org/articles/10.
24033/bsmf.1401/. → 68
[320] Riguet, J.: Les relations de Ferrers. CR Acad. Sci. Paris 232(1729), 116
(1951). → 281
[321] Roberts, F.S.: Measurement theory. Addison-Wesley, Reading, Mass.
(1979). → 99
[322] Rouane-Hacene, M., Huchard, M., Napoli, A., Valtchev, P.: A proposal for
combining formal concept analysis and description logics for mining re-
lational data. In [251] pp. 51–65 (2007). doi:10.1007/978-3-540-70901-5_4.
→ 334
[323] Rudolph, S.: Relational Exploration - Combining Description Logics and
Formal Concept Analysis for Knowledge Specification. Universitätsverlag
Karlsruhe (2006). → 326, 331, 333
[324] Rudolph, S.: Some notes on managing closure operators. In [102] pp. 278–
291 (2012). doi:10.1007/978-3-642-29892-9_25. → 21
[325] Rudolph, S.: On the succinctness of closure operator representations.
In [173] pp. 15–36 (2014). doi:10.1007/978-3-319-07248-7_2. → 21
[326] Rudolph, S., Săcărea, C., Troancă, D.: Towards a navigation paradigm for
triadic concepts. In: Formal Concept Analysis: 13th International Con-
ference, ICFCA 2015, Nerja, Spain, June 23-26, 2015, Proceedings 13, pp.
252–267. Springer (2015). doi:10.1007/978-3-319-19545-2_16. → 73
[327] Rusch, A., Wille, R.: Knowledge spaces and formal concept analysis. In
[56] p. 427ff. (1986). doi:10.1007/978-3-642-80098-6_36. → 113
[328] Ryssel, U., Distel, F., Borchmann, D.: Fast algorithms for implication bases
and attribute exploration using proper premises. Annals of Mathemat-
ics and Artificial Intelligence pp. 101–113 (2014). doi:10.1007/s10472-013-
9355-9. → 115
[329] Săcărea, C.: Towards a theory of contextual topology. Shaker Verlag
(2001). → 75
[330] Sakurai, T.: On formal concepts of random formal contexts. Information
Sciences 578, 615–620 (2021). doi:10.1016/j.ins.2021.07.065. → 69
[331] Schader, M. (ed.): Analysing and modelling data and knowledge.
Springer–Verlag (1992). → 356
[332] Schaffert, S.: Zerlegungssätze für symmetrische Kontexte. Master’s thesis,
TH Darmstadt (1985). → 70, 71
[333] Scheich, P., Skorsky, M., Vogt, F., Wachter, C., Wille, R.: Conceptual data
systems. In [295] pp. 72–84 (1993). doi:10.1007/978-3-642-50974-2_8.
→ 71
References 355

[334] Schlemmer, T.: Annotating lattice orbifolds with minimal acting automor-
phisms. In: CLA’12, vol. CEUR-WS 972, pp. 57–68 (2012). → 308
[335] Schmid, J.: Bialgebraic contexts for distributive lattices–revisited. In [149]
pp. 403–407 (2005). doi:10.1007/978-3-540-32262-7_28. → 357
[336] Schmidt, B.: Ein Zusammenhang zwischen Graphentheorie und Verband-
stheorie. Diplomarbeit, TH Darmstadt (1982). → 71
[337] Schmidt, J.: Zur Kennzeichnung der Dedekind-MacNeilleschen Hülle
einer geordneten Menge. Archiv der Mathematik 7(4), 241–249 (1956).
doi:10.1007/bf01900297. → 68, 71, 281
[338] Schmidt, R.: Verbandstheoretische Analyse von Liniendiagrammen.
Diplomarbeit, TH Darmstadt (1988). → 280
[339] Schütt, D.: Abschätzungen für die Anzahl der Begriffe von Kontexten.
Diplomarbeit, TH Darmstadt (1987). → 112
[340] Sertkaya, B.: Formal concept analysis methods for desciption logics. Ph.D.
thesis, TU Dresden (2007). Qucosa Dresden → 331, 333
[341] Sertkaya, B., Oğuztüzün, H.: Proof of the basic theorem on concept lattices
in Isabelle/HOL. In: International Symposium on Computer and Informa-
tion Sciences, pp. 976–985. Springer (2004). doi:10.1007/978-3-540-30182-
0_98. → 68
[342] Shmuely, Z.: The structure of Galois connections. Pacific Journal of Math-
ematics 54(2), 209–225 (1974). doi:10.2140/pjm.1974.54.209. → 214
[343] Skorsky, M.: How to draw a concept lattice with parallelograms. In [401]
(1989). → 113, 280
[344] Skorsky, M.: Endliche Verbände: Diagramme und Eigenschaften. Ph.D.
thesis, TH Darmstadt (1992). → 112, 113, 279
[345] Soldano, H., Ventos, V.: Abstract concept lattices. In [372] pp. 235–250
(2011). doi:10.1007/978-3-642-20514-9_18. → 74
[346] Sowa, J.F.: Conceptual graphs summary. Conceptual Structures: current
research and practice 3, 66 (1992). → 333
[347] Spangenberg, N., Wolff, K.E.: Comparison of biplot analysis and formal
concept analysis in the case of a repertory grid. In: Classification, Data
Analysis, and Knowledge Organization: Models and Methods with Appli-
cations, pp. 104–112. Springer (1991). doi:10.1007/978-3-642-76307-6_15.
→ 72
[348] Stahl, J., Wille, R.: Preconcepts and set representations of contexts. In [166]
pp. 431–438 (1986). → 281, 332
[349] Stanley, R.P.: Enumerative combinatorics volume 1 second edi-
tion. Cambridge studies in advanced mathematics (2011).
doi:10.1017/cbo9781139058520. → 14
[350] Stephan, J.: Substitution decomposition of lattices. Ph.D. thesis, TH Darm-
stadt (1991). → 214
[351] Stephan, J.: Substitution products of lattices. In [106] pp. 321–336 (1991).
→ 214
[352] Stern, M.: Semimodular lattices: theory and applications, vol. 73. Cam-
bridge University Press (1999). doi:10.1017/cbo9780511665578. → 280
356 References

