Boolean Functions for Cryptography and Error

Correcting Codes
Claude Carlet∗

∗
LAGA, University of Paris 8, France; e-mail: [email protected].

1
Contents
1 Introduction 5

2 Generalities on Boolean functions 8

2.1 Representation of Boolean functions . . . . . . . . . . . . . . 9
2.2 The discrete Fourier transform on pseudo-Boolean and on
Boolean functions . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2.1 Fourier transform and NNF . . . . . . . . . . . . . . . 31
2.2.2 The size of the support of the Fourier transform and
its relationship with Cayley graphs . . . . . . . . . . . 32

3 Boolean functions and coding 33

3.1 Reed-Muller codes . . . . . . . . . . . . . . . . . . . . . . . . 36

4 Boolean functions and cryptography 42

4.1 Cryptographic criteria for Boolean functions . . . . . . . . . . 47
4.1.1 The algebraic degree . . . . . . . . . . . . . . . . . . . 48
4.1.2 The nonlinearity . . . . . . . . . . . . . . . . . . . . . 50
4.1.3 Balancedness and resiliency . . . . . . . . . . . . . . . 56
4.1.4 Strict avalanche criterion and propagation criterion . . 59
4.1.5 Non-existence of nonzero linear structure . . . . . . . 59
4.1.6 Algebraic immunity . . . . . . . . . . . . . . . . . . . 61
4.1.7 Other criteria . . . . . . . . . . . . . . . . . . . . . . . 65

5 Classes of functions for which restrictions on the possible

values of the weights, Walsh spectra and nonlinearities can
be proved 69
5.1 Affine functions . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2 Quadratic functions . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 Indicators of flats . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.4 Normal functions . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5 Functions admitting partial covering sequences . . . . . . . . 74
5.6 Functions with low univariate degree . . . . . . . . . . . . . . 77

6 Bent functions 78
6.1 The dual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6.2 Bent functions of low algebraic degrees . . . . . . . . . . . . . 82
6.3 Bound on algebraic degree . . . . . . . . . . . . . . . . . . . . 84
6.4 Constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.4.1 Primary constructions . . . . . . . . . . . . . . . . . . 85

2
 6.4.2 Secondary constructions . . . . . . . . . . . . . . . . . 91
6.4.3 Decompositions of bent functions . . . . . . . . . . . . 99
6.5 On the number of bent functions . . . . . . . . . . . . . . . . 99
6.6 Characterizations of bent functions . . . . . . . . . . . . . . . 100
6.6.1 characterization through the NNF . . . . . . . . . . . 100
6.6.2 Geometric characterization . . . . . . . . . . . . . . . 101
6.6.3 characterization by second-order covering sequences . 102
6.7 Subclasses: hyper-bent functions . . . . . . . . . . . . . . . . 103
6.8 Superclasses: partially-bent functions, partial bent functions
and plateaued functions . . . . . . . . . . . . . . . . . . . . . 105
6.9 Normal and non-normal bent functions . . . . . . . . . . . . . 109
6.10 Kerdock codes . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.10.1 Construction of the Kerdock code . . . . . . . . . . . 111

7 Resilient functions 113

7.1 Bound on algebraic degree . . . . . . . . . . . . . . . . . . . . 113
7.2 Bounds on the nonlinearity . . . . . . . . . . . . . . . . . . . 115
7.3 Bound on the maximum correlation with subsets of N . . . . 117
7.4 Relationship with other criteria . . . . . . . . . . . . . . . . . 117
7.5 Constructions . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.5.1 Primary constructions . . . . . . . . . . . . . . . . . . 119
7.5.2 Secondary constructions . . . . . . . . . . . . . . . . . 124
7.6 On the number of resilient functions . . . . . . . . . . . . . . 131

8 Functions satisfying the strict avalanche and propagation

criteria 133
8.1 P C(l) criterion . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.1.1 Characterizations . . . . . . . . . . . . . . . . . . . . . 133
8.1.2 Constructions . . . . . . . . . . . . . . . . . . . . . . . 134
8.2 P C(l) of order k and EP C(l) of order k criteria . . . . . . . 134

9 Algebraic immune functions 135

9.1 General properties of the algebraic immunity and its relation-
ship with some other criteria . . . . . . . . . . . . . . . . . . 136
9.1.1 Algebraic immunity of random functions . . . . . . . . 136
9.1.2 Algebraic immunity of monomial functions . . . . . . 136
9.1.3 Functions in odd numbers of variables with optimal
algebraic immunity . . . . . . . . . . . . . . . . . . . . 136
9.1.4 Relationship between normality and algebraic immunity137

3
 9.1.5 Relationship between algebraic immunity, weight and
nonlinearity . . . . . . . . . . . . . . . . . . . . . . . . 138
9.2 The problem of finding functions achieving high algebraic im-
munity and high nonlinearity . . . . . . . . . . . . . . . . . . 139
9.3 The functions with high algebraic immunity found so far and
their parameters . . . . . . . . . . . . . . . . . . . . . . . . . 139

10 Symmetric functions 143

10.1 Representation . . . . . . . . . . . . . . . . . . . . . . . . . . 143
10.2 Fourier and Walsh transforms . . . . . . . . . . . . . . . . . . 145
10.3 Nonlinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
10.4 Resiliency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
10.5 Algebraic immunity . . . . . . . . . . . . . . . . . . . . . . . 148
10.6 The super-classes of rotation symmetric and Matriochka sym-
metric functions . . . . . . . . . . . . . . . . . . . . . . . . . 149

4
1 Introduction
A fundamental objective of cryptography is to enable two persons to commu-
nicate over an insecure channel (a public channel such as internet) in such
a way that any other person is unable to recover their message (called the
plaintext) from what is sent in its place over the channel (the ciphertext).
The transformation of the plaintext into the ciphertext is called encryption,
or enciphering. Encryption-decryption is the most ancient cryptographic
activity (ciphers already existed four centuries B. C.) but its nature has
deeply changed with the invention of computers, because the cryptanalysis
(the activity of the third person, the eavesdropper, who aims at recovering
the message) can use their power.
The encryption algorithm takes as input the plaintext and an encryption
key KE , and it outputs the ciphertext. If the encryption key is secret, then
we speak of conventional cryptography, of private key cryptography or of
symmetric cryptography. In practice, the principle of conventional cryptog-
raphy relies on the sharing of a private key between the sender of a message
(often called Alice in cryptography) and its receiver (often called Bob). If,
on the contrary, the encryption key is public, then we speak of public key
cryptography. Public key cryptography appeared in the literature in the late
seventies.
The decryption (or deciphering) algorithm takes as input the ciphertext and
a secret1 decryption key KD . It outputs the plaintext.

plaintext - ciphertext - plaintext

Encryption Decryption -
b public b
channel
KE KD

Public key cryptography is preferable to conventional cryptography, since

it allows to securely communicate without having previously shared keys in
a secure way: every person who wants to receive secret messages can keep
secret a decryption key and publish an encryption key; if n persons want to
secretly communicate pairwise using a public key cryptosystem, they need n
1
According to principles already stated in 1883 by A. Kerckhoffs [212], who cited a still
more ancient manuscript by R. du Carlet [50], only the secret keys must be kept secret –
the confidentiality should not rely on the secrecy of the encryption method – and a cipher
cannot be considered secure if it can be decrypted by the designer himself.

5
encryption keys and n decryption keys, when conventional cryptosystems
will need n2 = n(n−1)


2 keys. But all known public key cryptosystems are
much less efficient than conventional cryptosystems (they allow a much lower
data throughput) and they also need much longer keys to ensure the same
level of security. This is why conventional cryptography is still widely used
and studied nowadays. Thanks to public key cryptosystems, the share-out
of the necessary secret keys can be done without using a secure channel (the
secret keys for conventional cryptosystems are strings of a few hundreds of
bits only and can then be encrypted by public key cryptosystems). Proto-
cols specially devoted to key-exchange can also be used.

The objective of error correcting codes is to enable digital communication

over a noisy channel in such a way that the errors in the transmission of bits
can be detected2 and corrected by the receiver. This aim is achieved by using
an encoding algorithm which transforms the information before sending it
over the channel. In the case of block coding3 , the original message is
treated as a list of binary words (vectors) of the same length – say k –
which are encoded into codewords of a larger length – say n. Thanks to
this extension of the length, called redundancy, the decoding algorithm can
correct the errors of transmission (if their number is, for each sent word,
smaller than or equal to the so-called correction capacity of the code) and
recover the correct message. The set of all possible codewords is called
the code. Sending words of length n over the channel instead of words of
length k slows down the transmission of information in the ratio of nk . This
ratio, called the transmission rate, must be as high as possible, to allow fast
communication.

message - Encoding codeword - -message

Decoding
noisy
channel

In both cryptographic and error correcting coding activities, Boolean

functions (that is, functions from the vectorspace Fn2 of all binary vectors of
2
If the code is used only to detect errors, then when an error is detected, the information
must be requested and sent again in a so-called “automatic request” procedure.
3
We shall not address convolutional coding here.

6
length n, to the finite field with two elements4 F2 ) play roles:
- every code of length 2n , for some positive integer n, can be interpreted
as a set of Boolean functions, since every n-variable Boolean function can
be represented by its truth table (an ordering of the set of binary vectors
of length n being first chosen) and thus associated with a binary word of
length 2n , and vice versa; important codes (Reed-Muller, Kerdock codes)
can be defined this way as sets of Boolean functions;
- the role of Boolean functions in conventional cryptography is even more
important: cryptographic transformations (pseudo-random generators in
stream ciphers, S-boxes in block ciphers) can be designed by appropriate
composition of nonlinear Boolean functions.
In both frameworks, n is rarely large, in practice. The error correcting
codes derived from n-variable Boolean functions have length 2n ; so, tak-
ing n = 10 already gives codes of length 1024. For reason of efficiency, the
S-boxes used in most block ciphers are concatenations of sub S-boxes on at
most 8 variables. In the case of stream ciphers, n was in general at most
equal to 10 until recently. This has changed with the algebraic attacks (see
[113, 117, 150] and see below) but the number of variables is now most often
limited to 20.
Despite the fact that Boolean functions are currently used in cryptog-
raphy and coding with low numbers of variables, determining and studying
those Boolean functions satisfying the desired conditions (see Subection 4.1
below) is not feasible through an exhaustive computer investigation: the
n
number |BF n | = 22 of n-variable Boolean functions is too large when n ≥ 6.
We give in table 1 below the values of this number for n ranging between 4
and 8.

n 4 5 6 7 8
|BF n | 216 232 264 2128 2256
≈ 6 · 104 4 · 109 1019 1038 1077

Table 1: Number of n-variable Boolean functions

Assume that visiting an n-variable Boolean function, and determining whe-

ther it has the desired properties, needs one nano-second (10−9 seconds),
then it would need millions of hours to visit all functions in 6 variables, and
about one hundred billions times the age of the universe to visit all those
in 7 variables. The number of 8-variable Boolean functions approximately
equals the number of atoms in the whole universe! We see that trying to find
4
Denoted by B is some chapters of the present collection.

7
functions satisfying the desired conditions by simply picking up functions
at random is also impossible for these values of n, since visiting a non-
negligible part of all Boolean functions in 7 or more variables is not feasible,
even when parallelizing. The study of Boolean functions for constructing or
studying codes or ciphers is essentially mathematical. But clever computer
investigation is very useful to imagine or to test conjectures, and sometimes
to generate interesting functions.

2 Generalities on Boolean functions

In this chapter and in the chapter “Vectorial Boolean Functions for Cryp-
tography” which follows, the set {0, 1} will be most often endowed with the
structure of field (and denoted by F2 ), and the set Fn2 of all binary vectors5
of length n will be viewed as an F2 -vectorspace. We shall denote simply by 0
the null vector in Fn2 . The vectorspace Fn2 will sometimes be also endowed
with the structure of field – the field F2n (also denoted by GF (2n )); indeed,
this field being an n-dimensional vectorspace over F2 , each of its elements
can be identified with the binary vector of length n of its coordinates relative
to a fixed basis. The set of all Boolean functions f : Fn2 → F2 will be denoted
as usual by BF n . The Hamming weight wH (x) of a binary vector x ∈ Fn2
being the number of its nonzero coordinates (i.e. the size of {i ∈ N/ xi 6= 0}
where N denotes the set {1, · · · , n}, called the support of the codeword ),
the Hamming weight wH (f ) of a Boolean function f on Fn2 is (also) the
size of the support of the function , i.e. the set {x ∈ Fn2 / f (x) 6= 0}. The
Hamming distance dH (f, g) between two functions f and g is the size of the
set {x ∈ Fn2 / f (x) 6= g(x)}. Thus it equals wH (f ⊕ g).
Note. Some additions of bits will be considered in Z (in characteristic
0) and denoted then by +, and some will be computed modulo 2 and de-
noted by ⊕. These two different notations will be necessary because some
representations of Boolean functions will live in characteristic 2 and some
representations of the same functions will live in characteristic 0. But the
additions of elements of the finite field F2n will be denoted by +, as it is
usual in mathematics. So, for simplicity (since Fn2 will often be identified
with F2n ) and because there will be no ambiguity, we shall also denote by +
the addition of vectors of Fn2 when n > 1.
5
Coders say “words”

8
2.1 Representation of Boolean functions
Among the classical representations of Boolean functions, the one which is
most usually used in cryptography and coding is the n-variable polynomial
representation over F2 , of the form
!
M Y M
f (x) = aI xi = aI xI , (1)
I∈P(N ) i∈I I∈P(N )

where P(N ) denotes the power set of N = {1, · · · , n}. Every coordinate xi
appears in this polynomial with exponents at most 1, because every bit in F2
equals its own square. This representation belongs to F2 [x1 , · · · , xn ]/(x21 ⊕
x1 , · · · , x2n ⊕ xn ). It is called the Algebraic Normal Form (in brief the ANF).

Example: let us consider the function f whose truth-table is

x1 x2 x3 f (x)
0 0 0 0
0 0 1 1
0 1 0 0
0 1 1 0
1 0 0 0
1 0 1 1
1 1 0 0
1 1 1 1

It is the sum (modulo 2 or not, no matter) of the atomic functions f1 , f2

and f3 whose truth-tables are

x1 x2 x3 f1 (x) f2 (x) f3 (x)

0 0 0 0 0 0
0 0 1 1 0 0
0 1 0 0 0 0
0 1 1 0 0 0
1 0 0 0 0 0
1 0 1 0 1 0
1 1 0 0 0 0
1 1 1 0 0 1

The function f1 (x) takes value 1 if and only if 1 ⊕ x1 = 1, 1 ⊕ x2 = 1 and

x3 = 1, that is if and only if (1 ⊕ x1 )(1 ⊕ x2 ) x3 = 1. Thus the ANF of f1

9
can be obtained by expanding the product (1 ⊕ x1 )(1 ⊕ x2 ) x3 . After similar
observations on f2 and f3 , we see that the ANF of f equals (1 ⊕ x1 )(1 ⊕
x2 ) x3 ⊕ x1 (1 ⊕ x2 ) x3 ⊕ x1 x2 x3 = x1 x2 x3 ⊕ x2 x3 ⊕ x3 . 2

Another possible representation of this same ANF uses an indexation by

means of vectors of Fn2 instead of subsets of N ; if, for any such vector u,
we denote by au what is denoted by asupp(u) in Relation (1) (where supp(u)
denotes the support of u), we have the equivalent representation:
 
M n
Y
f (x) = au  xj uj  .
u∈Fn
2 j=1
Qn uj
The monomial j=1 xj is often denoted by xu .

Existence and uniqueness of the ANF By applying the Lagrange in-

terpolation method described in the example above, it is a simple matter to
show the existence of the ANF of every Boolean function. This implies that
the mapping, from every polynomial P ∈ F2 [x1 , · · · , xn ]/(x21 ⊕ x1 , · · · , x2n ⊕
xn ) to the corresponding function x ∈ Fn2 7→ P (x), is onto BF n . Since
the size of BF n equals the size of F2 [x1 , · · · , xn ]/(x21 ⊕ x1 , · · · , x2n ⊕ xn ), this
correspondence is one to one6 . But more can be said.

Relationship
Q between a Boolean function and its ANF The prod-
uct xI = i∈I xi is nonzero if and only if xi is nonzero (i.e. equals 1) for
every i ∈ I, thatL
is, if I is included in the support of x; hence, the Boolean
function f (x) = I∈P(N ) aI xI takes value
M
f (x) = aI , (2)
I⊆supp(x)

where supp(x) denotes the support of x. IfLwe use the notation f (x) =
au xu , we obtain the relation f (x) = ux au , where u  x means
L
u∈Fn
2
that supp(u) ⊆ supp(x) (we say that u is covered by x). A Boolean func-
tion f ◦ can be associated to the ANF of f : for n ◦
Levery x ∈u F2 ,◦ we set f (x) =
asupp(x) , that is, with the notation f (x) = u∈Fn au x : f (u) = au . Re-
2
◦
lation (2) shows that f is the image of f by the so-called binary Möbius
6
Another argument is that this mapping is a linear mapping from a vectorspace over F2
of dimension 2n to a vectorspace of the same dimension.

10
transform.
The converse is also true:

Proposition 1 Let f be a Boolean function on Fn2 and let xI

L
I∈P(N ) aI
be its ANF. We have:
M
∀I ∈ P(N ), aI = f (x). (3)
x∈Fn
2/ supp(x)⊆I
L
Proof. Let us denote x∈Fn f (x) by bI and consider the func-
2 / supp(x)⊆I
I
L
tion g(x) = I∈P(N ) bI x . We have
 
M M M
g(x) = bI =  f (y)
I⊆supp(x) I⊆supp(x) y∈Fn
2/ supp(y)⊆I

and thus  
M M
g(x) = f (y)  1 .
y∈Fn
2 I∈P(N )/ supp(y)⊆I⊆supp(x)
L
The sum I∈P(N )/ supp(y)⊆I⊆supp(x) 1 is null if y 6= x, since the set {I ∈
P(N )/ supp(y) ⊆ I ⊆ supp(x)} contains 2wH (x)−wH (y) elements if supp(y) ⊆
supp(x), and none otherwise. Hence, g = f and, by uniqueness of the ANF,
bI = aI for every I. 2

Algorithm There exists a simple divide-and-conquer butterfly algorithm

to compute the ANF from the truth-table (or vice-versa), that we can call the
Fast Möbius Transform. For every u = (u1 , · · · , un ) ∈ Fn2 , the coefficient au
of xu in the ANF of f equals
M
[f (x1 , · · · , xn−1 , 0)] if un = 0 and
(x1 ,···,xn−1 )(u1 ,···,un−1 )
M
[f (x1 , · · · , xn−1 , 0) ⊕ f (x1 , · · · , xn−1 , 1)] if un = 1.
(x1 ,···,xn−1 )(u1 ,···,un−1 )

Hence if, in the truth-table of f , the binary vectors are ordered in lexico-
graphic order, with the bit of higher weight on the right, the table of the
ANF equals the concatenation of the ANFs of the (n − 1)-variable functions
f (x1 , · · · , xn−1 , 0) and f (x1 , · · · , xn−1 , 0)⊕f (x1 , · · · , xn−1 , 1). We deduce the
following algorithm:

11
 1. write the truth-table of f , in which the binary vectors of length n are
in lexicographic order as decribed above;

2. let f0 and f1 be the restrictions of f to F2n−1 × {0} and F2n−1 × {1},

respectively7 ; replace the values of f1 by those of f0 ⊕ f1 ;

3. apply recursively step 2, separately to the functions now obtained in

the places of f0 and f1 .

When the algorithm ends (i.e. when it arrives to functions in one variable
each), the global table gives the values of the ANF of f . The complexity of
this algorithm is of n 2n XORs.

Remark.
The algorithm works the same if the vectors are ordered in standard lex-
icographic order, with the bit of higher weight on the left (indeed, this
corresponds to applying it to f (xn , xn−1 , · · · , x1 )).

The degree of the ANF is denoted by d◦ f and is called the algebraic

degree of the function (this makes sense thanks to the existence and unique-
ness of the ANF): d◦ f = max{|I|/ aI 6= 0}, where |I| denotes the size of I.
Some authors also call it the nonlinear order of f . According to Relation (3),
we have:

Proposition 2 The algebraic degree d◦ f of any n-variable Boolean function

f equals the maximum dimension of the subspaces {x ∈ Fn2 / supp(x) ⊆ I}
on which f takes value 1 an odd number of times.

The algebraic degree is an affine invariant (it is invariant under the action
x1
 
 x2 
of the general affine group): for every affine isomorphism L :   ...  ∈


xn
x1 a1
   
 x2   a2 
Fn2 →
7 M × n
 ...  ⊕  ...  ∈ F2 (where M is a nonsingular n × n matrix
  

xn an
7
The truth-table of f0 (resp. f1 ) corresponds to the upper (resp. lower) half of the
table of f .

12
over F2 ), we have d◦ (f ◦ L) = d◦ f . Indeed, the composition by L clearly
cannot increase the algebraic degree, since the coordinates of L(x) have de-
gree 1. Hence we have d◦ (f ◦ L) ≤ d◦ f (this inequality is more generally
valid for every affine homomorphism). And applying this inequality to f ◦ L
in the place of f and to L−1 in the place of L shows the inverse inequality.
Two functions f and f ◦ L where L is an F2 -linear automorphism of Fn2 (in
the case case a1 = a2 = · · · = an = 0 above) will be called linearly equivalent
and two functions f and f ◦ L, where L is an affine automorphism of Fn2 ,
will be called affinely equivalent.
The algebraic degree being an affine invariant, Proposition 2 implies that
it also equals the maximum dimension of all the affine subspaces of Fn2 on
which f takes value 1 an odd number of times.
It is shown in [297] that, for every nonzero n-variable Boolean function f ,
denoting by g the binary Möbius transform of f , we have d◦ f + d◦ g ≥ n.
This same paper deduces characterizations and constructions of the func-
tions which are equal to their binary Möbius transform, called coincident
functions.

Remarks.
1. Every atomic function has algebraic degree n, since its ANF equals (x1 ⊕
1 )(x2 ⊕ 2 ) · · · (xn ⊕ n ), where i ∈ F2 . Thus, a Boolean function f has
algebraic degree n if and only if, in its decomposition as a sum of atomic
functions, the number of these atomic functions is odd, that is, if and only
if wH (f ) is odd. This property will have an important consequence on the
Reed-Muller codes and it will be also useful in Section 3.
2. If we know that the algebraic degree of an n-variable Boolean func-
tion f is bounded above by d < n, then the whole function can be recovered
from some of its restrictions (i.e., a unique function corresponds to this
partially defined Boolean function). Precisely, according to the existence
and uniqueness of the ANF, the knowledge of the restriction f|E of the
Boolean function f (of algebraic degree at most d) to a set E implies the
knowledge Lof the whole function if and only if the system of the equations
f (x) = a x I , with indeterminates a ∈ F , and where x
I∈P(N )/ |I|≤d I I 2
ranges over E (this makes |E| equations), has a unique solution8 . This
happens with the set Ed of all words of Hamming weights smaller than or
equal to d, since Relation (3) gives the value of aI (when I ∈ P(N ) has size
8
Taking f|E null leads to determining the so-called annihilators of the indicator of E
(the function 1E , also called characteristic function of E, defined by 1E (x) = 1 if x ∈ E
and 1E (x) = 0 otherwise); this is the core analysis of Boolean functions from the viewpoint
of algebraic attacks, see Subsection 4.1.

13
|I| ≤ d). Notice that Relation (2) allows then to express the value of f (x)
for every x ∈ Fn2 by means of the values taken by f at all words of Hamming
weights smaller than or equal to d. We have (using the notation au instead
of aI , see above):
M M M
f (x) = au = au = f (y) |{u ∈ Ed / y  u  x|
ux ux yx
u∈Ed y∈Ed
  
d−wH (y)  
M X wH (x) − wH (y) 
= f (y)  [mod 2] .
i
yx i=0
y∈Ed

More generally, the whole function f can be recovered from f|E for every
set E affinely equivalent to Ed , according to the affine invariance of the
algebraic degree. This also generalizes to “pseudo-Boolean” (that is, real-
valued) functions, if we consider the numerical degree (see below) instead of
the the algebraic degree, cf. [350]. 2

The simplest functions, from the viewpoint of the ANF, are those Boolean
functions of algebraic degrees at most 1, called affine functions:

f (x) = a1 x1 ⊕ · · · ⊕ an xn ⊕ a0 .

They are the sums of linear and constant functions. Denoting by a · x the
usual inner product a · x = a1 x1 ⊕ · · · ⊕ an xn in Fn2 , or any other inner
product (symmetric and such that, for every a 6= 0, the function x → a · x
is a nonzero linear form on Fn2 ), the general form of an n-variable affine
function is a · x ⊕ a0 (with a ∈ Fn2 ; a0 ∈ F2 ).
Affine functions play an important role in coding (they are involved in the
definition of the Reed-Muller code of order 1, see Subsection 3.1) and in cryp-
tography (the Boolean functions used as “nonlinear functions” in cryptosys-
tems must behave as differently as possible from affine functions, see Sub-
section 4.1).

Trace representation(s) A second kind of representation plays an im-

portant role in sequence theory, and is also used for defining and study-
ing Boolean functions. It leads to the construction of the Kerdock codes
(see Subsection 6.10). Recall that, for every n, there exists a (unique up to
isomorphism) field F2n (also denoted by GF (2n )) of order 2n (see [248]). The
vectorspace Fn2 can be endowed with the structure of this field F2n . Indeed,
we know that F2n has the structure of an n-dimensional F2 -vectorspace; if

14
we choose an F2 -basis (α1 , · · · , αn ) of this vectorspace, then every element
x ∈ Fn2 can be identified with x1 α1 + · · · + xn αn ∈ F2n . We shall still denote
by x this element of the field.
1. It is shown in the chapter “Vectorial Boolean Functions for Cryptogra-
phy” (see another proof below) that every mapping from F2n into F2n admits
a (unique) representation as a polynomial
n −1
2X
f (x) = δi xi (4)
i=0

over F2n in one variable and of (univariate) degree at most 2n − 1. Any

Boolean function on F2n is a particular case of a vectorial function from F2n
to F2n (since F2 is a subfield of F2n ) and admits therefore such a unique
representation, that we shall call the univariate representation of f . For
n
every u, v ∈PF2n we have (u + v)2 = u2 + v 2 and u2 = u. A univariate
n
polynomial 2i=0−1 δi xi , δi ∈ F2n , is then the univariate representation of a
P n 2 P n P2n −1
2 −1 i 2 −1 2 2i
Boolean function if and only if i=0 δi x = i=0 δi x = i=0 δ i xi
n
[mod x2 +x], that is, δ0 , δ2n −1 ∈ F2 and, for every i = 1, · · · , 2n −2, δ2i = δi2 ,
where the index 2i is taken mod 2n − 1.
2 n−1
2. The function defined on F2n by trn (u) = u + u2 + u2 + · · · + u2 is F2 -
linear and satisfies (trn (u))2 = trn (u2 ) = trn (u); it is therefore valued in F2 .
This function is called the trace function from F2n to its prime field F2 or the
absolute trace function on F2n . The function (u, v) 7→ trn (u v) is an inner
product in F2n (that is, it is symmetric and for every v 6= 0, the function
u → trn (u v) is a nonzero linear form on F2n ). Every Boolean function can
be written in the form f (x) = trn (F (x)) where F is a mapping from F2n
into F2n (an example of such mapping F is defined by F (x) = λ f (x) where
trn (λ) = 1 and f (x) is the univariate representation). Thus, every Boolean
function can be also represented in the form
n −1
2X
!
trn βi xi , (5)
i=0

where βi ∈ F2n . Such a representation is not unique. Now, thanks to the

fact that trn (u2 ) = trn (u) for every u ∈ F2n , we can restrict the exponents
i with nonzero coefficients βi so that there is at most one such exponent in
each cyclotomic class {i × 2j [ mod (2n − 1)] ; j ∈ N} of 2 modulo 2n − 1
(but this still does not make the representation unique). We shall call this
expression the absolute trace representation of f .

15
3. We come back to the univariate representation. Let us see how it can
be obtained from the truth table of the function and represented in a con-
venient way by using the notation trn . Denoting by α a primitive element
n
of the field F2n (that is, an element such that F2n = {0, 1, α, α2 , · · · , α2 −2 },
which always exists [248]), the Mattson-Solomon polynomial of the vector
n
(f (1), f (α), f (α2 ), · · · , f (α2 −2 )) is the polynomial [258]
n −1
2X n −2
2X
2n −1−j
A(x) = Aj x = A−j xj
j=1 j=0

with: n −2
2X
Aj = f (αk )αkj .
k=0

Note that the Mattson Solomon transform is a discrete Fourier transform.

