Reducing Matroid Optimization to Basis Search

Robert Streit∗ Vijay K. Garg∗,†

August 9, 2024
arXiv:2408.04118v1 [cs.DS] 7 Aug 2024

Abstract
In combinatorial optimization, matroids provide one of the most elegant structures for algo-
rithm design. This is perhaps best identified by the Edmonds-Rado theorem relating the success
of the simple greedy algorithm to the anatomy of the optimal basis of a matroid [Edm71; Rad57].
As a response, much energy has been devoted to understanding a matroid’s favorable compu-
tational properties. Yet surprisingly, not much is understood where parallel algorithm design is
concerned. Specifically, while prior work has investigated the task of finding an arbitrary basis
in parallel computing settings [KUW88], the more complex task of finding the optimal basis
remains unexplored. We initiate this study by reexamining Borůvka’s minimum weight span-
ning tree algorithm in the language of matroid theory, identifying a new characterization of the
optimal basis by way of a matroid’s cocircuits as a result. Furthermore, we then combine such
insights with special properties of binary matroids to reduce optimization in a binary matroid
to the simpler task of search for an arbitrary basis, with only logarithmic asymptotic overhead.
Consequentially, we are able to compose our reduction with a known basis search method of
[KUW88] to obtain a novel algorithm for finding the optimal basis of a binary matroid with
only sublinearly many adaptive rounds of queries to an independence oracle. To the authors’
knowledge, this is the first parallel algorithm for matroid optimization to outperform the greedy
algorithm in terms of adaptive complexity, for any class of matroid not represented by a graph.

1 Introduction
Within the study of algorithm design the greedy method is a glistening jewel, capturing the ad-
miration of classrooms and practitioners alike for its simplicity and unreasonable versatility. For
context, the term “greedy algorithm” loosely applies to any computation which iteratively pro-
gresses a solution by applying a fixed heuristic in each step. Characterizing when – or even why
– such a method would achieve success has fascinated researchers for decades, and has led to an
energetic interplay between algorithm design and the study of discrete structures. One of the more
well known objects of study is the matroid, describing the idea of mathematical independence in
an abstract way. In particular, matroids abstract both concepts of linear independence in a vector
space and cyclical dependence in a graph in a unified manner by way of their exchange property.
We give the full definition below.

Definition 1.1 (Matroid). A matroid (E, M), or simply M when the ground set E is understood
from context, is a set family with members called the independent sets, which satisfy the following:

1. (Non-Emptiness) There exists at least one independent set, i.e. M =

6 ∅,
∗
Department of Electrical and Computer Engineering at The University of Texas at Austin.
†
Partially supported by NSF CNS-1812349, and the Cullen Trust for Higher Education Endowed Professorship.

1
 2. (Hereditary) For all independent X ∈ M and subsets Y ⊆ E, if Y ⊆ X then Y ∈ M,
3. (Exchange) For all independent X, Y ∈ M, if |X| > |Y | then there exists an element x ∈ X\Y
such that Y + x ∈ M.
The discovery of matroids is generally attributed to Whitney [Whi35], however this discovery
was concurrent to Nakasawa’s lost work [NK09] and Birkhoff essentially arrived at the same notions
through lattice theory [Bir35]. In the realm of combinatorial optimization, matroids define the
boundary for which the greedy algorithm successfully finds a basis X ∈ M, Pthat is an independent
set of maximal size, minimizing an additive function given by the sum y∈X w(y) for a weight
function w : E → R in a non-empty and hereditary set family. In the case where (E, M) is
a matroid, the problem of finding this optimal basis X is called matroid optimization. Then, a
solution given by the greedy algorithm is the following: Maintain a partial solution X, and in each
time step choose the element of minimum weight among those y ∈ / X such that X + y ∈ M, i.e.
among those points which maintain independence after augmentation. This is repeated until X is
a basis, which is witnessed by the inexistence of any point y ∈ E which can make X into a larger
independent set. Then, the Edmonds-Rado theorem states that the greedy algorithm correctly
solves the former optimization over a non-empty and hereditary feasibility structure M if and only
if M is a matroid [Edm71; Rad57]. Such a sweeping characterization naturally led to the study of
the greedy algorithm in a variety of generalizations overviewed in Section 2.
That said, in recent times the need for scalable computations in the age of big data has motivated
a renewed interest in the design of parallel algorithms for combinatorial optimization. This interest
began in part with the work of [BS18] examining parallel computation in the related problem of
submodular maximization, and has continued to expand to other related problems overviewed in
Section 2. A key insight of [BS18] is that minimizing the adaptivity present in an algorithm is a
necessary step towards achieving efficient parallel designs. Loosely speaking, the adaptivity used
by an algorithm is the longest sequence of questions, or queries, it must ask about a problem
instance to obtain a solution. Hence, in a query model of computation this notion measures the
complexity in terms of computational depth, representing an informational dependence on a chosen
oracle. Then, given its close connection to the greedy algorithm, the setting of parallel matroid
optimization presents an intriguing challenge. This is because the greedy algorithm is highly –
even maximally – adaptive, as it asks an independence oracle for the set of elements which can
be feasibly added to the partial solution in a step by step fashion. By examination, one sees that
the adaptive complexity of the greedy algorithm is, in fact, linear. We see this as a computational
barrier, one ready to be broken. So, is it possible to limit this adaptivity to outperform the greedy
algorithm in terms of efficiency?
For a certain subclass of matroids the answer is already known to be “yes.” This is the class
of graphic matroids, those possessing independent sets given by the edges corresponding to acyclic
subgraphs of an undirected graph. Therefore, the minimum spanning tree problem (see for example
Figure 1) is an instance of graphic matroid optimization. Well, Borůvka’s algorithm (surveyed in
Section 4) is a well known deterministic parallel solution for computing minimum spanning trees
in time polylogarithmic in the number of vertices. This algorithm is especially notable for its
simplicity, as deterministic parallel algorithms often have a complex description due to a (usually
necessary) appeal to sophisticated algebraic techniques for the sake of symmetry breaking. As a
consequence, this indicates a parallel solution to optimization in the restricted class of graphic
matroids. Fascinatingly so, because Borůvka’s analysis is the first known to employ exchange
arguments connecting linear independence to graph theory, even predating Whitney’s inception of
the field [KLS12]. This is ironic, being that the algorithmic properties of matroids were first utilized
in the design of a parallel method, since the greedy algorithm’s high degree of adaptivity makes

2
its principles somewhat orthogonal to those of parallel algorithm design. More so, when it comes
to general matroids which are not necessarily encoded by graphs, the literature so far appears to
have largely ignored aspects of parallelism in matroid optimization. In light of this discussion our
central question is now revised: Is there a parallel optimization algorithm outperforming the greedy
algorithm in terms of complexity for non-graphic matroids?
Towards addressing our central question, we initiate the study of parallel computation in the
task of finding the optimal basis of a matroid. Our goal is identifying structural matroid properties
enabling a parallel solution to matroid optimization, ideally one with a straightforward description
motivated by intuition and observation. As such, our analyses take a principled approach heavily
leveraging the mathematical properties of matroids, their duality, and their lattice of flats. This
will allow us to identify that binary matroids, i.e. those matroids representable on Galois fields
of size two, possess many favorable properties allowing for parallel matroid optimization. Binary
matroids are a rich generalization of graphic matroids. For example, this class includes those
matroids representable via totally unimodular matrices, and has seen application in the design of
linear codes [Kas08], extremal combinatorics [Luo19], and polyhedral techniques in combinatorial
optimization [GT89; Gou+12]. Influential to our methodology, we point out here that due to
√
[KUW88] it is already known that a non-optimal basis can be found in O( n) adaptive rounds
of queries to an independence oracle. This indicates that one economical approach of achieving
our goal is reducing optimization to the simpler task of basis search. In fact, this is our chosen
approach. With that, we now overview our work.

1.1 Contributions
Given its position as a parallel matroid optimization method, we begin by examining Borůvka’s
algorithm. Specifically, we do so through the lens of matroid theory to identify general proper-
ties enabling the design of parallel algorithms. Perhaps owing to its simplicity, while Borůvka’s
algorithm is a parallel method its description still exhibits some “greedy” traits. In particular,
the algorithm operates in rounds by selecting edges in parallel using a fixed, yet local, heuristic
on each vertex in a graph contraction formed with respect to the current partial solution. We
believe such a principle is the source of the algorithm’s intuitive and compact description, and so
the goal of our reexamination is to identify similarly local heuristics for application to more general
structures in our own methods. Concretely, our focus will be properties translating local attributes
of a matroid’s substructures to global guarantees on a basis. Indeed, such properties will result in
a novel reduction of the task of optimization to the simpler task of search within the class of binary
matroids, with only logarithmic overhead in terms of parallel rounds of queries to a basis search
oracle. Our reduction is given in Algorithm 2, and is the key technical contribution of our work.
We remark here that our reduction is deterministic, simple in its description, and designed using
intuition arising from our reexamination of Borůvka’s algorithm. Then, as corollary we are able to
compose our techniques with the parallel basis search method of [KUW88] to ultimately obtain a
novel method for parallel matroid optimization. Our composition correctly finds the optimal basis
of a binary matroid in sublinear adaptive rounds of query to an independence oracle, which to our
knowledge is the first such algorithm of its kind for any matroid not represented by a graph. As
a result, we find that it is possible to outperform the greedy algorithm on non-graphic matroids, at
least in complexity terms.
Speaking more quantitatively, our contributions are the following:

1. We provide an alternative analysis of Borůvka’s algorithm using the language and tools of
matroid theory in Section 4. As a result, we obtain Theorem 4.2 stating that a point is a

3
 member of the optimal basis if and only if it locally minimizes the weight function within any
cocircuit. We remark that this characterization applies to any class of matroid.

2. Next, in Section 5 we utilize the combinatorial properties of binary matroids to show that our
certificate of optimality implies that optimization is as easy as search within a deterministic
parallel computation, specifically on the binary matroid. By “as easy,” we mean that our
reduction adds only logarithmic overhead in terms of complexity. This novel reduction is
given in Algorithm 2 of Section 5.2, whose analysis requires new perspectives on modular
pairs in the lattice of flats given in Section 5.1.

3. With that, we identify a novel method for finding the optimal basis of a binary matroid in
√
O (log r · n) adaptive rounds of queries to an independence oracle, where n gives the size of
the matroid’s ground set and r the rank. We do so by composing our reduction with a simple
parallel basis search method of [KUW88], overviewed in Section 5.3.

4. Finally, in Section 6 we conclude by providing exciting open questions for future work.

As a byproduct, our work has a number of secondary contributions of consequence. For exam-
ple, our argument verifying the complexity of our reduction utilizes properties of flats generating
sublattices isomorphic to Boolean algebras, and as such our examination of these phenomena in
Section 5.1 reveals a new understanding of the occurrence of Boolean algebras in geometric lat-
tices. Furthermore, Appendix A also provides a simplified direct proof of the equivalence between
White’s circuit based characterization of binary matroids [Whi71; Whi87], and Tutte’s forbidden
minor theorem [Tut58]. Our reduction relies on this characterization, so we provide this proof
not only to assist other researchers but to also concretely identify the challenges associated with
generalizing our ideas to other matroids (challenges which we report in Section 6). However, we
remark that our efforts employ a technique making use of the lattice of flats, resulting in a direct
proof which is significantly simpler than that currently available in the literature (see, e.g., Chapter
2 of [Whi87]). Furthermore, we provide proofs of many “folklore” results concerning relationships
arising from matroid duality in Appendix B, as well as a proof of certain details omitted from the
analysis of Algorithm 3 in [KUW88] in Appendix C.

2 Related Work
As indicated in the introduction, since the pioneering work of [Edm71; Rad57] many succeeding
works have studied the greedy algorithm in a variety of settings obtained by relaxing matroid op-
timization. One natural generalization is that of relaxing the matroid structure, which is to say
considering different types of set families or relational structures representing “feasibility” alto-
gether. This approach has included polymatroids [DW73; Edm03], matroid embeddings [HMS93],
and greedoid languages [KLS12]. On other side, there are those generalizations obtained by relax-
ing the assumption that the objective function is additive. This is primarily exemplified through
the closely related problem of submodular optimization [NWF78; Fuj05] and its connections to the
greedy algorithm, as well as optimization problems defined over M ♮ -concave functions (a subclass
of submodular functions containing additive functions) in the context of discrete convex analysis
[Mur98; DW92]. In particular, for submodular optimization the greedy algorithm obtains a 1 − 1e -

approximation to monotone submodular maximization in the unconstrained setting [NWF78], and

this approximation is still achievable via different algorithmic means in the presence of matroid
constraints [Cal+07]. Furthermore, by [Fei98] the 1 − 1e -approximation is the best possible with-

4
out utilizing any other instance dependent structure. We provide this background because most
related contemporary work examines facets of the latter problems, which we now outline.
In the context of parallel algorithm design, submodular optimization has recently seen some-
what of a renaissance. This began with the breakthrough work of [BS18] showing an efficient
parallel algorithm for approximate unconstrained monotone submodular maximization. However,
this initial attempt left a gap between its approximation ratio and the best 1 − 1e -approximation.

After a sequence of work [BRS19a; FMZ19b; CQ19b; CQ19a] a method obtaining essentially the
best approximation, sans a small ǫ > 0, with polylogarithmic adaptivity in the presence of matroid
constraints was ultimately obtained in [BRS19b]. This line of work was subsequently extended to
non-monotone submodular maximization [FMZ19a], submodular minimization [BS20], M ♮ -concave
maximization with cardinality constraints [Kup+20], and the streaming setting [Kaz+19]. Rele-
vantly, because [BRS19b] worked in the setting of matroid constraints one should question whether
their techniques can be applied towards matroid optimization. Well there is a glaring difference in
our settings, being where these work obtain approximation algorithms we are interested in exact
solutions. Furthermore, it appears that the analysis of these methods is currently limited to the
approximation setting. For instance, [Kup+20] examine the applications of methods of [BRS19b]
to M ♮ -concave maximization in the presence of constraints given by a uniform matroid to achieve
a (1 − ǫ)-approximation for fixed ǫ > 0. This setting is well known for the greedy algorithm’s
ability to obtain an exact solution [DT95], and includes additive objective functions. However, the
complexity of [Kup+20] varies in 1ǫ , and so setting ǫ to guarantee an exact solution would induce
a level of adaptivity polynomial in the range of the objective function. Hence, the search for exact
algorithms in the case of parallel matroid optimization is warranted. Moreover, for the sake of
further comparison we remark that our methods are deterministic, whereas those of the preceding
work rely on randomized sampling procedures.
Focusing specifically on matroids, there is little work for comparison in the context of parallel
computation. Notably, [KUW88] examines the relative power of various oracles associated with
non-empty hereditary set families in the task of searching for a basis. However, they do not ex-
amine the task of optimization, that of searching for the optimal basis. The related problem of
matroid intersection has seen some recent attention in the parallel computing setting. Recently,
[CCK23] showed a lower bound stating that an algorithm solving matroid intersection in polyloga-
rithmic adaptive rounds of queries to a rank oracle is not possible. In response, [Bli22] obtained a
deterministic solution with superpolylogarithmic, yet sublinear, adaptivity. Furthermore, towards
understanding the power of a deterministic parallel matroid intersection algorithm, [GGR24] gave
a reduction from search to decision for matroid intersection in the rank oracle model. That said,
given the difference in problem setting our work is not immediately comparable. However, on
the topic of reduction, [Kar98] gives a reduction from searching for the optimal basis to verifying
whether a basis is optimal. This is similar in spirit to our reduction in Section 5, however a subtle
difference is that we reduce to the task of searching for an arbitrary basis, rather than verification.
Furthermore, the reduction presented in [Kar98] utilizes sampling on a general matroid, and does
not examine a parallel computing setting. In contrast, our reduction is deterministic, and restricts
its attention to binary matroids to obtain progress in parallel algorithm design. Finally, we point
out that [Daw80] reexamines the greedy algorithm for matroid optimization using cocircuits as
well. However, due to the difference in purpose [Daw80] does not make similar claims or show the
relationship between the local optimality within a cocircuit and membership in the optimal basis
(i.e. Theorem 4.2 of the present manuscript), and is likewise not applicable to our work.

5
 e9

e1 e2 e1 e2
e8

e7 e7
e6 e6

e3 e3
e4 e4

e5 e5
(a) A connected graph over 7 unlabeled vertices and (b) The minimum weight spanning tree when
9 numerically labelled edges. weights increase with edge label.

Figure 1: A graph is shown left and its minimum weight spanning tree is shown on the right.
For matroids corresponding to acyclic subgraphs of a connected graph, the spanning trees give the
bases. Hence, the minimum weight spanning tree is the solution to an optimization instance over
such a matroid.