[353] Stöhr, B., Wille, R.: Formal concept analysis of data with tolerances. In:
Analyzing and Modeling Data and Knowledge: Proceedings, pp. 117–127.
Springer (1991). doi:10.1007/978-3-642-46757-8_13. → 114
[354] Strahringer, S.: Direct products of convex-ordinal scales. Order 11, 361–383
(1994). doi:10.1007/bf01108768. → 72, 247
[355] Strahringer, S.: Dimensionality of ordinal structures. Discrete Mathemat-
ics 144(1-3), 97–117 (1995). doi:10.1016/0012-365x(94)00289-u. → 72
[356] Strahringer, S., Wille, R.: Convexity in ordinal data. In [54] pp. 113–120
(1991). doi:10.1007/978-3-642-76307-6_16. → 72
[357] Strahringer, S., Wille, R.: Towards a structure theory for ordinal data. In
[331] pp. 129–139 (1992). doi:10.1007/978-3-642-46757-8_14. → 72
[358] Strahringer, S., Wille, R.: Conceptual clustering via convex-ordinal struc-
tures. In [295] pp. 85–98 (1993). doi:10.1007/978-3-642-50974-2_9. → 72
[359] Stumme, G.: Attribute exploration with background implications and ex-
ceptions. In [56] pp. 457–469 (1996). doi:10.1007/978-3-642-80098-6_39.
→ 114
[360] Stumme, G.: The concept classification of a terminology extended by con-
junction and disjunction. In: PRICAI’96: Topics in Artificial Intelligence:
4th Pacific Rim International Conference on Artificial Intelligence Cairns,
Australia, August 26–30, 1996 Proceedings 4, pp. 121–131. Springer (1996).
doi:10.1007/3-540-61532-6_11. → 333
[361] Stumme, G.: Distributive concept exploration—a knowledge acquisition
tool in formal concept analysis. In: Annual Conference on Artificial Intel-
ligence, pp. 117–128. Springer (1998). doi:10.1007/bfb0095433. → 114,
247
[362] Stumme, G., Taouil, R., Bastide, Y., Pasquier, N., Lakhal, L.: Computing
iceberg concept lattices with titanic. Data & knowledge engineering 42(2),
189–222 (2002). doi:10.1016/s0169-023x(02)00057-5. → 279
[363] Stumme, G., Wille, R.: A geometrical heuristic for drawing concept lattices.
In: R. Tamassia, I. Tollis (eds.) Graph Drawing: DIMACS GD’84, pp. 452–
459. Springer (1995). doi:10.1007/3-540-58950-3_399. → 113
[364] Stumme, G., Wille, R.: Begriffliche Wissensverarbeitung: Methoden und
Anwendungen. Springer-Verlag (2000). doi:10.1007/978-3-642-57217-3.
→ 343, 353
[365] Szathmary, L., Valtchev, P., Napoli, A., Godin, R., Boc, A., Makarenkov, V.:
A fast compound algorithm for mining generators, closed itemsets, and
computing links between equivalence classes. Ann. Math. Artif. Intell.
70(1-2), 81–105 (2014). doi:10.1007/s10472-013-9372-8. → 115
[366] Takàcs, V.: Two applications of Galois graphs in pedagogical research
(Feb. 8, 1984). Manuscript of a lecture given at TH Darmstadt. → 68
[367] Tamari, D.: The algebra of bracketings and their enumeration. Nieuw
Arch. Wisk 3(10), 131–146 (1962). → 235, 247
[368] Teo, S.: Representing finite lattices as complete congruence lattices of com-
plete lattices. Ann. Univ. Sci. Budapest. Eötvös Sect. Math. 33, 177–182
(1990). → 160
References 357