We have, for every 0 ≤ i ≤ 2n − 2:
n −1
2X n −1 2n −2
2X X
i −ij
A(α ) = Aj α = f (αk )α(k−i)j = f (αi )
j=1 j=1 k=0

Pn P2n −2 (k−i)j (k−i)(2n −1)

(since 2j=1−1 α(k−i)j = j=0 α = α αk−i +1 +1 equals 0 if 1 ≤ k 6=
i ≤ 2nP− 2), and A is therefore the univariate representation of f , if f (0) =
n
A0 = 2i=0−2 f (αi ) (note that this works also for functions from F2n to F2n )
that is, if f has even weight, i.e. has algebraic degree strictly less than n.
n n
Otherwise, we have f (x) = A(x) + 1 + x2 −1 , since 1 + x2 −1 takes value 1
at 0 and 0 at every nonzero element of F2n .
Note that A2j = A2j . Denoting by Γ(n) the set obtained by choosing one
element in each cyclotomic class of 2 modulo 2n − 1 (the most usual choice
for k is the smallest element in its cyclotomic class, called the coset leader
of the class), this allows representing f (x) in the form
n
X
trnj (A−j xj ) + (1 + x2 −1 ), (6)
j∈Γ(n)

where  = wH (f ) [mod 2] and where nj is the size of the cyclotomic class

containing j. Note that, for every j ∈ Γ(n) and every x ∈ F2n , we have
nj
Aj ∈ F2nj (since A2j = Aj ) and xj ∈ F2nj as well. We shall call this
expression the trace representation of f . Obviously, it is nothing more than
an alternate expression for the univariate representation. For this reason,
it is unique (if we restrict the coefficient of xj to live in F2nj ). But it is

16
useful to distinguish the different expressions by different names. We shall
call globally “trace representations” the three expressions (4), (5) and (6).
Trace representations and the algebraic normal form are closely related. Let
us see how the ANF can be Pnobtained from the univariate representation:
we express x in the form i=1 xi αi , where (α1 , · · · , αn ) is a basis of the
F2 -vectorspace F2 . Recall P
n that, for every j ∈ Z/(2n − 1)Z, the binary
expansion of j has the form s∈E 2s , where E ⊆ {0, 1, · · · , n − 1}. The size
of E is often called the 2-weight of j and written w P2 (j). We write more
n−1 s
conveniently the binary expansion of j in the form: s=0 js 2 , js ∈ {0, 1}.
We have then:
n −1
2X n
!j
X
f (x) = δj xi αi
j=0 i=1

2n −1 n
!Pn−1
s=0 js 2
s
X X
= δj xi αi
j=0 i=1
n −1
2X n−1 n
!js
s
Y X
= δj xi αi2 .
j=0 s=0 i=1

Expanding these last products and simplifying gives the ANF of f .

Function f has then algebraic degree maxj=0,···,2n −1/ δj 6=0 w2 (j). Indeed,
according to the above equalities, its algebraic degree is clearly bounded
above by this number, and it can not be strictly smaller, because the number
of Boolean n-variable functions P n of algebraic degrees at most d equals the
number of the polynomials 2j=0−1 δj xj such that δ0 , δ2n −1 ∈ F2 and δ2j =
δj2 ∈ F2n for every j = 1, · · · , 2n − 2 and maxj=0,···,2n −1/ δj 6=0 w2 (j) ≤ d.
We have also:

Proposition 3 [51] Let a be any element of F2n and k any integer [mod
2n − 1]. If f (x) = trn (axk ) is not the null function, then it has algebraic
degree w2 (k).

Proof. Let nk be again the size of the cyclotomic class containing k. Then
the univariate representation of f (x) equals
 nk 2nk n−nk
  n 2n n−n
2
a + a2 + a2 + · · · + a2 xk + a + a2 k + a2 k + · · · + a2 k x2k

 n 2n n−n
2nk −1 n −1
+ · · · + a + a2 k + a2 k + · · · + a2 k x2 k k .

17
All the exponents of x have 2-weight w2 (k) and their coefficients are nonzero
if and only if f is not null. 2

Remark. Another (more complex) way of showing Proposition 3 is used

in [51] as follows: let r = w2 (k); we consider the r-linear function φ over
the field F2n whose value at (x1 , ..., xr ) equals the sum of the images by
f of all the 2r possible linear combinations of the xj ’s. Then φ(x1 , ..., xr )
equals the sum, for all bijective mappings σ from {1, · · · , r} onto E (where
σ(j)
k = s∈E 2s ) of trn (a rj=1 xj2 ). Proving that f has degree r is equivalent
P Q
to proving that φ is not null, and it can be shown that if φ is null, then f is
null.

The representation over the reals has proved itself to be useful for
characterizing several cryptographic criteria [63, 87, 88] (see Sections 6 and
7). It represents Boolean functions, and more generally real-valued functions
on Fn2 (that are called n-variable pseudo-Boolean functions) by elements of
R [x1 , · · · , xn ]/(x21 −x1 , · · · , x2n −xn ) (or of Z [x1 , · · · , xn ]/(x21 −x1 , · · · , x2n −xn )
for integer-valued functions). We shall call it the Numerical Normal Form
(NNF).
The existence of this representation for every pseudo-Boolean function is
easy to show with the same arguments as for the ANFs of Boolean functions
(writing 1 − xi instead of 1 ⊕ xi ). The linear mapping from every element of
the 2n -th dimensional R-vectorspace R [x1 , · · · , xn ]/(x21 − x1 , · · · , x2n − xn ) to
the corresponding pseudo-Boolean function on Fn2 being onto, it is therefore
one to one (the R-vectorspace of pseudo-Boolean functions on Fn2 having
also dimension 2n ). We deduce the uniqueness of the NNF.
We call the degree of the NNF of a function its numerical degree. Since
the ANF is the mod 2 version of the NNF, the numerical degree is always
bounded below by the algebraic degree. It is shown in [286] that, if a Boolean
function f has no ineffective variable (i.e. if it actually depends on each of
its variables), then the numerical degree of f is greater than or equal to
log2 n − log2 log2 n.
The numerical degree is not an affine invariant. But the NNF leads to an
affine invariant (see a proof of this fact in [88]; see also [191]) which is more
discriminant than the algebraic degree:

Definition 1 Let f be a Boolean function on Fn2 . We call generalized de-

gree of f the sequence (di )i≥1 defined as follows:
for every i ≥ 1, di is the smallest integer d > di−1 (if i > 1) such that, for
every multi-index I of size strictly greater than d, the coefficient λI of xI in

18
the NNF of f is a multiple of 2i .

Example: the generalized degree of any nonzero affine function is the se-
quence of all positive integers.

xI
P
Similarly as for the ANF, a (pseudo-) Boolean function f (x) = I∈P(N ) λI
takes value: X
f (x) = λI . (7)
I⊆supp(x)

But, contrary to what we observed for the ANF, the reverse formula is not
identical to the direct formula:

Proposition 4 Let f be a pseudo-Boolean function on Fn2 and let its NNF

be I∈P(N ) λI xI . Then:
P

X
∀I ∈ P(N ), λI = (−1)|I| (−1)wH (x) f (x). (8)
x∈Fn
2 | supp(x)⊆I

Thus, function f and its NNF are related through the Möbius transform
over integers. X
Proof. Let us denote the number (−1)|I| (−1)wH (x) f (x) by µI
x∈Fn2 | supp(x)⊆I
I
P
and consider the function g(x) = I∈P(N ) I x . We have
µ
 
X X X
g(x) = µI = (−1)|I| (−1)wH (y) f (y)
I⊆supp(x) I⊆supp(x) y∈Fn
2 | supp(y)⊆I

and thus
 
X X
g(x) = (−1)wH (y) f (y)  (−1)|I|  .
y∈Fn
2 I∈P(N )/ supp(y)⊆I⊆supp(x)

X
The sum (−1)|I| is null if supp(y) 6⊆ supp(x). It
I∈P(N )/ supp(y)⊆I⊆supp(x)
is also null if supp(y) is included in supp(x), but different. Indeed, de-
PwH (x)−wH (y) wH (x)−wH (y)

noting |I| − wH (y) by i, it equals ± i=0 i (−1)i =
±(1 − 1)wH (x)−wH (y) = 0. Hence, g = f and, by uniqueness of the NNF, we
have µI = λI for every I. 2

19
We have seen that the ANF of any Boolean function can be deduced from
its NNF by reducing it modulo 2. Conversely, the NNF can be deduced from
the ANF since we have
I
M Y
f (x) = aI xI ⇐⇒ (−1)f (x) = (−1)aI x
I∈P(N ) I∈P(N )
Y
⇐⇒ 1 − 2 f (x) = (1 − 2 aI xI ).
I∈P(N )

Expanding this last equality gives the NNF of f (x) and we have [87]:
2n
X X
λI = (−2)k−1 aI1 · · · aIk , (9)
k=1 {I1 ,...,Ik } |
I1 ∪···∪Ik =I

where “{I1 , . . . , Ik } | I1 ∪· · ·∪Ik = I” means that the multi-indices I1 , . . . , Ik

are all distinct, in indefinite order, and that their union equals I.
A polynomial P (x) = J∈P(N ) λJ xJ , with real coefficients, is the NNF of
P

some Boolean function if and only if we have P 2 (x) = P (x), for every x ∈ Fn2
(which is equivalent to P = P 2 in R [x1 , · · · , xn ]/(x21 − x1 , · · · , x2n − xn )), or
equivalently, denoting supp(x) by I:
 2
X X
∀I ∈ P(N ),  λJ  = λJ . (10)
J⊆I J⊆I

Remark.
Imagine that we want to generate a random Boolean function through its
NNF (this can be useful, since we will see below that the main cryptographic
criteria, on Boolean functions, can be characterized, in simple ways, through
their NNFs). Assume that we have already chosen the values λJ for every
J ⊆ I (whereP I ∈ P(N ) is some multi-index) except for I itself. Let us de-
note the sum J⊆I | J6=I λJ by µ. Relation (10) gives (λI +µ)2 = λI +µ. This
equation of degree 2 has two solutions (it has same discriminant as the equa-
tion λI 2 = λI , that is 1). One solution corresponds to the choice P (x) = 0
(where I = supp(x)) and the other one corresponds to the choice P (x) = 1.
2
Thus, verifying that a polynomial P (x) = I∈P(N ) λI xI with real coeffi-
P
cients represents a Boolean function can be done by checking 2n relations.
But it can also be done by verifying a simple condition on P and checking
a single equation.

20
Proposition 5 Any polynomial P ∈ R [x1 , · · · , xn ]/(x21 − x1 , · · · , x2n − xn )
is the NNF of an integer-valued function if and only if all of its coefficients
are integers. Assuming that thisPcondition is satisfied, P is the NNF of a
2
P
Boolean function if and only if: x∈Fn P (x) = x∈Fn P (x).
2 2

Proof. The first assertion is a direct consequence of Relations (7) and (8).
If all the coefficients of P are integers, then we have P 2 (x) ≥ P (x) for
every x; this implies that the 2n equalities, expressing that P
the correspond-
ing function is Boolean, can be reduced to the single one x∈Fn P 2 (x) =
2
2
P
x∈F2n P (x).
The translation of this characterization in terms of the coefficients of P is
given in Relation (32) below.

2.2 The discrete Fourier transform on pseudo-Boolean and

on Boolean functions
Almost all the characteristics needed for Boolean functions in cryptography
and for sets of Boolean functions in coding can be expressed by means of
the weights of some related Boolean functions (of the form f ⊕ `, where `
is affine, or of the form Da f (x) = f (x) ⊕ f (x + a)). In this framework,
the discrete Fourier transform is a very efficient tool: for a given Boolean
function f , the knowledge of the discrete Fourier transform of f is equivalent
with the knowledge of the weights of all the functions f ⊕ `, where ` is linear
(or affine). Also called Hadamard transform, the discrete Fourier transform
is the linear mapping which maps any pseudo-Boolean function ϕ on Fn2 to
the function ϕ b defined on Fn2 by
X
ϕ(u)
b = ϕ(x) (−1)x·u (11)
x∈Fn
2

where x · u is some chosen inner product (for instance the usual inner prod-
uct x · u = x1 u1 ⊕ · · · ⊕ xn un ).

Algorithm There exists a simple divide-and-conquer butterfly algorithm

to compute ϕ, b called the Fast Fourier Transform (FFT). For every a =
(a1 , · · · , an−1 ) ∈ Fn−1
2 and every an ∈ F2 , the number ϕ(a b 1 , · · · , an ) equals
X
(−1)a·x [ϕ(x1 , · · · , xn−1 , 0) + (−1)an ϕ(x1 , · · · , xn−1 , 1)] .
x=(x1 ,···,xn−1 )∈Fn−1
2

21
Hence, if in the tables of values of the functions, the vectors are ordered
in lexicographic order with the bit of highest weight on the right, the table
of ϕ
b equals the concatenation of those of the discrete Fourier transforms of
the (n−1)-variable functions ψ0 (x) = ϕ(x1 , · · · , xn−1 , 0)+ϕ(x1 , · · · , xn−1 , 1)
and ψ1 (x) = ϕ(x1 , · · · , xn−1 , 0)−ϕ(x1 , · · · , xn−1 , 1). We deduce the following
algorithm:
1. write the table of the values of ϕ (its truth-table if ϕ is Boolean), in
which the binary vectors of length n are in lexicographic order as de-
cribed above;

2. let ϕ0 be the restriction of ϕ to Fn−1

2 × {0} and ϕ1 the restriction of ϕ
n−1 9
to F2 × {1} ; replace the values of ϕ0 by those of ϕ0 + ϕ1 and those
of ϕ1 by those of ϕ0 − ϕ1 ;

3. apply recursively step 2, separately to the functions now obtained in

the places of ϕ0 and ϕ1 .
When the algorithm ends (i.e. when it arrives to functions in one variable
each), the global table gives the values of ϕ.
b The complexity of this algo-
rithm is of n 2n additions/substractions.

Application to Boolean functions For a given Boolean function f , the

discrete Fourier transform can be applied to f itself, viewed as a function
valued in {0, 1} ⊂ Z. We denote by fb the corresponding discrete Fourier
transform of f . Notice that fb(0) equals the Hamming weight of f . Thus,
the Hamming distance dH (f, g) = |{x ∈ Fn2 / f (x) 6= g(x)}| = wH (f ⊕ g)
between two functions f and g equals f[ ⊕ g(0).
The discrete Fourier transform can also be applied to the pseudo-Boolean
function fχ (x) = (−1)f (x) (often called the sign function 10 ) instead of f
itself. We have X
fbχ (u) = (−1)f (x)⊕x·u .
x∈Fn
2

9
The table of values of ϕ0 (resp. ϕ1 ) corresponds to the upper (resp. lower) half of the
table of ϕ.
10
The symbol χ is used here because the sign function is the image of f by the non-
trivial character over F2 (usually denoted by χ); to be sure that the distinction between
the discrete Fourier transforms of f and of its sign function is easily done, we change
the font when we deal with the sign function; many other ways of denoting the discrete
Fourier transform can be found in the literature.

22
 x1 x2 x3 x4 x1 x2 x3 x1 x4 f (x) fχ (x) fbχ (x)
0 0 0 0 0 0 0 1 2 4 0 0
1 0 0 0 0 0 0 1 0 0 0 0
0 1 0 0 0 0 1 -1 -2 -4 8 8
1 1 0 0 0 0 1 -1 0 0 0 8
0 0 1 0 0 0 0 1 2 0 0 0
1 0 1 0 0 0 0 1 0 0 0 0
0 1 1 0 0 0 1 -1 -2 0 0 0
1 1 1 0 1 0 0 1 0 0 0 0
0 0 0 1 0 0 0 1 0 0 0 4
1 0 0 1 0 1 1 -1 2 4 4 -4
0 1 0 1 0 0 1 -1 0 0 0 4
1 1 0 1 0 1 0 1 -2 0 4 -4
0 0 1 1 0 0 0 1 0 0 0 -4
1 0 1 1 0 1 1 -1 2 0 -4 4
0 1 1 1 0 0 1 -1 0 0 0 4
1 1 1 1 1 1 1 -1 2 -4 4 -4

Table 2: truth table and Walsh spectrum of f (x) = x1 x2 x3 ⊕ x1 x4 ⊕ x2

We shall call Walsh transform 11 of f the Fourier transform of the sign func-
tion fχ . We give in Table 2 an example of the computation of the Walsh
transform, using the algorithm recalled above.
Notice that fχ being equal to 1 − 2f , we have

fbχ = 2n δ0 − 2fb (12)

where δ0 denotes the Dirac symbol , i.e. the indicator of the singleton {0}, de-
fined by δ0 (u) = 1 if u is the null vector and δ0 (u) = 0 otherwise; see Propo-
sition 7 for a proof of the relation b 1 = 2n δ0 . Relation (12) gives conversely
f
fb = 2n−1 δ0 − χ and in particular:
b
2

fbχ (0)
wH (f ) = 2n−1 − . (13)
2
11
The terminology is not much more settled in the literature than is the notation; we take
advantage here of the fact that many authors, when working on Boolean functions, use the
term of Walsh transform instead of discrete Fourier transform: we call Fourier transform
the discrete Fourier transform of the Boolean function itself and Walsh transform (some
authors write “Walsh-Hadamard transform”) the discrete Fourier transform of its sign
function.

23
Relation (13) applied to f ⊕ `a , where `a (x) = a · x, gives:
fbχ (a)
dH (f, `a ) = wH (f ⊕ `a ) = 2n−1 − . (14)
2
The mapping f 7→ fbχ (0) playing an important role, and being applied in the
sequel to various functions deduced from f , we shall also use the specific
notation X
F(f ) = fbχ (0) = (−1)f (x) . (15)
x∈Fn
2

Properties of the Fourier transform The discrete Fourier transform,

as any other Fourier transform, has very nice and useful properties. The
number of these properties and the richness of their mutual relationship are
impressive. All of these properties are very useful in practice for studying
Boolean functions (we shall often refer to the relations below in the rest of
the chapter). Almost all properties can be deduced from the next lemma
and from the next two propositions.
Lemma 1 Let
P E be any vectorspace over F2 and ` any nonzero linear form
on E. Then x∈E (−1)`(x) is null.
Proof. The linear form ` being not null, its support is an affine hyperplane
of E and has 2dimE−1 = |E| 12 . Thus, `(x) being the sum
P
2 elements x∈E (−1)
of 1’s and -1’s in equal numbers, it is null. 2
Proposition 6 For every pseudo-Boolean function ϕ on Fn2 and every el-
ements a, b and u of Fn2 , the value at u of the Fourier transform of the
function (−1)a·x ϕ(x + b) equals (−1)b·(a+u) ϕ(a
b + u).
Proof. ThePvalue at u of the Fourier transform of the function (−1)a·x ϕ(x +
b) equals x∈Fn (−1)(a+u)·x ϕ(x + b) = x∈Fn (−1)(a+u)·(x+b) ϕ(x) and thus
P
2 2
equals (−1)b·(a+u) ϕ(a
b + u). 2
Proposition 7 Let E be any vector subspace of Fn2 . Denote by 1E its indi-
cator (recall that it is the Boolean function defined by 1E (x) = 1 if x ∈ E
and 1E (x) = 0 otherwise). Then:
E = |E| 1E ⊥ ,
1c (16)
where E ⊥ = {x ∈ Fn2 / ∀y ∈ E, x · y = 0} is the orthogonal of E.
In particular, for E = Fn2 , we have b
1 = 2n δ0 .
12
Another way of seeing this is as follows: choose a ∈ E such that `(a) = 1; then the
mapping x 7→ x + a is one to one between `−1 (0) and `−1 (1).

24
Proof. For every u ∈ Fn2 , we have 1c u·x . If the linear
P
E (u) = x∈E (−1)
form x ∈ E 7→ u · x is not null on E (i.e. if u 6∈ E ⊥ ) then 1cE (u) is null,
⊥
according to Lemma 1. And if u ∈ E , then it clearly equals |E|. 2

We deduce from Proposition 7 the Poisson summation formula, which

has been used to prove many cryptographic properties in [243], [253], [54]
and later in [41, 42], and whose most general statement is:

Corollary 1 For every pseudo-Boolean function ϕ on Fn2 , for every vector

subspace E of Fn2 , and for every elements a and b of Fn2 , we have:
X X
(−1)b·u ϕ(u)
b = |E| (−1)a·b (−1)a·x ϕ(x). (17)
u∈a+E x∈b+E ⊥
P
Proof. Let us first assume that a = b = 0. The sum u∈E ϕ(u),
b by defini-
u·x
P P P
tion, equals u∈E x∈Fn ϕ(x)(−1) = x∈Fn ϕ(x) 1E (x). Hence, accord-
c
2 2
ing to Proposition 7:
X X
ϕ(u)
b = |E| ϕ(x). (18)
u∈E x∈E ⊥

We apply this last equality to the function (−1)a·x ϕ(x + b), whose Fourier
transform’s value at (−1)b·(a+u) ϕ(a
P u equalsb·(a+u) b + u),Paccording to Proposi-
tion 6. We deduce u∈E (−1) b + u) = |E| x∈E ⊥ (−1)a·x ϕ(x + b),
ϕ(a
which is equivalent to Equality (17). 2

Relation (17) with a = 0 and E = Fn2 gives:

Corollary 2 For every pseudo-Boolean function ϕ on Fn2 :
bb = 2n ϕ.
ϕ (19)

Thus, the Fourier transform is a permutation on the set of pseudo-Boolean

functions on Fn2 and is its own inverse, up to division by a constant. In order
to avoid this
√ division, the Fourier transform is often normalized, that is, di-
vided by 2n = 2n/2 so that it becomes its own inverse. We do not use this
normalized transform here because the functions we consider are integer-
valued, and we want their Fourier transforms to be also integer-valued.
Corollary 2 allows to show easily that some properties, valid for the Fourier
transform of any function ϕ having some specificities, are in fact necessary
and sufficient conditions for ϕ having these specificities. For instance, ac-
cording to Proposition 7, the Fourier transform of any constant function ϕ

25
takes null value at every nonzero vector; since the Fourier transform of a
function null at every nonzero vector is constant, Corollary 2 implies that
a function is constant if and only if its Fourier transform is null at every
nonzero vector. Similarly, ϕ is constant on Fn2 \ {0} if and only if ϕ b is
constant on Fn2 \ {0}.
A classical property of the Fourier transform is to be an isomorphism
from the set of pseudo-Boolean functions on Fn2 , endowed with the so-called
convolutional product (denoted by ⊗), into this same set, endowed with the
usual (Hadamard) product (denoted by ×). We recall the definition of the
convolutional product between two functions ϕ and ψ:
X
(ϕ ⊗ ψ)(x) = ϕ(y)ψ(x + y)
y∈Fn
2

(adding here is equivalent to substracting since the operations take place in

Fn2 ).

Proposition 8 Let ϕ and ψ be any pseudo-Boolean functions on Fn2 . We

have:
ϕ\⊗ ψ = ϕ×
b ψ.
b (20)
Consequently:
b ⊗ ψb = 2n ϕ×ψ.
ϕ [ (21)

Proof. We have
X X X
\
ϕ ⊗ ψ(u) = (ϕ ⊗ ψ)(x) (−1)u·x = ϕ(y)ψ(x + y) (−1)u·x
x∈Fn
2 x∈Fn n
2 y∈F2

X X
= ϕ(y)ψ(x + y) (−1)u·y⊕u·(x+y) .
x∈Fn
2 y∈Fn
2

Thus
 
X X
\
ϕ ⊗ ψ(u) = ϕ(y)(−1)u·y  ψ(x + y) (−1)u·(x+y) 
y∈Fn
2 x∈Fn
2
  
X X
= ϕ(y)(−1)u·y   ψ(x) (−1)u·x  = ϕ(u)
b ψ(u).
b
y∈Fn
2 x∈Fn
2

This proves the first equality. Applying it to ϕ

b and ψb in the places of ϕ and
\b
ψ, we obtain ϕb ⊗ ψ = 22n ϕ×ψ, according to Corollary 2. Using again this

26
same corollary, we deduce Relation (21). 2

Relation (21) applied at 0 gives

X
b ⊗ ψ(0)
ϕ b = 2n ϕ×ψ(0)
[ = 2n ϕ(x)ψ(x) = 2n ϕ ⊗ ψ(0). (22)
x∈Fn
2

Taking ψ = ϕ in (22), we obtain Parseval’s relation:

Corollary 3 For every pseudo-Boolean function ϕ, we have:

X X
b 2 (u) = 2n
ϕ ϕ2 (x).
u∈Fn
2 x∈Fn
2

If ϕ takes values ±1 only, this becomes:

X
b 2 (u) = 22n .
ϕ (23)
u∈Fn
2

This is why, when dealing with Boolean functions, we shall most often pre-
fer using the Walsh transform of f (that is, the Fourier transform of the
function fχ = (−1)f (x) ) instead of the Fourier transform of f .

Relation (20) leads to another relation involving the derivatives of a Boolean

function.

Definition 2 Let f be an n-variable Boolean function and let b be any vector

in Fn2 . We call derivative of f in the direction of b the Boolean function
Db f (x) = f (x) ⊕ f (x + b).

For instance, the derivative of a function of the form g(x1 , · · · , xn−1 ) ⊕

xn h(x1 , · · · , xn−1 ) in the direction of (0, · · · , 0, 1) equals h(x1 , · · · , xn−1 ).
Relation (20) applied with ψ = ϕ = fχ implies the so-called Wiener-
Khintchine Theorem:
b2
χ ⊗ fχ = fχ .
f\ (24)
We have (fχ ⊗ fχ )(b) = x∈Fn (−1)Db f (x) = F(Db f ) (the notation F was
P
2
defined at Relation (15)). Thus Relation (24) shows that fb 2 is the Fourier
χ
transform of the so-called auto-correlation function b 7→ ∆f (b) = F(Db f )
(this property was first used in the domain of cryptography in [53]):
X 2
∀u ∈ Fn2 , F(Db f )(−1)u·b = fbχ (u). (25)
b∈Fn
2

27
Applied at vector 0, this gives
X
F(Db f ) = F 2 (f ). (26)
b∈Fn
2

Corollary 1 and Relation (25) imply that, for every vector subspace E of Fn2
and every vectors a and b (cf. [42]):
X 2 X
(−1)b·u fbχ (u) = |E|(−1)a·b (−1)a·e F(De f ) . (27)
u∈a+E e∈b+E ⊥

Another interesting relation has been also shown in [42] (see also [250]):

Proposition 9 Let E and E 0 be subspaces of Fn2 such that E ∩ E 0 = {0}

and whose direct sum equals Fn2 . For every a ∈ E 0 , let ha be the restriction
of f to the coset a + E (ha can be identified with a function on Fk2 where k
is the dimension of E). Then
X 2 X
fbχ (u) = |E ⊥ | F 2 (ha ) . (28)
u∈E ⊥ a∈E 0

Proof. Every element of Fn2 can be written in a unique way in the form x +
a where x ∈ E and a ∈ E 0 . For P every e ∈ E, we have F(De f ) =
f (x+a)⊕f (x+e+a) =
P
x∈E;a∈E 0 (−1) a∈E 0 F(De ha ). We deduce from Rela-
⊥
tion (27), applied with E instead of E, and with a = b = 0, that
!
X 2 X X X
⊥ ⊥
f (u) = |E |
b
χ F(De f ) = |E | F(De ha )
u∈E ⊥ e∈E e∈E a∈E 0
!
X X
= |E ⊥ | F(De ha ) .
a∈E 0 e∈E

Thus, according to Relation (26) applied with E in the place of Fn2 (re-
call that E can be identified with Fk2 where k is the dimension of E):
P b2 ⊥
P 2 2
u∈E ⊥ fχ (u) = |E | a∈E 0 F (ha ).

Fourier transform and linear isomorphisms A last relation that must

be mentioned shows what the composition with a linear isomorphism implies
on the Fourier transform of a pseudo-Boolean function:

28
Proposition 10 Let ϕ be any pseudo-Boolean function on Fn2 . Let M be a
x1
 
 x2 
 ...  7→
nonsingular n×n binary matrix and L the linear isomorphism L :  

xn
x1
 
 x2  0 −1 and by L0 the
M ×
 ... . Let us denote by M the transpose of M


xn
x1 x1
   
x 2 x 
linear isomorphism L0 :   7→ M 0 ×  .2  (note that L0 is the adjoint
  
.
 ..   .. 
xn xn
operator of L−1 , that is, satisfies u · L−1 (x) = L0 (u) · x for every x and u).
Then
ϕ[◦L=ϕ b ◦ L0 . (29)

Proof. For every u ∈ Fn2 , we have ϕ[ ϕ(L(x))(−1)u·x =

P
◦ L(u) = x∈Fn
2
−1 (x) 0
u·L = x∈Fn ϕ(x)(−1)L (u)·x . 2
P P
x∈Fn ϕ(x)(−1)
2 2

A relationship between algebraic degree and Walsh transform was

shown in [229] (see also [54]):

Proposition 11 Let f be an n-variable Boolean function (n ≥ 2), and

let 1 ≤ k ≤ n. Assume that the Walsh transform of f takes values divisible
by 2k (i.e., according to Relation (12), that its Fourier transform takes values
divisible by 2k−1 , or equivalently, according to Relation (14), that all the
Hamming distances between f and affine functions are divisible by 2k−1 ).
Then f has algebraic degree at most n − k + 1.

Proof. Let us suppose that f has algebraic degree d > n − k + 1 and,

consider a term xI of degree d in its algebraic normal form. The Poisson
summation formula (18) applied to ϕ = fχ and to the vectorspace E = {u ∈
Fn2 / ∀i ∈ I, ui = 0} gives u∈E fbχ (u) = 2n−d x∈E ⊥ fχ (x). The orthogo-
P P
nal E ⊥ of E equals {u ∈ Fn2 / ∀i 6∈ I, ui P
= 0} = {u ∈ Fn2 / supp(u) ⊆ I}.
According to Proposition 2, we have that x∈E ⊥ f (x) is not even and there-
P P
fore x∈E ⊥ fχ (x) is not divisible by 4. Hence, u∈E fbχ (u) is not divisible
by 2n−d+2 and it is therefore not divisible by 2k . A contradiction. 2
The converse of Proposition 11 is obviously valid if k = 1. It is also valid

29
if k = 2, since the n-variable Boolean functions of degrees at most n − 1
are those Boolean functions of even Hamming weights, and f (x) ⊕ u · x
has degree at most n − 1 too for every u, since n ≥ 2. It is finally also
valid for k = n, since the affine functions are characterized by the fact that
their Walsh transforms take values ±2n and 0 only (more precisely, their
Walsh transforms take value ±2n once, and all their other values are null,
because of Parseval’s relation). The converse is false for any other value of
k. Indeed, we shall see below that it is false for k = n − 1 (n ≥ 4), since
there exist quadratic functions f whose Walsh transforms take values ±2n/2
for n even, ≥ 4, and ±2(n+1)/2 for n odd, ≥ 5. It is then an easy task to
deduce that the converse of Proposition 11 is also false for any value of k
such that 3 ≤ k ≤ n − 1: we choose a quadratic function g in 4 variables,
whose Walsh transform value at 0 equals 22 , that is, whose weight equals
23 − 2 = 6, and we take f (x) = g(x1 , x2 , x3 , x4 ) x5 · · · xl (5 ≤ l ≤ n). Such
function has algebraic degree l − 2 and its weight equals 6; hence its Walsh
transform value at 0 equals 2n − 12 and is therefore not divisible by 2k with
k = n − (l − 2) + 1 = n − l + 3 ≥ 3.
It is possible to characterize the functions whose Walsh transform values are
divisible by 2n−1 : they are the affine functions and the sums of an indicator
of a flat – an affine space – of co-dimension 2 and of an affine function (they
have degree at most 2 according to Proposition 11 and the characteriza-
tion follows from the results of subsection 5.2). Determining those Boolean
functions whose Walsh transform is divisible by 2k is an open problem for
3 ≤ k ≤ n − 2.
Note that it is possible to characterize the fact that a Boolean function
has degree at most d by means of its Fourier or Walsh transform: since a
Boolean function has algebraic degree at most d if and only if its restriction
to any (d + 1)-dimensional flat has an even weight, we can apply Poisson
summation formula (17).