3 Preliminaries
Our central problem of focus is that of matroid optimization. That is, given a matroid M on a
ground set E and a weight function w : E → R find a basis of minimum weight, i.e. solving,
X
min w(y). (1)
X∈M
y∈X

In this work we assume that w is injective, which is to say that weights are distinct, to ensure there
is a unique optimal solution. For algorithmic purposes this is of no concern, as one can enforce
our assumption by choosing deterministic tie-breaking rules a priori. Now, we remark that matroid
optimization is a generalization of the well known task of identifying a minimum spanning tree in
a graph, or spanning forest when the graph is not connected. See Figure 1 for example. We point
this out because we will use Borůvka’s algorithm for computing a minimum weight spanning tree
to identify general properties enabling parallel solutions to matroid optimization. In particular, our
focus will be on identifying suitable structures to enable parallel matroid optimization by reduction
to the search for any arbitrary basis. Since our end goal is developing new parallel algorithms, it
should be noted that these forementioned “suitable structures” will consist of means of translating
local properties to global guarantees. In particular, we say that a point z ∈ E is the locally optimal
point in a set X ⊆ E if it is the point of least weight in X. Then, one way to interpret the success of
the greedy algorithm is that it identifies a locally optimal point in the set of allowable continuations
of a partial solution step by step to build a globally optimal basis, i.e. the solution to Eq. 1. This
leads to a powerful solution only emboldened by its simplicity, a mantra which we’d like to at least
approximate if not maintain in our own methods. Hence, our investigation will identify conditions
allowing one to use similar local heuristics of selecting a locally optimal point, but now from many
various sets simultaneously to obtain a solution in parallel.

6
 Our model of computation will be the following: We assume an oracle query model divided
into rounds, wherein each round the algorithm is allowed to make polynomially many simultaneous
queries to prespecified oracles. To be precise, by simultaneous queries we mean that any query made
in some round cannot be constructed using information obtained from responses made in that same
round. The complexity is then the number of rounds needed for the algorithm to terminate. Since
the number of rounds taken provides a measure of informational dependence an algorithm has on
the oracle, this provides a measure of adaptivity needed by the algorithm and has been referred to
as the adaptive complexity in a body of recent literature examining the related task of submodular
optimization. See, for example [BS18] and its succeeding works given in Section 2. Chiefly, the
adaptive complexity with respect to independence oracles will be our final measure of performance.
However, we will also measure the overhead of our reduction using the number of adaptive rounds
of queries to both independence oracles and oracles abstracting a basis search routine. These are
the primary oracles considered in this work. Specifically, the latter is what we denote as a basis
search oracle which accepts a matroid M as query and responds with an arbitrary basis of M.
Considering such an oracle is nonstandard, and we do so only for the purposes of abstracting the
task of identifying a basis in our reduction. The formerly mentioned oracle, called the independence
oracle in the literature, accepts any subset X ⊆ E of the ground set and responds with a “yes”
answer whenever X is independent, and a “no” answer otherwise. The representation of a matroid
is a tricky subject because the number of independent sets is often exponential in the cardinality
of the ground set, and so a succinct description of the matroid may not exist. See any of [JK82;
HK78; HK81; May08] for research more deeply examining this phenomenon. As a consequence, in
the analysis of algorithms interacting with matroids usually an oracle of some kind is used. For
example, even Edmond’s seminal work appeals to the independence oracle to describe the greedy
algorithm [Edm71]. Here we make use of independence oracles, even though we largely limit our
attention to matroids representable over fields, because there are only a few known classes for which
a linear representation can be effectively computed at this time [Mis+20]. In fact, the independence
oracle is the weakest among the typical choices of matroid oracle, and by instantiating the action
of finding a basis using a parallel method of [KUW88] we will identify a novel method for binary
matroid optimization with sublinear adaptive complexity in the independence query model.
In what follows we detail the combinatorial tools employed in our work. Specifically, we begin
by more fully espousing the structures and operations associated with matroids. Succeeding this,
we will give a brief overview of the lattice theoretic devices we rely upon in this work. Then, we
conclude this section by giving examples of matroids which will become of increasing importance
throughout the manuscript. For the interested reader, the standard references for matroid theory
are [Wel10] and [Oxl06], whereas [Whi87] contains more advanced topics. On the other hand,
standard references for lattice theory are [Gra11] and [DP02], along with [Gar15] and [CLM12].
We remark that the latter references are notable for their respective work compiling applications
of lattices and partial orders to computer science topics. We assume some basic familiarity with
discrete math, particularly graphs and partially ordered sets. Finally, we adopt the commonly
abused notation X + x for taking the union of a set X with a singleton x, translating to X + x ,
X ∪ {x}. And likewise, we take a similar point of view for subtraction by letting X − x , X \ {x}.

3.1 Matroids
Matroids define an amorphous object, and as a result there are an overwhelming number of ways
one can interpret them. Consequently, there are many matroid theoretic concepts used in this work
which we now attempt to overview. To begin, as given in Definition 1.1 a matroid’s independent
sets can be thought of as a description of how points can be arranged to make lines, triangles, and

7
higher order simplices. Specifically, for a matroid M over a ground set E we can assign a rank
to each subset X ⊆ E corresponding to the cardinality of the largest independent set it contains.
Syntactically, this is described by the rank function ρ : 2E → Z+ defined as,

ρ(X) , max |X ∩ Y |.
Y ∈M

Furthermore, the rank function is submodular, meaning that,

ρ(X ∩ Y ) + ρ(X ∪ Y ) ≤ ρ(X) + ρ(Y ),

for all X, Y ⊆ E. In light of this, whenever X and Y are subsets of E witnessing the above with
equality they are called a modular pair. Now, connecting ρ back to our prior description, the rank
of a set corresponds to its dimension and so we call ρ(E) the rank of the matroid. Hence, those
elements of E forming singleton sets of rank 1 are called points, which form a line when arranged
into a set of rank 2, and so on. We point out here that there technically can exist elements of rank
0, which are referred to as loops. Continuing on, we see that rank 2 sets can be related to a set of
collinear points in an affine space by analogy, whereas sets of higher rank correspond to a collection
of points in a subspace. Then, in the presence of such an analogy it should be no surprise that
matroids possess a notion of span. This is, the span of a set X ⊆ E is given by those elements of
the ground set which can be augmented into X without changing the rank. The span is defined
precisely by the rank closure operator σ : 2E → 2E ,

σ(X) , {y ∈ E | ρ(X) = ρ(X + y)}.

We point out here that the function σ is in fact a closure operator, and as implication it is monotone
in that X ⊆ Y implies σ(X) ⊆ σ(Y ) for any choice of X, Y ⊆ E. Furthermore, an independent
set B ∈ M spans a subset X ⊆ E whenever their spans are in alignment, i.e. σ(B) = σ(X).
Correspondingly, in this scenario we would call B a basis of X. Hence, the bases of the matroid
span the ground set E.
Complementary (in a quite literal sense) to the span of a set X ⊆ E is that of its continuations,
given by the operator Γ : 2E → 2E ,
Γ(X) , E \ σ(X).
In English, the continuations of a set are points which can increase the rank by augmentation.
Therefore, adding a continuation to an independent set leads to another independent set. A flat
is a set that is equal to its own span, i.e. a set F ⊆ E such that F = σ(F ). Equivalently then, a
flat is a set such that the addition of any new point will increase the rank. Furthermore, using the
properties of the rank closure σ one can show that the flats are closed under intersection, see [Wel10]
for example. In the next section, we will see that the flats of a matroid correspond to a special
lattice ordering with many favorable properties. The coatoms of this lattice, i.e. those elements
lying directly below the top element, are given by the hyperplanes. Specifically, the hyperplanes
are those flats of rank one less than that of the matroid. As such, hyperplanes are closely related
to the bases of a matroid. For example, let M be a matroid of rank r. Then, because the rank of
a basis B less one of its elements is simply r − 1 it follows that σ(B − x) is always a hyperplane
for all x ∈ B. The fact that we can associate hyperplanes with bases in this way will be important
in our algorithmic methods. Finally, the reader should observe that in the absence of loops the
singleton of every element of the ground set is equal to its own span, and is hence a flat.
The previous concepts were defined by a matroid’s independent sets. In contrast, we now
examine properties defined by the dependent sets. Specifically, a circuit is a dependent set of

8
minimal size. As a result, if C ⊆ E is a circuit of M then the removal of any element will lead to
an independent set, i.e.,
C − x ∈ M, ∀x ∈ C.
This language originates in graph theory since the circuits of a graph are given by cycles of minimal
size, in contrast to the bases of the corresponding matroid given by those acyclic subgraphs of
maximal size. The combinatorial properties of circuits on binary matroids will be critical to the
correctness of our reduction in Section 5.2.
Finally, we examine the operations of a matroid and matroid duality. These operations are best
compared to that of edge contraction and deletion in a graph – the former collapsing a component
into a single vertex and the latter simply removing edges from the graph. To be precise, the
contraction of a matroid by a set X ⊆ E is given by removing independent sets which do not
contain a subset spanning X. Specifically, this is given by M/X and defined as,

M/X , {Y \ σ(X) ∈ M | Y ∈ M}.

In essence then, for any basis B of X and basis B ′ ∈ M/X of the contraction, the union B ∪ B ′
is a basis of the original matroid M. This corresponds to the intuition that a spanning tree of
a connected graph can be decomposed into two parts: One which spans some component and
another that spans the contraction formed by that component. The other matroid operation is
that of deletion. Specifically, given a set X ⊆ E the deletion by X is given by the independent sets
containing only elements which are not members of X. This is written as M \ X, defined by,

M \ X , {Y ∈ M | X ∩ Y = ∅}.

Contraction and deletion can be used to obtain new matroids which are, in some sense, a substruc-
ture of the original matroid. Because of this, any matroid which can be obtained through successive
deletion and contraction operations on M is called a minor of M [Wel10; Oxl06]. In Section 5
and Appendix A we will discuss more about how minors are a useful concept for characterizing a
specific matroid’s structure. Finally, matroid duality will play a prominent role in our work. Given
a matroid M, its dual matroid M∗ is given by the subsets of the complements of bases of M, i.e.,

M∗ , {X ⊆ E | (∃B ∈ M) ρ(B) = ρ(E) and X ⊆ E \ B}.

We point out here that (M∗ )∗ = M, i.e. taking the dual is an involution. Furthermore, a class of
matroid is self dual if the class is closed under taking duals. Binary matroids are an example of
such a class. Central to our algorithmic reduction, showing that optimization is as easy as search in
the case of binary matroids, is that the continuations of a matroid’s hyperplanes correspond to the
circuits of its dual. Because of its importance we state the following lemma, proven in Appendix B
using the dual rank function.

Lemma 3.1 (Folklore). Let M be a matroid on a ground set E. Then, H ⊆ E is a hyperplane if

and only if its continuations Γ(H) form a circuit in the dual matroid M∗ .

Traditionally, a circuit of the dual matroid M∗ is called a cocircuit of M. Cocircuits alone are
paramount in our work due to our characterization of optimality given in Theorem 4.2. However,
the analysis of our reduction in Section 5 will also make heavy use of representation of cocircuits as
the continuations of hyperplanes, and accompanying properties of the lattice of flats. Now, towards
approaching such ideas we will overview a small amount of lattice theory used in our methodology.

9
3.2 The Lattice of Flats
Lattices are a special algebraic structure that are most conveniently defined as a partially ordered
set. In the context of our work, lattices are a useful tool for answering “analytic” questions in
discrete settings (for example see [Sta11]), one employed in the analysis of our reduction in Section 5.
Specifically, let (Ω, ⊏) be any partially ordered set, with the symbol ‘⊑’ taking the reflexive closure
of the ordering relation. Furthermore, we will let the symbol ‘≺’ give the covering relation. That
is, x ≺ y if and only if x ⊏ y and there exists no z ∈ Ω such that x ⊏ z ⊏ y. Recall, given x, y
and z in Ω we say that z is a lower bound of x and y if z ⊏ x and z ⊏ y. Taking this further, z is
the greatest lower bound, or meet, of x and y whenever z is a unique element greater than all lower
bounds of x and y, i.e.,
z ′ ⊏ x and z ′ ⊏ y =⇒ z ′ ⊑ z.
Observe, a meet may not exist. However, in the case it does the meet of x and y is written as x ⊓ y.
The concepts of least upper bound and join are defined analogously, and given by x ⊔ y when exists.
A partially ordered set such that a join and meet are always guaranteed existence is a lattice.
Lattices are relevant because the flats form a lattice when ordered by inclusion. Specifically,
given two flats F1 and F2 of a matroid M the corresponding meet and join are given by the span
of the intersection and span of the union respectively, i.e.,

F1 ⊓ F2 = σ(F1 ∩ F2 ), and, F1 ⊔ F2 = σ(F1 ∪ F2 ).

Since the flats are closed under intersection, it is helpful to observe that the meet simplifies to
F1 ⊓ F2 = F1 ∩ F
2 . For
a matroid M, we symbolically refer to its corresponding lattice of flats by
L (M) = σ 2E , ⊓, ⊔ . It is well known that the lattice of flats of a matroid always possesses
two special properties: Firstly, L (M) is atomistic. To understand this, an element is an atom of a
lattice if it covers the bottom element. A lattice then is atomistic if every element can be written as
a join of atoms. In the case of a matroid without loops, the bottom element of L (M) is the empty
set ∅ and so the atoms are given by singletons formed from the points. Therefore, the atomistic
property follows from the fact that every flat F is spanned by the points within any one of its
bases. The second special property is that L (M) is semimodular. This means that for all flats
F1 , F2 ∈ σ 2E , if F1 ≻ F1 ⊓ F2 then F2 ≺ F1 ⊔ F2 . A lattice which is atomistic and semimodular

is referred to as a geometric lattice, and it is well known that matroids and geometric lattices are
isomorphic representations of each other (see [Gra11] for more information). Viewing matroids
through their lattice of flats can reveal many interesting properties. For example, a principal filter
generated by an element x consists of x and all of its upper bounds. One fact which we will employ
at times is that contraction yields a principal filter of L (M), as one can verify,

L (M/X) ∼ = F ∈ σ 2E | F ⊇ σ(X) ,


(2)

for all X ⊆ E. Similar statements can be made about deletion by a flat (but not an arbitrary
subset) to show certain minors correspond to intervals of the lattice. See [Wel10] for details. In
the next section, when introducing examples of matroids we will also describe their corresponding
lattice of flats to help the reader obtain some familiarity with these structures.
Finally, before continuing we must introduce a few final pieces of terminology. Given a collection
of flats F1 , . . . , Fm , the corresponding sublattice they generate will be denoted by hF1 , . . . , Fm i.
This is the smallest substructure of L (M) closed under taking meets and joins containing the flats
F1 , . . . , Fm . A sublattice is a Boolean algebra if and only if it is representable as a powerset over
elements given by the atoms. Using the language we’ve already defined, it follows that a Boolean
algebra is equivalently an atomistic lattice where every join over a set of atoms is associated with

10
a unique element. It is also worth noting that this is then a complemented structure which, in
particular, makes every atom associate with a unique coatom. To see this, suppose that for the
sake of example hF1 , . . . , Fm i is a Boolean algebra with atoms given by F1 , . . . , Fm . Then, because
every join of atoms leads to a unique element it follows that the coatoms are given by each possible
join of m − 1 of the generating flats. Hence, the mapping,
G
Fi 7→ Fj , (3)
j6=i

is a bijection associating each atom with a unique coatom. In particular, this mapping implies
that an atom lies below all coatoms except its dual, so the mapping is inverted from a coatoms by
taking the meet of the other coatoms. For the interested reader, we remark that this is a particular
example of a duality associated with distributive lattices, a class of lattices isomorphic to a powerset
in a more general way (see [DP02]). Furthermore, from semimodularity one can show that L (M)
is graded by ρ, in the sense that the rank of a flat is consistent with the covering relation as F1 ≺ F2
implies ρ(F1 ) = ρ(F2 ) − 1. This is because a flat F1 is covered by F2 if and only if there exists a
continuation x ∈ Γ(F1 ) making B + x spans F2 for all bases B of F1 . One insightful consequence
is that two flats are a modular pair (as defined in the last section) if they cover their meet, as
semimodularity implies they are covered by their join. This is to say,

F1 , F2 ≻ F1 ⊓ F2 =⇒ ρ(F1 ∩ F2 ) + ρ(F1 ∪ F2 ) = ρ(F1 ) + ρ(F2 ),

since then ρ(F1 ) = ρ(F2 ), ρ(F1 ∪ F2 ) = ρ(F1 ) + 1, and ρ(F1 ∩ F2 ) = ρ(F1 ) − 1. The properties of
modular pairs of flats which we construct in this way will be so important to the analysis of our
optimization methods that Section 5.1 is devoted to their favorable properties on binary matroids.
Remark 3.2. The term modular pair generally has a different meaning in the context of lattice
theory, in that it is those pairs of elements which satisfy half of the diamond isomorphism theorem
associated with modular lattices. However, it is well known that on semimodular lattices (and hence
geometric lattices) these two usages coincide, see [Sta71] for example.