[369] Thiele, S.: Ordnungsdiagramme der Begriffsverbände von Skalenproduk-

ten. Master’s thesis, TH Darmstadt (1991). → 50, 240, 247
[370] Trotter, W.T.: Combinatorics and partially ordered sets: dimension the-
ory. Johns Hopkins University Press (1992). doi:10.56021/9780801844256.
→ 281
[371] Umbreit, S.: Formale Begriffsanalyse mit unscharfen Begriffen. Ph.D. the-
sis, Universität Halle (1994). See [302]. → 73, 114
[372] Valtchev, P., Jäschke, R. (eds.): Proceedings of the 9th ICFCA, vol.
LNAI 6628. Springer Science & Business Media (2011). doi:10.1007/978-
3-642-20514-9. → 341, 349, 352, 355
[373] Valverde-Albacete, F.J., Peláez-Moreno, C.: Extending conceptualisation
modes for generalised formal concept analysis. Information Sciences
181(10), 1888–1909 (2011). doi:10.1016/j.ins.2010.04.014. → 74
[374] Ventos, V., Soldano, H.: Alpha galois lattices: An overview. In [149] pp.
299–314 (2005). doi:10.1007/978-3-540-32262-7_21. → 279
[375] Vityaeva, E.E., Degtiarevb, V., Pakb, B., Meisterb, Y.: Formalization of “nat-
ural” classification and “natural” concepts by probabilistic generalization
of formal concepts analysis. In: CLA’20, vol. CEUR-WS 2648, pp. 59–73
(2020). → 74
[376] Vogt, F.: Subgroup lattices of finite abelian groups: Structure and cardi-
nality. In [16] pp. 241–259 (1993). → 75, 160, 214
[377] Vogt, F.: Bialgebraic contexts. Ph.D. thesis, TH Darmstadt (1994). → 75
[378] Vogt, F.: Bialgebraic contexts for finite distributive lattices. Algebra Uni-
versalis 35, 151–165 (1996). doi:10.1007/bf01190975. See also [335]. → 75
[379] Vogt, F., Wachter, C., Wille, R.: Data analysis based on a conceptual file. In
[54] (1991). doi:10.1007/978-3-642-76307-6_18. → 71
[380] Vogt, F., Wille, R.: Ideas of algebraic concept analysis. In [55] pp. 191–203
(1994). doi:10.1007/978-3-642-46808-7_17. → 75
[381] Vormbrock, B., Wille, R.: Semiconcept and protoconcept algebras: the basic
theorems. In [157] pp. 34–48 (2005). doi:10.1007/11528784_2. → 113,
322, 323, 332
[382] Voutsadakis, G.: Polyadic concept analysis. Order 19, 295–304 (2002).
→ 73
[383] Waterman, A.: Colloquium lecture at McMaster 1966. Unpublished, see
[289]. → 214
[384] Weinheimer, T.: Merkmalsprodukte formaler Kontexte und ihre Be-
griffsverbände. Diplomarbeit, TH Darmstadt (1994). → 70
[385] Wild, M.: Covering order sublattices of Boolean lattices. Preprint 244
(1991). → 280
[386] Wild, M.: The joy of implications, aka pure Horn formulas: mainly
a survey. Theoretical Computer Science 658, 264–292 (2017).
doi:10.1016/j.tcs.2016.03.018. → 113, 115
[387] Wille, R.: Subdirekte Produkte und konjunkte Summen. J. reine u. ange-
wandte Math. 239/240, 333–338 (1969). doi:10.1515/crll.1969.239-240.333.
→ 247
358 References