Characterizing the Fourier transforms of integer-valued pseudo-

Boolean functions and of Boolean functions Obviously, according
to the inverse Fourier transform property (19), the Fourier transforms of
integer-valued functions (resp. the Walsh transforms of Boolean functions)
are those integer-valued functions over Fn2 whose Fourier transforms take
values divisible by 2n (resp. equal to ±2n ). Also, the Walsh transforms of
Boolean functions being those integer-valued functions ϕ over Fn2 such that
b2 equals the constant function 22n , they are those integer-valued functions
ϕ
ϕ such that ϕ \ ⊗ ϕ = 22n (according to Relation (20) applied with ψ = ϕ),

30
that is ϕ ⊗ ϕ = 22n δ0 . But these characterizations are not easy to use
mathematically and they are neither easily computable: they need to check
2n divisibilities by 2n for the Fourier transforms of integer-valued functions,
and 2n equalities for the Walsh transforms of Boolean functions.
Since the main cryptographic criteria on Boolean functions will be char-
acterized below as properties of their Walsh transforms, it is important to
have characterizations which are as simple as possible. We have seen that
characterizing the NNFs of integer-valued (resp. Boolean) functions is easy
(resp. easier than with Fourier transform). So it is useful to clarify the
relationship between these two representations.

2.2.1 Fourier transform and NNF

There is a similarity between the Fourier transform and the NNF:
- the functions (−1)u·x , u ∈ Fn2 , constitute an orthogonal basis of the space
of pseudo-Boolean functions, and the Fourier transform corresponds, up to
normalization, to a decomposition over this basis;
- the NNF is defined similarly with respect to the (non-orthogonal) basis of
monomials.
Let us see now how each representation can be expressed by means of the
other.
I
P
Let ϕ(x) be any pseudo-Boolean function and let I∈P(N ) λI x be its NNF.
For every word x ∈ Fn2 , we have: ϕ(x) =
P
I⊆supp(x) λI . Setting b =
X
(1, · · · , 1), we have ϕ(x+b) = λI (since the support of x+b
I∈P(N )/ supp(x)∩I=∅
equals Fn2 \ supp(x)).
For every I ∈ P(N ), the set {x ∈ Fn2 / supp(x) ∩ I = ∅} is an (n − |I|)-
dimensional vector subspace of Fn2 . Let us denote it
Pby EI . Its orthogonal
equals {u ∈ Fn2 / supp(u) ⊆ I}. We have ϕ(x + b) = I∈P(N ) λI 1EI . Apply-
ing Propositions 6 (with a = 0) and 7, we deduce:
X
ϕ(u)
b = (−1)wH (u) 2n−|I| λI . (30)
I∈P(N ) | supp(u)⊆I

Using the same method as for computing λI by means of the values of f , it

is an easy task to deduce:
X
λI = 2−n (−2)|I| ϕ(u).
b (31)
u∈Fn
2 | I⊆supp(u)

Note that if ϕ has numerical degree at most D, then, according to Rela-

tion (30), we have ϕ(u)
b = 0 for every vector u of weight strictly greater

31
than D and that the converse is true, according P to Relation (31)
Applying Relation (30) to ϕ(x) = P (x) = I∈P(N ) λI xI and to ϕ(x) =
P 
P 2 (x) = I∈P(N ) xI , with u = 0, we deduce
P
0
J,J ∈P(N ) | I=J∪J 0 λ J λJ 0

from Proposition 5 that a polynomial P (x) = I∈P(N ) λI xI , with integer

P
coefficients, is the NNF of a Boolean function if and only if
X X X
2n−|I| λJ λJ 0 = 2n−|I| λI . (32)
I∈P(N ) J,J 0 ∈P(N ) | I=J∪J 0 I∈P(N )

Remark. The NNF presents the interest of being a polynomial represen-

tation, but it can also be viewed as the transform which maps any pseudo-
Boolean function f (x) = I∈P(N ) λI xI to the pseudo-Boolean function g
P
defined by g(x) = λsupp(x) . Let us denote this mapping by Φ. Three other
transforms have also been used for studying Boolean functions:
- the mapping Φ−1 (the formulae relating this mapping and the Walsh trans-
form are slightly simpler than for Φ; see [306]);
- a mapping defined by a formula similar to Relation (8), but in which
supp(x) ⊆ I is replaced by I ⊆ supp(x); see [171];
- the inverse of this mapping. 2

2.2.2 The size of the support of the Fourier transform and its
relationship with Cayley graphs
Let f be a Boolean function and let Gf be the Cayley graph associated
to f : the vertices of this graph are the elements of Fn2 and there is an edge
between two vertices u and v if and only if the vector u + v belongs to the
support of f . Then (see [18]), if we multiply by 2n the values fb(a), a ∈ Fn2 ,
of the Fourier spectrum of f , we obtain the eigenvalues of the graph Gf
(that is, by definition, the eigenvalues of the adjency matrix (Mu,v )u,v∈Fn2
of Gf , whose term Mu,v equals 1 if u + v belongs to the support of f , and
equals 0 otherwise).
As a consequence, the cardinality Nfb of the support {a ∈ Fn2 / fb(a) 6= 0} of
the Fourier transform of any n-variable Boolean function f is greater than
or equal to the cardinality Ngb of the support of the Fourier transform of
any restriction g of f , obtained by keeping constant some of its input bits.
Indeed, the adjency matrix Mg of the Cayley graph Gg is a submatrix of
the adjency matrix Mf of the Cayley graph Gf ; the number Ngb equals the
rank of Mg , and is then smaller than or equal to the rank Nfb of Mf .
This property can be generalized to any pseudo-Boolean function ϕ. More-
over, a simpler proof is obtained by using the Poisson summation for-

32
mula (17): let I be any subset of N = {1, · · · , n}; let E be the vector sub-
space of Fn2 equal to {x ∈ Fn2 / xi = 0, ∀i ∈ I}; we have E ⊥ = {x ∈ Fn2 / xi =
0, ∀i ∈ N \ I} and the sum ofP E and of E ⊥ is direct; then, for every a ∈ E ⊥
and every b ∈ E, the equality u∈a+E (−1)b·u ϕ(u) b = |E| (−1)a·b ψ(a),
b where
⊥
ψ is the restriction of ϕ to b + E , implies that, if Nfb = k, that is, if ϕ(u)
b
is nonzero for exactly k vectors u ∈ Fn , then clearly ψ(a)
2
b is nonzero for at
most k vectors a ∈ E ⊥ .
If ϕ is chosen to be a Boolean function of algebraic degree d and if we choose
for I a multi-index of size d such that xI is part of the ANF of ϕ, then the re-
striction ψ has odd weight and its Fourier transform takes therefore nonzero
values only. We deduce (as proved in [18]) that Nϕb ≥ 2d . Notice that Nϕb
equals 2d if and only if at most one element (that is, exactly one) satisfying
ϕ(u)
b 6= 0 exists in each coset of E, that is, in each set obtained by keeping
constant the coordinates xi such that i ∈ I.
The number Nϕb is also bounded above by D n
P

i=0 i , where D is the numer-
ical degree of ϕ. This is a direct consequence of Relation (30) and of the
observation which follows Relation (31).
The graph viewpoint also gives insight on the Boolean functions whose
Fourier spectra have at most three values (see [18]).
A hypergraph can also be related to the ANF of a Boolean function f . A
related (weak) upper bound on the nonlinearity of Boolean functions (see
definition in Subsection 4.1) has been pointed out in [364].

3 Boolean functions and coding

We explained in the introduction how, in error correcting coding, the mes-
sage is divided into vectors of the same length k, which are transformed into
codewords of length N > k, before being sent over a noisy channel, in order
to enable the correction of the errors of transmission (or of storage, in the
case of CD, CD-ROM and DVD) at their reception. A choice of the set of
all possible codewords (called the code – let us denote it by C) allows to
correct up to t errors (in the transmission of each codeword) if and only if
the Hamming distance between any two different codewords is greater than
or equal to 2t + 1 (so, if d is the minimum distance between two codewords,
the code can enable to correct up to d−1

2 errors, where “b c” denotes the
integer part). Indeed, the only information the receiver has, concerning the
sent word, is that it belongs to C. In order to be always able to recover the
correct codeword, he needs that, for every word y at distance at most t from
a codeword x, there does not exist another codeword x0 at distance at most t

33
from y, and this is equivalent to saying that the Hamming distance between
any two different codewords is greater than or equal to 2t+1. This necessary
condition is also sufficient13 . Thus, the problem of generating a good code
consists in finding a set C of binary words of the same length whose mini-
mum distance mina6=b∈C dH (a, b) (where dH (a, b) = |{i/ ai 6= bi }|) is high14 .

A code is called a linear code if it has the structure of a linear subspace of

FN2 where N is its length. The minimum distance of a linear code equals the
minimum Hamming weight of all nonzero codewords, since the Hamming
distance between two vectors equals the Hamming weight of their difference
(i.e. their sum since we reduce ourselves here to binary vectors). We shall
write that a linear code is an [N, k, d]-code if it has length N , dimension k
and minimum distance d. It can then be described by a generator matrix G,
obtained by choosing a basis of this vectorspace and writing its elements as
the rows of this matrix. The code equals the set of all the vectors of the form
u × G, where u ranges over Fk2 (and × is the matrix product) and a possible
encoding algorithm is therefore the mapping u ∈ Fk2 7→ u × G ∈ FN 2 . The
generator matrix is well suited for generating the codewords, but it is not for
checking if a received word of length N is a codeword or not. A characteri-
zation of the codewords is obtained thanks toLthe generator matrix H of the
N
dual code C ⊥ = {x ∈ FN 2 / ∀y ∈ C, x · y = i=1 xi yi = 0} (such a matrix
is called a parity-check matrix ): we have x ∈ C if and only if x × H t is the
null vector. It is a simple matter to prove that the minimum distance of the
code equals the minimum number of linearly dependent columns of H. For
instance, the Hamming code, which has by definition for parity-check matrix
the n × (2n − 1) matrix whose columns are all the non-zero vectors of Fn2
in some order, has minimum distance 3. This code depends, stricto sensu,
on the choice of the order, but we say that two binary codes are equivalent
codes if they are equal, up to some permutation of the coordinates of their
codewords.

We shall use in the sequel the notion of covering radius of a code: it

is the smallest integer ρ such that the spheres of radius ρ centered at the
13
In practice, we still need to have an efficient decoding algorithm to recover the sent
codeword; the naive method consisting in visiting all codewords and keeping the nearest
one from the received word is inefficient because the number 2k of codewords is too large,
in general.
14
High with respect to some known bounds giving the necessary trade-offs between
the length of the code, the minimum distance between codewords and the number of
codewords, see [258, 298])

34
codewords cover the whole space, i.e. the minimum integer t such that every
binary word of length N lies at Hamming distance at most t from at least one
codeword, i.e. the maximum multiplicity of errors that have to be corrected
when maximum likelihood decoding is used on a binary symmetric channel.
The covering radius of a code is an important parameter [111], which can be
used for analyzing and improving the decoding algorithms devoted to this
code.

A linear code C is a cyclic code if it is invariant under cyclic shifts of the

coordinates (see [258]). Cyclic codes have been extensively studied in coding
theory. They have useful properties, that we briefly recall: representing each
codeword (c0 , · · · , cN −1 ) by the polynomial c0 + c1 X + · · · + cN −1 X N −1 , we
obtain an ideal of the quotient algebra F2 [X]/(X N + 1) (viewed as a set of
polynomials of degrees at most N − 1, each element of the algebra being
identified to its minimum degree representent). This algebra is a principal
domain, and any (linear) cyclic code has a unique element having minimal
degree, called its generator polynomial . To simplify the presentation, we
shall assume now that N = 2n − 1 (which will be the case in the sequel).
n
The generator polynomial being (as easily shown) a divisor of X 2 −1 + 1,
its roots all belong to F∗2n . The code equals the set of all those polynomials
which include the roots of the generator polynomial among their own roots.
The generator polynomial having all its coefficients in F2 , its roots are of
the form {αi , i ∈ I} where I ⊆ Z/(2n − 1)Z is a union of cyclotomic classes
of 2 modulo 2n − 1. The set I is called the defining set of the code. The
elements αi , i ∈ I are called the zeroes of the code, which has dimension
N − |I|. The generator polynomial of C ⊥ is the reciprocal of the quotient of
n
X 2 −1 + 1 by the generator polynomial of C, and its defining set therefore
equals {2n − 1 − i; i ∈ Z/(2n − 1)Z \ I}.
A very efficient bound on the minimum distance of cyclic codes is the BCH
bound [258]: if I contains a string {l+1, · · · , l+k} of length k in Z/(2n −1)Z,
then the cyclic code has minimum distance greater than or equal to k + 1.
A proof of this bound (in the framework of Boolean functions) is given in
the proof of Theorem 15. This bound is valid for cyclic codes over any finite
field as well. When the length of such a cyclic code equals the order of the
underlying field less 1, the set of zeros can be any set of nonzero elements of
the field; when it is constituded of consecutive powers of a primitive element,
the code is called a Reed-Solomon code.
A cyclic code C of length N being given, the extended code of C is the set
of vectors (c−∞ , c0 , · · · , cN −1 ), where c−∞ = c0 ⊕ · · · ⊕ cN −1 . It is a linear
code of length N + 1 and of the same dimension as C.

35
 Cyclic codes over F2 can also be considered in terms of the trace function
and therefore viewed as sets of Boolean functions (when their length is 2n −1,
recall we assume this). Any codeword of a cyclic code with non-zeroes αi
for
Pl i in the cyclotomic classes containing u1 , · · · , ul can be represented as
tr (a x −ui ), a ∈ F n .
i=1 n i i 2

3.1 Reed-Muller codes

As explained in the introduction, every code whose length equals 2n , for
some positive integer n, can be interpreted as a set of Boolean functions.
The existence of Reed-Muller codes comes from the following observation:

Theorem 1 Any two distinct n-variable functions f and g of algebraic de-

grees at most r have mutual distances at least 2n−r .

Proof. In order to prove this property, it is necessary and sufficient to show

that any nonzero Boolean function f of algebraic degree d ≤ r has weight at
least 2n−r (since the difference between two Boolean functions of algebraic
degrees at most r has algebraic degree at most r). This can be proved by a
double
Q induction over r and n (see [258]), but there exists a simpler proof.
Let i∈I xi be a monomial of degree d in the ANF of f ; consider the 2n−d
restrictions of f obtained by keeping constant the n − d coordinates of x
whose indices lie outside I. Each of these restrictions, viewed as a function
on Fd2 , has an ANF of Q degree d because, when fixing these n − d coordi-
nates,Qthe monomial i∈I xi is unchanged and all the monomials different
from i∈I xi in the ANF of f give monomials of degrees strictly less than
d. Thus any such restriction has an odd (and hence a nonzero) weight (see
Subsection 2.1). The weight of f being equal to the sum of the weights of
its restrictions, f has weight at least 2n−d , which completes the proof. 2

The functions of Hamming weight 2n−r and degree r have been charac-
terized, see a proof in [258]. We give below a proof which brings a little
more insight on the reasons of this characterization.

Proposition 12 The Boolean functions of algebraic degree r and of Ham-

ming weight 2n−r are the indicators of (n − r)-dimensional flats (i.e. the
functions whose supports are (n − r)-dimensional affine subspaces of Fn2 ).

Proof. The indicators of (n − r)-dimensional flats have clearly Hamming

weight 2n−r and they have degree r, since every (n − r)-dimensional flat
equals {x ∈ Fn2 / `i (x) = 1, ∀i = 1, · · · , r} where the `i ’s are affine and

36
have
Qr linearly independent linear parts, and the ANF of its indicator equals
i=1 `i (x). Conversely, Qlet f be a function of algebraic degree r and of Ham-
ming weight 2n−r . Let i∈I xi be a monomial of degree r in the ANF of f
and let J = {1, · · · , n} \ I. For every vector α ∈ FJ2 , let us denote by fα the
restriction of f to the flat {x ∈ Fn2 ; ∀j ∈ J, xj = αj }. According to the proof
of Theorem 1, and since f has Hamming weight 2n−r , each function fα is
the indicator of a singleton {aα }. Let us prove that the mapping a : α → aα
is affine, i.e. that, for every α, β, γ ∈ FJ2 , we have aα+β+γ = aα + aβ + aγ
(this will complete the proof of the proposition since, denoting by xJ the
vector of FJ2 whose coordinates match the corresponding coordinates of x,
the support of f equals the set {x ∈ Fn2 / xI = axJ } and that the equality
xI = axJ is equivalent to r linearly independent linear equations). Proving
this is equivalent to proving that the function of Hamming weight at most
4 equal to fα+β+γ ⊕ fα ⊕ fβ ⊕ fγ has algebraic degree at most r − L2. But
more generally, for every k-dimensional flat A of FJ2 , the function α∈A fα
has degree at most r − k (this can be easily proved by induction on k, using
that f has degree r). 2

Remark. Q
1. The proof of Theorem 1 shows in fact that, if a monomial Q i∈I xi has
coefficient 1 in the ANF of f , and if every other monomial i∈J xi such that
I ⊂ J has coefficient 0, then the function has weight at least 2n−|I| . Apply-
ing this observation to the Möbius transform f ◦ of f - whose definition has
been given after Relation (2) - shows that, if there exists a vector x ∈ Fn2
such that f (x) = 1 and f (y) = 0 for every vector y 6= x whose support
contains supp(x), then the ANF of f has at least 2n−wH (x) terms (this has
been first observed in [364]). Indeed, the Möbius transform of f ◦ is f .
2. The d-dimensional subspace E = {x ∈ Fn2 / xi = 0, ∀i 6∈ I}, in the proof
of Theorem 1, is a maximal odd weighting subspace: the restriction of f to
E has odd weight, and the restriction of f to any of its proper superspaces
has even weight (i.e. the restriction of f to any coset of E has odd weight).
Similarly as above, it can be proved, see [364], that any Boolean function
admitting a d-dimensional maximal odd weighting subspace E has weight
at least 2n−d .

The Reed-Muller code of order r is by definition the set of all Boolean

functions of algebraic degrees at most r (or more precisely the set of the bi-
nary words of length 2n corresponding to (last columns of) the truth-tables
of these functions). Denoted by R(r, n), it is an F2 -vectorspace of dimen-

37
 n n



sion 1+n+ 2 +· · ·+
(since this is the number of monomials of degrees at
r
n n
most r, which constitute a basis of R(r, n)) and thus, it has 21+n+( 2 )+···+( r )
elements.
For r = 1, it equals the set of all affine functions. Notice that the weight of
any non-constant affine function being equal to the size of an affine hyper-
plane, it equals 2n−1 .

Historic note: the Reed-Muller code R(1, 5) was used in 1972 for trans-
mitting the first black-and-white photographs of Mars. It has 26 = 64 words
of length 25 = 32, with mutual distances at least 24 = 16. Each codeword
corresponded to a level of darkness (this made 64 different levels). Up to
= 7 errors could be corrected in the transmission of each codeword. 2
 16−1 
2

R(r, n) is a linear code, i.e. an F2 -vectorspace. Thus, it can be described by

a generator matrix G. For instance, a generator matrix of the Reed-Muller
code R(1, 4) is:
 
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
 0 1 0 0 0 1 1 1 0 0 0 1 1 1 0 1 
 
G=  0 0 1 0 0 1 0 0 1 1 0 1 1 0 1 1 

 0 0 0 1 0 0 1 0 1 0 1 1 0 1 1 1 
0 0 0 0 1 0 0 1 0 1 1 0 1 1 1 1

(the first row corresponds to the constant function 1 and the other rows
correspond to the coordinate functions x1 , · · · , x4 )15 .
The duals of Reed-Muller codes are Reed-Muller codes:
Theorem 2 The dual
M
R(r, n)⊥ = {f ∈ BF n / ∀g ∈ R(r, n), f · g = f (x) g(x) = 0}
x∈Fn
2

equals R(n − r − 1, n).

Proof. We have seen in Subsection 2.1 that the n-variable Boolean functions
of even weights are the elements of R(n − 1, n). Thus, R(r, n)⊥ is the set of
those functions f such that, for every function g of algebraic degree at most
15
We have chosen to order the words of length 4 by increasing weights; we could have
chosen other orderings; this would have led to other codes, but equivalent ones, having
the same parameters (a binary code C of length N is said to be equivalent to another
binary code C 0 of the same length if there exists a permutation σ on {1, · · · , N } such that
C = {(xσ(1) , · · · , xσ(N ) )/ x ∈ C 0 ).

38
r, the product function f g (whose value at any x ∈ Fn2 equals f (x)g(x)) has
algebraic degree at most n − 1. This is clearly equivalent to the fact that f
has algebraic degree at most n − r − 1. 2
n j
If the vector-space F2 is identified with the field F2n , the functions trn (ax )
such that w2 (j) ≤ n − r − 1 being a generating family of R(n − r − 1, n)
(according to what we have seen on the trace representation of Boolean
functions), we have that a Boolean function f belongsP to R(r, n) if and only
if, for every j such that w2 (j) ≤ n − r − 1, we have x∈F2n f (x) trn (axj ) = 0
for every a ∈ F2n , that is, x∈F2n f (x) xj = 0.
P

The Reed-Muller codes are invariant under the action of the general
affine group. More precisely, it is a simple matter to show that, for any
1 ≤ r ≤ n − 1, the automorphism group of R(r, n) (that is, the group of
all permutations σ of Fn2 such that f ◦ σ ∈ R(r, n) for every f ∈ R(r, n))
equals the general affine group. The sets R(r, n) or R(r, n)/R(r0 , n) have
been classified under this action for some values of r, of r0 < r and of n,
see [183, 185, 30, 259, 341, 342].

The Reed-Muller code R(r, n) is an extended cyclic code for every r <
n (see [258]): the zeroes of the corresponding cyclic code (R∗ (r, n), the
punctured Reed-Muller code of order r) are the elements αi such that 1 ≤ i ≤
2n −2 and such that the 2-weight of i is at most equal to n−r−1. Indeed, the
n
codewords of R∗ (r, n) are the vectors of the form (g(1), g(α), · · · , g(α2 −2 ))
where g is a Boolean function ofPalgebraic degree at most r; such function
has univariate polynomial form 0≤j≤2n −2 gj xj and we have
w2 (j)≤r

 
X X X
g(αl )αli = gj  αl(i+j) 
0≤l≤2n −2 0≤j≤2n −2 0≤l≤2n −2
w2 (j)≤r

and 0≤l≤2n −2 αl(i+j) equals 0 when w2 (i) ≤ n − r − 1 and w2 (j) ≤ r since

P
i + j cannot be null and it cannot equal 2n − 1 either since w2 (i + j) ≤
w2 (i) + w2 (j). Hence, the αi ’s such that 1 ≤ i ≤ 2n − 2 and w2 (i) ≤ n − r − 1
are zeroes of the code. Since their number equals the co-dimension of the
code, they are the only zeroes of the code.

The problem of determining the weight distributions of the

Reed-Muller codes, the MacWilliams identity and the notion of
dual distance: What are the possible distances between the words of R(r, n),

39
or equivalently the possible weights in R(r, n) (or better, the weight distri-
bution of R(r, n))? The answer, which is useful for improving the efficiency
of the decoding algorithms and for evaluating their complexities, is known
for every n if r ≤ 2: see Subsection 5.2. For r ≥ n − 3, it can also be
deduced from the very nice relationship, due to F. J. MacWilliams, existing
between every linear code and its dual: let C beP any binary linear code of
length N ; consider the polynomial WC (X, Y ) = N i=0 Ai X
N −i Y i where A
i
is the number of codewords of weight i. This polynomial is called the weight
enumerator of C and describes16 the weight distribution (Ai )0≤i≤N of C.
Then (see [258, 298])

WC (X + Y, X − Y ) = |C| WC (X, Y ). (33)

We give a sketch of Pproof Qof this1−xMacWilliams’ identity: we observe first

that WC (X, Y ) = x∈C N i=1 X i Y xi ; we deduce W (X + Y, X − Y ) =
C
P QN xi Y ); applying a classical method of expansion, we
x∈C i=1 (X + (−1)
derive WC (X + Y, X − Y ) = x∈C b∈FN N 1−bi ((−1)xi Y )bi (choos-
P P Q

2 i=1 X
ing X in the i-th factor X + (−1)xi Y for bi = 0 and (−1)xi Y for bi =
1; all the different possible choices are taken into account by consider-
ing allPbinary words b of length
N −w (b) w (b)
P N ). We b·x

obtain then WC (X + Y, X −
Y ) = b∈FN X H Y H
x∈C (−1) and we conclude by using Re-
2
lation (16) with E = C.
The MacWilliams identity allows, theoretically, to deduce the weight distri-
bution of R(n − r − 1, n) from the weight distribution of R(r, n) (in fact, to
actually determine this weight distribution, it is necessary to be able to ex-
plicitely expand the factors (X +Y )N −i (X −Y )i and to simplify the obtained
expression for WC (X + Y, X − Y ); this is possible by running a computer up
to some value of n). But this gives no information for the cases 3 ≤ r ≤ n−4
which remain unsolved (except for small values of n, see [17], and for n = 2r,
because the code is then self-dual, see [258, 298]). McEliece’s theorem [272]
(or Ax’s theorem [12]; see also the Stickelberger theorem, e.g. in [232, 236])
shows that the weights (and thus the distances) in R(r, n) are all divisi-
n−1
ble by 2d r e−1 = 2b r c , where due denotes the ceiling - the smallest integer
n

greater than or equal to u - and buc denotes the integer part (this can also be
shown by using the properties of the NNF, see [87]). Moreover,l if f mhas de-
h n−d0 i
gree d and g has degree d0 ≤ d, then dH (f, g) ≡ wH (f ) mod 2 d [209]
(see also [195]). In [36], A. Canteaut gives further properties of the weights
16
WC is a homogeneous version of the classical generating series for the weight distri-
bution of C.

40
in f ⊕ R(1, n). Kasami and Tokura [207] have shown that the only weights
in R(r, n) occuring in the range [2n−r ; 2n−r+1 [ are of the form 2n−r+1 −2i for
some i; and they have completely characterized the codewords with these
weights (and computed their number). The functions whose weights are
between the minimum distance 2n−r and 2.5 times the minimum distance
have also been characterized, in [208].
The principle of MacWilliams’ identity can also be applied to nonlinear
codes. When C is not linear, the weight distribution of C has no great
relevance. The distance distribution has morePinterest. We consider the
1 N N −i Y i , where B
distance enumerator of C: DC (X, Y ) = |C| i=0 Bi X i
is the size of the set {(x, y) ∈ C 2 / dH (x, y) = i}. Note that, if C is
linear, then D QCN = 1−(x
WC . Similarly as above, we see that DC (X, Y ) =
1 P
X i ⊕yi ) Y xi ⊕yi ; we deduce that the polynomial D (X+
|C| (x,y)∈C 2 i=1 C
1 P QN x ⊕y
Y, X − Y ) equals |C| (x,y)∈C 2 i=1 (X + (−1) i i Y ). Expanding these
QN
1 P
X 1−bi ((−1)xi ⊕yi Y )bi , that
P

products, we obtain |C| (x,y)∈C 2 b∈FN i=1
2
is
!2
1 X N −wH (b) wH (b) X b·x
DC (X + Y, X − Y ) = X Y (−1) (34)
|C|
b∈F2 N x∈C

Hence, DC (X + Y, X − Y ) has non-negative coefficients.

The minimum exponent of Y with nonzero coefficient in P the polynomial
DC (X + Y, X − Y ), that is, the number min{wH (b); b 6= 0, x∈C (−1)b·x 6=
0}, is usually denoted by d⊥ and is called theP
dual distance of C. Note that
the maximum number j such that the sum x∈C (−1)b·x is null, for every
nonzero vector b of weight at most j, equals d⊥ − 1 (see more in [129, 130]).
This property will be useful in Subsection 4.1.

It is shown in [52] (see the remark of Subsection 5.2 in the present chap-
ter) that for every Boolean function f on Fn2 , there exists an integer m and
a Boolean function g of algebraic degree at most 3 on Fn+2m2 whose Walsh
transform satisfies: gc (0) = 2 m fb (0). This means that the weight of f is
χ χ
related to the weight of a function of degree at most 3 (but in a number of
variables which can be exponentially larger) in a simple way. This shows
that the distances in R(3, n) can be very diverse, contrary to those in R(2, n).
2

41
4 Boolean functions and cryptography
Stream ciphers are based on the so-called Vernam cipher (see Figure 1) in
which the plaintext (a binary string of some length) is bitwise added to a
(binary) secret key of the same length, in order to produce the ciphertext.
The Vernam cipher is also called the one time pad because a new random
secret key must be used for every encryption. Indeed, the bitwise addition
of two ciphertexts corresponding to the same key equals the addition of the
corresponding plaintexts, which gives much information on these plaintexts
(it is often enough to recover both plaintexts; some secret services and spies
learned this at their own expenses!).