3.3 Examples
In this section we give some examples of prominent matroids and their corresponding lattice of flats.
Beyond giving the reader more familiarity with matroids and associated structures, our examples
will also allow us to introduce the classes of matroids featured in this paper. As indicated, Whitney’s
initial foray, i.e. [Whi35], revealed a notion of independence which abstracts the properties of
linearly independent sets in vector spaces and forests in graphs. Coincidentally, our first two
examples consist of the graphic and linear matroids respectively, making this connection explicit.
Example 3.3 (Graphic Matroid). A graphic matroid represents the structure of a graph’s acyclic
subgraphs. Specifically, let M be a matroid on a finite set E. Then, M is graphic if and only if
there exists a graph G with edges given by E such that M consists of edge sets inducing the acyclic
subgraphs of G. To see that this structure is a matroid, first fix such a graph G and note that M is
nonempty as the empty subgraph contains no cycles. Furthermore, because a subgraph of an acyclic
subgraph of G is also acyclic, M is also hereditary. Finally, let X and Y be two sets of edges
both inducing acyclic subgraphs where |X| > |Y |. Since both X and Y induce acyclic subgraphs,
there must exist an edge e incident to some vertex of G which is not incident to any edge in Y . In
particular, there is an e ∈ X \Y , and it is impossible for Y + e to encode a cycle. Hence, Y + e ∈ M
and the exchange property follows. A graph and the Hasse diagram of the flats of the corresponding
graphic matroid are given in Figure 2.

11
 {a, b, c, d}

d {a, b} {a, c} {a, d} {b, c, d}

a
b

c {a} {b} {c} {d}

∅
(a) A graph with edges {a, b, c, d}, which forms (b) The Hasse diagram of the lattice of flats of the cor-
the ground set of a graphic matroid. responding graphic matroid.

Figure 2: Shown above is a graph (left) and the lattices of flats of the associated graphic matroid
(right). The independent sets consist of the acyclic subgraphs, and a basis is given by any spanning
tree, which the reader can verify are given by {a, b, c}, {a, b, d}, and {a, c, d}. Observe, that among
the rank 2 flats the three containing the edge a are themselves acyclic subgraphs. These acyclic
subgraphs are flats as the addition of any remaining edge forms a spanning tree. However, there is
one rank 2 flat of cardinality 3, consisting of the triangle {b, c, d}. There are three independent sets
spanning this triangle ({b, c}, {b, d}, and {c, d} to be precise), and the addition of the remaining
edge to one of these independent sets cannot increase the rank because a cycle forms.

Example 3.4 (Linear Matroid). A linear matroid is a matroid having a representation over a
vector space. To be specific, suppose that M is a finite matroid on a ground set E, and let V be a
vector space over some field F which we leave unspecified for now. Then, M is a linear matroid with
representation over F whenever there exists a function f : 2E → V , called a representation, from
the subsets of E to the vector space V such that a set X is independent if and only if the restriction
f |X is injective and the vectors comprising the direct image f (X) are linearly independent. By
requiring f |X to be injective, we ensure that each independent set is mapped to its own unique
vector (at least among the other independent sets), which imposes some amount of “soundness” in
the encoding given by f . However, we point out that this detail of no concern if one were to make
the entirety of f injective by simply assigning each atom its own unique vector. Furthermore, we
see that the imposition for linear independence within the image of an independent set realizes the
exchange axioms and hereditary axioms, due to the Steinitz lemma from linear algebra and that
linear independence is closed under taking subsets, respectively. We point out the simplest linear
matroid is essentially that given by the independent sets of a vector space. That said, a Hasse
diagram of the lattice of flats corresponding to the subspaces of F32 is given in Figure 3.

While the definition of a linear matroid at first glance appears quite abstract, we will see specific
examples of linear matroids and corresponding mappings in the forthcoming sections. In fact, within
the context of this paper those linear matroids representable over the Galois field of order two F2 ,
referred to in the literature as binary matroids [Whi87], will be an important subject of focus.
Interestingly, graphic matroids themselves are a subclass of binary matroids. We will provide more
intuition for this relationship in Section 5, but for now the reader can examine Figure 3 to see a
correspondence between F32 and the graphic matroid of Figure 2 encoded in their respective lattices
of flats. That said, in Section 5 we will exploit certain combinatorial properties of binary matroids

12
  
1, 0, 0, 1, 1, 0, 1
0, 1, 0, 1, 0, 1, 1
0, 0, 1, 0, 1, 1, 1

             
1, 0, 1 1, 0, 1 1, 0, 1 0, 0, 0 0, 1, 1 0, 1, 1 1, 0, 1
0, 1, 1 0, 0, 0 0, 1, 1 1, 0, 1 1, 0, 1 0, 1, 1 1, 1, 0
0, 0, 0 0, 1, 1 0, 1, 1 0, 1, 1 0, 1, 1 1, 0, 1 0, 1, 1

             
1 0 0 1 1 0 1
0 1 0 1 0 1 1
0 0 1 0 1 1 1

∅
h0i
Figure 3: The Hasse diagram corresponding to the lattice of flats given by F32 − 0 . We omit the
0
zero vector because its span is the empty set, which makes it a loop. Linearly independent sets
of vectors of a vector space form the most “full” linear matroidhin isome sense, and it follows that
0
every matroid of rank 3 representable on F2 is a minor of F32 − 0 . Since the graphic matroid in
0
Figure 2 is of rank 3, this concept
nh should apply. Inio
fact, one can work out that the lattice of flats
0 0 0 1
i h i h i h
obtained from the deletion F32 \ 0 , 0 , 1 , 0 is isomorphic to that in Figure 2.
0 1 1 1

{a, b, c, d}

{a} {b} {c} {d}

Figure 4: The lattice of flats of U42 , the four point line, expressed over the ground set {a, b, c, d}.
Observe that the hyperplanes are given by the elements of the ground set, which implies that the
basic sets are precisely those subsets of {a, b, c, d} with cardinality 2.

algorithmically for parallel matroid optimization.

Finally, another important class of matroid making appearance in this paper is that of the
uniform matroid – a class with basis elements so uniform in their description.

Example 3.5 (Uniform Matroid). A uniform matroid of rank r is a matroid such that a subset of
E is independent if and only if its cardinality is less than or equal to r. Syntactically, a uniform
matroid of rank r over a ground set of n elements is referred to as Unr . Whenever the rank r = 2,
then we refer to Un2 as the n point line as its lattice of flats consists of a single line which happens
to cover n points. An illustration of the Hasse diagram of the flats of U42 is given in Figure 4.

Beyond their conceptual simplicity, uniform matroids are a particularly useful class because of

13
 Algorithm 1: Borůvka’s Minimum Weight Spanning Tree [Bor26a; Bor26b; NMN01]
input : Graph G and Weight Function w : E(G) → R
output : Set of Edges X ⊆ E(G) Giving the Minimum Weight Spanning Tree

1 X←∅
2 G′ ← G
3

4 while edges E(G′ ) 6= ∅ do

5 forall vertices v ∈ V (G) in parallel do
6 X ← X + arg min{w(e) | e ∈ δ(v)}
7
8 G′ ← G/X
9

10 output X

their use as forbidden minors. In particular, one combinatorial description of binary matroids given
by Tutte [Tut58] is that a matroid is representable over F2 if and only if it does not contain U42 as
a minor. By referencing Figure 4, we see that in the language of lattice theory this is equivalent to
stating that the lattice of flats of a matroid does not contain an interval of height two with six or
more elements if and only if it is binary. We survey these ideas more fully in Appendix A, as their
implications form the foundation for our reduction.

4 Borůvka’s Algorithm
In this section we will introduce Borůvka’s minimum weight spanning tree algorithm and provide
an alternative analysis of its correctness. See [NMN01] for a translation of Borůvka’s original work
[Bor26a; Bor26b]. Furthermore, we remark that we are not the first to employ Borůvka’s algorithm
in modern research. Indeed, portions of his method are used both in Karger, Klein, and Tarjan’s
randomized minimum spanning tree algorithm running in time linear in the edges [KKT95], and
the fastest known deterministic minimum spanning tree algorithm due to Chazelle [Cha00]. For us,
because Borůvka’s algorithm holds the prestigious position of being a decisively simple method for
parallel optimization of a graphic matroid we provide an alternative analysis of its correctness, the
goal being the identification of general principles allowing for parallel computation in more abstract
structures. Our approach will reveal a novel connection between the cocircuits of a matroid (of any
kind) and the optimality of a basis. In particular, we will see that local optimality of points within
cocircuits will translate to global guarantees on a solution. In essence, this connection will then
imply that a parallel optimization scheme can be obtained simply by the simultaneous identification
of many cocircuits. Following this, Section 5 will fully develop the algorithmic implications of this
insight in the case of binary matroids.
Now, assume we are given a graph G with vertices V
(G) and edges E(G). Then Borůvka’s
algorithm, given in Algorithm 1, is a simple O log2 |V (G)| time parallel method operating in the
following way: Let one round correspond to an iteration of the while loop (line 4). Then, one round
consists of every vertex v ∈ V (G) examining their incident edges δ(v), and selecting the incident
edge of minimum weight (that which is locally optimal among the incident edges of the vertex)
to be independently added to an ongoing partial solution X (lines 5 and 6). Following this, once
all vertices have augmented the partial solution X by their selection, the graph is contracted by
X (line 8) and this procedure is repeated. To be specific, because the edges chosen in a round

14
 e9

e1 e2 e9
e8

e7
e6
e8
e3
e4 {e1 } {e3 , e4 , e5 } {e2 }
e6
e7
e5

(a) The first step of the algorithm selects the edges (b) After contraction, we obtain the above with ver-
e1 , e1 , e3 , and e4 , which partitions the graph into tices labelled by the edges captured in the corre-
three components identified by the dashed lines. Ob- sponding component. These vertices participate in
serve that these edges correspond to those of mini- the second round of the algorithm, and select e6 and
mum weight incident to each vertex. e7 to form a completed spanning tree.

Figure 5: An illustration of the evolution of a graph and edges selected by Borůvka’s algorthm
(shown in red dotted lines) when executed on the graph of Figure 1. The minimum weight spanning
tree is computed in two steps, corresponding to the left and right subfigures.

partition the graph into a number of connected components, one sees that the vertices in any
round of Algorithm 1 correspond to the components induced by the edges selected by those of the
preceding round. This relationship is illustrated in Figure 5, which shows the selected edges and
corresponding evolution of a graph throughout a parallel execution of Borůvka’s algorithm.
For context, we give a short sketch of the arguments certifying the time complexity and correct-
ness of this method: Where time complexity is concerned, note that because every edge is incident
to two vertices it follows that the size of the vertex set of the contraction at the end of a round
is no more than half the size of the vertex set of the graph at the start of a round. Hence, we
can have at most log |V (G)| rounds until we have contracted to a single vertex, at which point the
edge set is empty and the algorithm terminates. Furthermore, the simultaneous computations done
within each round, i.e. each vertex selecting its best edge in line 6, can each be accomplished via
parallel sorts each taking O(log |V (G)|) time. Finally, the action of taking a graph contraction at
the end is largely symbolic, and real implementations require vertices to detect the induced com-
ponent they correspond to. This can be accomplished via some standard atomics in logarithmic
time (see Chapter 5 of [JáJ92] for a technique using pointer jumping). With that done, we move
on to correctness. Traditionally speaking, correctness is typically proven by a kind of greedy ex-
change argument. First, one identifies that the edge of minimum weight of those incident to some
vertex v must be a member of the minimum spanning tree. Indeed, let e = arg mine′ ∈δ(v) w(e′ ) and
consider any spanning tree not containing e. If we add e to the spanning tree, then we will induce
a (single) cycle. Since this cycle goes through v, it must also contain another edge e′ incident to v.
Hence, removing e′ will break the cycle, thus making another spanning tree of lesser weight in turn.
However, the reader should see that this exchange argument generalizes to say that those edges

15
of minimum weight incident to any component must too be a member of the minimum spanning
tree. To be specific, any cycle going through a component is preserved under contraction, and so
one may reenact our previous argument on the graph contraction. Therefore, it follows that the
minimum weight spanning tree of any contraction G/X, for X ⊆ E(G), must be a subgraph of the
minimum weight spanning tree of G. And so, correctness follows.
Though we were sparing in some details for the sake of brevity this is, in essence, the workings
behind the success of Borůvka’s algorithm. Our goal now is to identify more general principles
underlying this success which can be used in the construction of more general parallel optimization
methods for matroids. Towards doing so, in the next section we reanalyze the correctness of
Borůvka’s algorithm using the tools and language of matroid theory. Our analysis will show a
connection between the cocircuits and optimality, and will reveal a general procedure for the design
of parallel matroid algorithms in turn.

4.1 Bonds, Cocircuits, and an Alternative Analysis of Correctness

So, to what extent can these ideas be generalized? Unfortunately, where the complexity is concerned
the answer is not very much as the argument relied on the properties of binary relations. However,
as far as correctness is concerned, the argument towards the end of the last section reveals a subtle
point: Edges belong to the minimum weight spanning tree if they are the unique incident edge of
minimum weight for some component. Furthermore, because some edge of a suboptimal spanning
tree can always be exchanged for an edge of lower weight from an optimal tree, this characterization
also goes the other way. To be precise, our sketch identifies that an edge belongs to the minimum
weight spanning tree if and only if they are the unique edge of minimum weight for some component.
Before we can understand the significance of this insight, we introduce the concept of a bond which
we will see is a specialization of the cocircuits of a matroid to graphs.

Definition 4.1 (Bond). Let G be a graph. A set of edges X ⊆ E(G) is a cut set if X consists of
edges incident to a single component of G. Then, the cut set X is a bond if it is maximal in the
sense that X is contained in no other cut set.

A cut set is named so because those edges incident to a component go between that component
and its complement formed from the leftover vertices. So, such edges form a cut of the graph. A
bond then is simply the entire set of edges incident to a single component. We also observe that
the bond generated by a component is the same generated by the component’s complement. An
illustrative example of bonds generated in the execution of Borůvka’s algorithm is given at a later
point in Figure 6. So, all said the sketch we gave shows that edges belong to the minimum weight
spanning tree if and only if they are of minimum weight among other edges in some bond.
To see how we can generalize this insight, firstly observe that in a certain sense one can associate
graph components with the hyperplanes of the corresponding graphic matroid. Indeed, suppose
we fix edges F ⊆ E(G) forming the entirety of any component in a graph G. Then, see that F
is the flat of the graphic matroid as the addition of any other edge will connect another vertex,
thereby increasing the rank. Furthermore, if we let F ′ consist of the edges corresponding to the
complementary component (i.e. those edges within the subgraph obtained by deleting the vertices
in the component associated with F ) then F ⊔ F ′ must give another flat with edges given exactly
by the union F ∪ F ′ . This is because edges beyond this union connect the components induced
by F and F ′ , thereby increasing the rank. It follows then by our preceding discussion that the
continuations Γ(F ⊔ F ′ ) are precisely the bond generated by the components corresponding to F
and F ′ . However, because F and F ′ are complementary components the addition of any edge in the
bond Γ(F ⊔ F ′ ) to F ⊔ F ′ will span the matroid. Putting this in graph theoretic terms, a basis of

16
F ⊔ F ′ is simply the union of two trees spanning the components given by F and F ′ , respectively, in
the graph. Then, the addition of any edge in the bond Γ(F ⊔ F ′ ) connects the two components in a
way that creates one tree spanning the entire graph, i.e. a basis of the graphic matroid. Therefore,
F ⊔ F ′ is a hyperplane and it follows that the bond Γ(F ⊔ F ′ ) is a cocircuit in the graphic matroid.
Because this construction is generic, we find that bonds correspond to cocircuits. Hence, while
we remark that we haven’t yet proven anything, our intuition relating bonds to the minimum
spanning tree indicates that edges belong to the optimal basis of a graphic matroid if and only
if they are all locally optimal in some cocircuit. And so, if true then it follows that Borůvka’s
algorithm is correct because it is able to identify many such locally optimal choices of edges from
varying cocircuits sufficiently fast. This insight will form the foundation for our approach to parallel
matroid optimization. However, besides the fact that we must verify this intuition, there is also a
glaring question: Do locally optimal points within cocircuits always comprise those of the optimal
basis, or is there some miraculous property of graphs causing this phenomenon in the case of a
graphic matroid? The next sequence of arguments shows that it is, in fact, the former. Specifically,
the main result of this section is the following, showing a connection between the global optimality
of a basis and the local optimality of point within cocircuits holding for a matroid of any kind.

Theorem 4.2. Let M be a matroid, w : E → R an injective weight function, and X ⊆ E be

any subset of the ground set. Then, X is a basis of minimum cumulative weight if and only if it
contains all points which are of minimum weight in some cocircuit.

Towards proving this result, we first show that composing locally optimal points of many co-
circuits together always leads to an independent set.