[388] Wille, R.: Subdirekte Produkte vollständiger Verbände. J. reine u. ange-

wandte Math. 239/240, 53–70 (1976). doi:10.1515/crll.1969.239-240.333.
→ 160, 247
[389] Wille, R.: Restructuring lattice theory: an approach based on hierarchies of
concepts. In: I. Rival (ed.) Ordered sets, pp. 445–470. Reidel, Dordrecht–
Boston (1982). doi:10.1007/978-94-009-7798-3_15. Reprinted 2009 in the
proceedings of the 7th ICFCA conference. → 68, 71, 113, 309
[390] Wille, R.: Subdirect decomposition of concept lattices. Algebra Universalis
17, 275–287 (1983). doi:10.1007/bf01194537. → 70, 160, 214
[391] Wille, R.: Line diagrams of hierarchical concept systems. Int. Classif. 11,
77–85 (1984). doi:10.5771/0943-7444-1984-2-77. Engl. version of [392].
→ 161
[392] Wille, R.: Liniendiagramme hierarchischer Begriffssysteme. In: H.H. Bock
(ed.) Anwendungen der Klassifikation: Datenanalyse und numerische
Klassifikation, pp. 32–51. Indeks-Verlag, Frankfurt (1984). → 358
[393] Wille, R.: Sur la fusion des contextes individuels. Mathématiques et Sci-
ences humaines 85, 57–71 (1984). → 148, 247
[394] Wille, R.: Complete tolerance relations of concept lattices. In: G. Eigen-
thaler, H.K. Kaiser, W.B. Müller, W. Nöbauer (eds.) Contributions to gen-
eral algebra, vol. 5, pp. 213–220. Hölder–Pichler–Tempsky, Wien (1985).
→ 160, 214
[395] Wille, R.: Finite distributive lattices as concept lattices. Atti Inc. Logica
Mathematica 2, 635–648 (1985). → 71, 245, 280
[396] Wille, R.: Tensorial decomposition of concept lattices. Order 2, 81–95
(1985). doi:10.1007/bf00337926. → 70, 72, 214, 280
[397] Wille, R.: Bedeutungen von Begriffsverbänden. In [165] pp. 161–211
(1987). → 114, 161
[398] Wille, R.: Dependencies between many-valued attributes. In [53] pp. 581–
586 (1987). → 114, 247
[399] Wille, R.: Subdirect product construction of concept lattices. Discrete
Mathematics 63(2-3), 305–313 (1987). doi:10.1016/0012-365x(87)90019-7.
→ 161, 247
[400] Wille, R.: Geometric representation of concept lattices. In: O. Opitz (ed.)
Conceptual and numerical analysis of data, pp. 239–255. Springer–Verlag
(1989). doi:10.1007/978-3-642-75040-3_17. → 113
[401] Wille, R. (ed.): Klassifikation und Ordnung. Indeks-Verlag, Frankfurt
(1989). → 355
[402] Wille, R.: Knowledge acquisition by methods of formal concept analysis.
In: E. Diday (ed.) Proceedings of the conference on data analysis, learning
symbolic and numeric knowledge, pp. 365–380. Nova Science Publishers
(1989). → 114
[403] Wille, R.: Lattices in data analysis: how to draw them with a computer. In:
I. Rival (ed.) Algorithms and order, pp. 33–58. Kluwer, Dordrecht–Boston
(1989). doi:10.1007/978-94-009-2639-4_2. → 113, 281
References 359