Key Key

Plaintext ? Ciphertext Ciphertext ? Plaintext

-⊕ - ··· -⊕ -

Figure 1: Vernam cipher

The Vernam cipher, which is the only known cipher offering uncondi-
tional security (see [332]) if the key is truly random and if it is changed for
every new encryption, was used for the communication between the heads of
USA and USSR during the cold war (the keys being carried by diplomats)
and by some secret services.
In practice, since the length of the private key must be equal to the
length of the plaintext, pseudo-random generators are most often used in
order to minimize the size of the private key (but the unconditional security
is then no longer ensured): a method is chosen for producing long pseudo-
random sequences from short random secret keys (only the latter are actually
shared; the method is supposed to be public; according to the Kerckhoffs
principle, only the parameters which can be used by the attacker to break
the system must be kept secret). The pseudo-random sequence is used in
the place of the key in a Vernam cipher. For this reason, it is also called
the keystream. If the keystream only depends on the key (and not on the
plaintext), the cipher is called synchronous 17 . Stream ciphers, because they
operate on data units as small as a bit or a few bits, are suitable for fast
17
There also exist self-synchronous stream ciphers, in which each keystream bit depends

42
telecommunication applications. Having also a very simple construction,
they are easily implemented both in hardware and software.
The first method for generating a pseudo-random sequence from a secret
key has used Linear Feedback Shift Registers (LFSR). In such an LFSR

L L L

×c1 ×cL−1 ×cL

sn 6 6 6

- sn−1 ··· sn−L+1 sn−L -

Figure 2: LFSR

(see Figure 2, where × means multiplication), at every clock-cycle, the bits

sn−1 , · · · , sn−L contained in the flip-flops of the LFSR move to the right.
The right-most bit is the current
LL output and the left-most flip-flop is feeded
with the linear combination i=1 ci sn−i , where the ci ’s are bits. Thus, such
an LFSR outputs a recurrent sequence satisfying the relation
L
M
sn = ci sn−i .
i=1

Such sequence is always ultimately periodic18 (if cL = 1, then it is periodic;

we shall assume that cL = 1 in the sequel, because otherwise, the same
sequence can be output by an LFSR of a shorter length, except for its first
bits, and this can be exploited
L in attacks) with period at most 2L − 1. The
i
generating series s(X) = i≥0 si X of the sequence can be expressed in a
nice way (see the chapter by Helleseth and Kumar in [298]): s(X) = G(X)F (X) ,
LL−1 i Li 
where G(X) = i=0 X j=0 ci−j sj is a polynomial of degree smaller
than L and F (X) = 1 ⊕ c1 X ⊕ · · · ⊕ cL X L is the feedback polynomial .
The short secret key contains the initialization s0 , · · · , sL−1 of the LFSR and
on the n preceding ciphertext bits, which allows re-synchronising after n bits if an error
of transmission occurs between Alice and Bob
18
Conversely, every ultimately periodic sequence can be generated by an LFSR.

43
the values of the feedback coefficients ci (these must be kept secret; otherwise,
the observation of L consecutive bits of the key would allow recovering all
the subsequent sequence).
But these LFSRs are cryptographically weak because of the Berlekamp-
Massey algorithm [269]: let L be the length of a minimum length LFSR
producing the same sequence (this length, called the linear complexity of
the sequence, is assumed to be unknown from the attacker; note that it
equals L if and only if the polynomials F and G above are co-prime), then
if we know at least 2L consecutive bits, Berlekamp-Massey algorithm recov-
ers the values of L and of the feedback coefficients of an LFSR of length
L generating the sequence, and the initialization of this LFSR in O(L2 )
elementary operations. A modern way of avoiding this attack is by using
Boolean functions. The first model which appeared in the litterature for
using Boolean functions is the combiner model (see Figure 3).

x1
LFSR 1
@
@
x2 R
@
LFSR 2 - output si
f -
..
. 
xn
LFSR n

Figure 3: Combiner model

Notice that the feedback coefficients of the n LFSRs used in such a generator
can be public. The Boolean function is also public, in general, and the short
secret key gives only the initialization of the n LFSRs: if we want to use
for instance a 128 bit long secret key, this allows using n LFSRs of lengths
L1 , · · · , Ln such that L1 + · · · + Ln = 128.
Such system clearly outputs a periodic sequence whose period is at most
the LCM of the periods of the sequences output by the n LFSRs (assuming
that cL = 1 in each LFSR; otherwise, the sequence is ultimately periodic).
So, this sequence is also recurrent and can therefore be produced by a single
LFSR. However, as we shall see, well-chosen Boolean functions allow the
linear complexity of the sequence to be much larger than the sum of the
lengths of the n LFSRs. Nevertheless, choosing LFSRs producing sequences

44
of large periods19 , choosing these periods pairwise co-prime so that to have
then the largest possible global period, and choosing f such that the lin-
ear complexity is large enough too are not sufficient. As we shall see, the
combining function should also not leak information about the individual
LFSRs and behave as differently as possible from affine functions, in several
different ways.
The combiner model is only a model, useful for studying attacks and related
criteria. In practice, the systems are more complex (see for instance at URL
http://www.ecrypt.eu.org/stream/ how are designed the stream ciphers of
the eSTREAM Project).
An alternative model is the filter model , which uses a single LFSR (of a
longer length). A filtered LFSR outputs f (x1 , · · · , xn ) where f is some n-
variable Boolean function, called a filtering function, and where x1 , · · · , xn
are the bits contained in some flip-flops of the LFSR, see Figure 4.

L L L
6 6 6
- si+L−1 ··· si+1 si
x1 xi xn
? ? ?
f (x1 , x2 , · · · , xn )
output
?

Figure 4: Filter model

Such system is equivalent to the combiner model using n copies of the

LFSR. However, the attacks, even when they apply to both systems, do not
work similarly (a first obvious difference is that the lengths of the LFSRs
are different in the two models). Consequently, the criteria that the involved
Boolean functions must satisfy because of these attacks may be different for
19
e.g. m-sequences also called maximum length sequences, that is, sequences of period
2L − 1 where L is the linear complexity – assuming that L = L, this corresponds to taking
a primitive feedback polynomial – which can be represented in the form si = trn (aαi )
where α is a primitive element of F2n , and which have very strong properties; see the
chapter by Helleseth and Kumar in [298].

45
the two models and we shall have to distinguish between the two models
when describing the attacks and the related criteria.

Other pseudo-random generators exist. A Feedback Shift Register has

the same structure as an LFSR, but the left-most flip-flop is feeded with
f (xi1 , · · · , xin ) where n ≤ L and xi1 , · · · , xin are bits contained in the flip-
flops of the FSR, and where f is some n-variable Boolean function. The
linear complexity of the produced sequence can then be near 2L , see [199]
for general FSRs and [98] for FSRs with quadratic feedback function f . The
linear complexity is difficult to study in general. Nice results similar to
those on the m-sequences exist in the case of FCSRs (Feedback with Carry
Shift-Registers), see [218, 167, 8, 168].

Boolean functions also play an important role in block ciphers. Every

block cipher admits as input a binary vector (x1 , · · · , xn ) (a block of plain-
text) and outputs a binary vector (y1 , · · · , ym ); the coordinates y1 , · · · , ym
are the outputs to Boolean functions (depending on the key) whose common
input is (x1 , · · · , xn ), see Figure 5.

Plaintext: x1 xn
···
? ?
Key
- E

···
? ?
Ciphertext: y1 ym

Figure 5: Block cipher

But the number n of variables of these Boolean functions being large (most
often, more than a hundred), these functions can not be analyzed. Boolean
functions on fewer variables are in fact involved in the ciphers. All known
block ciphers are the iterations of a number of rounds.
We give in Figures 6 and 7 a description of the rounds of the DES and of
the AES. The input to a DES round is a binary string of length 64, divided
into two strings of 32 bits each (in the figure, they enter the round, from
above, on the left and on the right); confusion (see below what this term

46
 Round key

+l +l
? ?
 P  S   E  ?

? ?

Figure 6: A DES round

means) is achieved by the S-box, which is a nonlinear transformation of a

binary string of 48 bits20 into a 32 bit long one. So, 32 Boolean functions on
48 variables are involved. But, in fact, this nonlinear transformation is the
concatenation of eight sub-S-boxes, which transform binary strings of 6 bits
into 4 bit long ones. So, 32 (that is, 8 × 4) Boolean functions on 6 variables
are involved.
In the (standard) AES round, the input is a 128 bit long string, divided into
16 strings of 8 bits each; the S-box is the concatenation of 16 sub-S-boxes
corresponding to 16 × 8 Boolean functions on 8 variables.
A block cipher being considered, the individual properties of all the
involved Boolean functions can be studied (see Subsection 4.1), but this
is not sufficient. The whole sub-S-boxes must be globally studied (see the
chapter “Vectorial Boolean Functions for Cryptography”).

4.1 Cryptographic criteria for Boolean functions

The design of conventional cryptographic systems relies on two fundamental
principles introduced by Shannon [332]: confusion and diffusion. Confusion
aims at concealing any algebraic structure in the system. It is closely related
to the complexity21 of the involved Boolean functions. Diffusion consists in
spreading out the influence of any minor modification of the input data or of
the key over all outputs. These two principles were stated more than half a
century ago. Since then, many attacks have been found against the diverse
known cryptosystems, and the relevance of these two principles has always
been confirmed. The known attacks on each cryptosystem lead to criteria
[276, 300, 336] that the implemented cryptographic functions must satisfy.
More precisely, the resistance of the cryptosystems to the known attacks can
20
The E-box has expanded the 32 bit long string into a 48 bit long one.
21
That is, the cryptographic complexity, which is different from circuit complexity, for
instance.

47
 ? ?
S1 S16
···
? ?

linear permutation

?
Round key - +

?

Figure 7: An AES round

be quantified through some fundamental characteristics (some, more related

to confusion, and some, more related to diffusion) of the Boolean functions
used in them; and the design of these cryptographic functions needs to con-
sider various characteristics simultaneously. Some of these characteristics
are affine invariants, i.e. are invariant under affine equivalence (recall that
two functions f and g on Fn2 are called affinely equivalent if there exists a lin-
ear isomorphism L from Fn2 to Fn2 and a vector a such that f (x) = g(L(x)+a)
for every input x ∈ Fn2 ) and some are not. Of course, all characteristics can-
not be optimum in the same time, and trade-offs must be considered (see
below).

4.1.1 The algebraic degree

Cryptographic functions must have high algebraic degrees. Indeed, all cryp-
tosystems using Boolean functions for confusion (combining or filtering func-
tions in stream ciphers, functions involved in the S-boxes of block ciphers,
...) can be attacked if the functions have low degrees. For instance, in

48
the case of combining functions, if n LFSRs having lengths L1 , · · · , Ln are
combined by the function
!
M Y
f (x) = aI xi ,
I∈P(N ) i∈I

where P(N ) denotes the power set of N = {1, · · · , n}, then (see [318]) the
sequence produced by f has linear complexity
!
X Y
L≤ aI Li
I∈P(N ) i∈I

(and L equals this number under the sufficient condition that the sequences
output by the LFSRs are m-sequences and the lengths of these LFSRs are
pairwise co-prime), see [367]. In the case of the filter model, we have a
less precise result [317]: if L is the length of the LFSR and if the feedback
polynomial is primitive, then the linear complexity of the sequence satisfies:
d◦ f  
X L
L≤ .
i
i=0
Moreover, if L is a prime, then
 
L
L≥ ,
d◦ f
and the fraction of functions f of given Pdalgebraic
◦f
degree which output a
L
sequence of linear complexity equal to i=0 i is at least e−1/L . In both

models, the algebraic degree of f (recall that this is the largest size of I such
that aI = 1) has to be high so that L can have high value (the number of
those nonzero coefficients aI , in the ANF of f , such that I has large size,
can also play a role, but clearly a less important one). In the case of block
ciphers, using Boolean functions of low degrees makes the higher order dif-
ferential attack [215, 227] effective.
When n tends to infinity, random Boolean functions have almost surely
algebraic degrees at least n − 1 since the number of Boolean functions of
Pn−2 n
algebraic degrees at most n − 2 equals 2 i=0 ( i ) = 22 −n−1 and is negligible
n

n
with respect to the number 22 of all Boolean functions. But we shall see
that the functions of algebraic degrees n − 1 or n do not allow achieving
some other characteristics (balancedness, resiliency, ...).
We have seen in Subsection 2.1 that the algebraic degree is an affine invari-
ant.

49
4.1.2 The nonlinearity
In order to provide confusion, cryptographic functions must lie at large Ham-
ming distance to all affine functions. Let us explain why. We shall say that
there is a correlation between a Boolean function f and a linear function ` if
dH (f, `) is different from 2n−1 . Because of Parseval’s Relation (23) applied
to the sign function fχ and of Relation (14), any Boolean function has cor-
relation with some linear functions of its input. But this correlation should
be small: the existence of affine approximations of the Boolean functions in-
volved in a cryptosystem allows in various situations (block ciphers, stream
ciphers) to build attacks on this system (see [173, 271]).
In the case of stream ciphers, these attacks are the so-called fast correla-
tion attacks [47, 109, 156, 200, 201, 202, 275]: let g be a linear approximation
of f (or f ⊕ 1, then we change f into f ⊕ 1) whose distance to f is smaller
than 2n−1 . Then, denoting by P r[E] the probability of an event E:

dH (f, g) 1
p = P r[f (x1 , · · · , xn ) 6= g(x1 , · · · , xn )] = n
= − ,
2 2
where  > 0. The pseudo-random sequence s corresponds then to the trans-
mission with errors of the sequence σ which would be produced by the
same model, but with g instead of f . Attacking the cipher can be done
by correcting the errors as in the transmission of the sequence σ over a
noisy channel. Assume that we have N bits su , · · · , su+N −1 of the pseudo-
random sequence s, then P r[si 6= σi ] ≈ p. The set of possible sequences
σu , · · · , σu+N −1 is a vectorspace, that is, a linear code of length N and dimen-
sion at most L. We then use a decoding algorithm to recover σu , · · · , σu+N −1
from su , · · · , su+N −1 and since g is linear, the linear complexity of the se-
quence σ is small and we obtain by the Berlekamp-Massey algorithm the
initialisation of the LFSR. We can then compute the whole sequence s.
There are several ways for performing the decoding. The method exposed in
[275] and improved by [109] P is as follows. We call a parity check polynomial
any polynomial a(x) = 1 + rj=1 aj xj (ar 6= 0) which is a multiple of the
feedback polynomial of an LFSR P generating the sequence σi . Denoting by
i
σ(x) the generating function i≥0 σi x , the product a(x) σ(x) is a polyno-
mial of degree less than r. We use for the decoding a set of parity check
polynomials satisfying three conditions: their degrees are bounded by some
integer m, the number of nonzero coefficients aj in each of them is at most
some number t (i.e., each polynomial has Hamming weight at most t+1) and
for every j = 1, · · · , m, at most one polynomial has nonzero coefficient aj .
Each parity check polynomial a(x) = 1 + rj=1 aj xj gives a linear relation
P

50
σi = rj=1 aj σi−j = j=1,···,r / aj 6=0 σi−j for every i ≥ m and the relations
P P
corresponding to different polynomials involve different indices i − j. If we
replace the (unknown) σi ’s by the si ’s then some of these relations become
false but it is possible by using the method of Gallager [160] to compute a
sequence zi such that P r(zi = σi ) > 1 − p. Then it can be proved that iter-
ating this process converges to the sequence σ (with a speed which depends
on m, t and p).
In the case of block ciphers, we shall see in the chapter “Vectorial Boolean
Functions for Cryptography” that the Boolean functions involved in their S-
boxes must also lie at large Hamming distances to affine functions, to allow
resistance to the linear attacks.
The nonlinearity of f is the minimum Hamming distance between f and
affine functions. The larger is the nonlinearity, the larger is p in the fast
correlation attack and the less efficient is the attack. Hence, the nonlinearity
must be high (in a sense that will be clarified below) and we shall see that
this condition happens to be necessary against other attacks as well. A high
nonlinearity is surely one of the most important cryptographic criteria.
The nonlinearity is an affine invariant, by definition, since dH (f ◦ L, ` ◦ L) =
dH (f, `), for every functions f and `, and for every affine automorphism L,
and since ` ◦ L ranges over the whole set of affine functions when ` does.
It can be computed through the Walsh transform: let `a (x) = a1 x1 ⊕ · · · ⊕
an xn = a · x be any linear function; according to Relation (14), we have
dH (f, `a ) = 2n−1 − 12 fbχ (a) and we deduce dH (f, `a ⊕ 1) = 2n−1 + 12 fbχ (a); the
nonlinearity of f is therefore equal to:
1
nl(f ) = 2n−1 − max |fb (a)|. (35)
2 a∈Fn2 χ

Hence a function has high nonlinearity if all of its Walsh values have low
magnitudes.
Parseval’s Relation (23) applied to fχ gives a∈Fn fbχ 2 (a) = 22n , and implies
P
2
that the mean of fbχ 2 (a) equals 2n . The maximum of fbχ 2 (a) being greater
than or equal to its mean (equality occurs if and only if fbχ 2 (a) is constant),
we deduce that maxa∈Fn |fb (a)| ≥ 2n/2 . This implies
2 χ

nl(f ) ≤ 2n−1 − 2n/2−1 . (36)

This bound, valid for every Boolean function and tight for every even n,
as we shall see, will be called the covering radius bound (since this is the
value of the covering radius of the Reed-Muller code of order 1 if n is even;

51
indeed, in the case of the Reed-Muller code of order 1, the covering radius
coincides with the maximum nonlinearity of Boolean functions). The cover-
ing radius bound can be improved when we restrict ourselves to sub-classes
of functions (e.g. resilient and correlation-immune functions, see Section 7).
A Boolean function will be considered as highly nonlinear if its nonlinearity
lies near the upper bound corresponding to the class of functions to which it
belongs. The meaning of “near” depends on the framework, see [203]. Olejár
and Stanek [289] have shown that, when n tends to infinity, random Boolean
√ n−1
functions on Fn2 have almost surely nonlinearity greater than 2n−1 − n 2 2
(this is easy to prove by counting – or more precisely by upper bound-
ing – the number of functions whose nonlinearities are bounded above by
a given number, see [66]). Rodier [311] has shown later more precisely
that, asymptotically, almost all Boolean functions have nonlinearity between
√ √ 4 ln n
 √ √ 
2 n−1 −2 n/2−1 n 2 ln 2 + n and 2 n−1 −2 n/2−1 n 2 ln 2 − 5 ln
n
n
√
and therefore located in the neighbourhood of 2n−1 − 2n/2−1 2n ln 2.
Equality occurs in (36) if and only if |fbχ (a)| equals 2n/2 for every vector a.
The corresponding functions are called bent functions. They exist only for
even values of n, because 2n−1 − 2n/2−1 must be an integer (in fact, they
exist for every n even, see Section 6). The whole Section 6 is devoted to
bent functions.
For n odd, Inequality (36) cannot be tight. The maximum nonlinearity
n−1
of n-variable Boolean functions lies then between 2n−1 − 2 2 (which can
always be achieved e.g. by quadratic functions, see Subsection 5.2) and
n−1
2n−1 − 2n/2−1 . It has been shown in [177, 284] that it equals 2n−1 − 2 2
when n = 1, 3, 5, 7, and in [295, 296], by Patterson and Wiedemann22 , that
n−1
it is strictly greater than 2n−1 −2 2 if n ≥ 15 (a review on what was known
in 1999 on the best nonlinearities of functions on odd numbers of variables
n−1
was given in [154], see also [29, 237]). This value 2n−1 − 2 2 is called the
quadratic bound because, as we already mentioned, such nonlinearity can
be achieved by quadratic functions. It is also called the bent-concatenation
bound since it can also be achieved by the concatenation of two bent func-
tions in n − 1 variables. Very recently it has been proved in [210] (see also
[262]) that the best nonlinearity of Boolean functions in odd numbers of
variables is strictly greater than the quadratic bound for any n > 7.

22
It has been later proved (see [328, 141] and [267, 216]) that balanced functions with
n−1
nonlinearity strictly greater than 2n−1 −2 2 , and with algebraic degree n−1, or satisfying
P C(1), exist for every odd n ≥ 15.

52
 The nonlinearity of a Boolean function f equals the minimum distance
of the linear code R(1, n) ∪ (f ⊕ R(1, n)). More generally, the minimum
distance of a code defined as the union of cosets f ⊕ R(1, n) of the Reed-
Muller code of order 1, where f ranges over a set F, equals the minimum
nonlinearity of the functions f ⊕ g, where f and g are distinct and range
over F. This observation allows constructing good nonlinear codes such as
Kerdock codes (see Subsection 6.10).
Bent functions being not balanced (i.e. their values being not uniformly
distributed, see below), they are improper for use in cryptosystems23 (see
below). For this reason, even when they exist (for n even), it is also necessary
to study those functions which have large but not optimal nonlinearities, say
n−1
between 2n−1 − 2 2 and 2n−1 − 2n/2−1 , among which some balanced func-
tions exist. The maximum nonlinearity of balanced functions is unknown
for any n ≥ 8.
Two relations have been first observed in [360, 363] between the nonlinear-
ity and the derivatives of Boolean functions (we give here simpler proofs):
applying Relation (27), relating the values of the Walsh transform of a func-
tion on a flat a + E to the autocorrelation coefficients of the function on a
flat b + E ⊥ , to all linear hyperplanes E = {0, e}⊥ , e 6= 0, to all vectors a
2 2
and to b = 0, and using that maxu∈E fb (u) ≥ 1
P
χ |E|fb (u), we deduce:
u∈E χ

1r n
nl(f ) ≤ 2n−1 − 2 + max |F(De f )|.
2 e6=0

And the obvious relation wH (f ) ≥ 12 wH (De f ), valid for every e ∈ Fn2 , leads
when applied to the functions f ⊕ `, where ` is affine, to the lower bound:
1
nl(f ) ≥ 2n−2 − min |F(De f )|. (37)
4 e6=0

Another lower bound on the nonlinearity is a consequence of Remark 2 after

Theorem 1: if f admits a maximal odd weighting subspace E of dimen-
sion d ≥ 2, then for every affine function `, the function f ⊕ ` also admits E
as maximal odd weighting subspace (since the restriction of ` to E and to
any of its superspaces has an even weight) and thus has nonlinearity at least
23
As soon as n is large enough (say n ≥ 20), the difference 2n/2−1 between their weights
and the weight 2n−1 of balanced functions is very small with respect to this weight.
However, according to [13, Theorem 6], 2n bits of the pseudo-random sequence output
by f are enough to distinguish it from a random sequence. Nevertheless, we shall see in
Section 6 that highly nonlinear functions can be built from bent functions.

53
2n−d .

The r-th order nonlinearity: changing one or a few bits in the output
to a low degree Boolean function (that is, in its truth-table) gives a function
with high degree and does not fundamentally modify the robustness of the
system using this function (explicit attacks using approximations by low de-
gree functions exist for block ciphers but not for all stream ciphers however,
see e.g. [219]). A relevant criterion is the nonlinearity profile, that is, the
sequence of the Hamming distances to the Reed-Muller code of order r, for
all values of r < n. This distance is called the r-th order nonlinearity (and if
r is not specified, the higer order nonlinearity) of f and denoted by nlr (f ).
This criterion is related to the maximum correlation of the Boolean function
with respect to a subset of variables, or equivalently, to the minimal distance
of the function to functions depending on a subset of variables (which plays
a role with respect to the correlation attack, see below in Subsection 4.1.7)
since a function depending on k variables has algebraic degree at most k.
Hence the r-th order nonlinearity is a lower bound to the distance to func-
tions depending of at most k variables. The former is much more difficult
to study than the latter.
The best known asymptotic upper bound on nlr (f ) is
√
n−1 15 √
2 − · (1 + 2)r−2 · 2n/2 + O(nr−2 )
2
(see [92], where a non-asymptotic - and more complex - bound is also
given). Counting the number of functions whose r-th order nonlinearities
are bounded above by a given number allows proving that, when n tends
to infinity,
qPthere exist functions with r-th order nonlinearity greater than
r n n−1
n−1


2 − i=0 i 2 2 . But this does not help obtaining explicit func-
tions with non-weak r-th order nonlinearity.
Computing the r-th order nonlinearity of a given function with algebraic
degree strictly greater than r is a hard task for r > 1 (in the case of the first
order, we have seen that much is known in theory and also algorithmically
since the nonlinearity is related to the Walsh transform, which can be com-
puted by the algorithm of the Fast Fourier Transform; but for r > 1, very
little is known). Even the second order nonlinearity is known only for a few
peculiar functions and for functions in small numbers of variables. A nice
algorithm due to G. Kabatiansky and C. Tavernier and improved and im-
plemented by Fourquet and Tavernier [157] works well for r = 2 and n ≤ 11
(in some cases, n ≤ 13), only. It can be applied for higher orders, but it is

54
then efficient only for very small numbers of variables. No better algorithm
is known. Proving lower bounds on the r-th order nonlinearity of functions
(and therefore proving their good behavior with respect to this criterion) is
also a quite difficult task, even for the second order. Until recently, there
had been only one attempt, by Iwata-Kurosawa [198], to construct func-
tions with lower bounded r-th order nonlinearity. But the obtained value,
2n−r−3 (r + 5), of the lower bound was small. Also, lower bounds on the r-th
order nonlinearity by means of the algebraic immunity of Boolean functions
have been derived (see Section 9) but they are small too. In [73] is introduced
a method for efficiently bounding below the nonlinearity profile of a given
function in the case lower bounds exist for the (r − 1)-th order nonlinearities
of the derivatives of f :

Proposition 13 Let f be any n-variable function and r a positive integer

smaller than n. We have:
1
nlr (f ) ≥ max nlr−1 (Da f )
2 a∈Fn2

and
1
s X
n−1
nlr (f ) ≥ 2 − 22n − 2 nlr−1 (Da f ).
2
a∈Fn
2

The first bound is easy to prove and  the second one comes from the equalities
 
n−1 1 X f (x)⊕h(x) 

nlr (f ) = 2 − max  (−1)  and:
2 h∈BF n / d f ≤r  n
◦ 
x∈F 2

 2
X X X
 (−1)f (x)⊕h(x)  = (−1)Da f (x)⊕Da h(x) .
x∈Fn
2 a∈Fn n
2 x∈F2

Lower bounds for the second order nonlinearities of some functions (known
for being highly nonlinear) are deduced in [73], as well as bounds for the
n
whole nonlinearity profile of the multiplicative inverse function trn (x2 −2 )
(used in the S-box of the AES with n = 8, see the chapter “Vectorial Boolean
Functions for Cryptography”): the r-th order nonlinearity of this function
−r
is approximately bounded below by 2n−1 − 2(1−2 ) n and therefore asymp-
totically equivalent to 2n−1 , for every fixed r. Note that the extension of the
Weil bound recalled in Subsection 5.6 is efficient for lower bounding the r-th
order nonlinearity of the inverse function only for r = 1. Indeed, already for
r = 2, the univariate degree of a quadratic function in trace representation

55
form can be bounded above by 2bn/2c + 1 only and this gives a bound in
2n on the maximum magnitude of the Walsh transform and therefore no
information on the nonlinearity.

4.1.3 Balancedness and resiliency

Cryptographic functions must be balanced functions (their output must be
uniformly – that is, equally – distributed over {0, 1}) for avoiding statistical
dependence between the plaintext and the ciphertext. Notice that f is
balanced if and only if fbχ (0) = F(f ) = 0.
A stronger condition is necessary in the filtering model of pseudo-random
generators, in order to avoid so-called distinguishing attacks. These attacks
are able to distinguish the pseudorandom sequence (si )i∈N from a random
sequence. A way of doing so is to observe that the distribution of the
sequences (si+γ1 , · · · , si+γn ) is not uniform, where γ1 , · · · , γn are the positions
where the input bits to the filtering function are chosen. J. Golić [163] has
observed that if the feedback polynomial of the LFSR is primitive and if the
filtering function has the form x1 ⊕ g(x2 , · · · , xn ) or g(x1 , · · · , xn−1 ) ⊕ xn ,
then the property of uniformity is satisfied. A. Canteaut [40] has proved that
this condition on the function is also necessary for having uniformity. For
choosing a filtering function, we shall have to choose a function g satisfying
the cryptographic criteria listed in the present section, and use f defined by
means of g in one of the two ways above.
There is an additional condition to balancedness in the case of the combiner
model: any combining function f (x) must stay balanced if we keep constant
some number of coordinates xi of x.

Definition 3 Let n be a positive integer and m < n a non-negative integer.

An n-variable function f is called an m-resilient function24 if any of its
restrictions obtained by fixing at most25 m of its input coordinates xi is
24
More generally, a (non necessarily balanced) combining function whose output distri-
bution probability is unaltered when any m (or, equivalently, at most m) of the inputs
are kept constant is called an m-th order correlation-immune function. Similarly with
resiliency, correlation immunity is characterized by the set of zero values in the Walsh
spectrum of the function: f is m-th order correlation-immune if and only if fbχ (u) = 0, i.e.
fb(u) = 0, for all u ∈ Fn 2 such that 1 ≤ wH (u) ≤ m. The notion of correlation-immune
function is related to the notion of orthogonal array (see [35]). Only resilient functions
are of interest as cryptographic functions (but Boolean correlation-immune functions play
a role with respect to vectorial resilient functions, see the chapter “Vectorial Boolean
Functions for Cryptography”).
25
Or exactly m, this is equivalent.