Lemma 4.3. Let M be a matroid, w : E → R an injective weight function, and C1∗ , . . . , Cm ∗ be any

collection of cocircuits. Then, the set composed of the minimum weight point in each of the given
cocircuits is independent. Symbolically, this is {x1 , . . . , xm } ∈ M for xi , arg miny∈Ci∗ w(y).

Proof. Let the points x1 , . . . xm be as defined above, and assume without loss of generality that the
cocircuits are selected so that the chosen xi are distinct (i.e. i 6= j implies xi 6= xj ). Furthermore,
we will also impose that the cocircuits are labelled in accordance to their locally optimal point’s
weight, so that i < j implies w(xi ) < w(xj ). Notice then, this means that a point xi of lesser label
must be in the span of a hyperplane corresponding to a cocircuit of greater label, as w(xi ) < w(xj ).
To make this observation precise, let H1 , . . . , Hm be the unique selection of hyperplanes making,

Γ(Hi ) = Ci∗ , ∀i ∈ {1, . . . , m},

whose existence is guaranteed by Lemma 3.1. Then, we’ve already identified that,

i < j =⇒ xi ∈ Hj , ∀i, j ∈ {1, . . . , m},

due to our labelling scheme. Therefore, for labels i < j it follows by the monotonicity of the rank
closure that σ({x1 , . . . , xi }) ⊆ Hj .
We can turn this observation into an inductive argument: To begin, suppose by hypothesis
that {x1 , . . . , xi−1 } is independent. Recall that xi ∈ Ci∗ = Γ(Hi ), and hence xi ∈ / σ(Hi ). Then,
because σ({x1 , . . . , xi−1 }) ⊆ Hi , by monotonicity of the rank closure we must also have xi ∈ /
σ({x1 , . . . , xi−1 }). Therefore, {x1 , . . . , xi } is independent as well. Finally, our base case is given by
observing that {x1 } itself is independent. Indeed, because x1 ∈ C1∗ it must be that x1 is not a loop,
and so the lemma follows.

17
 We can now show the theorem. It will follow from an elementary exchange argument which is
common to combinatorial analyses within matroids, one which is even reminiscent in spirit of that
we sketched for the initial correctness of Borůvka’s algorithm.

Proof of Theorem 4.2. To begin, we examine the “only if” direction. This is, we prove that if a
subset X ⊆ E does not contain all points which are locally optimal in some cocircuit, then X
cannot be a basis. Fix such a set X, and suppose that X is a basis as otherwise the desired
conclusion follows in an uninteresting way. By assumption, there exists a point y ∈ X which does
not minimize w in any cocircuit. This allows us to do the following: Since X − y is rank one less
than the matroid, it follows that σ(X − y) is a hyperplane and so Γ(X − y) gives a cocircuit by
Lemma 3.1. Then, by assumption it must be that,

y 6= arg min w(z),

z∈Γ(X−y)

and so there exists z ∈ Γ(X − y) such that w(z) < w(y). Therefore, X − y + z is a basis and,
X X
w(x) < w(x).
x∈(X−y+z) x∈X

Thus, X is not an optimal basis.

Now for the “if” direction, suppose that a subset X ⊆ E contains all elements which are
locally optimal in some cocircuit. It follows from Lemma 4.3 that X is independent. Furthermore,
by applying Lemma 3.1 to our assumption, it follows that for each hyperplane H the subset X
contains some element lying beyond H. Hence, it cannot be that the span σ(X) is included in any
hyperplane, and X spans the matroid. This makes X a basis, and so X is a feasible solution for our
optimization. What is left is verifying that X is optimal. Well, suppose that the optimal solution
O ∈ M is such that O 6= X. By assumption, O contains at least one point z not locally optimal in
any cocircuit. However, this means we can apply the argument we gave for the “only if” direction
to find that O cannot actually be the optimal basis, a contradiction.

To reiterate, Theorem 4.2 shows a local to global phenomenon, where those points which are
locally optimal within a cocircuit compose together to form a globally optimal basis. For the goal
of parallel matroid optimization, this implies a suitable method is identifying many cocircuits and
applying the local heuristic of selecting their respective points of minimum weight in a simultaneous
fashion. Furthermore, if one can guarantee that the identified cocircuits have many differing locally
optimal points, then this idea will naturally lead to an efficient parallel algorithm. So, given that
we have simplified the task of optimization to that of identifying a suitable subset of the cocircuits,
our task is now resolving how one can find such a subset with our desired guarantees.
In the context of graphic matroids, Borůvka’s algorithm reveals that graphs provide some kind
of convenient encoding for doing so. Specifically, recall that by the nature of graph contraction
the vertices in each step of Algorithm 1 correspond to components of the given graph. Hence, the
action of a vertex selecting its best incident edge corresponds to selecting the locally optimal edge
in the associated bond. In particular, this means that Borůvka’s method identifies cocircuits by
examining the bonds associated with each component given by a vertex in a step of the procedure.
This principle is illustrated in Figure 6, associating bonds with their components corresponding
to a step of Borůvka’s algorithm given in Figure 5. It is now clear how Theorem 4.2 implies the
correctness of Borůvka’s algorithm, and so we conclude this section with the following corollary.

18
 e9 e9

e1 e2 e1 e2
e8 e8

e7 e7
e6 e6

e3 e3
e4 e4

e5 e5

(a) In the first round of the algorithm the upper (b) In the second round of the algorithm, the vertex
right vertex selects the best edge among its incident formed by contracting {e3 , e4 }, selected in the previ-
edges {e2 , e8 , e9 } forming the cut shown by the red ous round, selects the best edge among its incident
dotted line. This set is a bond, as it is the full set edges {e6 , e7 , e8 } forming the cut shown by the red
of incident edges to the component consisting of the dotted line. This vertex corresponds to the compo-
upper right vertex alone, or equivalently the comple- nent comprised of {e3 , e4 , e5 }, and so the contracted
mentary component formed by {e1 , e3 , e4 , e5 , e6 , e7 }. vertex’s set of incident edges is a bond.

Figure 6: An illustration of the bonds corresponding to the incident edges of certain vertices in the
graph corresponding to the execution of Borůvka’s algorithm given by Figure 5. The left examines
those edges corresponding to the upper right vertex in the first round, whereas the right examines
those corresponding to the contracted vertex given by the selected {e3 , e4 } edges in the second
round. The induced components and their complements in the original graph are circled by a
dashed line. The edges under consideration form maximal sized cut sets shown in red, i.e. a bond.

Corollary 4.4. Let M be a graphic matroid, and w : E → R an injective weight function. Then,
Algorithm 1, i.e. Borůvka’s minimum weight spanning tree algorithm, correctly outputs a basis of
minimum weight.

Proof. Since the incident edges of a component form a bond, it follows that in every step the
algorithm is only capable of augmenting the partial solution X (line 6 of Algorithm 1) with an edge
corresponding to a locally optimal point in the associated cocircuit of the graphic matroid. Upon
termination then, X is a basis composed only of points locally optimal in some cocircuit. Therefore,
because X is full rank Theorem 4.2 and the hereditary property imply that X consists of all such
locally optimal points, and so correctness follows by another application of Theorem 4.2.

4.2 Implications Beyond Graphic Matroids

The results of the last section show that a parallel method for matroid optimization is attainable
if one has the ability to quickly identify cocircuits with distinct locally optimal points. Our goal
in the sections which follow is to examine and generalize this notion in the context of matroids
more general than that of graphic matroids. However, when moving beyond graphic matroids it

19
is not immediately clear how one can easily do this for more general structures. Specifically the
vertices of a graph encode a connectivity structure between the points of the corresponding graphic
matroid, which we posit is the reason why Borůvka’s algorithm is able to identify the locally optimal
points of cocircuits so quickly. Unfortunately for us, moving beyond graphic matroids removes this
structure, and we must rely upon other means to identify cocircuits.
Towards identifying such means, let M be an arbitrary matroid and B ∈ M a basis. Then,
recall that for all points x ∈ B the span σ(B − x) is a hyperplane, which by Lemma 3.1 makes
Γ(B −x) a cocircuit in turn. So, given a basis of a matroid of rank r we are able to identify r distinct
cocircuits, at the very least. This observation motivates our reduction of matroid optimization to
basis search in the context of binary matroids. However, we remark that the existence of a parallel
basis search method is not sufficient for implying a fast parallel method for optimization, as the
r cocircuits identified in this way might collide on the same locally optimal point. This is why
we appeal to a few helpful properties afforded by the binary matroid. These properties and our
reduction are detailed in the next section, which culminate into a novel method for binary matroid
optimization with sublinear adaptivity in the independence query model.

5 Optimizing Binary Matroids

Recall the definition of a linear matroid given in Section 3.3. A binary matroid is a linear matroid
with a representation over the Galois field of order two, which we denote by F2 . Graphic matroids
themselves are a prime example of binary matroids. To see this, let M be a graphic matroid of
rank r and V be the vector space V = Fr+1 E
2 . Then, a suitable representation f : 2 → V of M
is given by labeling the coordinates of a vector by the vertices (of which there will be r + 1) and
letting the coordinates of f (e) take a “1” if its label is incident to the edge e, and a “0” otherwise.
Such a representation is injective, and one will find that summing the vectors corresponding to a
set of edges forming a cycle will give the zero vector. It follows that linearly independent vectors
correspond to acyclic subgraphs, that is, independent sets of M. For example, below we give a
representation of the running example, given initially by the graph in Figure 1, in matrix form
where the ith column corresponds to f (ei ) and the coordinates (i.e., the rows) are labeled in an
arbitrary manner consistent with the vertices:

1 0 0 0 0 1 0 0 0
 
1 0 0 0 0 0 0 0 1
 
0 1 0 0 0 0 0 1 1
 
0 0 1 1 0 1 1 1 0
 
0 1 0 0 0 0 1 0 0
 
0 0 0 1 1 0 0 0 0
0 0 1 0 1 0 0 0 0

Observe that the third, fourth, and fifth columns sum up to the zero vector, which aligns with our
desired behavior as {e3 , e4 , e5 } gives a cycle of the graph given in Figure 1.
Of course, binary matroids are much more general constructions than those induced by graphs.
A prominent example of a more general binary matroid is that of regular matroids, which are those
representable over every (not necessarily finite) field. Interestingly, regular matroids are equivalent
to the class of totally unimodular matrices (see [Tut58; Ger87] for a proof and [Whi87] for a broader
introduction to this relationship), which are those matrices for which every square submatrix has
a determinant of zero or positive/negative one. This is to say, a sufficient condition for a binary
representation of a matroid is that arranging the image of a representation f into the columns of

20
a matrix gives one which is totally unimodular. As a result, our methods imply that one can find
an optimally weighted basis of the column space of a totally unimodular matrix with complexity
sublinear in the number of columns, at least when the mechanisms for testing for independence are
abstracted away. Given the connections between totally unimodular matrices and the integrality
of linear programming relaxations of integer programs (see [Sch+03]), we remark that this is an
interesting avenue for future work exploring broader application of our schemes to parallel algorithm
design for combinatorial optimization. However, we point out that the class of binary matroids is
much larger than just these two examples.
While interesting, the representation theoretic formulation of the binary matroid is a bit opaque
for our algorithmic desires. That said, there is a simple combinatorial description given by way
of forbidden minors [Whi87; Wel10]. Specifically, for a class of matroids K the matroid N is a
forbidden minor whenever M ∈ K if and only if M does not contain a minor isomorphic to N .
Forbidden minors thus provide constructive descriptions of the identities of matroids within a class.
And, where binary matroids are concerned, Tutte showed that the four point line U42 is a forbidden
minor [Tut58]. We discuss this more as well as its implications in Appendix A. One such implication
is White’s characterization, stating that a matroid is binary if and only if the symmetric difference
C1 △ C2 of two distinct circuits with nonempty intersection yields another circuit, whenever C1
and C2 are a modular pair [Whi71]. Recall, the challenge in trying to use a basis to identify a
selection of cocircuits for use in a parallel matroid optimization is that these cocircuits may collide
on the same locally optimal point. Well, since binary matroids are self dual we can utilize White’s
characterization to generate further cocircuits via symmetric difference in the presence of collisions.
In what follows, we will develop a better understanding of this characterization and its algorithmic
uses by further examining the properties of modular pairs of flats in the lattice L (M). In particular,
we will identify conditions sufficient for the correct application of White’s characterization in our
reduction. Following this, in Section 5.2 we will finally introduce our reduction in Algorithm 2.
Then, by composing our reduction with the parallel basis search algorithm of [KUW88] we will
obtain a novel optimization algorithm for binary matroids breaking the linear complexity barrier
set by the greedy algorithm in Section 5.3.

5.1 Modular Pairs and (Co-)Free Sets

In this section, we explore the properties of modular pairs of circuits and hyperplanes, as well
as their connections and algorithmic implications. This begins with a certain characterization of
binary matroids by way of their circuits. In particular, White proved that the symmetric difference
of a modular pair of circuits always yields another circuits on a binary matroid [Whi71; Whi87].
This is stated below, along with Tutte’s forbidden minor characterization [Tut58].

Theorem 5.1 ([Tut58; Whi71; Whi87]). Let M be a matroid. Then, the following are equivalent:

1. M is representable over F2 ,

2. M does not contain a minor isomorphic to U42 ,

3. And, for all distinct circuits C1 and C2 , if C1 ∩ C2 6= ∅ and C1 and C2 form a modular pair
then C1 △ C2 is another circuit.

To aid in understanding the workings and limitations of our algorithmic methods, we provide
a simplified proof of White’s characterization of binary matroids in Appendix A. We note here
that our approach only uses Tutte’s forbidden minor theorem and properties of the lattice of flats,
resulting in a proof significantly more direct than available in the current literature (e.g. [Whi87]).

21
 As identified, we can use Theorem 5.1 to generate further cocircuits in the case of collisions
on the same locally optimal point. Specifically, in this case a symmetric difference clearly will
not contain such a point. As a result, the theorem provides a reasonable tool for use in making
a parallel algorithm using Theorem 4.2 as a certificate of optimality terminate quickly. However,
applying Theorem 5.1 in this way requires the premise that these cocircuits colliding upon the
same locally optimal point form a modular pair. Hence, our algorithmic methods must ensure that
this property holds. To do so, we will use connections between cocircuits and their corresponding
hyperplanes, connections being stronger than those previously given in Section 3. Specifically, the
technical details behind the success of our reduction in Section 5.2 begin with the observation
that two cocircuits are a modular pair if and only if their corresponding hyperplanes are also a
modular pair. This correspondence has been invoked without proof in other literature, for example
in [Whi87], so we defer its proof to Appendix B.
Lemma 5.2 (Folklore). Let M be a matroid, C1∗ and C2∗ be two cocircuits, and H1 and H2 the
corresponding hyperplanes such that Γ(H1 ) = C1∗ and Γ(H2 ) = C2∗ . Then, we have that C1∗ and C2∗
are a modular pair if and only if H1 and H2 are a modular pair.
The statement of the previous lemma is rooted in matroid duality. The astute reader must
imagine then that a suitable approach towards proving this statement lies in relating the rank
function to its dual, and such a reader would be correct. This is the method that we take in
Appendix B. However, Lemma 5.2 is still not enough to enforce the guarantees we need, being
that we need colliding cocircuits to always make modular pairs. Our solution lies in making an
association between independent sets and cofree sets, which are flats that satisfy a strong version
of pairwise modularity. Following [KLS12], we first define free sets as flats presenting as the atoms
of a Boolean algebra lying upon the lattice of flats and then assign cofree sets to the dual concept.
Definition 5.3 (Free Sets). Let M be a matroid. Then, a set of flats F1 , . . . , Fm is free if the
sublattice hF1 , . . . , Fm i has atoms given by F1 , . . . , Fm and is isomorphic to a Boolean algebra.
Similarly, flats are cofree if sublattice hF1 , . . . , Fm i has coatoms given by F1 , . . . , Fm and is, once
again, isomorphic to a Boolean algebra.
By combining this definition with Lemma 5.2, one observes that cocircuits corresponding to
hyperplanes forming a cofree set will always make modular pairs. Hence, within such a context, we
are free (pardon the pun) to apply Theorem 5.1 with reckless abandon to resolve collisions between
cocircuits possessing the same locally optimal points as we please. Toward achieving such abandon
algorithmically, we use the next sequence of lemmas to identify that cofree sets of hyperplanes can
be easily generated using a single basis. This insight relies on first connecting free sets of flats
and independent sets. Intuitively, whenever we are given an independent set {x1 , . . . , xm } ∈ M
of a matroid without loops, it follows that the sublattice h{x1 }, . . . , {xm }i given by the singletons
is a Boolean algebra. This is because, by the hereditary property every subset of {x1 , . . . , xm } is
independent, and every such subset also spans a distinct flat because the continuations must differ.
Of course, since {x1 }, . . . , {xm } is a sequence of points it follows that the points give the atoms
of the corresponding sublattice, as the points must all cover the empty flat. So, independent sets
can be associated with free sets formed by points of the matroid, where the identity of the free set
is given by the points in the independent set. The next lemma generalizes this notion to give the
identities of the flats forming any free set, using the properties of matroid contraction. In fact, this
association is biconditional.