[404] Wille, R.: The skeletons of free distributive lattices. Discrete Mathematics
88(2-3), 309–320 (1991). doi:10.1016/0012-365x(91)90017-v. → 71
[405] Wille, R.: Tensor products of complete lattices as closure systems. In [106]
pp. 381–385 (1991). → 214
[406] Wille, R.: Begriffliche Datensysteme als Werkzeug der Wissenskommu-
nikation. In: H.H. Zimmermann, H.D. Luckhardt, A. Schulz (eds.) Men-
sch und Maschine – Informationelle Schnittstellen der Kommunikation,
pp. 63–73. Universitätsverlag Konstanz (1992). → 113
[407] Wille, R.: Concept lattices and conceptual knowledge systems. Com-
puters & mathematics with applications 23(6-9), 493–515 (1992).
doi:10.1016/0898-1221(92)90120-7. → 71, 113
[408] Wille, R.: Plädoyer für eine philosophische Grundlegung der Begrifflichen
Wissensverarbeitung. In [419] pp. 11–25 (1994). → 68
[409] Wille, R.: The basic theorem of triadic concept analysis. Order 12, 149–158
(1995). doi:10.1007/bf01108624. → 73
[410] Wille, R.: Begriffsdenken: Von der griechischen Philosophie bis zur Kün-
stlichen Intelligenz heute. Dilthey-Kastanie pp. 11–25 (1995). → 68
[411] Wille, R.: Restructuring mathematical logic: an approach based on Peirce’s
pragmatism. In: P.A. Aldo Ursini (ed.) Logic and algebra, pp. 267–281.
Marcel Dekker, New York (1996). doi:10.1201/9780203748671-12. → 75
[412] Wille, R.: Boolean concept logic. In [152] pp. 317–331 (2000).
doi:10.1007/10722280_22. → 311, 322, 324, 332
[413] Wille, R.: Boolean judgment logic. In [92] pp. 115–128 (2001).
doi:10.1007/3-540-44583-8_9. → 311
[414] Wille, R.: Contextual logic summary. Math. Preprint 2098, TH Darmstadt
(2001). → 333
[415] Wille, R.: Existential concept graphs of power context families. In: Interna-
tional Conference on Conceptual Structures, pp. 382–395. Springer (2002).
doi:10.1007/3-540-45483-7_29. → 334
[416] Wille, R.: Truncated distributive lattices: Conceptual structures
of simple-implicational theories. Order 20(3), 229–238 (2003).
doi:10.1023/b:orde.0000026494.22248.85. → 332
[417] Wille, R.: Preconcept algebras and generalized double Boolean algebras.
In [114] pp. 1–13 (2004). doi:10.1007/978-3-540-24651-0_1. → 318, 321,
332
[418] Wille, R.: Formal concept analysis and contextual logic. In: P. Hitzler,
H. Schärfe (eds.) Conceptual Structures in Practice, pp. 155–192. Chap-
man and Hall/CRC (2016). doi:10.1201/9781420060638-16. → vii
[419] Wille, R., Zickwolff, M.: Begriffliche Wissensverarbeitung: Grundfragen
und Aufgaben. B.I. Wissenschaftsverlag (1994). → 348, 359
[420] Wille, U.: Eine Axiomatisierung bilinearer Kontexte. Mitt. Math. Sem.
Gießen 200, 71–112 (1991). Diplomarbeit, → 75
[421] Wille, U.: Geometric representation of ordinal contexts. Ph.D. thesis, Uni-
versität Giessen (1995). → 75
360 References

[422] Wolff, K.E.: Concepts in fuzzy scaling theory: order and granularity. Fuzzy
Sets and Systems 132(1), 63–75 (2002). doi:10.1016/s0165-0114(02)00106-
9. → 73
[423] Wolff, K.E.: Ordnung, Wille und Begriff. Geschichten aus dem Ernst
Schröder Zentrum (2003). www.ernst-schroeder-zentrum.de → vii
[424] Wolff, K.E.: Applications of temporal conceptual semantic systems. In:
International Conference on Knowledge Processing in Practice, pp. 59–78.
Springer (2007). doi:10.1007/978-3-662-55433-3_11. → 75
[425] Xia, W.: Morphismen als formale Begriffe: Darstellung und Erzeugung.
Ph.D. thesis, TH Darmstadt (1993). → 161, 296, 308
[426] Yanagimoto, T., Okamoto, M.: Partial orderings of permutations and
monotonicity of a rank correlation statistic. Annals of the Institute of Sta-
tistical Mathematics 21, 489–506 (1969). doi:10.1007/bf02532273. → 55
[427] Zhi, H., Li, J.: Granule description of incomplete data: a cogni-
tive viewpoint. Cognitive Computation 14(6), 2108–2119 (2022).
doi:10.1007/s12559-021-09918-6. → 74
[428] Zickwolff, M.: Darstellung symmetrischer Strukturen durch Transversale.
In [106] pp. 391–403 (1991). → 308
[429] Zickwolff, M.: Rule exploration: first order logic in formal concept analy-
sis. Ph.D. thesis, TH Darmstadt (1991). → 114, 308
[430] Zschalig, C.: An fdp-algorithm for drawing lattices. In: CLA’07,
vol. CEUR-WS 331 (2007). → 113
Formal contexts and concept lattices in this
book

Living beings and water . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

Developing countries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Foundedness conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Order relation of B(K) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Free distributive lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Polarity lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Partitions of a set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Integer partitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Permutations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Projective geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
P G(2, 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Drive concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Standard scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Forum Romanum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Maximal antichains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Properties of triangles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Properties of binary relations . . . . . . . . . . . . . . . . . . . . . . . . 104
Sublattice to a closed relation . . . . . . . . . . . . . . . . . . . . . . . . 143
Computer generated diagram vs. atlas decomposition . . . . . . . . . . 180
Substitution product . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Tamari lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Generalizations of the distributive law . . . . . . . . . . . . . . . . . . . 268
Additively saturated subsets . . . . . . . . . . . . . . . . . . . . . . . . . 269
Subgroups of A4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
Graphs with four vertices . . . . . . . . . . . . . . . . . . . . . . . . . . 288
Closure systems on {a, b, c} . . . . . . . . . . . . . . . . . . . . . . . . . 291
Preconcepts of K . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
Media player . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 361
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2
Index