56
balanced.

This definition of resiliency was introduced by Siegenthaler26 in [336]; it

is related to an attack on the combiner model27 , called correlation attack :
if f is not m-resilient, then there exists a correlation between the output
to the function and (at most) m coordinates of its input; if m is small, a
divide-and-conquer attack due to Siegenthaler [337] uses this weakness for
attacking a system using f as combining function; in the original attack by
Siegenthaler, all the possible initializations of the m LFSRs corresponding
to these coordinates are tested (in other words, an exhaustive search of
the initializations of these specific LFSRs is done); when we arrive to the
correct initialization of these LFSRs, we observe a correlation (before that,
the correlation is negligible, as for random pairs of sequences); now that the
initializations of the m LFSRs are known, those of the remaining LFSRs
can be found with an independent exhaustive search. The fast correlation
attacks that we saw above can be more efficient if the Boolean combining
function is not highly nonlinear. More precisely, Canteaut and Trabbia
in [47] and Canteaut in [38] show that, to make the correlation attack on
the combiner model with an m-resilient combining function as inefficient
as possible, the coefficient fbχ (u) of the function has to be small for every
vector u of Hamming weight higher than but close to m. This condition is
satisfied if the function is highly nonlinear. Hence we see that nonlinearity
plays a role with respect to all the main attacks.
Note that, when we say that a function f is m-resilient, we do not mean
that m is the maximum value of k such that f is k-resilient. We will call
this maximum value the resiliency order of f .
Resiliency has been characterized by Xiao and Massey through the Fourier
and the Walsh transforms:

Theorem 3 [174] Any n-variable Boolean function f is m-resilient if and

only if fbχ (u) = 0 for all u ∈ Fn2 such that wH (u) ≤ m. Equivalently, f is
m-resilient if and only if it is balanced and fb(u) = 0 for all u ∈ Fn2 such that
0 < wH (u) ≤ m.

We give here a first direct proof of this fact: we apply Relation (28) to
E = {x ∈ Fn2 / xi = 0, ∀i ∈ I} where I is any set of indices of size m; the
sum of E and E ⊥ = {x ∈ Fn2 / xi = 0, ∀i 6∈ I} is direct and equals Fn2 ; hence
26
The term of resiliency was, in fact, introduced in [110], in relationship with another
cryptographic problem.
27
This attack has no equivalent for the filter model, where first order resiliency seems
sufficient; see more precisely in [170] the status of resiliency in the filter generator.

57
we can take E 0 = E ⊥ and we get
P b2 ⊥
P 2
u∈E ⊥ fχ (u) = |E | a∈E ⊥ F (ha ),
where ha is the restriction of f to a + E, that is, the restriction obtained by
fixing the coordinates of x whose indices belong to I to the corresponding
coordinates of a. The number F(ha ) is null if and only if ha is balanced
and clearly, all the numbers F(ha ), a ∈ E ⊥ are null if and only if all the
numbers fbχ (u), u ∈ E ⊥ are null. Since this is valid for every muti-index I
of size m, this completes the proof.
An alternate proof of this same result is obtained by applying the Poisson
summation formula (17) to ϕ = fχ , a = 0 and E = {x ∈ Fn2 / xi = 0, ∀i 6∈ I},
b ranging over Fn2 . We obtain that f is m-resilient if and only if, for every b
and every I of size m, we have u∈Fn / ui =0, ∀i6∈I (−1)b·u fbχ (u) = 0 and it can
P
2
easily be shown that this is equivalent to fb (u) = 0 for every u of weight at
χ
most m.
Theorem 3 shows that f is m-resilient if and only if its support has size 2n−1
and dual distance at least m + 1. Indeed, if C denotes the Psupport of f , the
dual distance of C equals the number min{wH (b); b 6= 0, x∈C (−1)b·x 6= 0},
according to RelationP(34) and to the observation which follows it; we have,
for every vector b: b·x = fb(b) and therefore, for every b 6= 0:
x∈C (−1)
b·x = − 1 fb (b). More generally, f is m-th order correlation im-
P
x∈C (−1) 2 χ
mune if and only if its support has dual distance at least m + 1. This had
been observed by Delsarte in [129, 130] (see also in a paper by J. Massey [270]
a generalization of this result to arrays over finite fields and other related
nice results).
An easily provable related property is that, if G is the generator matrix of
an [n, k, d] linear code, then for every k-variable balanced function g, the
n-variable function f (x) = g(x × Gt ) is (d − 1)-resilient [128] (but such func-
tion has nonzero linear structures, see below).
Contrary to the algebraic degree, to the nonlinearity and to the balanced-
ness, the resiliency order is not an affine invariant, except for the null order
(and for the order n, but the set of n-resilient functions is empty, because
of Parseval’s relation). It is invariant under any translation x 7→ x + b,
according to Proposition 6 and Theorem 3. The symmetry group of the set
of m-resilient functions and the orbits under its action have been studied
in [194]).
The whole Section 7 is devoted to resilient functions.

58
4.1.4 Strict avalanche criterion and propagation criterion
The Strict Avalanche Criterion (SAC) was introduced by Webster and
Tavares [352] and this concept was generalized into the Propagation Cri-
terion (P C) by Preneel et al. [300] (see also [301]). The SAC, and its gen-
eralizations, are based on the properties of the derivatives of Boolean func-
tions. These properties describe the behavior of a function whenever some
coordinates of the input are complemented. Thus, they are related to the
property of diffusion of the cryptosystems using the function. They concern
the Boolean functions involved in block ciphers. Let f be a Boolean function
on Fn2 and E ⊂ Fn2 . The function f satisfies the propagation criterion P C
with respect to E if, for all a ∈ E, the derivative Da f (x) = f (x) ⊕ f (a + x)
(see Definition 2) is balanced. It satisfies P C(l) if it satisfies P C with re-
spect to the set of all nonzero vectors of weights at most l. In other words,
f satisfies P C(l) if the auto-correlation coefficient F(Da f ) is null for every
a ∈ Fn2 such that 1 ≤ wH (a) ≤ l. Criterion SAC corresponds to P C(1).
It is needed, for some cryptographic applications, to have Boolean functions
which still satisfy P C(l) when a certain number k of coordinates of the input
x are kept constant (whatever are these coordinates and whatever are the
constant values chosen for them). We say that such functions satisfy the
propagation criterion P C(l) of order k. This notion, introduced in [300], is
a generalization of the strict avalanche criterion of order k, SAC(k) (which
is equivalent to P C(1) of order k), introduced in [155]. Obviously, if a func-
tion f satisfies P C(l) of order k ≤ n − l, then it satisfies P C(l) of order k 0
for any k 0 ≤ k.
There exists another notion, which is similar to P C(l) of order k, but
stronger [300, 302] (see also [61]): a Boolean function satisfies the extended
propagation criterion EP C(l) of order k if every derivative Da f , with a 6= 0
of weight at most l, is k-resilient.
All of these criteria are not affine invariants, in general.
A weakened version of the PC criterion has been studied in [222].

4.1.5 Non-existence of nonzero linear structure

We shall call the linear kernel of f the set of those vectors e such that
De f is a constant function. The linear kernel of any Boolean function is
an F2 -subspace of Fn2 . Any element e of the linear kernel of f is said to
be a linear structure of f . Nonlinear cryptographic functions used in block
ciphers should have no nonzero linear structure (see [148]). The existence

59
of nonzero (involuntary) linear structures, for the functions implemented in
stream ciphers, is a potential risk that should also be avoided, despite the
fact that such existence could not be used in attacks, so far.

Proposition 14 An n-variable Boolean function admits a nonzero linear

structure if and only if it is linearly equivalent to a function of the form
f (x1 , · · · , xn ) = g(x1 , · · · , xn−1 ) ⊕ ε xn where ε ∈ F2 . More generally, its
linear kernel has dimension at least k if and only if it is linearly equivalent
to a function of the form f (x1 , · · · , xn ) = g(x1 , · · · , xn−k ) ⊕ εn−k+1 xn−k+1 ⊕
· · · ⊕ εn xn where εn−k+1 , · · · , εn ∈ F2 .

Indeed, if we compose f on the right with a linear automorphism L such that

L(0, · · · , 0, 1) = e is a nonzero linear structure, we have then D(0,···,0,1) (f ◦
L)(x) = f ◦L(x)⊕f ◦L(x+(0, · · · , 0, 1)) = f ◦L(x)⊕f (L(x)+e) = De f (L(x)).
The case of dimension k is similar.
Note that, according to Proposition 14, if f admits a nonzero linear struc-
n−1
ture, then the nonlinearity of f is bounded above by 2n−1 − 2 2 (this
implies that the functions obtained by Patterson and Wiedemann cannot
have nonzero linear structure), since it equals twice that of g and since,
g being an (n − 1)-variable function, it has nonlinearity bounded above
n−1
by 2n−2 − 2 2 −1 . Similarly, if k is the dimension of the linear kernel of f ,
n+k−2
we have straighforwardly nl(f ) ≤ 2n−1 − 2 2 [41].
Another characterization of linear structures [228, 146] (see also [43]) is
a direct consequence of Relation (27), relating the values of the Walsh
transform of a function on a flat a + E to the autocorrelation coefficients
of the function on a flat b + E ⊥ , with b = 0 and E = {0, e}⊥ , that is
2
fb (u) = 2n−1 (2n + (−1)a·e F(De f )).
P
u∈a+E χ

Proposition 15 Let f be any n-variable Boolean function. The deriva-

tive De f equals the null function (resp. the function 1) if and only if the
support Sfb = {u ∈ Fn2 / fbχ (u) 6= 0} of fbχ is included in {0, e}⊥ (resp. its
χ
complement).

This is a direct consequence of the relation above deduced from (27), with
a · e = 1 if De f is null and a · e = 0 if De f = 1. Notice that, if De f is
the constant function 1 for some e ∈ Fn2 , then f is balanced (indeed, the
relation f (x + e) = f (x) ⊕ 1 implies that f takes the values 0 and 1 equally
often). Thus, a non-balanced function f has no nonzero linear structure if
and only if there is no nonzero vector e such that De f is null. According
to Proposition 15, this is equivalent to saying that the support of its Walsh

60
transform has rank n. A similar characterization exists for balanced func-
tions by replacing the function f (x) by a non-balanced function f (x) ⊕ b · x.
It is deduced in [105] (see more in [347]) that resilient functions of high or-
ders must have linear structures.
The existence/non-existence of nonzero linear structures is clearly an affine
invariant. But, contrary to the other criteria, it is an all-or-nothing crite-
rion. Meier and Staffelbach introduced in [276] a related criterion, leading
to a characteristic (that is, a criterion which can be satisfied at levels quan-
tified by numbers): a Boolean function on Fn2 being given, its distance to
linear structures is its distance to the set of all Boolean functions admitting
nonzero linear structures (among which we have all affine functions – hence,
this distance is bounded above by the nonlinearity – but also other functions,
such as all non bent quadratic functions). This distance is always bounded
above by 2n−2 . More precisely, it equals28 : 2n−2 − 14 maxe∈Fn2 ∗ |F(De f )|,
since a function g, which admits some nonzero vector e as a linear structure,
and which lies at minimum distance from f among all such functions, can
be obtained by choosing an affine hyperplane H such that Fn2 = H ∪ (e + H),
and defining g(x) = f (x) for every x ∈ H and g(x) = f (x + e) ⊕  for every
x ∈ (e + H), where  is chosen in F2 ; the Hamming distance between f and
this function g equals |{x ∈ e + H/ De f (x) =  ⊕ 1}| = 12 |{x ∈ Fn2 / De f (x) =

 ⊕ 1}| = 12 2n−1 − (−1) 2 F(De f ) . Recall that ∆f (e) = F(De f ) is the
auto-correlation function of f . We see (according to Theorem 8) that the
distance of f to linear structures equals 2n−2 if and only if f is bent.

4.1.6 Algebraic immunity

A new kind of attacks, called algebraic attacks, has been introduced recently
(see [117, 150, 113]). Algebraic attacks recover the secret key, or at least
the initialization of the system, by solving a system of multivariate algebraic
equations. The idea that the key bits can be characterized as the solutions
of a system of multivariate equations comes from C. Shannon [332]. In
practice, for cryptosystems which are robust against the usual attacks such
as the Berlekamp-Massey attack, this system is too complex to be solved (its
equations being highly nonlinear). However, in the case of stream ciphers,
we can get a very overdefined system (i.e. a system with a number of linearly
independent equations much greater than the number of unknowns). Let us
consider the combiner or the filter model, with a linear part (the n LFSRs
28
Note that this proves again Relation (37).

61
in the case of the combiner model, the single LFSR in the case of the filter
model) of size N and with an n-variable Boolean function f as combining
or filtering function; then there exists a linear permutation L : FN
2 7→ F2
N

and a linear mapping L0 : FN n

2 7→ F2 such that, denoting by u1 , · · · , uN
the initialisation of the LFSR and by (si )i≥0 the pseudo-random sequence
output by the generator, we have, for every i ≥ 0:

si = f (L0 ◦ Li (u1 , · · · , uN )).

The number of equations can then be much larger than the number of un-
knowns. This makes less complex the resolution of the system by using
Groebner basis (see [150]), and even allows linearizing the system (i.e. ob-
taining a system of linear equations by replacing every monomial of degree
greater than 1 by a new unknown); the resulting linear system has however
too many unkwnowns and cannot be solved. Nevertheless, Courtois and
Meier have had a simple but very efficient idea. Assume that there exist
functions g 6= 0 and h of low degrees (say, of degrees at most d) such that
f × g = h (where f × g denotes the Hadamard product of f and g, whose
support is the intersection of the supports of f and g, we shall omit writing
× in the sequel). We have then, for every i ≥ 0:

si g(L0 ◦ Li (u1 , · · · , uN )) = h(L0 ◦ Li (u1 , · · · , uN )).

This equation in u1 , · · · , uN has degree at most d, since L and L0 are linear,

and the system of equations obtained after linearization can then be solved
by Gaussian elimination.
Low degree relations have been shown to exist for several well known con-
structions of stream ciphers, which were immune to all previously known
attacks.
Note that if we only know the existence of a nonzero low degree multiple h of
f , then the support of h being included in that of f , we have (f ⊕ 1)h = 0,
and taking g = h, we have the desired relation f g = h (the paper [117]
mentioned the existence of low degree multiples of f for making the attack
feasible). It is a simple matter to see also that the existence of functions
g 6= 0 and h, of degrees at most d, such that f g = h is equivalent to the
existence of a function g 6= 0 of degree at most d such that f g = 0 or
(f ⊕ 1)g = 0. Indeed, f g = h implies f 2 g = f h, that is (since f 2 = f ),
f (g ⊕ h) = 0, which gives the desired equality if g 6= h by replacing g ⊕ h
by g, and if g = h then f g = h is equivalent to (f ⊕ 1)g = 0. A function g
such that f g = 0 is called an annihilator of f . The set of all annihilators is
equal to the ideal of all the multiples of f ⊕ 1.

62
 Pn
Let g be a function of degree
P P d. Let the ANF of g equal a0 + i=1 ai xi +
1≤i<j≤n ai,j xi xj + ... + 1≤i1 ≤...≤id ≤n ai1 ,...id xi1 ...xid . Note that g is an
annihilator of f if and only if f (x) = 1 implies g(x) = 0. Hence, g is an
annihilator of f if and only if the coefficients in its ANF satisfy the system
of homogeneous
Pd
linear equations which translates this fact. In this system,
n
we have i=0 i number of variables (the coefficients of the monomials of
degrees at most d) and wH (f ) many equations.
The minimum degree of g 6= 0 such that f g = 0 (i.e. such that g is an
annihilator of f ) or (f ⊕ 1)g = 0 (i.e. such that g is a multiple of f ) is
called the (standard) algebraic immunity of f and denoted by AI(f ). This
important characteristic is an affine invariant. More precisely, its auto-
morphism group (that is, the group of all permutations σ of Fn2 such that
AI(f ◦ σ) = AI(f ) for every Boolean function f ) equals the general affine
group. Indeed, denoting by AN (f ) the set of annihilators of f , we have
AN (f ◦ σ) = AN (f ) ◦ σ. Hence this automorphism group equals the auto-
morphism group of the Reed-Muller codes.
As shown in [117], the algebraic immunity of any n-variable function is
bounded above29 by dn/2e. Indeed, the sum of the number of monomials of
degrees at most dn/2e and of the (equal) number of the products between f
and these monomials being greater than 2n , these functions are necessarily
linearly dependent elements of the 2n -dimensional vectorspace of all Boolean
functions. This linear dependence gives two functions g and h of degrees at
most dn/2e such that f g = h and (g, h) 6= (0, 0), i.e. g 6= 0.
Let us see now what are the consequences of the existence of this attack
on the design of stream ciphers: let an n-variable function f , with alge-
braic immunity dn/2e be used for instance as a filtering function on an
LFSR) of length N ≥ 2k, where k is the length of the key (otherwise, it is
known that the system is not robust against an attack called time-memory-
data trade-off attack). Then the complexity of an algebraic attack using


log2 7
one annihilator of degree dn/2e is roughly 7 N0 + · · · + dn/2e N
≈

 2.8
7 N0 + · · · + dn/2e N


(see [117]). Let us choose k = 128 (which is usual)
and N = 256, then the complexity of the algebraic attack is at least 280
(which is considered nowadays as a just sufficient complexity) for n ≥ 13;
and it is greater than the complexity of an exhaustive search, that is 2128 ,
29
Consequently, it is bounded above by dk/2e if, up to affine equivalence, it depends
only on k variables, and by dk/2 + 1e if it has a linear kernel of dimension n − k, since it
is then equivalent, according to Proposition 14, to a function in k variables plus an affine
function.

63
for n ≥ 15. If the attacker knows several linearly independent annihilators
of degree dn/2e, then the number of variables must be enhanced! In prac-
tice, the number of variables will have to be near 20 (but this poses then a
problem of efficiency of the stream cipher).
A high value of AI(f ) is not a sufficient property for a resistance to
algebraic attacks, because of fast algebraic attacks, which work if one can
find g of low degree and h 6= 0 of reasonable degree such that f g = h, see
[113, 176] (note however that fast algebraic attacks need more data than
standard ones). This has been exploited in [115] to present an attack on
a stream cipher called SFINKS. Similarly as above, when the number of
monomials of degrees at most e, plus the number of monomials of degrees
at most d, is strictly greater than 2n – that is, when e + d ≥ n – there exist
g 6= 0 of degree at most e and h of degree at most d such that f g = h. An
n-variable function f is then optimal with respect to fast algebraic attacks if
there do not exist two functions g 6= 0 and h such that f g = h, d◦ g < dn/2e
and d◦ g + d◦ h < n. Since f g = h implies f h = f f g = f g = h, we see
that h is then an annihilator of f ⊕ 1, and if h 6= 0, its degree is then
at least equal to the algebraic immunity of f . This means that having a
high algebraic immunity is not only a necessary condition for a resistance to
standard algebraic attacks but also for a resistance to fast algebraic attacks.
The pseudo-random generator must also resist algebraic attacks on the
augmented function [153], that is (considering now f as a function in N
variables, to simplify description), the vectorial function F (x) whose output
equals the vector (f (x), f (L(x)), · · · , f (Lm−1 (x))), where L is the (linear)
update function of the linear part of the generator. Algebraic attacks can
be more efficient when applied to the augmented function rather than to the
function f itself. The efficiency of the attack depends not only on the func-
tion f , but also on the update function (and naturally also on the choice of
m), since for two different update functions L and L0 , the vectorial functions
F (x) and F 0 (x) = (f (x), f (L0 (x)), ..., f (L0 m−1 (x)) are not linearly equiva-
lent (neither equivalent in the more general sense called CCZ-equivalence,
that is, affine equivalence of the graphs of the functions, see the chapter
“Vectorial Boolean Functions for Cryptography”). Testing the behavior of
a function with respect to this attack is therefore a long term work (all
possible update functions have to be investigated).
Finally, a powerful attack on the filter generator has been introduced by
S. Rønjom and T. Helleseth in [313], which also adapts the idea of algebraic
attacks due to Shannon,
but in a different way. The complexity of the at-
tack is in about di=0 Ni operations, where d is the algebraic degree of the
P

64
filter function and N is the length of the LFSR. It needs about di=0 Ni
P

consecutive bits of the keystream output by the pseudo-random generator.

Since d is supposed toPbe close

to the number n Nof
variables of the filter
d N
function, the number i=0 i is comparable to n , while in the case of a
standard algebraic attack with the method due  to
 Courtois and Meier, the
AI(f ) N
ω
P
complexity of the attack is in O i=0 i operations, where ω ≈ 3
is the exponent of the Gaussian reduction30 and AI(f ) is the algebraic im-
PAI(f )

munity of the filter function, and it needs about i=0 Ni consecutive bits
of the keystream. Since AI(f ) is supposed to be close to dn/2e, we can see
that denoting by C the complexity of the Courtois-Meier attack and by C 0
the amount of data it needs, the complexity of the Rønjom-Helleseth attack
roughly equals C 2/3 and the amount of data it needs is roughly C 0 2 . From
the viewpoint of complexity, it is more efficient and from the viewpoint of
data it is less efficient.
The whole Section 9 is devoted to the algebraic immunity of Boolean func-
tions.

4.1.7 Other criteria

- The second moment of the auto-correlation coefficients:
X
V(f ) = F 2 (De f ) (38)
e∈Fn
2

has been introduced by Zhang and Zheng [359] for measuring the global ava-
lanche criterion (GAC), and also called the sum-of-squares indicator . The
absolute indicator is by definition maxe∈Fn2 , e6=0 | F(De f ) |. Both indicators
are clearly affine invariants. In order to achieve good diffusion, cryptographic
functions should have low sum-of-squares indicators and absolute indicators.
Obviously, we have V(f ) ≥ 22n , since F 2 (D0 f ) = 22n . Note that every lower
bound of the form V(f ) ≥ V straightforwardly
q implies that the absolute
V −22n
indicator is bounded below by 2n −1 . The functions achieving V(f ) =
22nare those functions whose derivatives De f (x), e 6= 0, are all balanced.
We shall see in Section 6 that these are the bent functions. If f has a k-
dimensional linear kernel, then V(f ) ≥ 22n+k (with equality if and only if f
30
As already seen, it can be taken equal to log2 7 ≈ 2.8 and the coefficient in the O
can be taken equal to 7, according to Strassen [340]; a still better exponent is due to
Coppersmith and Winograd but the multiplicative constant is then inefficiently high for
our framework.

65
is partially bent, see below).
Note that, according to Relation (26) applied to De f for every e, we have
X
V(f ) = F(Da De f ),
a,e∈Fn
2

where Da De f (x) = f (x) ⊕ f (x + a) ⊕ f (x + e) ⊕ f (x + a + e) is the second

order derivative of f .
Note also that, according to Relation (21) applied to ϕ(e) = ψ(e) = F(De f ),
we have, for any n-variable Boolean function f :
X 2 2 X
∀a ∈ Fn2 , fbχ (e)fbχ (a + e) = 2n F 2 (De f )(−1)e·a ,
e∈Fn
2 e∈Fn
2

2
as shown in [42] (indeed, the Fourier transform of ϕ equals fbχ , according to
Relation (25)), and thus, for a = 0:
X 4
fbχ (e) = 2n V(f ). (39)
e∈Fn
2

 
 
X 4 X 2 2 4
We have: fbχ (e) ≤  fχ (e)
b  max f (e) ≤ 2n max
b
n χ
fb (e).
n χ
e∈F2 e∈F2
e∈Fn
2 e∈Fn
2
2
fbχ (e) = 22n , we deduce, using Rela-
P
According to Parseval’s relation e∈Fn
2
2
tion (39): maxe∈Fn2 fbχ (e) ≥ V(f ) p
2n ≥ V(f ); thus, according to Relation (35)
relating the nonlinearity to the Walsh transform, we have (as first shown
in [360, 363]):
1p
nl(f ) ≤ 2n−1 − 2−n/2−1
p
V(f ) ≤ 2n−1 − 4 V(f ).
2

Denoting again by Nfb the cardinality of the support {a ∈ Fn2 / fbχ (a) 6= 0}
χ
of the Walsh transform of f , Relation (39) also implies the following re-
lation, first observed in [363]: V(f ) × Nfb ≥ 23n . Indeed, using for in-
χ
2
b 2 (a)
P
stance the Cauchy-Schwartz inequality, we see that n f
a∈F2 χ ≤
4 2
P 
2n
P
a∈Fn fχ (a) × Nfb and we have a∈Fn fχ (a) = 2 , according to Par-
b b
2 χ 2
seval’s Relation (23). Clearly, the functions satisfying nl(f ) = 2n−1 −
2−n/2−1 V(f ) (resp. V(f ) × Nfb = 23n ) are the functions whose Walsh
p
χ
transforms take at most one nonzero magnitude. These functions are called

66
plateaued functions (see Subsection 6.8 for further properties
p of plateaued
functions). The functions satisfying nl(f ) = 2n−1 − 12 4 V(f ) are (also
clearly) the bent functions.
Constructions of balanced Boolean functions with low absolute indicators
and high nonlinearities have been studied in [260].

- The maximum correlation of an n-variable Boolean function f with respect

to a subset I of N = {1, · · · , n} equals by definition (see [358]) Cf (I) =
F(f ⊕ g)
max , where BF I,n is the set of n-variable Boolean functions
g∈BF I,n 2n
depending on {xi , i ∈ I} only. According to Relation (13), the distance
from f to BF I,n equals 2n−1 (1 − Cf (I)). As we saw above, denoting the size
of I by r, this distance is bounded below by the r-th order nonlinearity.
The maximum correlation of any combining function with respect to any
subset I of small size should be small (i.e. its distance to BF I,n should be
high). It is straightforward to prove, by decomposing the sum F(f ⊕ g),
P |I| |F (h )|
that Cf (I) equals 2j=1 2nj , where h1 , · · · , h2|I| are the restrictions of f
obtained by keeping constant the xi ’s for i ∈ I, and to see that the distance
from f to BF I,n is achieved by the functions g taking value 0 (resp. 1) when
the corresponding value of F(hj ) is positive (resp. negative), and that we
have Cf (I) = 0 if and only if all hj ’s are balanced (thus, f is m-resilient if and
only if Cf (I) = 0 for every set I of size at most m). Also, according to the
P |I| 2 P |I|
2
Cauchy-Schwartz inequality, we have j=1 |F(hj )| ≤ 2|I| 2j=1 F 2 (hj ),
and Relation (28) directly implies the following inequality observed in [38]:
 1
2
−n 
X 2 |I|
Cf (I) ≤ 2 fbχ (u) ≤ 2−n+ 2 (2n − 2nl(f )) (40)
u∈Fn
2 / ui =0, ∀i6∈I

or equivalently:
 1
2

1 X 2 |I|
dH (f, BF I,n ) ≥ 2n−1 − n−1
− 2 2 −1 maxn fbχ (u)|.

 ≥2
fbχ (u)

2 u∈Fn
u∈F2
2 /
supp(u)⊆I

This inequality shows that, contrary to the case of approximations by func-

tions of algebraic degrees at most r (higher order nonlinearity), it is sufficient
that the first-order nonlinearity of a combining function be high for avoiding
close approximations of f by functions of BF I,n (when I has small size).

67
An affine invariant criterion related to the maximum correlation and also
related to the “distance to linear structures” is the following: the distance to
the Boolean functions g such that the space {e ∈ Fn2 / De g = 0} has dimen-
sion at least k (the functions of BF I,n can be viewed as n-variable functions
N \I
g such that the set {e ∈ Fn2 / De g = 0} contains F2 ). The results on the
maximum correlation above generalize to this criterion [38].

- The main cryptographic complexity criteria for a Boolean function are

the algebraic degree and the nonlinearity, but other criteria have also been
studied: the minimum number of terms in the algebraic normal forms of all
affinely equivalent functions, called the algebraic thickness (studied in [66]
and first evoked in [276]), the maximum dimension k of those flats E such
that the restriction of f to E is constant (f is then called a k-normal func-
tion) or is affine (f is called a k-weakly-normal function) [66] (see Sub-
section 5.4), the number of nonzero coefficients of the Walsh transform
[302, 316]. It has been shown in [66, 289, 316] that (asymptotically) al-
most all Boolean functions have high complexities with respect to all these
criteria.
For every even integer k such that 4 ≤ k ≤ 2n , the kth-order nonhomomor-
phicity [362] of a Boolean function equals the number of k-tuples (u1 , · · · , uk )
of vectors of Fn2 such that u1 + · · · + uk = 0 and f (u1 ) ⊕ · · · ⊕ f (uk ) = 0.
It is a simple matter to show (more directly than in [362]) that it equals
2(k−1)n−1 + 2−n−1
P b k (u). This parameter should be small (but no
n f
u∈F2 χ

related attack exists on stream ciphers). It is maximum and equals 2(k−1)n if

nk
and only if the function is affine. It is minimum and equals 2(k−1)n−1 +2 2 −1
if and only if the function is bent, and some relationship obviouly exists be-
tween nonhomomorphicity and nonlinearity.

Conclusion of this subsection: As we can see, there are numerous crypto-

graphic criteria for Boolean functions. The ones which must be necessarily
satisfied are balancedness, a high algebraic degree, a high nonlinearity, a
high algebraic immunity and a good resistance to fast algebraic attacks. It
is difficult but not impossible to find functions satisfying good trade-offs
between all these criteria (see Section 9). It is not clear whether it is possi-
ble to achieve additionally resiliency of a sufficient order, which is necessary
for the combiner model. Hence, the filter model may be more appropriate
(future research will determine this). Once we know the criteria above are
satisfied by some function f (except resiliency), it is a simple matter to
render f 1-resilient by composing it with a linear automorphism (we just

68
need for this that there exist n linearly independent vectors at which the
Walsh transform vanishes). First-order resiliency is useful for resisting some
distinguishing (less dreadful) attacks.