F
Lemma 5.4. Let M be a matroid and F1 , . . . , Fm a sequence of flats such that each element covers

F
the meet, i.e. Fi ≻ m j=1 Fj for all i ∈ {1, . . . , m}. Then, the flats are free if and only if S there exists
m
points x1 , . . . , xm such that {x1 , . . . , xm } ∈ M/ j 6=i j
F , and each x i satisfies x i ∈ Fi \ j=1 Fj .

22
Proof. For both directions, let F1 , . . . , Fm be flats such that each covers the total meet. We make
this assumption because if not true, then the flats fail to make a free set in a trivial way. Fur-
thermore, without loss of generality, assume the flats are distinct. Firstly, we examine the “if”
direction. Suppose that there exist points x1 , . . . , xm such that each xi lies in the span of Fi
alone and {x1 , . . . , xm } is an independent set in the contraction by the meet. Since the atoms of
hF1 , . . . , Fm i will always be given by the flats, we need only verify that our assumptions make the

F
sublattice a Boolean algebra. We do so by constructing an isomorphism between hF1 , . . . , Fm i and
the powerset on {x1 , . . . , xm }. Well, because we have the covering relation Fi ≻ m j=1 Fj for all
i ∈ {1, . . . , m}, it follows that,

G
 
m
σ Fj + xi  = Fi , ∀i ∈ {1, . . . , m}.
j=1

Sm
In particular, because of our assumption that each point satisfies xi ∈ Fi \ j=1 Fj , we have,

G
 
m G
σ Fj ∪ {xi }i∈I  = Fi , ∀I ⊆ {1, . . . , m}.
j=1 i∈I
F
Hence, the mapping {xi }i∈I 7→ j∈I Fj preserves order operations, and we need only show that
it is injective and surjective to witness an isormorphism. Well, injectivity follows from {xj }j ∈I / ⊆

F
′ ′
Γ ({xi }i∈I ) holding for all choice of indices I ⊆ {1, . . . , m}, where Γ gives the continuations on
the contraction M/ m j=1 Fj . Furthermore, recall L (M) is atomistic. Therefore, surjectivity holds
since a sublattice of an atomistic lattice is also atomistic, implying that F the only flats present in
hF1 , . . . , Fm i are joins of subsets of {F1 , . . . , Fm }. Hence, {xi }i∈I 7→ j∈I Fj is an isomorphism
between the powerset on {x1 , . . . , xm } and hF1 , . . . , Fm i.

F
Now for the “only if” direction, suppose there does not exist points satisfying the conditions of

F
the lemma. It follows that there exists no way to compose continuations of the meet m j=1 Fj to
get an independent set {x1 , . . . , xm′ } ∈ M/ m F
j=1 j where each point is a continuation leading to a
distinct flat from the sequence F1 , . . . , Fm . Well, for any such independent set it is not possible to

F
have any point xi lying in the span of two or more of the flats since we assumed they are distinct.
m Fm
Therefore, let {x1 , . . . , xm′ } ∈ M/ j=1 be an independent set spanning j=1 Fj in particular, i.e.
the top element of hF1 , . . . , Fm i. By atomicity the height of hF1 , . . . , Fm i can be no more than m,
and so m′ ≤ m. Hence, there must exist a flat Fi that does not contain any point from x1 , . . . , xm′
in its span. It follows then that the top element of hF1 , . . . , Fm i can be written as a join over only
m − 1 of its atoms, and thus cannot be a Boolean algebra.

Translating this into the dual concept, we obtain a corollary connecting the bases of matroid
contractions to cofree sets of hyperplanes. As a specific result, any basis B ∈ M of a minor obtained
via contraction gives a cofree set of hyperplanes via its spans σ(B − x), for each x ∈ B.

F
Corollary 5.5. Let M be a matroid and H1 , . . . , Hm a sequence of hyperplanes. Then, the hyper-
planes are cofree if and only if there exists a basis B = {x1 , . . . , xm } of the contraction M/ m
j=1 Hj
with a labelling satisfying,

G
 
m
σ Hj ∪ (B − xi ) = Hi , ∀i ∈ {1, . . . , m}. (4)
j=1

23
 {a, b, c, d}

d {a, b} {a, c} {a, d} {b, c, d}

a
b

c {a} {b} {c} {d}

∅
(a) Spanning tree {a, b, c} of the graph. (b) The sublattice h{a, b}, {a, c}, {b, c, d}i.

Figure 7: Shown left is a spanning tree of the graph in Figure 2, i.e. a basis of the corresponding
graphic matroid, while right is the Boolean algebra generated by the corresponding cofree set.
Specifically, it is evident that σ({a, b, c}−b) = {a, c} and σ({a, b, c}−c) = {a, b} whereas σ({a, b, c}−
a) = {b, c, d}. These hyperplanes are the coatoms of h{a, b}, {a, c}, {b, c, d}i, which we observe is a
Boolean algebra. Hence, {a, b}, {a, c}, {b, c, d} is cofree and generated by the basis {a, b, c}.

Proof. Let the points B = {x1 , . . . , xm } satisfy Eq. 4. Furthermore, for each i ∈ {1, . . . , m} let,

G
 
m
Fi , σ  H j + xi  .
j=1

As such, each Fi is a flat covering the bottom element of L (M/ m F

j=1 Hj ) by assumption, since
xi ∈ Γ(Hi ) implies that it is a continuation of the meet by monotonicity of rank closure and the
rank can only be one more than that of the meet. So, since B is a basis, and hence an independent
set, in this contraction we may apply Lemma 5.4 to find that hF1 , . . . , Fm i is a Boolean algebra
with atoms given by F1 , . . . , Fm . However, because,

G
 
G m
Fj = σ  Hj ∪ (B − xi ) , ∀i ∈ {1, . . . , m},
j6=i j=1

by construction we can apply the assumption made in Eq. 4 to find that the hyperplanes satisfy,
G
Hi = Fj , ∀i ∈ {1, . . . , m},
j6=i

thereby giving the coatoms of hF1 , . . . , Fm i. Examining the converse direction, if this construction
is not possible then it follows by the reverse direction of Lemma 5.4 that the atoms of hH1 , . . . , Hm i
cannot be a free set. As consequence, hH1 , . . . , Hm i cannot be a Boolean algebra.

Therefore, it is always true that taking the sets of continuations Γ(B −xi ) for each i ∈ {1, . . . , r}
from a basis B = {b1 , . . . , br } on a matroid M of rank r will yield r cocircuits corresponding to
a cofree set of hyperplanes. And, as a result, our reduction will be permitted to generate further
cocircuits using symmetric difference operations by way of Theorem 5.1. To more clearly define the
relationship between bases and cofree sets of hyperplanes, we say that a basis B = {x1 , . . . , xm }

24
of the contraction M/ m F j=1 Hj generates the sublattice hH1 , . . . , Hm i whenever H1 , . . . , Hm is a
cofree set of hyperplanes satisfying Eq. 4. We choose this terminology because the argument given
in Corollary 5.5 reveals that such a basis B identities the flats in the sublattice hH1 , . . . , Hm i.
Specifically, if B generates hH1 , . . . , Hm i then by application of Eq. 4 it follows that there is a
unique subset of points in B associated with every flat in the sublattice. In particular, because

F
the sublattice is a Boolean algebra, every one of its flats F is associated with a subset of indices
I ⊆ {1, . . . , m} uniquely determining F = i∈I Hi , and so it follows that B \ {xi }i∈I spans F in
the contraction. Therefore, since translating the left-hand side of Eq. 4 to lattice notation gives,

G G
     
m m G
σ Hj ∪ (B − xi ) =  Hj  ⊔  {xj } ,
j=1 j=1 j ∈I
/

it follows that every flat of the sublattice hH1 , . . . , Hm i can be associated with a join of the points
of B using the following relationship,

G G
 
m
!
G
F = Hi ⇐⇒ F =  Hj  ⊔ {xi } . (5)
i∈I j=1 i∈I
/

As an illustrative example, in Figure 7 we show the Boolean algebra generated by a spanning tree
of the graph given in Figure 2. With this, we are ready to proceed to our reduction, essentially
showing that where binary matroids are concerned optimization is as easy as basis search in parallel
models of computation.

5.2 Reducing Optimization to Search

We now use the properties of binary matroids to reduce optimization to the simpler task of iden-
tifying a basis. As a consequence, one can optimize matroids efficiently in parallel, so long as one
is given access to a parallel basis search method. In Section 5.3, we fully develop this insight by
composing our reduction, given in Algorithm 2, with the parallel basis search method of [KUW88].
The upshot of this composition is a novel optimization method terminating in sublinearly many
adaptive rounds in the independence query model. Doing so will break the linear adaptivity barrier
set by the greedy algorithm, comprising our ultimate contribution.
Now, we present our novel reduction in Algorithm 2. The basic concept is simple: The method
operates by identifying the locally optimal points of cocircuits associated with the cofree set of
hyperplanes generated by a basis, and uses symmetric differences to generate further cocircuits
with different locally optimal points in the presence of a collision. Specifically, this begins with an
empty solution X, and a matroid M′ equal to the input M. Then, in each round we obtain a basis
via query on M′ to a search oracle B (line 5) abstracting some search routine. Following this, for
each element xi ∈ B we first construct the set of continuations Γ(B −xi ) via simultaneous queries to
the independence oracle (lines 9 and 10), then compute the locally optimal point in each cocircuit
given by these continuations (line 12). Each such locally optimal point is then augmented into the
partial solution X in parallel (line 13). By Theorem 4.2 such actions are guaranteed to terminate
with the optimal basis, however there is no guarantee that examining the cocircuits given by the
basis B alone will lead to the algorithm terminating quickly. In particular, it is possible that many
of the cocircuits identified in this manner will possess the same locally optimal point. Hence, in
lines 15–17 our method identifies the sets of cocircuits colliding on the same locally optimal point,
and generates further cocircuits by taking symmetric differences. By combining Corollary 5.5 with
Theorem 5.1 it will follow that doing so is sound where correctness is concerned, whereas the

25
 Algorithm 2: Reducing Binary Matroid Optimization to Basis Search
input : Binary Matroid M, Weight Function w : E → R, Independence Oracle ϕ, and Basis
Search Oracle B
output : Basic Set X ∈ M of Minimum Weight

1 X←∅
2 M′ ← M
3

4 while Γ(X) 6= ∅ do
5 B ← query B(M′ ) for a basis of M′
6 label B = {x1 , x2 , . . .}
7
8 forall i ∈ {1, . . . , |B|} in parallel do
9 forall y ∈
/ B − xi in parallel do
10 query ϕ(X + B − xi + y) to construct sets of continuations
11

12 zi ← arg min{w(y) | y ∈ Γ(B − xi )}

13 X ← X + zi
14
15 forall i, j ∈ {1, . . . , |B|} such that i 6= j in parallel do
16 if zi = zj then
17 X ← X + arg min{w(y) | y ∈ Γ(B − xi ) △ Γ(B − xj )}
18
19 M′ ← M/X
20

21 output X

adaptive complexity will follow from arguments exploiting the Boolean algebraic structure of cofree
hyperplanes. Finally, a round ends by contracting the input by the current partial solution X, and
updating the operating matroid M′ with M′ = M/X. We remark that the contraction is largely
symbolic – in an independence query model this operation is realized by prepending all queries with
the set X, for example. This process repeats until X has no more continuations, at which point it
follows from Theorem 4.2 that we have an optimal basis.
In the remainder of the section, we will formally verify that our reduction is correct and ter-
minates in a number of adaptive rounds of queries logarithmic in the rank of the matroid M.
However, before continuing we make a few comparisons between our reduction and Borůvka’s algo-
rithm for graphic matroids since our method was inspired by the latter. Keeping our comparisons
in the language of matroids, both methods simultaneously identify the locally optimal points of
many cocircuits in parallel, which we remind the reader is a general principle for creating parallel
matroid optimization algorithms identified in Section 4.2. For doing so, both methods require a
contraction operation in each round to ensure new sets of cocircuits are revealed in succeeding
rounds. But, where Borůvka’s algorithm uses a convenient encoding provided by the vertices of a
graphic matroid’s corresponding graph (along with those vertices of the contractions) to identify
the locally optimal points of cocircuits, our method does not possess this convenience. Hence, to
solve this discrepancy, our method relies on a subroutine for basis search to identify cocircuits. Fur-
thermore, where complexity is concerned the problem of cocircuits colliding is somewhat moot in
the context of Borůvka’s algorithm, as it is only possible for a locally optimal element to be chosen
by at most two cocircuits in each round due to the structure of binary relations. In contrast, we
require an appeal to the special properties of binary matroids to handle this phenomenon so that

26
we can guarantee our method will terminate quickly.
With that out of the way, we begin our analysis. First, we verify that our method correctly
solves matroid optimization. To do so we take care in ensuring that every element added to the
partial solution X is always a locally optimal point in some cocircuit. Hence, correctness will follow
from a succeeding application of Theorem 4.2, along with our statements concerning modular pairs
of cocircuits and cofree sets made in the last section.
Theorem 5.6. Let M be a binary matroid, and w : E → R an injective weight function. Then,
Algorithm 2 correctly outputs a basis of minimum weight.

Proof. Call each iteration of the while loop in Algorithm 2 a round, and let Xt be the elements
present in X at the beginning of the tth round. We will argue the invariant that Xt is independent
in M and consists only of locally optimal points from cocircuits is maintained throughout each
round of the algorithm’s execution. The first conjunct implies that the algorithm terminates with
a basis, since there is always at least one element added into Xt in each round (lines 8–13). It
follows then that once Xt is of full rank it can have no more continuations and must span the
matroid, making Xt a basis at termination. Then, the second conjunct will conclude the theorem,
as it allows optimality to be witnessed by Theorem 4.2.
Now, fix some round t and examine any element z ∈ Xt+1 \ Xt added into the partial solution
Xt in that round. Furthermore, let B = {x1 , x2 , . . .} be the basis given by the search oracle.
Then, recall that for all suitable indices i the span σ(B − xi ) gives a hyperplane of the contraction
M/Xt . Since contractions correspond to taking filters on the lattice of flats L (M), it follows that
σ(B − xi ) is a hyperplane of M which in turn makes Γ(B − xi ) a cocircuit. Therefore, if z were
locally optimal in some set of continuations in line 9, then it was likewise locally optimal in some
cocircuit. Otherwise, if this is not the case then z was selected in line 14. Well, it follows that there
exists indices i and j such that,

z= arg min w(y),

y∈Γ(B−xi ) △ Γ(B−xj )

where Γ(B − xi ) and Γ(B − xj ) have nonempty intersection since the same point locally minimizes
the weight function w on these sets. Our prior observations identify that Γ(B −xi ) and Γ(B −xj ) are
cocircuits, and in particular these cocircuits correspond to hyperplanes in a cofree set generated by
B on L (M/Xt ) by Corollary 5.5. Hence, Γ(B−xi ) and Γ(B−xj ) are a modular pair by Lemma 5.2,
and thus by Theorem 5.1 the difference Γ(B − xi ) △ Γ(B − xj ) is once again a cocircuit. And so, in
any case z must have been the locally optimal point in some cocircuit to be added to Xt . Therefore,
by Lemma 4.3 it follows that Xt always remains an independent set throughout an execution of
Algorithm 2. Furthermore, upon termination it can only be that the solution consists entirely of
points locally optimal in some cocircuit, meaning that correctness follows by Theorem 4.2.

Arguing that Algorithm 2 terminates quickly is more involved. We will show that the algorithm
terminates in a number of adaptive rounds of queries logarithmic in the rank of the matroid. Our
method of doing so will consist of using the properties of the sublattices formed by cofree sets
of hyperplanes to argue that the locally optimal points identified by the algorithm vary enough
in their identities to make partial solution X grow by a constant fraction of size in each round.
Towards doing so, we first argue that a collection of cocircuits colliding on the same locally optimal
point yield symmetric differences which are all distinct. This fact will become useful in a counting
argument towards showing the partial solution grows by a constant fraction in each round. To
prove this, we first require a technical lemma showing that in this scenario each such cocircuit must
contain a distinct point unique to itself among the other cocircuits. Recalling that the cocircuits

27
generated in line 10 of Algorithm 2 correspond to a cofree set of hyperplanes generated by the basis
given by the search oracle, our approach consists of arguments using the Boolean algebra structure
given by cofree hyperplanes.