Θ-dependent, 112 attribute, 23

-dimension, 278 -reduced, 31
W
-rank, 278 compound, 312
W
concept, 29
abstract concept lattices, 74 extent, 28
additive line diagram, 88 many-valued, 59
additively saturated, 269 reducible, 31
adjoint, 20 terminological, 330
adjunction, 19 value, 59
ALC attribute, 331 attribute exploration, 99
algebraic, 39, 70 attribute-valence, 306
all-extensional automorphism, 284
compound attribute, 314
annotation, 287 background knowledge, 106
anti-exchange axiom, 265 base point
antichain, 2 of a lattice, 263
antimatroids, 280 of an extent, 263
antitone map, 4 basic level concept, 279
apposition, 41 Basic Theorem, 26
Armstrong rules, 94 for fuzzy concept analysis, 73
arrow relations, 33 for polyadic concept analysis, 73
arrow-closed, 121 for triadic concept analysis, 73
associative lattice operations, 6 on coherence networks, 334
atlas, 171 on concept algebras, 324
with overlapping neighbor maps, 175 on line diagrams, 113
atomistic on preconcept algebras, 321
context, 45 on preconcept lattices, 321
lattice, 7 on protoconcept algebras, 323
atoms, 7 on semiconcept algebras, 323

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024 363
B. Ganter, R Wille, Formal Concept Analysis, https://doi.org/10.1007/978-3-031-63422-2
364 Index

Begriffsverband, 68 scale, 300

biclique, 52, 71 scale measure, 309
binary relation, 1 coatoms, 7
biordinal lattice, 65 commutant, 58
biordinal scales, 64 comparable, 2
block of a tolerance relation, 132 compatible subcontext, 120
block relation, 134 complementary, 53
bond, 148, 297 complementary context, 41
concatenation of, 150 complementary set representation, 274
Boolean attribute logic, 314 complete
Boolean enrichment, 313 congruence relation, 124
Boolean factor analysis, 277 homomorphism, 8
Boolean matrix product, 44 lattice, 5
bottom of a lattice, 5 many-valued attribute, 59
boundary concepts, 30 many-valued context, 59
breadth, 251 set of implications, 94
sublattice, 8
canonical -representation, 263
W
tolerance relation, 131
canonical basis, 97 completely distributive, 8
canonical direct basis, 115 compound attributes, 312
canonical projection, 126, 164 concept algebra, 323
cardinal sum, 4 concept exploration, 114
centralizer concept faithful morphism, 292
in a group, 58 concept graphs, 333
chain, 2, 65 concept lattice, 26
chain finite, 41 concept-preserving, 292
Chu maps, 296 conceptual graphs, 333
CLA conferences, vii conceptual scale, 45
clarified, 30 conceptual scaling, 60
clause, 313 conclusion
cumulated, 316 of an implication, 93
of a formal context, 105 condition C
clause logic for closed relations, 142
of a formal context, 314 for semiconvexity, 268
clones, 58 congruence
closed relation, 140 induced by subcontext, 127
closed set, 9 congruence classes, 125
closed set of implications, 94 congruence relation, 125
closure, 9 conjunction
closure operator, 9 of attributes, 312
closure system, 9 context
number of, 21 clarified, 30
random, 70 complementary, 41
co-algebraic, 39 contrary, 41
coarser dual, 41
Index 365