5 Classes of functions for which restrictions on the

possible values of the weights, Walsh spectra
and nonlinearities can be proved
5.1 Affine functions
The weights and the Walsh spectra of affine functions are peculiar: the
Walsh transform of the function `(x) = a · x ⊕ ε takes null value at every
vector u 6= a and takes value 2n (−1)ε at a.
Concatenating affine functions gives the so-called Maiorana-McFarland func-
tions: for every n-variable function f , if we order all the binary words of
length n in lexicographic order, with the bit of higher weight on the right (for
instance), then the truth-table of f is the concatenation of the restrictions
of f obtained by setting the values of the (say) s last bits of the input and let-
ting the others freely range over F2 . If all these restrictions are affine then f
is called a Maiorana-McFarland function. These Maiorana-McFarland func-
tions will be studied in Section 6 (Subsection 6.4, for bent functions) and
Section 7 (Subsection 7.5, for resilient functions). The computation of their
weights, Walsh spectra and nonlinearities are easier than for general Boolean
functions, and in some cases can be completely determined.

5.2 Quadratic functions

The behavior of the functions of R(2, n), called quadratic functions, is also
peculiar. Recall that Relation (26) states that, for every Boolean function
f: X
F 2 (f ) = F(Db f ).
b∈Fn
2

If f is quadratic, then Db f is affine for every b ∈ Fn2 , and is therefore either

balanced or constant. Since F(g) = 0 for every balanced function g, we
deduce: X
F 2 (f ) = 2n (−1)Db f (0) , (41)
b∈Ef

where Ef is the set of all b ∈ Fn2

such that Db f is constant. The set Ef is the
linear kernel of f (see Subsection 4.1). In the case of quadratic functions,

69
it also equals the kernel {x ∈ Fn2 / ∀y ∈ Fn2 , ϕf (x, y) = 0} of the symplectic
(i.e. bilinear, symmetric, and null over the diagonal) form associated to f :
ϕf (x, y) = f (0) ⊕ f (x) ⊕ f (y) ⊕ f (x + y). The restriction of the function
b 7→ Db f (0) = f (b) ⊕ f (0) to this vectorspace is linear, as can be easily
checked; we deduce that F 2 (f ) equals 2n |Ef | if this linear form on Ef is
null, that is, if f is constant on Ef , and is null otherwise. According to
Relation (13), this proves the following:
Theorem 4 Any quadratic function f is balanced if and only if its restric-
tion to its linear kernel Ef ( i.e. the kernel of its associated symplectic form)
n+k
is not constant. If it is not balanced, then its weight equals 2n−1 ± 2 2 −1
where k is the dimension of Ef .
Note that Theorem 4 implies that f is balanced if and only if there exists
b ∈ Fn2 such that the derivative Db f (x) = f (x)⊕f (x+b) equals the constant
function 1 (take b in Ef such that f (b) 6= f (0)). For general Boolean func-
tions, this condition is sufficient for f being balanced, but it is not necessary.
Theorem 4 applied to f ⊕ `, where ` is a linear function such that f ⊕ ` is
not balanced (such function ` always exists, according to Parseval’s relation)
shows that the co-dimension of Ef must be even (this co-dimension is the
rank of ϕf ).
The weight of aquadratic function can be any element of the set {2n−1 } ∪
n−1 i n

{2 ± 2 ; i = 2 − 1, · · · , n − 1}. Its nonlinearity can be any element of
the set {2n−1 − 2i ; i = n2 − 1, · · · , n − 1}, and if f has weight 2n−1 ± 2i , then
for every affine function l, the weight of the function f ⊕ l belongs to the
set {2n−1 − 2i , 2n−1 , 2n−1 + 2i }.
Determining whether the weight is 2n−1 − 2i or 2n−1 + 2i (when the func-
tion is not balanced), and more generally studying the sign of the Walsh
transform is in general much more difficult than determining the value of i,
or equivalently the magnitude of the Walsh transform. In [226] is studied
the sign of the values of the Walsh transform of Gold and Kasami func-
tions. The former are quadratic (the latter are not but they are related to
quadratic functions, see the chapter “Vectorial Boolean Functions for Cryp-
tography”). In [164], the result of [226] is generalized: for every AB power
function xd over F2n (see definition in the chapter “Vectorial Boolean Func-
tions for Cryptography”) whose restriction to any subfield of F2n is also AB,
d n+1 n+1
the value x∈F2n (−1)trn (x +x) equals 2 2 if n ≡ ±1 [mod 8] and −2 2 if
P
n ≡ ±3 [mod 8].
Any quadratic non-affine function f having a monomial of degree 2 in
its ANF, we can assume without loss of generality that, up to a non-
singular linear transformation, this monomial is x1 x2 . The function has

70
then the form x1 x2 ⊕ x1 f1 (x3 , · · · , xn ) ⊕ x2 f2 (x3 , · · · , xn ) ⊕ f3 (x3 , · · · , xn )
where f1 , f2 are affine functions and f3 is quadratic. Then, f (x) equals
(x1 ⊕ f2 (x3 , · · · , xn ))(x2 ⊕ f1 (x3 , · · · , xn )) ⊕ f1 (x3 , · · · , xn )f2 (x3 , · · · , xn ) ⊕
f3 (x3 , · · · , xn ) and is therefore affinely equivalent to the function x1 x2 ⊕
f1 (x3 , · · · , xn )f2 (x3 , · · · , xn ) ⊕ f3 (x3 , · · · , xn ). Applying this method recur-
sively shows:
Theorem 5 Every quadratic non-affine function is affinely equivalent to
x1 x2 ⊕ · · · ⊕ x2l−1 x2l ⊕ x2l+1 (where l ≤ n−1 2 ) if it is balanced, to x1 x2 ⊕
· · · ⊕ x2l−1 x2l (where l ≤ n/2) if it has weight smaller than 2n−1 and to
x1 x2 ⊕ · · · ⊕ x2l−1 x2l ⊕ 1 (where l ≤ n/2) if it has weight greater than 2n−1 .
This allows describing precisely the weight distribution of R(2, n) [258].

Remark. Let f1 , f2 and f3 be any Boolean functions on Fn2 . Define the

function on Fn+2
2 : f (x, y1 , y2 ) = y1 y2 ⊕ y1 f1 (x) ⊕ y2 f2 (x) ⊕ f3 (x). Then we
have
X
F(f ) = (−1)(y1 ⊕f2 (x))(y2 ⊕f1 (x))⊕f1 (x)f2 (x)⊕f3 (x)
x∈Fn
2 / y1 ,y2 ∈F2

X X
= (−1)y1 y2 ⊕f1 (x)f2 (x)⊕f3 (x) = 2 (−1)f1 (x)f2 (x)⊕f3 (x) .
x∈Fn
2 / y1 ,y2 ∈F2
x∈Fn
2

So, starting with a function g = f1 f2 ⊕ f3 , we can relate F(g) to F(f ), on

two more variables, in which the term f1 f2 has been replaced by y1 y2 ⊕
y1 f1 (x) ⊕ y2 f2 (x). Applying this repeatedly (“breaking” this way all the
monomials of degrees at least 4), this allows showing easily (see [52]) that,
for every Boolean function g on Fn2 , there exists an integer m and a Boolean
function f of algebraic degree at most 3 on Fn+2m 2 whose Walsh transform
takes value fbχ (0) = 2m gc χ (0) at 0. As we already mentioned, this proves
that the functions of algebraic degree 3 can have weights much more diverse
than functions of degrees at most 2.  
P n−1
2 2i +1
The trace representation of quadratic functions is trn β∅ + i=0 βi x
 P n2 −1 i
 n/2
for n odd and trn β∅ + i=0 βi x2 +1 + tr n2 (γx2 +1 ) for n even, where
the βi ’s belong to F2n and γ belongs to F2n/2 . For n odd, the quadratic func-
n−1
tions of nonlinearity 2n−1 − 2 2 (called semi-bent functions because their
n+1
extended Walsh spectra only contain the values 0 and 2 2 ) of the form
P(n−1)/2 i
trn ( i=1 ci x2 +1 ) have been studied in [106] and cubic bent functions
have been deduced by concatenation of these semi-bent functions. Further

71
functions of this kind have been given and studied in [197, 217].
Concatenating quadratic functions gives a super-class of the class of Maiorana-
McFarland functions, studied in [64], and presented in Section 7 (Subsection
7.5.2) below.

5.3 Indicators of flats

As we have already seen, a Boolean function f is theQ indicator of a flat A of
co-dimension r if and only if it has the form f (x) = ri=1 (ai · x ⊕ εi ) where
a1 , · · · , ar ∈ Fn2 are linearly independent and ε1 , · · · , εr ∈ F2 . Then f has
weight 2n−r . Moreover, set a ∈ Fn2 . If a is linearly independent of a1 , · · · , ar ,
then the function f (x)⊕a·x is balanced (and hence fbχ (a) = 0), since it is lin-
early equivalent to a function of the Pform g(x1 , · · · , xr )⊕xr+1 . If a is linearly
dependent of a1 , · · · , ar , say a = ri=1 ηi ai , then a · x takes constant value
Lr Lr
η ( ⊕ 1) on the flat; hence, fb(a) = x∈A (−1)a·x
P
i=1 ηi (ai · x) = Lr i i
i=1
i (i ⊕1) . Thus, if a = r
equals then 2n−r (−1) i=1 ηL
P
i=1 ηi ai 6= 0, then we
r
have fχ (a) = −2
b n−r+1 (−1) i=1 ηi (i ⊕1) ; and we have fbχ (0) = 2n − 2|A| =
n
2 −2 n−r+1 .
Note that the nonlinearity of f equals 2n−r and is bad as soon as r ≥ 2. But
indicators of flats can be used to design Boolean functions with good non-
linearities: concatenating sums of indicators of flats and of affine functions
gives another super-class of the Maiorana-McFarland functions, studied in
[68] and presented in Section 7 (Subsection 7.5.2) below.

Note. As recalled in Section 3.1, the functions of R(r, n) whose weights

occur in the range [2n−r ; 2n−r+1 [ have been characterized by Kasami and
Tokura [207]; any such function is the product of the indicator of a flat and
of a quadratic function or is the sum (modulo 2) of two indicators of flats.
The Walsh spectra of such functions can also be precisely computed.

5.4 Normal functions

Let E and E 0 be subspaces of Fn2 such that E ∩ E 0 = {0} and whose direct
sum equals Fn2 . Denote by k the dimension of E. For every a ∈ E 0 , let ha be
the restriction of f to the coset a + E. Then, Relation (28) in Proposition 9
implies
2 X
maxn fbχ (u) ≥ F 2 (ha )
u∈F2
a∈E 0

72
 2
(indeed, the maximum of fbχ (u) is greater than or equal to its mean). Hence
2
we have: maxu∈Fn fb (u) ≥ F 2 (ha ) for every a. Applying this property to
2 χ
f ⊕ `, where ` is any linear function, and using Relation (35) relating the
nonlinearity of a function to the maximum magnitude of its Walsh transform,
we deduce:
∀a ∈ E 0 , nl(f ) ≤ 2n−1 − 2k−1 + nl(ha ). (42)
This bound was first proved (in a different way) by Zheng et al. in [364].
The present proof is from [42]. Relation (42) can also be deduced from the
Poisson summation formula (17) applied to the sign function of f , and in
P the roles
which of E and
 E ⊥ are exchanged: let us choose b ∈ Fn2
such that
f (x)⊕b·x  is maximum, that is, equals 2k − 2nl(h ) . Then
x∈a⊕E (−1) a

 
   
 X a·u b
 ⊥ k

 (−1) fχ (u)
 = |E | 2 − 2nl(ha ) .
u∈b⊕E ⊥ 

Then the mean of (−1)a·u fbχ (u), when u ranges over b ⊕ E ⊥ , is equal to
± 2k − 2nl(ha ) . Thus, the maximum magnitude of fbχ (u) is greater than

or equal to 2k − 2nl(ha ). This implies Relation (42). These two methods,

for proving (42), lead to two different necessary conditions for the case of
equality (see [66]).
Relation (42) implies in particular that, if the restriction of f to a k-
dimensional flat of Fn2 is affine (say equals `), then nl(f ) ≤ 2n−1 − 2k−1 ,
and that, if equality occurs, then f ⊕ ` is balanced on every other coset of
this flat.

Definition 4 A function is called k-weakly-normal (resp. k-normal) if its

restriction to some k-dimensional flat is affine (resp. constant).

H. Dobbertin introduced this terminology by calling normal the functions

that we call n/2-normal here (we shall also call normal the n/2-normal func-
tions, in the sequel). He used this notion for constructing balanced functions
with high nonlinearities (see Subsection 7.5.1). It is proved in [66] that, for
every α > 1, when n tends to infinity, random Boolean functions are almost
surely [α log2 n]-non-normal. This means that almost all Boolean functions
have high complexity with respect to this criterion. As usual, the proof
of existence of non-normal functions does not give examples of such func-
tions. Alon, Goldreich, Hastad and Peralta give in [2] several constructions
of functions which are nonconstant on flats of dimension n/2. This is not
explicitly mentioned in the paper. What is shown is that the functions are

73
not constant on flats defined by equations xi1 = a1 , ..., xin/2 = an/2 . As the
proof still works when composing the function by an affine automorphism,
it implies the result.
There are also explicit constructions which work for dimensions (1/2 − ) n,
for some small  > 0 very recently found by Jean Bourgain [24].
Functions which are nonconstant on flats of dimensions nδ for every δ > 0
are also given in [14]. These constructions are very good asymptotically (but
may not be usable to obtain functions in explicit numbers of variables).
As far as we know, no construction is known below nδ .

5.5 Functions admitting partial covering sequences

The notion of covering sequence of a Boolean function has been introduced
in [95].

Definition 5 Let f be an n-variable Boolean function. An integer-valued31

sequence P (λa )a∈Fn2 is called a covering sequence of f if the integer-valued
function a∈Fn λa Da f (x) takes a constant value. This constant value is
2
called the level of a covering sequence. If the level is nonzero, we say that
the covering sequence is a non-trivial covering sequence.
P
Note that the sum a∈Fn λa Da f (x) involves both kinds of additions: the
P 2
addition in Z and the addition ⊕ in F2 (which is concealed inside Da f ).
It was shown in [95] that any function admitting a non-trivial covering se-
quence is balanced (see Theorem 6 below for a proof) and that any balanced
function admits the constant sequence 1 as covering sequence (the level of
this sequence is 2n−1 ).
A characterization of covering sequences by means of the Walsh transform
was also given in [95]: denote again by Sfb the support {u ∈ Fn2 | fbχ (u) 6= 0}
χ

of fχ ; then f admits an integer-valued sequence λ = (λa )a∈Fn2 as covering

b
sequence if and only if the Fourier transform λ b of the function a 7→ λa takes
a constant value on Sfb . Indeed, replacing Da f (x) by 12 − 21 (−1)Da f (x) =
χ
1 1 f (x) (−1)f (x+a) in the equality
P
2 − 2 (−1) a∈Fn λa Da f (x) = ρ, we see that
2
f admits the covering sequence λ with level P ρ if and only
 if, for every
n f (x+a) λa − 2ρ (−1)f (x) ; since
P
x ∈ F2 , we have a∈Fn λa (−1) = a∈Fn
2 2
two integer-valued functions are equal if and only if their Fourier transforms
31
or real-valued, or complex-valued; but taking real or complex sequences instead of
integer-valued ones has no practical sense.

74
are equal, the characterization follows,
P thanks to the  straightforward rela-
f (x+a)+x·b a·b
P
tion a,x∈Fn λa (−1) = a∈Fn λa (−1) fbχ (b) = λ(b)
b fb (b).
χ
2 2
Knowing a covering sequence (trivial or not) of a function f allows knowing
that all the vectors a such that f (x) ⊕ a · x is non-balanced belong to the set
b−1 (µ), where µ = λ(0)−2ρ
λ b is the constant value of λ b on S b ; hence, if f ad-
χ f
mits a covering sequence λ = (λa )a∈Fn2 with level ρ (resp. with level ρ 6= 0),
then f is k-th order correlation-immune (resp. k-resilient) where k + 1 is
the minimum Hamming weight of nonzero b ∈ Fn2 such that λ(b) b = µ. Con-
versely, if f is k-th order correlation-immune (resp. k-resilient) and if it is
not (k + 1)-th order correlation-immune (resp. (k + 1)-resilient), then there
exists at least one (non-trivial) covering sequence λ = (λa )a∈Fn2 with level
ρ such that k + 1 is the minimum Hamming weight of b ∈ Fn2 satisfying
λ(b)
b = λ(0) b − 2ρ.
A particularly simple covering sequence is the indicator of the set of vectors
of weight one. The functions which admit this covering sequence are called
regular; they are (ρ − 1)-resilient (where ρ is the level); more generally, any
function, admitting as covering sequence the indicator of a set of vectors
whose supports are disjoint, has this same property. See further properties
in [95].
But knowing a covering sequence for f gives no information on the non-
linearity of f , since it gives only information on the support of the Walsh
transform, not on the nonzero values it takes. In [69] is weakened the def-
inition of covering sequence, so that it can help computing the (nonzero)
values of the Walsh transform.

Definition 6 Let f be a Boolean function on P Fn2 . A partial covering se-

quence for f is a sequence (λa )a∈Fn2 such that a∈Fn λa Da f (x) takes two
2
values ρ and ρ0 (distinct or not) called the levels of the sequence. The par-
tial covering sequence is called non-trivial if one of the constants is nonzero.

A simple example of non-trivial partial covering sequence is as follows: let E

be any set of derivatives of f . Assume that E contains a nonzero function
and is stable under addition (i.e. is a non-trivial F2 -vectorspace). Then
P |E|
g∈E g takes on values 0 and 2 . Thus, if E = {Da f / a ∈ E} (where
we choose E minimum, so that any two different vectors of the set E give
different functions of E), then 1E is a non-trivial partial covering sequence.
The interest of non-trivial partial covering sequences is that they allow
simplifying the computation of the weight and of the Walsh transform of f .

75
Theorem 6 Let (λa )a∈Fn2 be a partial covering sequence of a Boolean func-
tion f , of levels ρ and 0
P ρ.
Let A = {x ∈ F2 / a∈Fn λa Da f (x) = ρ0 } (assuming that ρ0 6= ρ; otherwise,
n
2
when λ is in fact a covering sequence of level ρ, we set A = ∅).
Then, for every vector b ∈ Fn2 , we have:
  X
b − λ(0)
λ(b) b + 2 ρ fbχ (b) = 2 (ρ − ρ0 ) (−1)f (x)⊕b·x .
x∈A

Proof. By definition, we have, for every x ∈ Fn2 :

X
λa Da f (x) = ρ0 1A (x) + ρ 1Ac (x)
a∈Fn
2

and therefore:
X X
λa (−1)Da f (x) = λa (1 − 2 Da f (x))
a∈Fn
2 a∈Fn
2
X
= λa − 2 ρ0 1A (x) − 2 ρ 1Ac (x).
a∈Fn
2

We deduce:
 
X X
λa (−1)f (x+a) = (−1)f (x)  λa − 2 ρ0 1A (x) − 2 ρ 1Ac (x) . (43)
a∈Fn
2 a∈Fn
2

The Fourier transform of the function (−1)f (x+a) maps every vector b ∈ Fn2
to the value x∈Fn (−1)f (x+a)⊕x·b = x∈Fn (−1)f (x)⊕(x+a)·b = (−1)a·b fbχ (b).
P P
2 2
Hence, taking the Fourier transform of both terms of equality (43), we get:
 
X
 λa (−1)a·b  fbχ (b) =
a∈Fn
2

 
X X X
 λa  fbχ (b) − 2 ρ0 (−1)f (x)⊕b·x − 2 ρ (−1)f (x)⊕b·x ,
a∈Fn
2 x∈A x∈Ac

that is
X
λ(b)
b fb (b) = λ(0)
χ
b fbχ (b) − 2 ρ fbχ (b) + 2 (ρ − ρ0 ) (−1)f (x)⊕b·x .
x∈A

76
Hence:
  X
b − λ(0)
λ(b) b + 2 ρ fbχ (b) = 2 (ρ − ρ0 ) (−1)f (x)⊕b·x . 2
x∈A

Hence, if ρ 6= 0, we have in particular an information on the weight of f :

ρ0 X
 
n
2 − 2wH (f ) = fχ (0) = 1 −
b (−1)f (x) .
ρ
x∈A

Examples are given in [69] of computations of the weights or Walsh

spectra of some Boolean functions (quadratic functions, Maiorana-McFar-
land’s functions and their extensions, and other examples of functions), using
Theorem 6.

5.6 Functions with low univariate degree

The following Weil’s Theorem is very well-known in finite field theory (cf.
[248, Theorem 5.38]):

Theorem 7 Let q be a prime power and f ∈ Fq [x] a univariate polynomial

of degree d ≥ 1 with gcd(d, q) = 1. Let χ be a non-trivial character of Fq .
Then  
X 
χ(f (x)) ≤ (d − 1) q 1/2 .
 


x∈Fq 

For q = 2n , this Weil’s  bound means that, for every nonzero a ∈ F2n :
tr (af (x))
 ≤ (d − 1) 2n/2 . And since adding a linear function
P
 x∈F2n (−1) n
 

trn (bx) to the function trn (af (x)) corresponds to adding (b/a) x to f (x) and
does not change its univariate degree, we deduce that, if d > 1 is odd and
a 6= 0, then:
nl(trn (af )) ≥ 2n−1 − (d − 1) 2n/2−1 .
An extension of the Weil bound to the character sums of functions of the
n −2
form f (x) + g(1/x) (where 1/x = x2 P takes value 0 at 0), among which
are the so-called Kloosterman sums trn (1/x+ax) , has been first
x∈F2n (−1)
obtained by Carlitz and Uchiyama [97] and extended by Shanbhag, Kumar
and Helleseth [330]: if f and g have odd univariate degrees, then
X
(−1)trn (f (1/x)+g(x)) ≤ (d◦ f + d◦ g)2n/2 .
x∈F2n

77
6 Bent functions
We recall the definition of bent functions:

Definition 7 A Boolean function f on Fn2 (n even) is called bent if its

Hamming distance to the set R(1, n) of all n-variable affine functions (the
nonlinearity of f ) equals 2n−1 − 2n/2−1 (the covering radius of the Reed-
Muller code of order 1).

Equivalently, as seen in Subsection 4.1, f is bent if and only if fbχ takes on

values ±2n/2 only (this characterization is independent of the choice of the
inner product on Fn2 , since any other inner product has the form hx, si = x ·
L(s), where L is an auto-adjoint linear automorphism, i.e. an automorphism
whose associated matrix is symmetric). Hence, f is bent if and only if its
distance to any affine function equals 2n−1 ± 2n/2−1 . Note that, for any
bent function f , half of the elements of the Reed-Muller code of order 1
lie at distance 2n−1 + 2n/2−1 from f and half lie at distance 2n−1 − 2n/2−1
(indeed, if ` lies at distance 2n−1 + 2n/2−1 from f , then ` ⊕ 1 lies at distance
2n−1 − 2n/2−1 and vice versa). In fact, the condition on fbχ can be weakened,
without losing the property of being necessary and sufficient:

Lemma 2 Any n-variable (n even ≥2) Boolean function f is bent if and

only if, for every a ∈ Fn2 , fbχ (a) ≡ 2n/2 mod 2n/2+1 , or equivalently fb(a) ≡


2n/2−1 mod 2n/2 .

 

Proof. This necessary condition is also sufficient, since, if it is satisfied, then

fbχ (a) = 2n/2 λa , where λa is odd for every a, Parseval’s Relation (23)
writing P
implies a∈Fn λ2a = 2n , which implies that λ2a = 1 for every a. 2
2

A slightly different viewpoint is that of bent sequences32 but we shall not

adopt it here because it most often gives no extra insight on the problems.
The nonlinearity being an affine invariant, so is the notion of bent function.
Clearly, if f is bent and ` is affine, then f ⊕` is bent. A class of bent functions
is called a complete class of functions if it is globally invariant under the
n
32
For each vector X in {−1, 1}2 , define: X̂ = √12n Hn X, where Hn is the Walsh-
Hadamard matrix, recursively defined by:
» –
Hn−1 Hn−1
Hn = , H0 = [1].
Hn−1 −Hn−1
n
The vectors X such that X̂ belongs to {−1, 1}2 are called bent sequences. They are the
images by the character χ = (−1)· of the bent functions on Fn
2.

78
action of the general affine group and the addition of affine functions.
The automorphism group of the set of bent functions is the general affine
group. This is a direct consequence of the property that, given a Boolean
function g, if for every bent function f , function f ⊕ g is also bent, then g
has degree at most 1 (which is easily proved).
Thanks to Relation (25) and to the fact that the Fourier transform of a
function is constant if and only if the function equals δ0 times some constant,
we see that any function f is bent if and only if, for any nonzero vector a,
the Boolean function Da f (x) = f (x) ⊕ f (x + a) is balanced. In other words:

Theorem 8 Any n-variable Boolean function (n even33 ) is bent if and only

if it satisfies P C(n).

For this reason, bent functions are also called perfect nonlinear functions 34 .
Equivalently, f is bent if and only if the 2n ×2n matrix H = [(−1)f (x+y) ]x,y∈Fn2
is a Hadamard matrix (i.e. satisfies H × H t = 2n I, where I is the identity
matrix), and if and only if the support of f is a difference set 35 of the ele-
mentary Abelian 2-group Fn2 [136, 204] (other types of difference sets exist,
see e.g. [139]). This implies that the Cayley graph Gf (see Subsection 2.2.2)
is strongly regular (see [18] for more precision).
The functions whose derivatives Da f , a ∈ H, a 6= 0 are all balanced, where
H is a linear hyperplane of Fn2 , are characterized in [41, 42] for every n; they
are all bent if n is even. The functions whose derivatives Da f , a ∈ E, a 6= 0
are all balanced, where E is a vector subspace of Fn2 of dimension n − 2, are
also characterized in these two papers.
Bent Pfunctions have the property that, for every even positive integer w,
the sum a∈Fn fbχ w (a) is minimum. Such sums (for even or odd w) play a
2
role with respect to fast correlation attacks [47, 40] (when these sums have
small magnitude for low values of w, this contributes to a good resistance
to fast correlation attacks).

33
In fact, according to the observations above, “n even” is implied by “f satisfies
P C(n)”; functions satisfying P C(n) do not exist for odd n.
34
The characterization of Theorem 8 leads to a generalization of the notion of bent
function to non-binary functions. In fact, several generalizations exist [3, 220, 257] (see [78]
for a survey); the equivalence between being bent and being perfect nonlinear is no more
valid if we consider functions defined over residue class rings (see [80]).
35
Thus, bent functions are also related to designs, since any difference set can be used
to construct a symmetric design, see [11], pages 274-278. The notion of difference set
is anterior to that of bent function, but it had not been much studied in the case of
elementary 2-groups before the introduction of bent functions.

79
A last way of looking at bent functions deals with linear codes: let f be any
n-variable Boolean function (n even). Denote its support {x ∈ Fn2 | f (x) = 1}
by Sf and write Sf = {u1 , · · · , uwH (f ) }. Consider a matrix G whose columns
are all the vectors of Sf , without repetition, and let C be the linear code
generated by the rows of this matrix. Thus, C is the set of all the vectors
Uv = (v · u1 , · · · , v · uwH (f ) ), where v ranges over Fn2 . Then:

Proposition 16 Let n be any even positive integer. Any n-variable Boolean

function f is bent if and only if the linear code C defined above has dimen-
sion n (i.e. G is a generator matrix of C) and has exactly two nonzero
Hamming weights: 2n−2 and wH (f ) − 2n−2 .
v·x
f (x) 1−(−1) fb(0)−fb(v)
P P
Indeed, wH (Uv ) equals x∈Fn f (x)×v·x = x∈Fn 2 = 2 .
2 2
fbχ (v)−fbχ (0)
Hence, according to Relation (12), wH (Uv ) equals 2n−2 + 4 , for ev-
ery nonzero vector v. Thus, C has dimension n and has the two nonzero
Hamming weights 2n−2 and wH (f )−2n−2 if and only if, for every v 6= 0, Uv is
nonzero and fbχ (v) = fbχ (0) or fbχ (v) = fbχ (0)+4wH (f )−2n+1 = fbχ (0)−2fbχ (0) =
−fbχ (0). If f is bent, then this condition is clearly satisfied. Conversely, ac-
cording to Parseval’s Relation (23), if this condition is satisfied, then fbχ (v)
equals ±2n/2 for every v, i.e. f is bent.

There exist two other characterizations [353] dealing with C:

1. C has dimension n and C has exactly two weights, whose sum equals
wH (f );
2. The length wH (f ) of C is even, C has exactly two weights, and one of
these weights is 2n−2 .