Lemma 5.7. Let M be a matroid and C1∗ , . . . , Cm ∗ a set of distinct cocircuits corresponding to a

cofree set of hyperplanes. Then, for each cocircuit Ci∗ there exists a point x ∈ E contained uniquely
in Ci∗ , among the given cociruits. To be precise, for all i ∈ {1, . . . , m},
[
(∃x ∈ E) x ∈ Ci∗ \ Cj∗ . (6)
j6=i

Proof. Before we begin, let H1 , . . . , Hm be the hyperplanes corresponding to the cocircuits, so that,

Γ(Hi ) = Ci∗ , ∀i ∈ {1, . . . , m}.

Our approach in proving the theorem will lie in manipulating these hyperplanes using the lattice
operations endowed upon the flats. In particular, the hyperplanes are cofree by assumption, and
so the sublattice hH1 , . . . , Hm i is a Boolean algebra with coatoms given by the generating hyper-
planes. Recall that in a Boolean algebra, there is always exactly one atom dual to every coatom.
This specific relationship is given in Eq. 3, which we apply to observe that there must exist flats
F1 , . . . , Fm such that,

G G
m
Fi ≻ Hj and Fi = Hj , ∀i ∈ {1, . . . , m}. (7)
j=1 j6=i

F F
Hence, the flats F1 , . . . , Fm are the atoms of the sublattice hH1 , . . . , Hm i. Well, for each atom Fi
there must exist a continuation xi of m j=1 Fj =
m
j=1 Hj , i.e. the bottom of hH1 , . . . , Hm i, making,

G
 
m
σ Hj + xi  = Fi , ∀i ∈ {1, . . . , m}.
j=1

Furthermore, because the atoms F1 , . . . , Fm are distinct it must also be that such xi are distinct.
Hence, using the second conjunct of Eq. 7 it follows that,
[
xi ∈ Hj \ Hi , ∀i ∈ {1, . . . , m}.
j6=i

Therefore, by complementation xi ∈ Γ(Hi ) \ j6=i Γ(Hj ) = Ci∗ \ j6=i Cj∗ for all i ∈ {1, . . . , m}.
S S
Hence, {x1 , . . . , xm } are the desired points satisfying Eq. 6, and the lemma follows.
∗ cor-
As consequence, we find that for every selection of two cocircuits from a set C1∗ , . . . , Cm
responding to cofree hyperplanes will result in a symmetric difference containing
two points not
present in any other. This implies that the given set of cocircuits will generate m
2 new and distinct
cocircuits upon taking symmetric differences, as summarized in the following corollary.
∗ a set of distinct cocircuits corresponding to a
Corollary 5.8. Let M be a matroid and C1∗ , . . . , Cm
cofree set of hyperplanes. Then, every possible pairwise symmetric difference is distinct, implying,
 
 ∗ ∗ m
Ci △ Cj | i 6= j = .
2

28
Proof. The last lemma guarantees that there exists a selection of atoms x1 , . . . , xm such that,
[
xi ∈ Ci∗ \ Cj∗ , ∀i ∈ {1, . . . , m}.
j6=i

Therefore, it follows that for each choice of i and j with i 6= j, that the symmetric difference of Ci∗
and Cj∗ is the only possible pairwise such difference among the cocircuits satisfying,

{xi , xj } ⊆ Ci∗ △ Cj∗ .

So, the symmetric differences must be distinct.

With that, the final technical result we need to verify the adaptive complexity of Algorithm 2
is the following dichotomy among points contained in the union of a set of cocircuits corresponding
to cofree hyperplanes: Any such point is either in a single cocircuit or in every cocircuit. This
fact will be used to show that the locally optimal points generated in line 17 of Algorithm 2 are
sufficiently distinct for the purposes of making the partial solution grow quickly.

Lemma 5.9. Let M be a matroid and C1∗ , . . . , Cm∗ a set of distinct cocircuits corresponding to a

cofree set of hyperplanes.

Tm Then, for any choice of point x ∈ E we have that x is contained either
in the intersection i=1 Ci or at most one choice of cocircuit from C1∗ , . . . , Cm
∗ ∗.

Proof. Our argument makes use of techniques similar to those employed in Lemma 5.7. Once again,
let H1 , . . . , Hm be the hyperplanes corresponding to the cocircuits in a way that makes,

Γ(Hi ) = Ci∗ , ∀i ∈ {1, . . . , m},

and recall that the sublattice hH1 , . . . , Hm i is isomorphicTto a Boolean algebra by assumption.
Continuing on, choose any x ∈ E satisfying both of x ∈ / m ∗ ∗
i=1 Ci and x ∈ Ci for some index i.
We point out here that if these conditions are not met, then either x is contained in all of the
cocircuits or none of the cocircuits, thereby satisfying the lemma in an uninteresting way. Well, by
construction it follows that x lies in some but not all of the hyperplanes. Then, define the flat F (x) ,

G
 
m
(x)
F , H j + x ,
j=1

F
F
and notice that our prior observation makes F (x) 6= m j=1 Hj . Taking this further, x must exist in
m
some basis B of M/ j=1 Hj because it does not lie in all of the hyperplanes. Or, more specifically,
since there exists Hi such that x ∈ Γ(Hi ), B is found by taking a basis of Hi and augmenting
it with x. Well, then B generates hH1 , . . . , Hm i, and in particular it follows by the relationship
between the flats of this sublattice and the points of B given in Eq. 5, that F (x) is an atom of this
sublattice since it corresponds to a join involving only a single point from B. But then, because
there is always exactly one coatom dual to every atom, there can exist only a single hyperplane
from H1 , . . . , Hm which does not lie above F (x) . It follows then that x is contained in the span of
all but one of the hyperplanes corresponding to the cocircuits, and therefore can only by contained
in a single cocircuit.

Now observe that for a set of cocircuits as described by Lemma 5.9, we lose the points contained
in the union upon taking pairwise symmetric differences. Hence, as a result points identified in
line 17 of Algorithm 2 can only belong to a single cocircuit, call it C ∗ , in a set of cocircuits

29
containing the same locally optimal point. It follows then that such a point can only appear in
symmetric differences involving C ∗ , which will be the key observation showing that our scheme for
resolving cocircuits colliding on the same locally optimal point will generate many new distinct
locally optimal points. Combining this in a counting argument with Corollary 5.8 will show that
the adaptive complexity of Algorithm 2 is logarithmic in the rank of the matroid. This is summed
up by the following theorem.

Theorem 5.10. Let M be a binary matroid of rank r, and w : E → R an injective weight

function. Then, Algorithm 2 terminates in O(log r) adaptive rounds of queries to basis search and
independence oracles.

Proof. Like in our argument for Theorem 5.6, call any iteration of the while loop in Algorithm 2 a
round. And, again, let Xt consist of the elements present in the partial solution X at the beginning
of the tth round as well. Then, we will argue that the rank of M/Xt shrinks by a factor of 2 in
each round, i.e.,
rt
rt+1 ≤ , (8)
2
where rt is the rank of M/Xt . Thus, we have rt = r − |Xt | by construction, because Theorem 5.6
guarantees Xt ∈ M. Furthermore, we also have that the basis B selected by oracle query is of rank
rt . Hence, by showing the recurrence given by Eq. 8 it will follow that there can be at most log r
rounds until the cardinality |Xt | = r, at which point the partial solution spans the matroid and the
algorithm terminates.
Now, examine the tth round and let B be the basis of M/Xt obtained by query. Recall that
for each xi ∈ B the span σ(B − xi ) corresponds to the identity of some hyperplane. In particular,
by Corollary 5.5 we see that B generates the Boolean algebra corresponding to a cofree set of |B|
hyperplanes in this manner. So, examine any point z ∈ E satisfying,

z = arg min w(y), ∃i ∈ {1, . . . , rt },

y∈Γ(B−xi )

and let K(z) give the cocircuits with indices corresponding to those i satisfying the above. Specif-
ically, assume that |K(z)| = m ≥ 2 so that there is some collision between two cocircuits in terms
of their locally optimal point under the weight function w. Our prior observation guarantees that
the hyperplanes corresponding to the cocircuits of K(z) are cofree, and so Corollary 5.8 implies
that K(z) generates m


2 distinct pairwise symmetric differences upon its elements. Furthermore,
let the point z ′ ∈ E be such that there exists C1∗ , C2∗ ∈ K(z) such that,

z ′ = arg min w(y).

y∈C1∗ △ C2∗

This means that z ′ is chosen to augment the partial solution by in line 17 of Algorithm 2. And so,
by properties of symmetric difference ′ ∗ ∗ ′
T z is contained in exactly one of CS 1 and C2 . Therefore, z is
not contained in the intersection K(z), but is contained in the union K(z). Hence, by way of
Lemma 5.9, z ′ can only be contained one cocircuit of K(z), so assume without loss of generality
that C1∗ ∋ z ′ . Because of this, the only pairwise symmetric differences from K(z) containing z ′
must involve C1∗ as an operand. Recalling that |K(z)| = m, it follows that there are at most m − 1
symmetric differences generated by the cocircuits of K(z) which can contain z ′ . And so, as our

30
choice of z ′ was arbitrary we have identified,
( )  
m
arg min w(y) (C1∗ , C2∗ ) ∈ K(z) × K(z) ≥ /(m − 1),
y∈C1∗ △ C2∗ 2
m · (m − 1)
= ,
2 · (m − 1)
m
= .
2
Now, observe that in any round Xt must be augmented by at least as many elements as it would
have been had the same point minimized w in the cocircuits given by the basis, i.e. if zi = zj for
all i 6= j in line 12 of Algorithm 2. This corresponds to the case where m, as above, follows m = rt .
With that we have,
rt rt
|Xt+1 | ≥ + |Xt | = + r − rt .
2 2
And so, finally, by plugging |Xt+1 | = r − rt+1 into the above and moving around terms, one finds,
rt rt
rt+1 ≤ rt − = ,
2 2
satisfying the recurrence of Eq. 8 as desired.

In plain language, we have successfully shown that for binary matroids a parallel optimization
algorithm is implied by the existence of an algorithm for parallel basis search. Specifically, since our
reduction only adds a logarithmic factor of overhead it follows that any such method terminating
in a sublinear number of rounds will imply an optimization method with sublinear complexity. For-
√
tunately for us, [KUW88] provides a general basis search method terminating in O( n) adaptive
rounds of queries to an independence oracle, where n is the size of the ground set. So, as corollary,
composing this algorithm with our reduction implies a matroid optimization algorithm with sub-
linear adaptivity in the independence query model. In the next section, we survey the methods of
[KUW88] to elaborate further and conclude these ideas.

5.3 Sublinear Round Matroid Optimization with Independence Queries

In this section, we combine our reduction with a method from [KUW88] to obtain a novel algo-
rithm for binary matroid optimization terminating in a sublinear number of adaptive rounds in
the independence query model. To the authors’ knowledge, this is the first such parallel algorithm
outperforming the greedy method in terms of complexity for any matroid more general than that
of a graphic matroid when limited to independence queries.
To begin, we briefly overview a basis search method from [KUW88] which we give in Algorithm 3.
For the reader, we note here that our presentation is slightly different, but the main ideas are the
same. Letting n = |E| be the size of the ground set, the algorithm first partitions the ground
√
set into n subsets labelled Z1 , Z2 , . . . of roughly equal size (line 3). These subsets represent the
points which the algorithm can consider for augmentation into the partial solution. With that,
the algorithm works by growing the partial solution X with augmentation by the entirety one of
the subsets Zi whenever possible, or by removing elements from the subsets posing a barrier to
independence when not possible. Specifically, in each round we query the independence oracle
on each possible union X ∪ Zi simultaneously (line 7), and as a result will either find that the
partial solution X can be augmented by one of these subsets to grow towards a basis, or that
no subset leads to an independent set in this manner. In case of the former we grow the partial

31
 Algorithm 3: Parallel Basis Search Method [KUW88]
input : Matroid M on Ground Set E with Independence Oracle ϕ
output : A Basic Set X ∈ M

1 X←∅
n jp ko
2 I ← 1, . . . , |E|
jp k jp k
3 Z1 , . . . , Zk ← partition of E into sets of size |E| or |E| + 1
4
5 while there exists i ∈ I such that Zi 6= ∅ do
6 forall i ∈ I in parallel do
7 query ϕ(X ∪ Zi ) and record those indices i making indepedent sets
8

9 if there exists i ∈ I such that Zi 6= ∅ and X ∪ Zi ∈ M then

10 i ← min{j ∈ I | X ∪ Zj ∈ M}
11 X ← X ∪ Zi
12 Zi ← ∅
13 else
14 forall i ∈ I in nparallel do o
(i) (i)
15 label Zi = z1 , z2 , . . .
16 forall j ∈ {1,
 . . . , |Z
n i |} in parallel
o do
(i) (i)
17 query ϕ X ∪ z1 , . . . , zj and record those indices j making independent sets
18

19 if Zi 6= ∅ then n n o o
(i) (i)
20 ℓ(i) ← arg min j ∈ I X ∪ z1 , . . . , zj ∈
/M
(i)
21 Zi ← Zi − zℓ(i)
22

23 output X

solution X by augmentation with one such subset, breaking ties by taking that with the least
index (line 10–11). However, we note that this tie breaking rule can be arbitrary. Following this,
the chosen subset is then assigned the empty set (line 12), so that it’s points are removed from
consideration in future rounds. Conversely, in the case that X cannot be augmented by any subset
without losing independence the method then queries the independence oracle for each subset and
its prefixes according to an arbitrary labelling simultaneously (line 15–17) so that some element
of each
n nonempty osubset can be identified for removal. Specifically, the first index ℓ(i) such that
(i) (i)
X ∪ z1 , . . . , zℓ(i) ∈
/ M will be removed from each Zi . For the interested reader, we’ll show that
doing so is safe in the sense that X and the subsets always contain a basis, i.e.,
!
[
σ X∪ Zi = E, (9)
i∈I

by using properties of circuits of the matroid in Appendix C. In particular, if Eq. 9 is maintained as

an invariant, it follows that after enough rounds the solution X must grow to become a basis since
points can only be added to the solution if independence is observed. This is the key to showing
Algorithm 3 is correct, which we more fully sketch out in Appendix C for the sake of completeness.
As an aside, our sketch towards the correctness of this method includes helpful details omitted from

32
 √
[KUW88]. For complexity, one can also see that this method terminates in O( n) adaptive rounds
√
of queries to the independence oracle, as in total the partial solution can only be augmented n
√
times and there can only be at most n + 1 rounds where elements are removed from the subsets.
Both of these observations follow from the way the partition was constructed, since there are only
√ √
n subsets initially and each is of size at most n + 1. Hence, the following.
Theorem 5.11 ([KUW88]). Let M be a matroid over a ground set E of n atoms. Then, Algo-
√
rithm 3 finds a basis in O ( n) adaptive rounds of queries to an independence oracle.
So, we see that composing Algorithm 3 with our reduction of optimization to search gives an
algorithm finding the optimal basis of a binary matroid in a sublinear number of adaptive rounds
of queries to an independence oracle.
Corollary 5.12. Let M be a binary matroid of rank r over a ground set E of n atoms, and
w : E → R an injective weight function. Then, composing Algorithm 2 with Algorithm 3 finds a
√
basis of minimum weight in O (log r · n) adaptive rounds of queries to an independence oracle.
Proof. One need only instantiate the search oracle used in Algorithm 2 with the procedure given
by Algorithm 3. By Theorem 5.10, our reduction adds an overhead only logarithmic in the rank.
Hence, for a rank r binary matroid M on an n element ground set, we find that this strategy
√
optimizes M in O(log r · n) adaptive rounds.

Remark 5.13. Before we depart, one should question if it is possible to transform Algorithm 3 into
an optimization method by first sorting the points according to their weight, and then constructing
the partition and corresponding labels in some strategic manner. If true, then all of our efforts
are in vain. Fortunately, there is a simple argument showing a matroid forcing Algorithm 3 to
fail to identify the optimal basis for any partition configuration and labelling of the points. This
matroid is constructed in the following way: M will have two bases. Choose two arbitrary indices
corresponding to the partition, say i and j, and compute the locally optimal points in Zi and Zj .
Without loss of generality let the locally optimal point of Zj be “heavier” under w than that of
Zi . Then, the first basis is given by the two computed locally optimal points, whereas the second
is given by the “heavier” point as well as any other arbitrary point from Zj . Of course, the first
basis is optimal, but one can observe that Algorithm 3 will necessarily output the second. This
construction is generic for any configuration of the partition and labelling among the elements.
Furthermore, the construction applies to binary matroids as one can verify that the lattice of flats
for our counterexample is isomorphic to that of U32 , the three point line, by observing that the unused
points make loops in M.

6 Discussion
In this work, we obtained a novel method for optimization in the binary matroid using sublinearly
many adaptive rounds of queries to an independence oracle. Underlying this success was our
reduction given in Algorithm 2, showing that optimization can be achieved in a number of rounds
logarithmic in the rank on a binary matroid when given access to a basis search routine. This
reduction relied upon the observation given in Theorem 4.2, stating a local to global certificate
of optimality for matroid optimization in that the points of the optimal basis must be the local
minimizers of the weight function within the cocircuits. In particular, this characterization of
optimality allowed us to give a deterministic and conceptually simple reduction using the propertries
of binary matroids, operating by using bases to identify cocircuits and generating further cocircuits
via symmetric differences in the presence of colliding local optima for the sake of fast termination.