many-valued, 59 disjoint union, 41

one-valued, 59 disjunction, 315
standard, 32 of attributes, 312
context Galois connections, 296 disjunctive attribute representation, 277
context morphism, 292 dismantling, 147
context-embedding, 293 distributed systems, 75
contextual logic, 333 distributive, 8
continuous, 292 domain
closure-, 15 of a many-valued attribute, 59
contracted implications, 96 dot notation, 41
contranominal scale, 46 double Boolean algebra, 322
contraordinal scale, 47 pure, 322
contrary context, 41 double coset, 287
convex geometries, 280 doubly founded, 38
convex set, 12 doubly irreducible, 146
convex-ordinal scale, 51 dual bond, 155
cross table, 23 regular, 157
cumulated clause, 316 dual order, 5
duality
data table, 59 for lattices, 6
database dependencies, 115 for ordered sets, 5
Dedekind cuts, 47 for concept lattices, 28
Dedekind-MacNeille completion, 46 dually adjoint, 16
dense, 299 dually isomorphic, 5
dense subcontext, 119 Duquenne-Guigues basis, 97
dependency
functional, 110 EL attribute, 331
ordinal, 110 equivalent scales, 300
with a tolerance, 112 exchange condition, 260
derivation operators, 24 Ext(K), 25
derived context, 62 extensionally closed, 292
description logics, 333 extensionally continuous, 292
DI-kernel, 147 extensionally equivalent, 313
dichotomic scale, 64 extensity, 9
dichotomization, 41, 72 extent, 24
dicomplemented, 324 extent closure, 80
dimension extremal point, 263
-dimension, 278
W
k-dimension, 270 face lattice, 12
order dimension, 270 factor lattice, 125, 134
direct product, 157 factors, 164
of P -contexts, 240 of a semiproduct, 107
of contexts, 44, 197 faulty data, 74
of lattices, 163 Ferrers dimension, 271
direct sum, 42 Ferrers relation, 271
366 Index

k-step, 273 generalizations of FCA, 72

filter, 49 generator
finer of closure, 14
scale, 300 geometrical diagram, 84
scale measure, 309 geometrical method, 84
FLE attribute, 331 globally equivalent, 313
flow formula, 309 glued, 173
folding, 287 gluing, 226
folding diagram, 288 graded, 258
follows (semantically), 94 Graph Theory and FCA, 71
formal conclusion, 333
formal context Hall-Dilworth gluing, 224
enriched, 313 Hasse diagram, 2
formal concept, 24 hierarchical order, 25
number of, 249 holds, 93
formal context, 23 homomorphic image, 121
category of, 152 Homomorphism Theorem, 126
number of, 69 horizontal sum, 42
random, 70
relational, 329 iceberg, 279
formal judgment, 333 ICFCA conferences, vii
free distributive lattice, 49 ideal, 49
free extent, 315 ideal-filter gluing, 224
frequent intent, 82 idempotency, 9
full column, 31 implication basis
full measure, 298 iteration free, 96
full row, 31 implication between attributes, 93
full sequent, 314 implication inference, 94
fully n-valued nominally measurable, implication of the context, 93
305 incidence relation, 23
fully dichotomially measurable, 305 incidence-preserving, 292
fully nominally measurable, 305 incidence-reflecting, 292
functionally dependent, 109 incomparable, 2
fusion, 222 incomplete knowledge, 113
fuzzy concept, 73 inconsistency, 315
infimum, 5
Galois condition, 15, 153 -dense, 7
Galois connection -founded, 39
between complete lattices, 15 -irreducible, 7
between formal contexts, 153 -morphism, 8
between ordered sets, 16 -preserving map, 8
between powersets, 18 infomorphism, 296
Galois theory, 16, 58 Int(K), 25
general interordinal scale, 49 integer partition, 54
general ordinal scale, 46 intensionally closed, 292
Index 367

intensionally continuous, 292 line diagram, 2, 29

intent, 24 linear extension, 5
frequent, 82 linear order, 2
interordinal lattice, 65 locally distributive, 262
interordinal scale Logical Concept Analysis, 311
general, 49 logical information systems, 74
one-dimensional, 64 lower bound, 5
interordinal scales, 64 lower neighbor, 2
interval, 3
interval doubling, 234 many-valued attribute, 59
interval orders, 281 many-valued context, 59
inverse relation, 1 pre-scaled, 65
inversion, 55 stratified, 65
irreducible, 7W measurable into a family of scales, 305
irredundant -representation, 262 measure, 298
irredundant join representation, 251 meet, 5
isomorphic meet subsemilattice, 8
closure systems, 15 meet-distributive, 262
formal contexts, 32, 283 meet-irreducible, 7
isomorphism between contexts, 283 meet-semidistributive, 262
iteration free minimal generator, 14
implication basis, 96 minimal separator, 71
minsupp, 82
join, 5 Möbius inversion, 14
join subsemilattice, 8 modal logic, 332
join-distributive, 262 modular, 258
join-irreducible, 7 monotone map, 3
join-semidistributive, 262 monotony, 9
Moore family, 9
k-dimension, 270 morphisms
kernel of a formal context, 279 of formal contexts, 292
knowledge spaces, 75 multi-adjoint concept lattices, 74
knowledge states, 75 mutually distributive, 203
lattice, 5 n-valent, 306
atomistic, 7 negation
of equivalence relations, 12, 53 of an attribute, 312
of permutations, 54 of formal concepts, 321
of scalings, 309 nested line diagram, 90
lattice folding, 287 Next Closure algorithm, 12
lattice isomorphism, 8 next closure system, 13
learning spaces, 75, 280 next extent, 80
lectic order, 12, 81 next intent, 80
lectically smaller, 12 nominal lattice, 65
length, 2 nominal scale, 63
of a Ferrers relation, 273
368 Index