6.1 The dual

If f is bent, then the dual function fe of f , defined on Fn2 by:

fbχ (u) = 2n/2 (−1)f (u)

e

is also bent and its own dual is f itself. Indeed, the inverse Fourier transform
property (19) applied to ϕ = fχ (the sign function of f ) gives, for every
vector a: u∈Fn (−1)f (u)⊕a·u = 2n/2 fχ (a) = 2n/2 (−1)f (a) .
P e
2
Let f and g be two bent functions, then Relation (22) applied with ϕ = fχ
and ψ = gχ shows that

F(fe ⊕ ge) = F(f ⊕ g). (44)

80
Thus, f ⊕ g and fe ⊕ ge have the same weight and the mapping f 7→ fe is an
isometry.
According to Proposition 6, for every a, b ∈ Fn2 and for every bent function
f , the dual of the function f (x + b) ⊕ a · x equals fe(x + a) ⊕ b · (x + a) =
fe(x + a) ⊕ b · x ⊕ a · b. Denoting b · x by `b (x), Relation (44), applied with
g(x) = f (x + b) ⊕ a · x, gives F(Da fe⊕ `b ) = (−1)a·b F(Db f ⊕ `a ), and applied
with g(x) = f (x) ⊕ `a (x) and with f (x + b) in the place of f (x), it gives the
following property, first observed in [61] (and rediscovered in [43]):

F(Da fe ⊕ `b ) = F(Db f ⊕ `a ) (45)

(from these two relations, we deduce that, if a · b = 1, then F(Da fe ⊕ `b ) =

F(Db f ⊕ `a ) = 0). Notice that, for every a and b, we have Db f = `a ⊕  if
and only if Da fe = `b ⊕ ).
Moreover, if a pair of Boolean functions f and f 0 satisfies the relation
F(Da f 0 ⊕ `b ) = F(Db f ⊕ `a ), then these functions are bent (indeed, taking
a = 0 shows that Db f is balanced for every b 6= 0 and taking b = 0 shows that
Da f 0 is balanced for every a 6= 0), and are then the duals of each other up to
the addition of a constant. Indeed, summing up the relation F(Da f 0 ⊕ `b ) =
F(Db f ⊕ `a ) for b ranging P over Fn2 shows that f 0 (0) ⊕ f 0 (a) = fe(0) ⊕ fe(a)
0 0 0 0
for every a, since we have x,b∈Fn (−1)f (x)⊕f (x+a)⊕b·x = 2n (−1)f (0)⊕f (a) ,
2
f (x)⊕f (x+b)⊕a·x = fb (0) × fb (a).
P
and x,b∈F2 n (−1) χ χ

The NNF of fe can be deduced from the NNF of f . Indeed, using equality
fe
fe = 1−(−1)
2 , we have fe = 21 − 2−n/2−1 fbχ = 12 − 2n/2−1 δ0 + 2−n/2 fb (according
to Relation (12)). Applying now Relation (30) (expressing the value of the
Fourier transform P by meansI of the coefficients of the NNF) to ϕ = f , we
deduce that if I∈P(N ) λI x is the NNF of f then:

1 X
fe(x) = − 2n/2−1 δ0 (x) + (−1)wH (x) 2n/2−|I| λI .
2
I∈P(N ) | supp(x)⊆I

Changing I into N \I in this relation, and observing that supp(x) is included

in N \ I if and only if xi = 0, ∀i ∈ I, we obtain the NNF of fe by expanding
the following relation:
n
1 Y X Y
f (x) = − 2
e n/2−1
(1 − xi ) + (−1)wH (x) 2|I|−n/2 λN \I (1 − xi ).
2
i=1 I∈P(N ) i∈I

We deduce (as shown in [87]):

81
Proposition 17 Let f be any n-variable bent function (n even). For every
I 6= N such that |I| > n/2, the coefficient of xI in the NNF of fe (resp. of f )
is divisible by 2|I|−n/2 .

Reducing this equality modulo 2 proves Rothaus’ bound (see below) and
that, for n ≥ 4 and |I| = n/2, the coefficient of xI in the ANF of f˜ equals
the coefficient of xN \I in the ANF of f . Using Relation (9), the equality
above can be related to the main result of [191] (but this result by Hou was
stated in a complex way).

The Poisson summation formula (17) applied to ϕ = fχ gives (see [54])

that for every vector subspace E of Fn2 , and for every elements a and b of
Fn2 , we have:
X X
(−1)f (x)⊕b·x = 2−n/2 |E| (−1)a·b (−1)f (x)⊕a·x . (46)
e

x∈a+E x∈b+E ⊥

Self-dual bent functions are studied in [77].

6.2 Bent functions of low algebraic degrees

Obviously, no affine function can be bent. All the quadratic bent functions
are known: according to the properties recalled in Subsection 5.2, any such
function
M
f (x) = ai,j xi xj ⊕ h(x) (h affine, ai,j ∈ F2 )
1≤i<j≤n

is bent if and only if one of the following equivalent properties is satisfied:

1. its Hamming weight is equal to 2n−1 ± 2n/2−1 ;

2. its associated symplectic form: ϕf : (x, y) 7→ f (0) ⊕ f (x) ⊕ f (y) ⊕

f (x + y) is non-degenerate (i.e. has kernel {0});

3. the skew-symmetric matrix M = (mi,j )i,j∈{1,···,n} over F2 , defined by:

mi,j = ai,j if i < j, mi,j = 0 if i = j, and mi,j = aj,i if i > j,
is regular (i.e. has determinant 1); indeed, M is the matrix of the
bilinear form ϕf ;

4. f (x) is equivalent, up to an affine nonsingular transformation, to the

function: x1 x2 ⊕ x3 x4 ⊕ · · · ⊕ xn−1 xn ⊕ ε (ε ∈ F2 ).

82
It is interesting to charaterize quadratic bent functions in the trace repre-
sentation. This leads for instance to the Kerdock code; see Subsection 6.10
where the bent functions leading to this code are given.
i
Let us study for example the case of the Gold function trn (vx2 +1 ), where
gcd(i, n) = 1. It is bent if and only if there is no nonzero x ∈ F2n such that
i i i n−i
trn (vx2 y+vxy 2 ) = 0 for every y ∈ F2n , i.e., the equation vx2 +(vx)2 =0
has no non-zero solution. Raising this equation to the 2i -th power gives
i 2i
v 2 x2 + vx = 0 and 2i − 1 being co-prime with 2n − 1, it is equivalent, after
i
dividing by vx (when x 6= 0) and taking the (2i − 1)th root, to vx2 +1 ∈ F2 .
i
Hence, the function trn (vx2 +1 ) is bent if and only if v is not the (2i + 1)-th
power of an element of F2n , that is (since gcd(2i +1, 2n −1) = 3), v is not the
third power of an element of F2n . The same result exists with the Kasami
2i i
function trn (x2 −2 +1 ), gcd(i, n) = 1 (this is proved in [139, Theorem 11]
for n not divisible by 3 and true also for n divisible by 3 as seen by Leander
[240]).
Another example of quadratic bent function in the trace representation uses
two trace functions, the trace function trn on the whole field F2n and the
P n2 −1 2i +1
trace function tr n2 on the subfield F2n/2 , is: f (x) = trn ( i=1 x )⊕
n/2
tr n2 (x2 +1 ).
A third example did not appear yet in the literature (as far as we know): let
n be coprime with 3 and i be coprime with n, then the function f (x, y) =
i i
tr n2 (x2 +1 + y 2 +1 + xy), x, y ∈ F2n/2 is bent. Indeed, its associated symplec-
tic form equals the function ((x, y), (x0 , y 0 )) → f (0, 0) ⊕ f (x, y) ⊕ f (x0 , y 0 ) ⊕
i i i i
f (x + x0 , y + y 0 ) = tr n2 (x2 x0 + xx0 2 + y 2 y 0 + yy 0 2 + xy 0 + x0 y). The kernel of
( ( i n−i
)
2 x2 + x2 +y =0
this symplectic form equals (x, y) ∈ F2n/2 / i n−i ; this
y2 + y2 +x=0
i n−i
set is reduced to {(0, 0)}, since denoting z = x+y we have z 2 +z 2 +z = 0
2i i 3i
which implies z 2 = z 2 + z and therefore z 2 = z, that is z ∈ F23i , and
therefore z ∈ F2 and since 1 is not solution z = 0. Then x and y must be null.

Open problem: characterize the bent functions of algebraic degrees at

least 3 (that is, classify them under the action of the general affine group).
This has been done for n ≤ 6 in [315] (see also [302] where the number
of bent functions is computed for these values of n). For n = 8, it has
been done in [190], for functions of algebraic degrees at most 3 only; all
of these functions have at least one affine derivative Da f , a 6= 0 (it has
been proved in [43] that this happens for n ≤ 8 only). The determination
of all bent 8-variable functions has been completed very recently, see [233].

83
Hans Dobbertin (with G. Leander) has presented in the posthumous paper
[143] a nice approach for generating new bent functions by recursively gluing
so-called Z-bent functions.

6.3 Bound on algebraic degree

The algebraic degree of any Boolean function f being equal to the maximum
size of the multi-index I such that xI has an odd coefficient in the NNF of f ,
Proposition 17 gives:

Proposition 18 Let n be any even integer greater than or equal to 4. The

algebraic degree of any bent function on Fn2 is at most n/2.

In the case that n = 2, the bent functions have degree 2, since they have
odd weight (in fact, they are the functions of odd weights).
The bound of Proposition 18 (which is obviously also true for fe) was first
proved in [315] and will be called Rothaus’ bound in the sequel. It can
also be proved (see below) by using a similar method as in the proof of
Proposition 11. This same method also allows obtaining a bound, shown
in [192], relating the gaps between n/2 and the algebraic degrees of f and
fe:

Proposition 19 The algebraic degrees of any n-variable bent function and

of its dual satisfy:
n/2 − d◦ fe
n/2 − d◦ f ≥ . (47)
d◦ fe − 1
A proof of Proposition 19 and a second proof of Proposition 18. Let us de-
note by d (resp. by d)
e the algebraic degree of f (resp. of fe) and consider a
I
term x of degree d in the ANF of f . The Poisson summation formula (18)
applied to ϕ = fχ (or Relation (46) with a = b = 0) and to the vectorspace
E = {u ∈ Fn2 / ∀i ∈ I, ui = 0} gives u∈E (−1)f (u) = 2n/2−d x∈E ⊥ fχ (x).
P e P
The orthogonal E ⊥ of E equals {u ∈ Fn2 / ∀i 6∈ I, ui = 0}. Accord-
ing to Relation (3), the restriction of f to E ⊥ has odd weight w, thus
d fe(u) is not
P P
x∈E ⊥ fχ (x) = 2 − 2w is not divisible by 4. Hence, u∈E (−1)
divisible by 2n/2−d+2 . We deduce the proof of Proposition 18: suppose that
d > n/2, then u∈E (−1)f (u) is not even, a contradiction with the fact that
P e

E has an even size. We prove now Proposition 19: according l to

m McEliece’s
n−d
fe(u) is divisible by 2 de . We deduce
P
theorem (or Ax’s theorem), u∈E l (−1)
m
the inequality n/2 − d + 2 > n−d e , that is, n/2 − d + 1 ≥
n−d
e , which is
d d

84
equivalent to (47). 2

Using Relation (7) instead of Relation (3) gives a more precise result than
Proposition 18, first shown in [87], which will be given in Subsection 6.6.
Proposition 19 can also be deduced from Proposition 17 and from some
divisibility properties, shown in [87], of the coefficients of the NNFs of
Boolean functions of algebraic degree d.
More on the algebraic degree of bent functions can be said for homoge-
neous functions (whose ANF contain monomials of fixed degree), see [279].

6.4 Constructions
There does not exist for n ≥ 10 a classification of bent functions under the
action of the general affine group. In order to understand better the struc-
ture of bent functions, we can try to design constructions of bent functions.
It is useful also to deduce constructions of highly nonlinear balanced func-
tions. Some of the known constructions of bent functions are direct, that is,
do not use as building blocks previously constructed bent functions. We will
call primary constructions these direct constructions. The others, sometimes
leading to recursive constructions, will be called secondary constructions.

6.4.1 Primary constructions

1. The Maiorana-McFarland original class M (see [136, 273]) is the set of
n/2
all the Boolean functions on Fn2 = {(x, y); x, y ∈ F2 }, of the form:

f (x, y) = x · π(y) ⊕ g(y) (48)

n/2 n/2
where π is any permutation on F2 and g any Boolean function on F2 (“·”
n/2
denotes here an inner product in F2 ). Any such function is bent. More pre-
cisely, the bijectivity of π is a necessary and sufficient condition36 for f being
bent, according to Relation (49) below, applied with r = n/2. Note that
for every function h(y), the function f (x, y) ⊕ h(y) is bent. This property
is characteristic of the functions of the form (48); indeed, taking h = δa ,
n/2
the indicator of the singleton {a}, we have P for every a, u, v ∈ F2 that
±2n/2 = x,y∈Fn/2 (−1)f (x,y)⊕u·x⊕v·y⊕δa (y) = x,y∈Fn/2 (−1)f (x,y)⊕u·x⊕v·y −
P
2 2
2 x∈Fn/2 (−1)f (x,a)⊕u·x⊕v·a = ±2n/2 ± 2 x∈Fn/2 (−1)f (x,a)⊕u·x . Hence for
P P
2 2

36
It is, because the input has been cut in two pieces x and y of the same length; it
is also possible to cut them in pieces of different lengths, see Proposition 20 below, and
bentness is then obviously not characterized by the bijectivity of π.

85
 n/2 P f (x,a)⊕u·x ∈ {0, ±2n/2 }. Clearly,
every a, u ∈ F2 , we have n/2 (−1)
x∈F2
having “ x∈Fn/2 (−1)f (x,a)⊕u·x = 0 for every u” for some a is impossi-
P
2
n/2
ble because of Parseval’s relation. Then, for every a ∈ F2 , there exists
n/2
u ∈ F2 such that x∈Fn/2 (−1)f (x,a)⊕u·x = ±2n/2 that is f (x, a) = u · x or
P
2
f (x, a) = u · x ⊕ 1.
The dual function fe(x, y) equals: y · π −1 (x) ⊕ g(π −1 (x)), where π −1 is the
inverse permutation of π. The completed class of M (that is, the small-
est possible complete class including M) contains all the quadratic bent
functions (according to Alinea 4 of the characterization of quadratic bent
functions given in Subsection 6.2; take π = id and g constant in (48)) and
all bent functions in at most 6 variables [135].
As we saw already in Subsection 5.1, the fundamental idea of Maiorana-
McFarland’s construction consists in concatenating affine functions. If we
order all the binary words of length n in lexicographic order, with the bit of
higher weight on the right, then the truth-table of f is the concatenation of
the restrictions of f obtained by setting the value of y and letting x freely
n/2
range over F2 . These restrictions are affine. In fact, Maiorana-McFarland’s
construction is a particular case of a more general construction of bent func-
tions [65] (see the next proposition), which is properly speaking a secondary
construction for r < n/2 and which is the original Maiorana-McFarland
construction for r = n/2 (this is why we give it in this subsection).

Proposition 20 Let n = r + s (r ≤ s) be even. Let φ be any mapping

from Fs2 to Fr2 such that, for every a ∈ Fr2 , the set φ−1 (a) is an (n − 2r)-
dimensional affine subspace of Fs2 . Let g be any Boolean function on Fs2
whose restriction to φ−1 (a) (viewed as a Boolean function on F2n−2r via an
affine isomorphism between φ−1 (a) and this vectorspace) is bent for every
a ∈ Fr2 , if n > 2r (no condition on g being imposed if n = 2r). Then the
function fφ,g = x · φ(y) ⊕ g(y) is bent on Fn2 .

Proof. This is a direct consequence of the equality (valid for every φ and
every g): X
r
fd
φ,g χ (a, b) = 2 (−1)g(y)⊕b·y , (49)
y∈φ−1 (a)

which comes from the fact that every function x 7→ fφ,g (x, y) ⊕ a · x ⊕ b · y
being affine, and P
thus constant or balanced, it contributes for a nonzero
value in the sum x∈Fr ,y∈Fs (−1)fφ,g (x,y)⊕x·a⊕y·b only if φ(y) = a. Accord-
2 2
ing to Relation (49), the function fφ,g is bent if and only if r ≤ n/2 and

86
 g(y)⊕b·y = ±2n/2−r for every a ∈ Fr2 and every b ∈ Fs2 . The
P
y∈φ−1 (a) (−1)
hypothesis in Proposition 20 is a sufficient condition for that (but it is not
a necessary one). 2
This construction is pretty general: the choice of any partition of F2 in 2r
s

flats of dimension (n − 2r) and of an (n − 2r)-variable bent function on each

of these flats leads to an n-variable bent function.
Obviously, every Boolean function can be represented (in several ways) in
the form fφ,g for some values of r ≥ 1 and s and for some mapping φ from Fs2
to Fr2 and Boolean function g on Fs2 . It has been observed in [257] that, if a
bent function has this form, then φ is balanced (i.e. is uniformly distributed
over Fr2 ). This is a direct consequence of the fact that, for every nonzero
a ∈ Fr2 , the Boolean function a · φ is balanced, since it equals the derivative
D(a,0) fφ,g , and of the characterization of balanced vectorial functions given
in the chapter “Vectorial Boolean Functions for Cryptography”.
It is shown in [25] that every bent function in 6 variables is affinely equiva-
lent to a function of the Maiorana-McFarland class.

Remark: There exist n/2-dimensional vector spaces of n-variable Boolean

functions whose non-zero elements are all bent. The Maiorana-McFarland
construction easily allows constructing such vector spaces. A result by
K. Nyberg (see the chapter “Vectorial Boolean Functions for Cryptogra-
phy”) shows that k-dimensional vector spaces of n-variable Boolean func-
tions whose non-zero elements are all bent cannot exist for k > n/2.

2. The Partial Spreads class PS, introduced in [136] by J. Dillon, is the set
of all the sums (modulo 2) of the indicators of 2n/2−1 or 2n/2−1 +1 “disjoint”
n/2-dimensional subspaces of Fn2 (“disjoint” meaning that any two of these
spaces intersect in 0 only, and therefore that their sum is direct and equals
Fn2 ). The bentness of such function is a direct consequence of Theorem 12
below. This is why we omit the proof of this fact here. According to this
same theorem, the dual of such a function has the same form, all the n/2-
dimensional spaces E being replaced by their orthogonals. Note that the
Boolean functions equal to the sums of the indicators of “disjoint” n/2-
dimensional subspaces of Fn2 share with quadratic functions the property of
being bent if and only if they have the weight of a bent function (which is
2n−1 ± 2n/2−1 ). J. Dillon denotes by PS − (resp. PS + ) the class of those
bent functions for which the number of n/2-dimensional subspaces is 2n/2−1
(resp. 2n/2−1 + 1). All the elements of PS − have algebraic degree n/2
exactly (indeed, by applying a linear isomorphism of Fn2 , we may assume

87
 n/2
that F2 × {0} is among the 2n/2−1 “disjoint” spaces defining the function,
and since the function vanishes at 0, Relation (3) shows that the monomial
x1 · · · xn/2 appears in its ANF), but not all those of PS + (which contains for
instance all the quadratic functions, if n/2 is even, see below). It is an open
problem to characterize the algebraic normal forms of the elements of class
PS, and it is not a simple matter to construct, practically, elements of this
class. J. Dillon exhibits in [136] a subclass of PS − , denoted by PS ap , whose
elements (that we shall call Dillon’s functions) are defined in an explicit
n/2
form: F2 is identified to the Galois field F2n/2 (an inner product in this field
being defined as x·y = tr n2 (xy), where tr n2 is the trace function from F2n/2 to
F2 ; we know that the notion of bent function is independent of the choice of
the inner product); the space Fn2 ≈ F2n/2 ×F2n/2 , viewed37 as a 2-dimensional
F2n/2 -vectorspace, is equal to the “disjoint” union of its 2n/2 +1 lines through
the origin; these lines are n/2-dimensional F2 -subspaces of Fn2 . Choosing any
2n/2−1 of the lines, and taking them different from those of equations x = 0
and y = 0, leads, by definition,
   of PS ap , that is, to a function
toan element
n/2 −2 x x
of the form f (x, y) = g x y 2 , i.e. g y with y = 0 if y = 0, where g is
n/2
a balanced
 Boolean function on F2 which vanishes at 0. The complements
x
g y ⊕ 1 of these functions are the functions g( xy ) where g is balanced and
does not vanish at 0; they belong to the class PS + . In both cases, the dual
of g( xy ) is g( xy ) (this is a direct consequence of Theorem 12). Any function
 n/2

f (x, y) = g x y 2 −2 can be represented as a function of a single variable
n/2 n/2
X belonging to F2n : we have x = aX + (aX)2 and y = bX + (bX)2
∗ n/2
 a, b ∈ F2n linearly
for some elements

independent over F2 , and
2n/2 −2 
we have
n/2 n/2 n/2 n/2
then f (X) = g a + a2 X 2 −1 b + b2 X 2 −1 , for every
X 6= 0.
Given a primitive element α of F2n , we have then for i = 0, · · · , 2n/2 and
j = 0, · · · , 2n/2 − 2:
 n/2
  n/2 n/2 n/2

f αi+j(2 +1) = g (a + a2 β i ) (b + b2 β i )2 −2 ,
n/2
where β = α2 −1 . The elements of the class PS # ap , of those Boolean func-
tions over F2n which can be obtained from those of PS ap by composition by
the transformations x ∈ F2n 7→ δx, δ 6= 0, and by addition of a constant38
37
Let ω be an element of F2n \ F2n/2 ; the pair (1, ω) is a basis of the F2n/2 -vectorspace
F2n ; hence, we have F2n = F2n/2 + ωF2n/2 .
38
The functions of PS ap are among them those satisfying f (0) = f (1) = 0.

88
are those Boolean functions f of weight 2n−1 ± 2n/2−1 on F2n such that,
n/2
denoting by α a primitive element of this field, f (α2 +1 x) = f (x) for every
x ∈ F2n . It is proved in [136, 82] that these functions
Pr are the functions of
weight 2 n−1 ±2 n/2−1 which can be written as i=1 trn (ai xji ) for ai ∈ F2n
and ji a multiple of 2 n/2 − 1 with ji ≤ 2n − 1.
Dillon [136] shows that, when n/2 is even, all quadratic bent functions
are equal to PS + functions or to their complements (while they cannot
be affinely equivalent to PS # ap functions because their degree does not equal
n/2): by affine equivalence we can restrict ourselves to the function (x, , y, η) ∈
F2n/2−1 × F2 × F2n/2−1 × F2 → tr(xy) + η + 1, where tr is the trace func-
tion from F2n/2−1 to F2 ; the support of this function equals the union of the
2n/2−1 + 1 n/2-dimensional vector spaces (very much related to the Ker-
dock code) S∞ = {0} × {0} × F2n/2−1 × F2 and Sa = {(x, , a2 x + atr(ax) +
a, tr(ax)); (x, ) ∈ F2n/2−1 × F2 } for a ∈ F2n/2−1 .
3. Dobbertin gives in [141] the construction of a class of bent functions which
contains both PS ap and M.  Theelements of this class are the functions f
defined by f (x, φ(y)) = g x+ψ(y)
y , where g is a balanced Boolean function
on F2n/2 and φ, ψ are two mappings from F2n/2 to itself such that, if T
denotes the affine subspace of F2n/2 spanned by the support of the function
gc g
χ (where gχ = (−1) ), then, for any a in F2n/2 , the functions φ and ψ are
affine on aT = {ax, x ∈ T }. The mapping φ must additionally be one to
one. The elements of this class do not have an explicit form, but Dobbertin
gives two explicit examples of bent functions constructed this way. In both,
φ is a power function (see below).

4. If n/2 is odd, then it is possible to deduce a bent Boolean function on Fn2

n/2 n/2
from any almost bent function from F2 to F2 . A vectorial Boolean func-
tion F : Fm m
2 → F2 is called almost bent if all of the component functions
m+1
v · F , v 6= 0 in Fm
2 , are plateaued with amplitude 2
2 (see in Subsection
6.8 the definition of these terms). The function γF (a, b), a, b ∈ Fm 2 , equal
to 1 if the equation F (x) + F (x + a) = b admits solutions, with a 6= 0 in
Fm2 , and equal to 0 otherwise is then bent (see the proof of this result in the
chapter “Vectorial Boolean Functions for Cryptography”). This gives new
bent functions related to the almost bent functions listed in this same chap-
ter. However, determining the ANF or the univariate representation of γF
is an open problem when F is a Kasami, Welch or Niho almost bent function.

5. Some infinite classes of bent functions have also been obtained, thanks
to the identification between the vectorspace Fn2 and the field F2n , as power

89
functions (which can also be called monomial functions), that is, functions
of the form trn (axi ), where trn is the trace function on F2n and where a 6= 0
and x belong to this same field. Obviously, a power function trn (axi ) can
be bent only if the mapping x → xi is not one to one (otherwise, the func-
tion would be balanced, a contradiction), that is, if i is not co-prime with
2n − 1. It has been proved in [240] that i must be co-prime either with
2n/2 − 1 or with 2n/2 + 1: it is a simple matter to show that fbχ (0) equals
1 modulo gcd(i, 2n − 1), and this implies that fbχ (0) = 2n/2 if and only if
gcd(i, 2n/2 + 1) = 1 and fbχ (0) = −2n/2 if and only if gcd(i, 2n/2 − 1) = 1
(this is easy to show by using that 2n/2 − 1 and 2n/2 + 1 are co-prime). The
known values of i for which there exists at least one a such that trn (axi ) is
bent are (up to conjugacy i → 2i [mod 2n − 1]):
n
- the Gold exponents i = 2j + 1, where gcd(j,n) is even (the corresponding
i i
function trn (ax ) is bent if and only if a 6∈ {x , x ∈ F2n }; the condition“
n
gcd(j,n) even” is for allowing existence of such a; the function belongs to the
Maiorana-McFarland class);
- the Dillon exponents [135] of the form j · (2n/2 − 1), where gcd(j, 2n/2 +
1) = 1 (the function trn (axi ), where a ∈ F2n/2 without loss of general-
ity, and i = j (2n/2 − 1) is then bent if and only if the Kloosterman sum
tr n (1/x+ax)
is null39 , where 1/0 = 0 and where tr n2 is the trace
P
x∈F2n/2 (−1)
2

function on the field F2n/2 ; this equivalence has been first proved by Dillon
[136]; more recently, Leander [240] has found another proof which gives more
insight; a small error in his proof has been corrected in [102]; the function
trn (axi ) belongs then to the PSap class);
- the Kasami exponents i = 22j − 2j + 1, where gcd(j, n) = 1 (the corre-
sponding function trn (axi ) is bent if and only if a 6∈ {x3 , x ∈ F2n }, see [139]
and [240]);
- and two exponents more recently found: i = (2n/4 + 1)2 where n is di-
visible by 4 but not by 8 (see [240], where the Gold and Dillon exponents
are also revisited, see also [104] where (at page 2) the set of all a’s such
that the corresponding function trn (axi ) is bent is determined: a = a0 bi ,
a0 ∈ wF2n/4 , w ∈ F4 \ F2 , b ∈ F2n ; the function belongs to the Maiorana-
McFarland class) and i = 2n/3 + 2n/6 + 1, where n is divisible by 6 [44] (the
corresponding function trn (axi ) is bent if and only if a = a0 bi , a0 ∈ F2n/2
39
The existence of a such that the Kloosterman sum is null had been conjectured by
Dillon. It has been proved by Lachaud and Wolfmann [225] who proved that the values
of such Kloosterman sums are all the numbers divisible by 4 in the range [−2n/4+1 +
1; 2n/4+1 + 1].

90
 n/6 n/6 2n/6
such that trn/2 (a0 ) := a0 + a02 + a02 = 0, b ∈ F2n ; it belongs to the
Maiorana-McFarland class).
Note that a still simpler bent function (but which is not expressed by
n/2
means of the function trn itself) is f (x) = tr n2 (x2 +1 ), that is, f (x) =
n/2 +1
 n/2 2  n/2 22  n/2 2n/2−1
x 2 + x2 +1 + x 2 +1 + · · · + x2 +1 . The symplectic
n/2
form ϕf (x, y) associated to f equals trn (y 2 x); its kernel is therefore trivial
and f is bent.
Some other functions are defined as the sums of a few power functions, see
[136, 139, 144, 145, 102, 197, 239, 242, 357].
Note that power functions and sums of power functions represent for the
designer of the cryptosystem using them the interest of being more eas-
ily computable than general functions (which allows using them with more
variables while keeping a good efficiency). Power functions have the pecu-
liarity that, denoting the set {xi ; x ∈ F∗2n } by U , two functions trn (axi ) and
trn (bxi ) such that a/b ∈ U are linearly equivalent. It is not clear whether
this is more an advantage for the designer or for the attacker of a system
using a nonlinear balanced function derived from such bent function.
Finally, bent functions have been also obtained by Dillon and McGuire [140]
as the restrictions of functions on F2n+1 , with n + 1 odd, to a hyperplane
2k k
of this field: these functions are the Kasami functions trn x2 −2 +1 and
the hyperplane has equation trn (x) = 0. The restriction is bent under the
condition that n + 1 = 3k ± 1.

Remark. The bent sequences given in [355] are particular cases of the
constructions given above (using also some of the secondary constructions
given below).
In [99] are constructed homogeneous bent functions (i.e. bent functions
whose ANFs are the sums of monomials of the same degree) on 12 (and less)
variables by using the invariant theory (which makes feasible the computer
searchs).

6.4.2 Secondary constructions

We have already seen in Proposition 20 a secondary construction based on
the Maiorana-McFarland construction. We describe now the others (which
have been found so far).

1. The first secondary construction given by J. Dillon and O. Rothaus

91
in [136, 315] is very simple: let f be a bent function on Fn2 (n even) and
n+m
g a bent function on Fm 2 (m even) then the function h defined on F2
by h(x, y) = f (x) ⊕ g(y) is bent. Indeed, we have clearly h c (a, b) =
χ

fbχ (a) × gc
χ (b). This construction, called the direct sum has unfortunately
no great interest from a cryptographic point of view, since it produces de-
composable functions (a Boolean function is called decomposable if it is
equivalent to the sum of two functions that depend on two disjoint subsets
of coordinates; such peculiarity is easy to detect and can be used for design-
ing divide-and-conquer attacks, as pointed out by J. Dillon in [137]).

2. A more interesting result, by the same authors, is the following: if g, h,

k and g ⊕ h ⊕ k are bent on Fn2 (n even), then the function defined at every
element (x1 , x2 , x) of Fn+2
2 (x1 , x2 ∈ F2 , x ∈ Fn2 ) by:
f (x1 , x2 , x) =
g(x)h(x) ⊕ g(x)k(x) ⊕ h(x)k(x) ⊕ [g(x) ⊕ h(x)]x1 ⊕ [g(x) ⊕ k(x)]x2 ⊕ x1 x2
is bent (this is a particular case of Theorem 10 below). No general class of
bent functions has been deduced from this Rothaus construction.