33
 Because our work initiates the study of parallel matroid optimization, there are many open
problems for future research. The most immediate of these concerns being what is achievable.
Specifically, be the overhead of our reduction is only logarithmic it may be possible to achieve a
method with polylogarithmic adaptive complexity. Towards this, we remark that the dominating
component of our asymptotic guarantee is from the basis search algorithm of [KUW88], taking
√
O( n) rounds of queries to an independence oracle. This method was not specialized to the binary
matroid, and so one may be able to identify a parallel basis search method using the properties of
binary matroids to achieve some speed up. Of course, it is always possible that such a method does
not exist, hence the following.
Open Question 6.1. Does there exist an algorithm optimizing a binary matroid in polylogarith-
mically many adaptive rounds of queries to an independence oracle?
Next, there is the task of generalizing our methods to other matroids. The next natural gener-
alization would be applying our method to linear matroids representable on Galois fields. At this
time, it does not appear possible to directly apply our methods to such a task. Specifically, while
the symmetric differences of cocircuits still generally contain points within some union of cocircuits
in nonbinary matroids, by referencing our proof of White’s characterization in Appendix A it be-
comes apparent that such differences may not reveal the locally optimal points of cocircuits. The
reasoning is technical and we provide further commentary on this limitation in Remark A.1. Hence,
parallel methods leveraging Theorem 4.2 as a certificate of optimality in more general settings will
require some new ideas to terminate quickly.
Open Question 6.2. Do there generally exist parallel optimization algorithms for matroids repre-
sentable on Galois fields?
Finally, it is worth noting that the complexity of both our reduction in Algorithm 2 and
Borůvka’s algorithm both vary in the rank of the matroid, and not the size of the ground set
(for latter, note that the size of a spanning tree is always one less than the number of vertices in
a graph). In some sense then, these methods both observe improved efficiency when given sparse
instances – that is instances with low rank. In linear algebra, such sparsity corresponds to the
increased presence of zero entries in a vector or matrix, and is known to control the complexity of
numerical optimizations like compressed sensing [Ba+10] and matrix completion [CT10]. In light
of this discussion, perhaps it is possible that a similar phenomenom exists in matroid optimization.
Open Question 6.3. What role does the rank function play in the adaptive complexity of matroid
optimization? Are sparser matroids easier to optimize than their dense counterparts?

References
[Ba+10] Ba, K. D. et al. “Lower bounds for sparse recovery”. In: Proceedings of the twenty-first
annual ACM-SIAM symposium on Discrete Algorithms. SIAM. 2010, pp. 1190–1197.
[Bir35] Birkhoff, G. “Abstract linear dependence and lattices”. In: American Journal of Math-
ematics 57.4 (1935), pp. 800–804.
[Bli22] Blikstad, J. “Sublinear-round parallel matroid intersection”. In: 49th International Col-
loquium on Automata, Languages, and Programming (ICALP 2022). Schloss-Dagstuhl-
Leibniz Zentrum für Informatik. 2022.
[Bor26a] Borůvka, O. “O jistém problému minimálnı́m (On a certain minimal problem)”. In:
Práce Mor. Přı́rodověd. Spol. V Brně III (1926), pp. 37–58.

34
[Bor26b] Borůvka, O. “Přı́spěvek k řešenı́ otázky ekonomické stavby elektrovodnı́ch sı́tı́ (Contri-
bution to the solution of a problem of economical construction of electrical networks)”.
In: Elektronický Obzor 15 (1926), pp. 153–154.
[BRS19a] Balkanski, E., Rubinstein, A., and Singer, Y. “An exponential speedup in parallel run-
ning time for submodular maximization without loss in approximation”. In: Proceedings
of the Thirtieth Annual ACM-SIAM Symposium on Discrete Algorithms. SIAM. 2019,
pp. 283–302.
[BRS19b] Balkanski, E., Rubinstein, A., and Singer, Y. “An optimal approximation for submod-
ular maximization under a matroid constraint in the adaptive complexity model”. In:
Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing.
2019, pp. 66–77.
[BS18] Balkanski, E. and Singer, Y. “The adaptive complexity of maximizing a submodular
function”. In: Proceedings of the 50th annual ACM SIGACT symposium on theory of
computing. 2018, pp. 1138–1151.
[BS20] Balkanski, E. and Singer, Y. “A lower bound for parallel submodular minimization”.
In: Proceedings of the 52nd annual ACM SIGACT symposium on theory of computing.
2020, pp. 130–139.
[Cal+07] Calinescu, G. et al. “Maximizing a submodular set function subject to a matroid con-
straint”. In: International Conference on Integer Programming and Combinatorial Op-
timization. Springer. 2007, pp. 182–196.
[CCK23] Chakrabarty, D., Chen, Y., and Khanna, S. “A polynomial lower bound on the number
of rounds for parallel submodular function minimization and matroid intersection”. In:
SIAM Journal on Computing (2023), FOCS21–1.
[Cha00] Chazelle, B. “A minimum spanning tree algorithm with inverse-Ackermann type com-
plexity”. In: Journal of the ACM (JACM) 47.6 (2000), pp. 1028–1047.
[CLM12] Caspard, N., Leclerc, B., and Monjardet, B. Finite ordered sets: concepts, results and
uses. 144. Cambridge University Press, 2012.
[CQ19a] Chekuri, C. and Quanrud, K. “Parallelizing greedy for submodular set function maxi-
mization in matroids and beyond”. In: Proceedings of the 51st Annual ACM SIGACT
Symposium on Theory of Computing. 2019, pp. 78–89.
[CQ19b] Chekuri, C. and Quanrud, K. “Submodular function maximization in parallel via the
multilinear relaxation”. In: Proceedings of the Thirtieth Annual ACM-SIAM Sympo-
sium on Discrete Algorithms. SIAM. 2019, pp. 303–322.
[CT10] Candès, E. J. and Tao, T. “The power of convex relaxation: Near-optimal matrix
completion”. In: IEEE transactions on information theory 56.5 (2010), pp. 2053–2080.
[Daw80] Dawson, J. E. “Optimal matroid bases: An algorithm based on cocircuits”. In: The
Quarterly Journal of Mathematics 31.1 (1980), pp. 65–69.
[DP02] Davey, B. A. and Priestley, H. A. Introduction to lattices and order. Cambridge uni-
versity press, 2002.
[DT95] Dress, A. W. and Terhalle, W. “Well-layered maps—A class of greedily optimizable set
functions”. In: Applied Mathematics Letters 8.5 (1995), pp. 77–80.
[DW73] Dunstan, F. and Welsh, D. “A greedy algorithm for solving a certain class of linear
programmes”. In: Mathematical Programming 5.1 (1973), pp. 338–353.

35
[DW92] Dress, A. W. and Wenzel, W. “Valuated matroids”. In: Advances in Mathematics 93.2
(1992), pp. 214–250.
[Edm03] Edmonds, J. “Submodular functions, matroids, and certain polyhedra”. In: Combi-
natorial Optimization—Eureka, You Shrink! Papers Dedicated to Jack Edmonds 5th
International Workshop Aussois, France, March 5–9, 2001 Revised Papers. Springer.
2003, pp. 11–26.
[Edm71] Edmonds, J. “Matroids and the greedy algorithm”. In: Mathematical programming 1
(1971), pp. 127–136.
[Fei98] Feige, U. “A threshold of ln n for approximating set cover”. In: Journal of the ACM
(JACM) 45.4 (1998), pp. 634–652.
[FMZ19a] Fahrbach, M., Mirrokni, V., and Zadimoghaddam, M. “Non-monotone submodular
maximization with nearly optimal adaptivity and query complexity”. In: International
Conference on Machine Learning. PMLR. 2019, pp. 1833–1842.
[FMZ19b] Fahrbach, M., Mirrokni, V., and Zadimoghaddam, M. “Submodular maximization with
nearly optimal approximation, adaptivity and query complexity”. In: Proceedings of
the Thirtieth Annual ACM-SIAM Symposium on Discrete Algorithms. SIAM. 2019,
pp. 255–273.
[Fuj05] Fujishige, S. Submodular functions and optimization. Elsevier, 2005.
[Gar15] Garg, V. K. Introduction to lattice theory with computer science applications. John
Wiley & Sons, 2015.
[Ger87] Gerards, A. M. H. “A short proof of Tutte’s characterization of totally unimodular
matrices”. In: (1987).
[GGR24] Ghosh, S., Gurjar, R., and Raj, R. “A deterministic parallel reduction from weighted
matroid intersection search to decision”. In: Algorithmica 86 (2024), pp. 1057–1079.
[Gou+12] Gouveia, J. et al. “A new semidefinite programming hierarchy for cycles in binary
matroids and cuts in graphs”. In: Mathematical programming 133.1 (2012), pp. 203–
225.
[Gra11] Gratzer, G. A. Lattice theory: foundation. Springer, 2011.
[GT89] Grötschel, M. and Truemper, K. “Decomposition and optimization over cycles in binary
matroids”. In: Journal of Combinatorial Theory, Series B 46.3 (1989), pp. 306–337.
[HK78] Hausmann, D. and Korte, B. “Lower bounds on the worst-case complexity of some
oracle algorithms”. In: Discrete Mathematics 24.3 (1978), pp. 261–276.
[HK81] Hausmann, D. and Korte, B. Algorithmic versus axiomatic definitions of matroids.
Springer, 1981.
[HMS93] Helman, P., Moret, B. M., and Shapiro, H. D. “An exact characterization of greedy
structures”. In: SIAM Journal on Discrete Mathematics 6.2 (1993), pp. 274–283.
[JáJ92] JáJá, J. Parallel algorithms. 1992.
[JK82] Jensen, P. M. and Korte, B. “Complexity of matroid property algorithms”. In: SIAM
Journal on Computing 11.1 (1982), pp. 184–190.
[Kar98] Karger, D. R. “Random sampling and greedy sparsification for matroid optimization
problems”. In: Mathematical Programming 82.1 (1998), pp. 41–81.

36
[Kas08] Kashyap, N. “A decomposition theory for binary linear codes”. In: IEEE Transactions
on Information Theory 54.7 (2008), pp. 3035–3058.
[Kaz+19] Kazemi, E. et al. “Submodular streaming in all its glory: Tight approximation, mini-
mum memory and low adaptive complexity”. In: International Conference on Machine
Learning. PMLR. 2019, pp. 3311–3320.
[KKT95] Karger, D. R., Klein, P. N., and Tarjan, R. E. “A randomized linear-time algorithm to
find minimum spanning trees”. In: Journal of the ACM (JACM) 42.2 (1995), pp. 321–
328.
[KLS12] Korte, B., Lovász, L., and Schrader, R. Greedoids. Vol. 4. Springer Science & Business
Media, 2012.
[Kup+20] Kupfer, R. et al. “The adaptive complexity of maximizing a gross substitutes valua-
tion”. In: Advances in Neural Information Processing Systems 33 (2020), pp. 19817–
19827.
[KUW88] Karp, R. M., Upfal, E., and Wigderson, A. “The complexity of parallel search”. In:
Journal of Computer and System Sciences 36.2 (1988), pp. 225–253.
[Luo19] Luo, S. “A counting lemma for binary matroids and applications to extremal problems”.
In: Journal of Combinatorial Theory, Series B 136 (2019), pp. 329–359.
[May08] Mayhew, D. “Matroid complexity and nonsuccinct descriptions”. In: SIAM Journal on
Discrete Mathematics 22.2 (2008), pp. 455–466.
[Mis+20] Misra, P. et al. “Linear representation of transversal matroids and gammoids parame-
terized by rank”. In: Theoretical Computer Science 818 (2020), pp. 51–59.
[Mur98] Murota, K. “Discrete convex analysis”. In: Mathematical Programming 83 (1998),
pp. 313–371.
[NK09] Nishimura, H. and Kuroda, S. A lost mathematician, Takeo Nakasawa: the forgotten
father of matroid theory. Springer, 2009.
[NMN01] Nešetřil, J., Milková, E., and Nešetřilová, H. “Otakar Borůvka on minimum spanning
tree problem Translation of both the 1926 papers, comments, history”. In: Discrete
mathematics 233.1-3 (2001), pp. 3–36.
[NWF78] Nemhauser, G. L., Wolsey, L. A., and Fisher, M. L. “An analysis of approximations for
maximizing submodular set functions—I”. In: Mathematical programming 14 (1978),
pp. 265–294.
[Oxl06] Oxley, J. G. Matroid theory. Vol. 3. Oxford University Press, USA, 2006.
[Rad57] Rado, R. “Note on independence functions”. In: Proceedings of the London Mathemat-
ical Society 3.1 (1957), pp. 300–320.
[Sch+03] Schrijver, A. et al. Combinatorial optimization: polyhedra and efficiency. Vol. 24. 2.
Springer, 2003.
[Sta11] Stanley, R. P. “Enumerative Combinatorics Volume 1 second edition”. In: Cambridge
studies in advanced mathematics (2011).
[Sta71] Stanley, R. P. “Modular elements of geometric lattices”. In: Algebra Universalis 1
(1971), pp. 214–217.
[Tut58] Tutte, W. “A homotopy theorem for matroids, II”. In: Transactions of the American
Mathematical Society 88.1 (1958), pp. 161–174.

37
[Wel10] Welsh, D. J. Matroid theory. Courier Corporation, 2010.
[Whi35] Whitney, H. “On the Abstract Properties of Linear Dependence”. In: American Journal
of Mathematics 57 (1935), pp. 63–87.
[Whi71] White, N. “The Bracket ring and combinatorial geometry.” Doctoral Dissertation. Har-
vard University, 1971.
[Whi87] White, N. Combinatorial geometries. 29. Cambridge University Press, 1987.

38
A Proof of White’s Characterization of Binary Matroids
In this section we will get more familiarity with Tutte’s forbidden minor theorem [Tut58], and its
relationship to White’s characterization of the binary matroid via modular pairs of circuits [Whi71;
Whi87]. Doing so will allow us to identify the limitations of Algorithm 2 towards generalizing
beyond binary matroids. Recall, Tutte showed that a matroid M is binary if and only if it does not
contain U42 , the four point line, as a minor. To get some intuition for why this is true, at least for
one direction, we start by pointing out that if a matroid M of rank r possesses a representation on
a field F, then there must exist a representation with image contained in Fr . Indeed, lets reexamine
the representation in Section 5 we gave of the graphic matroid given by the graph in Figure 1. By
observation, the rank of the graphic matroid is six, whereas the codomain of our representation was
F72 . Since we cannot have any dependent set map to a set of linearly independent vectors, it follows
that there must be some hidden subspace of F72 of rank six with which we can biject between the
image of our representation. This principle holds true not just for our specific example, but for any
representation of a linear matroid. Following this line of thought, because the rank of the four point
line is only two, if it does possess a binary representation it must also possess a representation with
F22 as a codomain. However, there are only three nonzero vectors in F22 and so it follows that for
any representation f of our choosing there will exist a selection of two points x and y in the ground
set such that f |{x,y} is not injective. But {x, y} ∈ U42 of course, and so f was never a representation
in the first place. Showing the other direction, that every matroid possessing no minors isomorphic
to U42 is binary, is a bit more involved. See [Whi87] for details.
Put in lattice theoretic terms, this implies that a matroid is binary if, and only if, its lattice of
flats does not possess any height two intervals consisting of six or more elements. Our argument will
employ straightforward consequences of this observation. In particular, we will relate the lattice of
flats of the dual matroid to the circuits of the primal matroid using the correspondence between
modular pairs of circuits and their hyperplanes from Lemma 5.2. However before doing so we will
introduce a little bit more terminology: A coline is a flat of rank two less than that of the matroid.
The name is chosen as flats presenting as colines are somewhat complementary to lines in the lattice
of flats. Furthermore, a principal ideal of a lattice is a set consisting of an element x and all other
elements below x. In the context of the lattice of flats, principal ideals
are generated by deletion
against a flat’s complement as one can verify that for all flats F ∈ σ 2E ,

L (M \ Γ(F )) = F ′ ∩ F | F ′ ∈ σ 2E ∼= F ′ ∈ σ 2E | F ′ ⊆ F .



We point out that the first equality follows from definition of deletion, whereas the second congru-
ence is due to flats being closed under intersection. See [Wel10] for more details on this construction.
With that, we are ready to give a direct proof of the equivalence between Tutte’s forbidden minor
theorem and White’s characterization of binary matroids.

Theorem 5.1 ([Tut58; Whi71; Whi87]). Let M be a matroid. Then, the following are equivalent:

1. M is representable over F2 ,

2. M does not contain a minor isomorphic to U42 ,

3. And, for all distinct circuits C1 and C2 , if C1 ∩ C2 6= ∅ and C1 and C2 form a modular pair
then C1 △ C2 is another circuit.