non-redundant, 96 Pol-Inv, 58
polarity, 52
object, 23 polarity lattice, 52
-reduced, 31 polyhedron, 12
concept, 29 Pontryagin duality, 57
intent, 28 positioning rule, 88
reducible, 31 power context families, 334
one-valued context, 59 powerset, 2
opposition of formal concepts, 321 preconcept, 276, 317
orbit-maximal, 285 preconcept algebra, 321
Order, 279 preconcepts
order, 2 ordering of, 318
-embedding, 3 premise
-isomorphism, 3 of an implication, 93
-preserving map, 3 principal filter, 3
-reversing map, 4 principal ideal, 3
of formal concepts, 25 product, 4, see also direct product
relation, 2 product measure, 302
as a lattice, 44 projective plane, 57
order dimension, 270 proper premise, 95
order extension, 5 of an attribute, 96
order filter, 48 protoconcept, 317
order folding, 287 pseudo-complement, 307
order ideal, 48 pseudo-intent, 97
order relation, 2
formal context of, 44 Q-atlas, 171
ordered set, 2
ordinal factor analysis, 281 random
ordinal lattice, 65 closure system, 70
ordinal scale formal context, 70
general, 46 rank function, 258
one-dimensional, 64 reduced, 31
ordinally dependent, 110 labelling, 29
ortholattice, 53 regular
orthomodular, 57 clause set, 314
overlapping neighborhoods, 173 relation
k-ary, 329
P -context, 219 binary, 1
P -fusion, 221 relational concept analysis, 334
P -lattice, 218 residual map, 20
P -product, 219 residuated map, 20
partial implications, 113 respect
partial order, 2 an implication, 93
pattern structures, 74 retract, 161
plain scaling, 60 rigging of a substitution sum, 186
Index 369

role, 329 stability, 21, 279

role depth, 330 standard context, 32
rough set, 74, 247 stem base, 97
rule exploration, 114 strong condition of semimodularity, 258
rule of lines, 87 subconcept, 25
rule of parallelograms, 86 subcontext, 117
saturated, 128
sails of a substitution sum, 186 subdirect decomposition
saturated subcontext, 128 of a context, 207
scale of a lattice, 164
abstract, 109 subdirect product, 164, 206
attributes, 60 subdirectly irreducible, 166
biordinal, 64 subgroup lattice, 11
coarser, 300 sublattice, 141
conceptual, 45, 60 subposition, 41
concrete, 109 subscale, 299
contranominal, 46, 63 subsemilattice, 8
contraordinal, 47 subspace lattice, 11
convex-ordinal, 51 substitution product, 186
dichotomic, 64 substitution sum, 184
finer, 300 proper, 184
general interordinal, 49 substitutionally indecomposable, 190
general ordinal, 46 subtensorial
interordinal, 64 decomposition, 205
nominal, 63 product, 205
ordinal, 64 sum
values, 60 of P -contexts, see P -fusion
scale measure, 298 of a Q-atlas, 171, 178
scaling, 60 of formal contexts, 42, 216
lattice of, 309 superconcept, 25
nominal, 64 supremum, 5
self-bond, 153 -dense, 7
self-dual, 8 -founded, 39
semiconcept, 317 -irreducible, 7
semiconvex, 268 -morphism, 8
semidistributive, 262 -preserving map, 8
semimodular, 258 surmise relationship, 315
semiproduct, 43
factors of, 107 table
sequent, 313 data-, 59
set dimension, 275 Tamari lattice, 235
set representation, 88, 274 temporal concept analysis, 75
simplified annotation, 288 tensor product, 200
skeleton, 173 tensorial operations, 239
skeleton tolerance, 173 terminological attributes, 330
370 Index

test context, 313

threshold scale, 301
Titanic algorithm, 279
tolerance relation, 131
top of a lattice, 5
topological space, 11
treillis de Galois, 68
triadic concept analysis, 73
trilattice, 73

unfolding, 290
union of formal contexts, 225
unit element, 5
upper bound, 5
upper neighbor, 2

valence
of a lattice element, 307
of an object, 306
vertical sum, 42
view
conceptual, 66
contextual, 66

weak condition of semimodularity, 258

weak negation, 323
weak opposition, 323
weakly dicomplemented lattice, 324
weakly distributive, 210
web ontology language, 333
width, 2

Xia’s context product, 284

zero element, 5