Proof. We’ll take the equivalence of the first two properties for granted and show that (3.) is
equivalent to (2.). We will first show that (3.) implies (2.). To do so, we will show there cannot

39
exist any height two interval of size six or more in the dual lattice of flats L (M∗ ). This suffices
because binary matroids are a self dual class. However, we can simplify this to only showing that
every coline is covered by no more than three hyperplanes. This is because, for any choice of height
two interval with top flat F⊤ and bottom flat F⊥ we can consider the deletion M∗ \ Γ(F⊤ ). This
is because L (M∗ \ Γ(F⊤ )) is simply the principal ideal generated by F⊤ , and so it follows that,

1. The interval containing all flats lying between F⊥ and F⊤ is preserved in the ideal corre-
sponding to L (M∗ \ Γ(F⊤ )),

2. F⊤ corresponds to the top element in L (M∗ \ Γ(F⊤ ),

3. And, F⊥ corresponds to a coline in L (M∗ \ Γ(F⊤ )) because it is of rank two less than F⊤ .

So in light of this, fix any coline F ∗ of the dual matroid M∗ and let H1∗ and H2∗ be two distinct
hyperplanes covering F ∗ . Note, such hyperplanes must exist because L (M∗ ) is atomistic meaning
that F ∗ cannot be an irreducible. Now, observe that the continuations of F ∗ leading to H1∗ must be
disjoint from those leading to H2∗ , as otherwise there would exist an independent set spanning both
H1∗ and H2∗ . With that, suppose Γ(F ∗ ) 6= (H1∗ \ F ∗ ) ∪ (H2∗ \ F ∗ ) as otherwise it follows that H1∗ and
H2∗ are the only hyperplanes covering F ∗ . Then, by complementation the “leftover” continuations
are given by,
Γ(F ∗ ) \ (H1∗ \ F ∗ ) ∪ (H2∗ \ F ∗ ) = Γ(H1∗ ) ∩ Γ(H2∗ ).

Our goal then is showing that F ∗ ∪(Γ(H1∗ )∩Γ(H2∗ )) is a flat, because then F ∗ is covered by only three
hyperplanes as the “leftover” continuations are accounted for. Well, let C1 and C2 be the circuits
of M such that Γ(H1∗ ) = C1 and Γ(H2∗ ) = C2 . Since H1∗ and H2∗ both cover F ∗ the hyperplanes
must be a modular pair, and hence C1 and C2 form a modular pair as well by Lemma 5.2. Then,
because (3.) holds by assumption, it follows that C1 △ C2 is a circuit of M. This, in turn, makes
C1 △ C2 the set of continuations for some hyperplanes of M∗ by Lemma 3.1. But,

C1 △ C2 = Γ(H1∗ ) △ Γ(H2∗ ) = E \ F ∗ ∪ Γ(H1∗ ) ∩ Γ(H2∗ ) ,

and so it can only be that F ∗ ∪ Γ(H1∗ ) ∩ Γ(H2∗ ) is the previously identified hyperplane of M∗

corresponding to C1 △ C2 . Therefore, M∗ can contain no height two interval of six or more elements.
As we assume the equivalence of (1.) and (2.), it follows that M∗ , as well as M because of self
duality, is a binary matroid. Hence, by once again applying our assumption that (1.) and (2.) are
equivalent, it follows that M cannot contain a minor isomorphic U42 .
Going the other direction, we now show that (2.) implies (3.). Select a distinct modular pair
circuits C1 and C2 of M such that C1 ∩ C2 6= ∅. We know that C1 and C2 give the continuations of
a modular pair of hyperplanes H1∗ and H2∗ on the dual matroid M∗ by Lemma 5.2. In particular,
by assumption we identify that Γ(H1∗ ) ∩ Γ(H2∗ ) 6= ∅. As a result, there is a third hyperplane
H3∗ covering the coline H1∗ ⊓ H2∗ , since the continuations of Γ(H1∗ ) ∩ Γ(H2∗ ) must lead to another
hyperplane when applied to the meet. Furthermore, due to the forbidden minor characterization
H3∗ is the only such hyperplane, and it must be that,

Γ(H3∗ ) = E \ (H1∗ ⊓ H2∗ ) ∪ (Γ(H1∗ ) ∩ Γ(H2∗ )) = Γ(H1∗ ) △ Γ(H2∗ ).

Hence Γ(H3∗ ) = C1 △ C2 , and so by Lemma 3.1 it follows that C1 △ C2 is a circuit of M.

Remark A.1. The prior argument reveals that we were able to use Theorem 5.1 in Algorithm 2, to
specifically generate new cocircuits in the presence of collisions, because of the following principle:
The intersection of the continuations corresponding to a modular pair of hyperplanes H1 and H2

40
must be exactly those points leading to a unique third hyperplane H3 when applied to any independent
set spanning the meet H1 ⊓ H2 . When moving beyond binary matroids, the points Γ(H1 ) ∩ Γ(H2 )
may now lead to a set of two or more flats when applied to sets spanning H1 ⊓ H2 . In the case
of this discrepancy, it would follow that for every x ∈ Γ(H1 ) ∩ Γ(H2 ) there exists a hyperplane
Hbad ≻ H1 ⊓ H2 covering the meet such that x ∈ Γ(Hbad ), and Hbad is not equal to H1 or H2 .
Then, note that in the case where the local minimizer z = arg miny∈Γ(H1 ) △ Γ(H2 ) w(y) is such
that w(z) > w(x) for all x ∈ Γ(H1 ) ∩ Γ(H2 ), Algorithm 2 fails to find a locally optimal point of
a cocircuit because the possible candidates will exist in the intersection Γ(H1 ) ∩ Γ(H2 ), not the
symmetric difference. At this time, using the intersection to identify the locally optimal points of
cocircuits appears to come with challenges, mainly where the complexity overhead is concerned.

B Matroid Duality
In this section we will discuss and verify certain “folklore” lemmas from the literature derived from
matroid duality. Helpful in understanding the dual matroid, one can verify that its rank function
ρ∗ : 2E → Z+ is given by,
ρ∗ (X) , ρ(E \ X) + |X| − ρ(E). (10)
And, because duality is an involution, the same equality holds when replacing ρ∗ by ρ and vice
versus in the above equation.
In this paper, we often referred to the connection between hyperplanes and cocircuits first given
in Lemma 3.1 of the preliminaries – specifically being that a corcircuit corresponds to the set of
continuations of some unique hyperplane. Of course, since the cocircuits of M are the circuits on
the dual matroid M∗ this is a statement best proven using structure connecting a matroid to its
dual. We do so in what follows, using the dual rank function.
Lemma 3.1 (Folklore). Let M be a matroid on a ground set E. Then, H ⊆ E is a hyperplane if
and only if its continuations Γ(H) form a circuit in the dual matroid M∗ .
Proof. Firstly, let C ∗ be a circuit in M∗ . Well, because C ∗ is a circuit of M∗ it follows that
ρ∗ (C ∗ ) = |C ∗ | − 1. Hence, using Eq. 10 we find,

ρ(E \ C ∗ ) = ρ∗ (C ∗ ) + |E \ C ∗ | − ρ∗ (E) = |E| − 1 − ρ∗ (E) = ρ(E) − 1,

where the last equality follows from another application of Eq. 10 along with the fact that ρ(∅) = 0.
Therefore, because E \ C ∗ has rank one less than that of the matroid, we see that E \ C ∗ spans a
hyperplane of M. Hence, let H be such a hyperplane, i.e. σ(E \ C ∗ ) = H. Well, once again using
the status of C ∗ as a circuit of M∗ , we also identify that for all x ∈ C ∗ ,

ρ(E \ C ∗ + x) = ρ∗ (C ∗ − x) + |E \ C ∗ | + 1 − ρ∗ (E) = |C ∗ | + |E \ C ∗ | − ρ∗ (E) = ρ(E).

Thus, the rank in M increases upon the addition of any element of C ∗ to E \ C ∗ . Because of this,
it can only be that E \ C ∗ equals its own span. It follows that H = E \ C ∗ , which in turn makes
Γ(H) = Γ(E \ C ∗ ) = C ∗ by complementation.
To argue the other direction, and conclude that this relationship holds biconditionally, first take
an arbitrary hyperplane H and observe that because for all x ∈ Γ(H),

ρ(H + x) = ρ(E) = |E| − ρ∗ (E),

it follows by interpretting this equality along with Eq. 10,

ρ∗ (Γ(H) − x) + |H| + 1 − ρ∗ (E) = |E| − ρ∗ (E).

41
Hence, because ρ∗ (Γ(H) − x) + |H| + 1 = |E| it follows that,

ρ∗ (Γ(H) − x) = |Γ(X)| − 1,

and so,
Γ(H) − x ∈ M∗ , ∀x ∈ Γ(H). (11)
However, we also have that ρ(H + x) = ρ(H) + 1. Then, combining this with Eq. 10 implies,

ρ∗ (Γ(H) − x) + |H| − ρ∗ (E) = ρ∗ (Γ(H)) + |H| − ρ∗ (E),

and so ρ∗ (Γ(H) − x) = ρ∗ (Γ(H)) for all x ∈ Γ(H). Combining this last insight with Eq. 11, it can
only be that Γ(H) is a circuit of M∗ .

Next, we examine the result stating that two cocircuits are a modular pair if and only if their
corresponding hyperplanes (given by Lemma 3.1) are also a modular pair. Doing so will once again
make use of the structure given by the dual rank function. For the proof, recall that as stated in
Appendix A, that a coline is a flat of rank two less than that of the matroid.
Lemma 5.2 (Folklore). Let M be a matroid, C1∗ and C2∗ be two cocircuits, and H1 and H2 the
corresponding hyperplanes such that Γ(H1 ) = C1∗ and Γ(H2 ) = C2∗ . Then, we have that C1∗ and C2∗
are a modular pair if and only if H1 and H2 are a modular pair.
Proof. Assume C1∗ 6= C2∗ . Then, |C1∗ ∩ C2∗ | < |C1∗ |, |C2∗ | and so by definition of circuits it follows
that,
ρ∗ (C1∗ ∩ C2∗ ) = |C1∗ ∩ C2∗ |,
i.e. the intersection is an independent set of M∗ . Now, recall that by the law of inclusion-exclusion,

|C1∗ ∩ C2∗ | + |C1∗ ∪ C2∗ | = |C1∗ | + |C2∗ |,

and so combining this with our prior observation and the fact that ρ∗ (C1∗ )+ρ∗ (C2∗ ) = |C1∗ |+|C2∗ |−2,
it follows that,
ρ∗ (C1∗ ∩ C2∗ ) + |C1∗ ∪ C2∗ | − 2 = ρ∗ (C1∗ ) + ρ∗ (C2∗ ).
Therefore, C1∗ and C2∗ are a modular pair if and only if ρ∗ (C1∗ ∪ C2∗ ) = |C1∗ ∪ C2∗ | − 2. Likewise,
because H1 ⊔ H2 = E and both ρ(H1 ) and ρ(H2 ) equal ρ(E) − 1, it follows that H1 and H2 are a
modular pair of L (M) if and only if their meet H1 ⊓ H2 is a coline. Because H1 ⊓ H2 = H1 ∩ H2 ,
this is witnessed by the identity,
ρ(H1 ∩ H2 ) = ρ(E) − 2.
Now, by keeping these observations in mind we can begin the real argument. Suppose C1∗ and
C2∗ are a modular pair. Using the definition of the dual rank function in Eq. 10 and our previously
identified certificate for these cocircuits forming a modular pair, it follows that,

ρ(H1 ∩ H2 ) + |C1∗ ∪ C2∗ | − ρ(E) = ρ∗ (C1∗ ∪ C2∗ ) = |C1∗ ∪ C2∗ | − 2, (12)

and so it is evident by rearranging terms that ρ(H1 ∩ H2 ) = ρ(E) − 2, making H1 and H2 a modular
pair as well. Going the other way, when H1 and H2 are a modular pair, by now using the dual
rank function and our certificate for hyperplanes forming a modular pair we see,

ρ∗ (C1∗ ∪ C2∗ ) − |C1∗ ∪ C2∗ | + ρ(E) = ρ(H1 ∩ H2 ) = ρ(E) − 2,

and so ρ(C1∗ ∪ C2∗ ) = |C1∗ ∪ C2∗ | − 2. This concludes the lemma.

42
C Sketch of Correctness of Parallel Basis Search from [KUW88]
In this section we sketch out an alternative proof of the correctness of Algorithm 3 which fills in
some details omitted from the initial source. Recall, in Section 5.3 we identified that Algorithm 3
is correct if, !
[
σ X∪ Zi = E, (13)
i∈I
always holds throughout the execution in the algorithm. To see why, first observe that maintaining
Eq. 13 amounts to imposing that the union of X and the subsets {Zi }i∈I always contain a basis.
Then, because the partial solution X can only be augmented by a set Zi if X ∪ Zi ∈ M (see lines
9–12 of Algorithm 3) it follows that X is always an independent set. Hence, since X is always an
independent set and any value of the sets {Zi }i∈I in some round represent the points which can
possibly be augmented into X in that round or future rounds, it follows that maintaining Eq. 13
guarantees that X is a basis once every set Zi follows Zi = ∅. So, one need only verify that Eq. 13
holds throughout the execution of Algorithm 3 to show that it is correct. To do so we will make
use of the following lemma, showing that for any set Y ⊆ E a selection of points can be removed
without perturbing the span so long as there exists a correspondence between such points and a
set of distinct circuits contained in Y .
Lemma C.1 (Folklore). Let M be a matroid over a ground set E, and Y ⊂ E any subset. We
have that for all Z ⊆ Y where |Z| = k, if there exists circuits C1 , . . . , Ck of Y such that each circuit
contains a single unique member of Z, i.e.,
 
[
Ci \ Cj  ∩ Z = 1, ∀i ∈ {1, . . . , k}.
j6=i

Then σ(Y ) = σ(Y \ Z), which is to say the span of Y remains unaltered after difference by Z.
To see how this lemma is useful, we first observe that for every independent set X ∈ M and
point y ∈ σ(X) \ X lying beyond X yet within its span, the set X + y is dependent by definition.
Specifically, X +y must contain a dependent set of minimal size, and we see that such a set would be
a circuit. In fact, one can show this circuit is unique by using various alternative axiomitizations of a
matroid [Oxl06]. Furthermore, because of the hereditary property all subsets of X are independent,
and so we observe that this unique minimal circuit contains y. This circuit is called the fundamental
circuit of X with respect to y. Then, we see that by dependence,
n o
(i) (i)
X ∪ z1 , . . . , zℓ(i) ,
(i)
as in line 20 of Algorithm 3 contains a fundamental circuit Ci with zℓ(i) as a member. Furthermore,
(i)
by construction we see that i 6= j implies zℓ(i) ∈
/ Cj because the initial subsets are disjoint. And of
course,
S by inclusion each of the fundamental circuits we generated in this way is also a circuit of
X ∪ i∈I Zi . Hence, correctness is a direct consequence of Lemma C.1 when letting it’s terms Y
and Z take,
n o
(1) (2)
[
Y =X∪ Zi , and, Z = zℓ(1) , zℓ(2) , . . . ,
i∈I

because then Eq. 13 is maintained throughout the execution of Algorithm 3 as corollary.

The preceding discussion provided a sketch towards the validity of Theorem 5.11, and all that
remains is showing Lemma C.1.

43
Proof of Lemma C.1. Firstly, observe the following fact: If X ⊆ E is dependent then the removal
of any of its atoms cannot change the span. Specifically, suppose X ∈/ M and let y ∈ X. Then,
there exists a circuit C ⊆ X such that y ∈ C. Hence, C − y is independent and y ∈ σ(C − y). But,
because the rank closure is monotonic and C − y ⊆ X − y, we have,

y ∈ σ(C − y) ⊆ σ(X − y).

Therefore, σ(X − y) = σ(X).

Now, to show the lemma suppose we remove elements of Y from X in the order dictated by an
injective labelling ℓ assigning the elements of Y to the unique circuit containing them,
[
ℓ(y) = i ⇐⇒ y ∈ Ci \ Cj .
j6=i

In particular, because i 6= ℓ(y) implies that y ∈

/ Ci it follows that,

i < ℓ(y) =⇒ Cℓ (y) ⊆ X \ ℓ−1 (1), . . . , ℓ−1 (i) ,


∀y ∈ Y. (14)

We emphasize that Eq. 14 follows precisely from the fact that each member of y only belongs to a
single given circuit, as if this were not the case then we could not maintain such a subset relation
iteratively. Then, as a result of Eq. 14 we find that the set resulting from removing an element of Y
at each step contains a circuit, and is hence dependent. Applying the argument from the previous
paragraph inductively thus concludes the lemma, as our approach holds up to permutation on ℓ.

44