3. Two classes of bent functions have been derived in [54] from Maiorana-
McFarland’s class, by adding to some functions of this class the indicators
of some vector subspaces:
- the class D0 whose elements are the functions of the form f (x, y) =
x · π(y) ⊕ δ0 (x) (recall that δ0 is the Dirac symbol; the ANF of δ0 (x) is
Qn/2 −1
i=1 (xi ⊕ 1)). The dual of such a function f is the function y · π (x) ⊕
δ0 (y). It is proved in [54] that this class is not included40 in the completed
versions M# and PS # of classes M and PS (i.e. the smallest possible
classes including them) and that every bent function in 6 variables is affinely
equivalent to a function of this class, up to the addition of an affine function.
Class D0 is a subclass of the class denoted by D, whose elements are the
functions of the form f (x, y) = x · π(y) ⊕ 1E1 (x)1E2 (y), where π is any
n/2 n/2
permutation on F2 and where E1 , E2 are two linear subspaces of F2
such that π(E2 ) = E1 ⊥ (1E1 and 1E2 denote their indicators). The dual of f
belongs to the completed version of this same class;
- the class C of all the functions of the form x · π(y) ⊕ 1L (x), where L is
n/2 n/2
any linear subspace of F2 and π any permutation on F2 such that, for
40
It is easy to show that a function f does not belong to M# by showing that there
does not exist an n/2-dimensional vector-subspace E of Fn 2 such that Da Db f is null for
every a, b ∈ E; it is much more difficult to show that it does not belong to PS # .

92
 n/2
any element a of F2 , the set π −1 (a + L⊥ ) is a flat. It is a simple matter
to see, as shown in [45], that, under the same hypothesis on π, if g is a
Boolean function whose restriction to every flat π −1 (a + L⊥ ) is affine, then
the function x · π(y) ⊕ 1L (x) ⊕ g(y) is also bent.
The fact that any function in class D or class C is bent comes from the
following theorem proved in [54], which has its own interest:

Theorem 9 Let b + E be any flat in Fn2 (E being a linear subspace of Fn2 ).

Let f be any bent function on Fn2 . The function f ? = f ⊕ 1b+E is bent if and
only if one of the following equivalent conditions is satisfied:

1. For any a in Fn2 \ E, the function Da f is balanced on b + E;

2. The restriction of the function fe(x) ⊕ b · x to any coset of E ⊥ is either

constant or balanced.

If f and f ? are bent, then E has dimension greater than or equal to n/2
and the algebraic degree of the restriction of f to b + E is at most dim(E) −
n/2 + 1.
If f is bent, if E has dimension n/2, and if the restriction of f to b + E has
algebraic degree at most dim(E) − n/2 + 1 = 1, i.e. is affine, then conversely
f ? is bent too.

Proof. Recall that a function is bent if and only if it satisfies P C(n). The
equivalence between Condition 1. and the bentness of f ? comes then from
the
P fact that D F(Da f ? ) equals F(Da f ) if a ∈ E, and equals F(Da f ) −
4 x∈b+E (−1) a f (x) otherwise.
We have fbχ (a)− fbχ? (a) = 2 x∈b+E (−1)f (x)⊕a·x . Using Relation (46), applied
P

with E ⊥ in the place of E, we deduce that for every a ∈ Fn2 :

X ⊥
 
(−1)f (u)⊕b·u = 2dim(E )−n/2−1 (−1)a·b fbχ (a) − fbχ? (a) ,
e

u∈a+E ⊥

?
and fbχ (a) − fbχ (a) takes value 0 or ±2n/2+1 for every a if and only if Con-
dition 2. is satisfied. So Condition 2. is necessary and sufficient, according
to Lemma 2 (at the beginning of Section 6).
Let us now assume that f and f ? are bent. Then 1b+E = f ? ⊕f has algebraic
degree at most n/2, according to Rothaus’ bound, and thus dim(E) ≥ n/2.
The values of the Walsh  transform of the restriction of f to b+E being equal
1 b ?
to those of 2 fχ − fχ , they are divisible by 2n/2 and thus the restriction
c
of f to b + E has algebraic degree at most dim(E) − n/2 + 1, according to

93
Proposition 11.
If f is bent, if E has dimension n/2, and if the restriction of f to b + E is
?
affine, then the relation fbχ (a) − fbχ (a) = 2 x∈b+E (−1)f (x)⊕a·x shows that
P
f ? is bent too, according to Lemma 2. 2

Remarks.
- Relation (46) applied to E ⊥ in the place of E, where E is some n/2-
dimensional subspace, shows straightforwardly that, if f is a bent function
on Fn2 , then f (x) ⊕ a · x is constant on b + E if and only if fe(x) ⊕ b · x
is constant on a + E ⊥ . The same relation shows that f (x) ⊕ a · x is then
balanced on every other coset of E and fe(x)⊕b·x is balanced on every other
coset of E ⊥ . Notice that Relation (46) shows also that f (x) ⊕ a · x cannot be
constant on a flat of dimension strictly greater than n/2 (i.e. that f cannot
be k-weakly-normal with k > n/2).
- Let f be bent on Fn2 . Let a and a0 be two linearly independent elements
of Fn2 . Let us denote by E the orthogonal of the subspace spanned by a and
a0 . According to condition 2. of Theorem 9, the function f ⊕ 1E is bent
if and only if Da Da0 fe is null (indeed, a 2-variable function is constant or
balanced if and only if it has even weight, and fe has even weight on any
coset of the vector subspace spanned by a and a0 if and only if, for every
vector x, we have f (x) ⊕ f (x + a) ⊕ f (x + a0 ) ⊕ f (x + a + a0 ) = 0). This
result has been restated in [43] and used in [45] to design (potentially) new
bent functions.

4. Other classes of bent functions have been deduced from a construction

given in [57], which generalizes the secondary constructions given in 1 and 2
above:

Theorem 10 Let n and m be two even positive integers. Let f be a Boolean

function on Fn+m
2 = Fn2 ×Fm m
2 such that, for any element y of F2 , the function
n
on F2 :
fy : x 7→ f (x, y)
is bent. Then f is bent if and only if, for any element s of Fn2 , the function

ϕs : y 7→ fey (s)

is bent on Fm 2 . If this condition is satisfied, then the dual of f is the function

fs (t) (taking as inner product in Fn2 ×Fm
f (s, t) = ϕ
e
2 : (x, y)·(s, t) = x·s⊕y ·t).

94
This very general result is easy to prove, using that, for every s ∈ Fn2 ,
X
(−1)f (x,y)⊕x·s = 2n/2 (−1)fy (s) = 2n/2 (−1)ϕs (y) ,
f

x∈Fn
2

and thus that X

fbχ (s, t) = 2n/2 (−1)ϕs (y)⊕y·t .
y∈Fm
2

This construction has also been considered in a particular case by Adams

and Tavares [1] under the name of bent-based functions, and later studied
by J. Seberry and X.-M. Zhang in [326] in special cases too.
A case of application of this construction is nicely simple:

Corollary 4 [67] Let f1 and f2 be two n-variable bent functions (n even)

and let g1 and g2 be two m-variable bent functions (m even). Define41

h(x, y) = f1 (x) ⊕ g1 (y) ⊕ (f1 ⊕ f2 )(x) (g1 ⊕ g2 )(y); x ∈ Fn2 , y ∈ Fm

2 .

Then h is bent and its dual is obtained from fe1 , fe2 , ge1 and ge2 by the same
formula as h is obtained from f1 , f2 , g1 and g2 .

Proof. For every y, the function hy (x) of Theorem 10 equals f1 (x) plus
the constant g1 (y) if g1 (y) = g2 (y) and f2 (x) plus the constant g1 (y) if
g1 (y) 6= g2 (y); thus it is bent and function ϕs (y) equals fe1 (s) ⊕ g1 (y) if
g1 (y) = g2 (y) and fe2 (s) ⊕ g1 (y) if g1 (y) 6= g2 (y), that is, equals fe1 (s) ⊕
g1 (y) ⊕ (fe1 ⊕ fe2 )(s) (g1 ⊕ g2 )(y). Hence, ϕs (y) is bent too and according to
Theorem 10, h is then bent and its dual equals:

h(s, t) = fe1 (s) ⊕ ge1 (t) ⊕ (fe1 ⊕ fe2 )(s)(e

e g1 ⊕ ge2 )(t).

2
What is interesting in this particular case of Theorem 10 (sometimes called
the indirect sum of bent functions) is that we only assume the bentness
of f1 , f2 , g1 , and g2 for deducing the bentness of h; no extra condition is
needed, contrary to the general construction.
Another simple application of Theorem 10, called the extension of Maiorana-
n/2
McFarland type is given in [79]: Let π be a permutation on F2 and
fπ,g (x, y) = x · π(y) ⊕ g(y) a related Maiorana-McFarland bent function.
41
h is the concatenation of the four functions f1 , f1 ⊕ 1, f2 and f2 ⊕ 1, in an order
controled by g1 (y) and g2 (y). This construction (f1 , f2 , g1 , g2 ) 7→ h will appear again
below to construct resilient functions; see Theorem 14.

95
Let (hy )y∈Fn/2 be a collection of bent functions on Fm
2 for some even integer
2
n/2 n/2
m. Then the function (x, y, z) ∈ F2 ×F2 ×Fm
2 → hy (z)⊕fπ,g (x, y) is bent.

Several classes have been deduced from Theorem 10 in [57], and later
in [192].
- Let n and m be two even positive integers. The elements of Fn+m 2 are
n/2
written (x, y, z, τ ), where x, y are elements of F2 and z, τ are elements
m/2 n/2 m/2
of F2 . Let π and π 0 be permutations on F2 and F2 (respectively)
m/2
and h a Boolean function on F2 . Then, the following Boolean function
on Fn+m
2 is bent:

f (x, y, z, τ ) = x · π(y) ⊕ z · π 0 (τ ) ⊕ δ0 (x)h(τ )

(recall that δ0 (x) equals 1 if x = 0 and is null otherwise). It is possible

to prove, see [57], that such a function does not belong, in general, to the
completed version of class M. It is also easy to prove that f does not belong,
in general, to the completed version of class D0 , since any element of D0 has
algebraic degree n+m 2 , and it is a simple matter to produce examples of
functions f whose algebraic degree is smaller than n+m 2 .
n/2 m/2
- Let n and m be two even positive integers. We identify F2 (resp. F2 )
with the Galois field F2n/2 (resp. with F2m/2 ). Let k be a Boolean function
on F2n/2 × F2m/2 such that, for any element x of F2n/2 , the function z 7→
k(x, z) is balanced on F2m/2 , and for any element z of F2m/2 , the function
x 7→ k(x, z) is balanced on F2n/2 . Then the function
 
x z
f (x, y, z, τ ) = k ,
y τ

is bent on Fn+m
2 .
- Let r be a positive integer. We identify Fr2 with F2r . Let π and π 0 be
two permutations on F2r and g a balanced Boolean function on F2r . The
following Boolean function on F4r r 4
2 = (F2 ) :
    
0 x x
f (x, y, z, τ ) = z · π τ + π ⊕ δ0 (z)g
y y

is a bent function.

5. X.-D. Hou and P. Langevin have made in [196] a very simple observation
which leads to potentially new bent functions:

96
Proposition 21 Let f be a Boolean function on Fn2 , n even. Let σ be a
permutation on Fn2 . We denote its coordinate functions by σ1 , · · · , σn and
we assume that, for every a ∈ Fn2 :
n
M
dH (f, ai σi ) = 2n−1 ± 2n/2−1 .
i=1

Then f ◦ σ −1 is bent.
−1 and the linear function
Indeed, the Hamming distanceLn between f ◦ σ
`a (x) = a · x equals dH (f, i=1 ai σi ).
Hou and Langevin deduced that, if h is an affine function on Fn2 , if f1 , f2
and g are Boolean functions on Fn2 , and if the following function is bent:

f (x1 , x2 , x) = x1 x2 h(x) ⊕ x1 f1 (x) ⊕ x2 f2 (x) ⊕ g(x)/ x ∈ Fn2 , x1 , x2 ∈ F2 ,

then the function

f (x1 , x2 , x) ⊕ (h(x) ⊕ 1) f1 (x)f2 (x) ⊕ f1 (x) ⊕ (x1 ⊕ h(x) ⊕ 1) f2 (x) ⊕ x2 h(x)

is bent.
They also deduced that, if f is a bent function on Fn2 whose algebraic degree
is at most 3, and if σ is a permutation on Fn2 such that, for every i = 1, · · · , n,
there exists a subset Ui of Fn2 and an affine function hi such that:
M
σi (x) = (f (x) ⊕ f (x + u)) ⊕ hi (x),
u∈Ui

then f ◦ σ −1 is bent.
n/2
Finally, X.-D. Hou [192] deduced that if f (x, y) (x, y ∈ F2 ) is a Maiorana-
McFarland’s function of the particular form x · y ⊕ g(y) and if σ1 , · · · , σn are
all of the form 1≤i<j≤n/2 ai,j xi yj ⊕ b · x ⊕ c · y ⊕ h(y), then f ◦ σ −1 is bent.
L
He gave several examples of application of this result.

6. Note that the construction of 5. does not increase the number of vari-
ables, contrary to most other secondary constructions. Another secondary
construction without extension of the number of variables was introduced
in [70]. It is based on the following result:

Proposition 22 Let f1 , f2 and f3 be three Boolean functions on Fn2 . Denote

by s1 the Boolean function equal to f1 ⊕f2 ⊕f3 and by s2 the Boolean function
equal to f1 f2 ⊕f1 f3 ⊕f2 f3 . Then we have f1 +f2 +f3 = s1 +2s2 . This implies

97
the following equality between the Fourier transforms: fb1 + fb2 + fb3 = sb1 + 2sb2
and the similar equality between the Walsh transforms:
fc
1 χ + f2 χ + f3 χ = sc
c c 1 χ + 2 sc
2χ . (50)

Proof. The fact that f1 + f2 + f3 = s1 + 2s2 (the sums being computed in Z

and not modulo 2) can be checked easily. The linearity of the Fourier trans-
form with respect to the addition in Z implies then fb1 +fb2 +fb3 = sb1 +2sb2 . The
equality f1 +f2 +f3 = s1 +2s2 also directly implies f1 χ +f2 χ +f3 χ = s1 χ +2s2 χ ,
thanks to the equality fχ = 1 − 2f valid for every Boolean function, which
implies Relation (50). 2

Proposition 22 leads to the following double construction of bent func-

tions:
Corollary 5 Let f1 , f2 and f3 be three n-variable bent functions, n even.
Denote by s1 the function f1 ⊕f2 ⊕f3 and by s2 the function f1 f2 ⊕f1 f3 ⊕f2 f3 .
Then:
- if s1 is bent and if s˜1 = f˜1 ⊕ f˜2 ⊕ f˜3 , then s2 is bent, and s˜2 = f˜1 f˜2 ⊕
f˜1 f˜3 ⊕ f˜2 f˜3 ;
- if sc n/2 for every a (e.g. if s is bent, or if it is
2 χ (a) is divisible by 2 2
quadratic, or more generally if it is plateaued; see the definition in Subsec-
tion 6.8), then s1 is bent.
Proof. - If s1 is bent and if s˜1 = f˜1 ⊕ f˜2 ⊕ f˜3 , then, for every a, Relation (50)
implies:
h i n−2
f˜1 (a) f˜2 (a) f˜3 (a) f˜1 (a)⊕f˜2 (a)⊕f˜3 (a)
sc
2 χ (a) = (−1) + (−1) + (−1) − (−1) 2 2
˜ ˜ ˜ ˜ ˜ ˜
= (−1)f1 (a)f2 (a)⊕f1 (a)f3 (a)⊕f2 (a)f3 (a) 2n/2 .
Indeed, as we already saw above with the relation f1 χ +f2 χ +f3 χ = s1 χ +2s2 χ ,
for every bits , η and τ , we have (−1) + (−1)η + (−1)τ − (−1)⊕η⊕τ =
2 (−1)η⊕τ ⊕ητ .
- If sc n/2 for every a, then the number sc (a), which
2 χ (a) is divisible by 2 1χ
h i
˜1 (a)
f f˜2 (a) f˜3 (a) n/2
is equal to (−1) + (−1) + (−1) 2 − 2 sc
2 χ (a), according to

Relation (50), is congruent with 2n/2 modulo 2n/2+1 for every a. This is
sufficient to imply that s1 is bent, according to Lemma 2 (at the beginning
of Section 6). 2

7. A construction related to the notion of normal extension of bent function

can be found in Proposition 31.

98
6.4.3 Decompositions of bent functions
The following theorem, proved in [42], is a direct consequence of Rela-
tion (28), applied to f ⊕ ` where ` is linear, and to a linear hyperplane
E of Fn2 , and of the well-known (easy to prove) fact that, for every even
integer n ≥ 4, the sum of the squares of two integers equals 2n (resp. 2n+1 )
if and only if one of these squares is null and the other one equals 2n (resp.
both squares equal 2n ):

Theorem 11 Let n be an even integer, n ≥ 4, and let f be an n-variable

Boolean function. Then the following properties are equivalent.

1. f is bent.

2. For every (resp. for some) linear hyperplane E of Fn2 , the Walsh trans-
forms of the restrictions h1 , h2 of f to E and to its complement (viewed
as Boolean functions on Fn−12 ) take values ±2n/2 and 0 only, and the
disjoint union of their supports equals the whole space Fn−1 2 .

Hence, a simple way of obtaining a plateaued function in an odd number of

variables and with optimal nonlinearity is to take the restriction of a bent
function to an affine hyperplane. Note that we have also (see [42]) that,
if a function in an odd number of variables is such that, for some nonzero
a ∈ Fn2 , every derivative Du f , u 6= 0, u ∈ a⊥ , is balanced, then its restriction
to the linear hyperplane a⊥ or to its complement is bent.
It is also proved in [42] that the Walsh transforms of the four restrictions of
a bent function to an (n − 2)-dimensional vector subspace E of Fn2 and to its
cosets have the same sets of magnitudes. It is a simple matter to see that,
denoting by a and b two vectors such that E ⊥ is the linear space spanned
by a and b, these four restrictions are bent if and only if Da Db fe takes on
constant value 1.
More on decomposing bent functions can be found in [42, 43, 101].

6.5 On the number of bent functions

The class of bent functions produced by the original Maiorana-McFarland’s
construction is far the widest class, compared to the classes obtained from
the other primary constructions.
n/2
The number of bent functions of the form (48) equals (2n/2 )! × 22 , which
 n/2+1 2n/2 √
is asymptotically equivalent to 2 e 2n/2+1 π (according to Stir-
ling’s formula) while the only other important construction of bent functions,

99
 2n/2 + 1
2n/2
≈ 2√ n/22 functions. However, the number of


PS ap , leads only to 2n/2−1
π2
provably bent Maiorana-McFarland’s functions seems negligible with respect
to the total number of bent functions. The number of (bent) functions which
are affinely equivalent to Maiorana-McFarland’s functions is unknown; it is
at most equal to the number of Maiorana-McFarland’s functions times the
number of affine automorphisms, which equals 2n (2n − 1)(2n − 2) · · · (2n −
2n−1 ). It seems also negligible with respect to the total number of bent func-
tions. The problem of determining an efficient lower bound on the number
of n-variable bent functions is open.
Rothaus’ inequality recalled in Subsection 6.3 (Proposition 18) states that
any bent function has algebraic degree at most n/2. Thus, the number of
bent functions is at most
n n
2
1+n+...+(n/2 ) = 22n−1 + 12 (n/2 ).

We shall call this upper bound the naive bound . For n = 6, the number
of bent functions is known and is approximately equal to 232.3 (see [302]),
which is much less than what gives the naive bound: 242 . For n = 8, the
number is also known: it has been first shown in [234] that it is inferior to
2129.2 ; it has been very recently calculated by Langevin, Leander et al. [233]
and equals approximately 2106.3 (the naive bound gives 2163 ). Hence picking
at random an 8-variable Boolean function of algebraic degree bounded above
by 4 does not allow obtaining bent functions (but more clever methods exist,
see [127, 82]). An upper bound improving upon the naive bound has been
found in [90]. It is exponentially better than the naive bound since it divides
n/2
it by approximately 22 −n/2−1 . But it seems to be still far from the exact
number of bent functions: for n = 6 it gives roughly 238 (to be compared
with 232.3 ) and for n = 8 it gives roughly 2152 (to be compared with 2106.3 ).

6.6 Characterizations of bent functions

6.6.1 characterization through the NNF
Proposition 23 Let f (x) = I∈P(N ) λI xI be the NNF of a Boolean func-
P
tion f on Fn2 . Then f is bent if and only if:
1. for every I such that n/2 < |I| < n, the coefficient λI is divisible
by 2|I|−n/2 ;
2. λN (with N = {1, · · · , n}) is congruent with 2n/2−1 modulo 2n/2 .
Proof. According to Lemma 2, f is bent if and only if, for every a ∈ Fn2 ,
fb(a) ≡ 2n/2−1 mod 2n/2 . We deduce that, according to Relation (30) ap-
 

plied with ϕ = f , Conditions 1. and 2. imply that f is bent.

100
Conversely, Condition 1. is necessary, according to Proposition 17. Condi-
tion 2. is also necessary since fb(1, · · · , 1) = (−1)n λN (from Relation (30)).
2
Proposition 23 and Relation (9) imply some restrictions on the coefficients
of the ANFs of bent functions, observed and used in [90] (and also partially
observed by Hou and Langevin in [196]).
Proposition 23 can be seen as a (much) stronger version of Rothaus’ bound,
since the algebraic degree of a Boolean function whose NNF is f (x) =
I
P
I∈P(N ) λI x equals the maximum size of I, such that λI is odd.

6.6.2 Geometric characterization

Proposition 23 also allows proving the following characterization:

Theorem 12 [85] Let f be a Boolean function on Fn2 . Then f is bent if and

only if there exist n/2-dimensional subspaces E1 , . . . , Ek of Fn2 (there is no
constraint on the number k) and integers m1 , . . . , mk (positive or negative)
such that, for any element x of Fn2 :
k
X h i
f (x) ≡ mi 1Ei (x) − 2n/2−1 δ0 (x) mod 2n/2 . (51)
i=1

If we have f (x) = ki=1 mi 1Ei (x) − 2n/2−1 δ0 (x) then the dual of f equals
P

fe(x) = ki=1 mi 1E ⊥ (x) − 2n/2−1 δ0 (x).

P
i

Proof (sketch of). Relation (51) is a sufficient condition for f being bent,
according to Lemma 2 (at the beginning of Section 6) and to Relation (16).
This same Relation (16) also implies the last sentence of Theorem 12. Con-
versely, if f is bent, then Proposition 23 allows to deduce Relation (51), by
expressing all the monomials xI by means of the indicators of subspaces of
dimension at least n − |I| (indeed, the NNFQ of the indicator
P of the subspace
{x ∈ Fn2 / xi = 0, ∀i ∈ I} being equal to i∈I (1 − xi ) = J⊆I (−1)|J| xJ , the
monomial xI can be expressed by means of this indicator and of the monomi-
als xJ , where J is strictly included in I) and by usingQLemma 3 below (note
that d ≥ n − |I| implies |I| − n/2 ≥ n/2 − d and that i∈N (1 − xi ) = δ0 (x)).
2

Lemma 3 Let F be any d-dimensional subspace of Fn2 . There exist n/2-

dimensional subspaces E1 , · · · , Ek of Fn2 and integers m, m1 , · · · , mk such

101
that, for any element x of Fn2 :
k
X h i
2n/2−d 1F (x) ≡ m + mi 1Ei (x) mod 2n/2 if d < n/2, and
i=1

k
X h i
1F (x) ≡ mi 1Ei (x) mod 2n/2 if d > n/2.
i=1

The class of those functions f which satisfy the relation obtained from (51)
by withdrawing “[mod 2n/2 ]” is called Generalized Partial Spread class and
denoted by GPS (it includes
Pk PS), see [55]. The dual fe of such function f
of GPS equaling f (x) = i=1 mi 1E ⊥ (x) − 2
e n/2−1 δ0 (x), it belongs to GPS
i
too.

There is no uniqueness of the representation of a given bent function

in the form (51).PThere exists another characterization, shown in [86], in
the form f (x) = ki=1 mi 1Ei (x) ± 2n/2−1 δ0 (x), where E1 , . . . , Ek are vector
subspaces of Fn2 of dimensions n/2 or n/2 + 1 and where m1 , . . . , mk are
integers (positive or negative). There is not a unique way, either, to choose
these spaces Ei . But it is possible to define some subclass of n/2-dimensional
and (n/2 + 1)-dimensional spaces such that there is uniqueness, if the spaces
Ei are chosen in this subclass.
P. Guillot has proved subsequently in [171] that, up to composition by a
translation x 7→ x + a, every bent function belongs to GPS.

6.6.3 characterization by second-order covering sequences

Proposition 24 [93] A Boolean function f defined on Fn2 is bent if and
only if: X
∀x ∈ Fn2 , (−1)Da Db f (x) = 2n . (52)
a,b∈Fn
2

Proof. If we multiply both terms of Relation (52) by fχ (x) = (−1)f (x) , we

obtain the (equivalent) relation:

∀x ∈ Fn2 , fχ ⊗ fχ ⊗ fχ (x) = 2n fχ (x);

P 
f (a)⊕f (a+b) (−1)f (b+x) =
P
indeed, we have fχ ⊗ fχ ⊗ fχ (x) = b∈Fn a∈F n (−1)
2 2
f (a+x)⊕f (a+b+x)⊕f (b+x) . According to the bijectivity of the Fourier
P
n
a,b∈F2 (−1)
transform and to Relation (20), this is equivalent to :

102
 3
∀u ∈ Fn2 , fbχ (u) = 2n fbχ (u) .

Thus, we have a,b∈Fn (−1)Da Db f (x) = 2n if and only if, for every u ∈ Fn2 ,
P
√ 2
fbχ (u) equals ± 2n or 0. According to Parseval’s relation, the value 0 can-
not be achieved by fb and this is therefore equivalent to the bentness of f . 2
χ

Relation (52) is equivalent to the relation a,b∈Fn (1−2Da Db f (x)) = 2n ,

P
2
that is a,b∈Fn Da Db f (x) = 22n−1 − 2n−1 , and hence to the fact that f ad-
P
2
mits the second order covering sequence with all-1 coefficients and with level
22n−1 − 2n−1 .

It is shown similarly in [93] that the relation similar to (52) but with any
integer in the place of 2n characterizes the class of plateaued functions (see
Subsection 6.8).

A characterization of bent functions through Cayley graphs also exists,

see [18].

6.7 Subclasses: hyper-bent functions

In [356], A. Youssef and G. Gong study the Boolean functions f on the field
F2n (n even) such that f (xi ) is bent for every i co-prime with 2n − 1. These
functions are called hyper-bent functions. The condition seems difficult to
satisfy. However, A. Youssef and G. Gong show in [356] that hyper-bent
functions exist. Their result is equivalent to the following (see [82]):

Proposition 25 All the functions of class PS #

ap are hyper-bent.

Let us give here a direct proof of this fact.

Proof. We can restrict ourselves without loss of generality to the functions
of class PS ap . Let ω be any element in F2n \ F2n/2 . The pair (1, ω) is a
basis of the F2n/2 -vectorspace F2n . Hence, we have F2n = F2n/2 + ωF2n/2 .
n/2
Moreover, every element y of F2n/2 satisfies y 2 = y and therefore trn (y) =
n/2−1 n/2−1
y + y2 + · · · + y2 + y + y2 + · · · + y2 = 0. Consider the inner
product in F2n defined by: y · y 0 = trn (y y 0 ); the subspace F2n/2 is then its
own
P orthogonal; hence, according to Relation (16), any sum of the form
tr (λy) is null if λ 6∈ F2n/2 and equals 2n/2 if λ ∈ F2n/2 .
y∈F2n/2 (−1)
n

Let us consider any element of the class PS ap , choosing a balanced  Boolean


n/2 y0
function g on F2 , vanishing at 0, and defining f (y 0 + ω y) = g y , with

103
y0
y = 0 if y = 0. For every a ∈ F2n , we have

y0
“ ”
f (x)⊕trn (a xi ) g ⊕trn (a (y 0 +ωy)i )
X X
(−1) = (−1) y
.
x∈F2n y,y 0 ∈F2n/2

y0
Denoting y by z, we see that:

y0
“ ”
g ⊕trn (a (y 0 +ωy)i ) i (z+ω)i )
X X
(−1) y
= (−1)g(z)⊕trn (a y .
y∈F∗n/2 ,y 0 ∈F2n/2 z∈F2n/2 ,y∈F∗n/2
2 2

0i ) 0
X X
The remaining sum (−1)g(0)⊕trn (a y = (−1)trn (a y ) (this equal-
y 0 ∈F2n/2 y 0 ∈F2n/2
ity being due to the fact that the mapping x → xi is one-to-one) equals 2n/2
if a ∈ FX
2n/2 and is null otherwise.
i
Thus, (−1)f (x)⊕trn (a x ) equals:
x∈F2n

i
X X X
(−1)g(z) (−1)trn (a(z+ω) y)
− (−1)g(z) + 2n/2 1F (a).
2n/2
z∈F2n/2 y∈F2n/2 z∈F2n/2

g(z) is null since g is balanced.

P
The sum z∈F2n/2 (−1)
g(z) trn (a(z+ω)i y)
equals ±2n/2 if a 6∈
P P
The sum z∈F2n/2 (−1) y∈F2n/2 (−1)
F2n/2 , since we prove in the next Lemma that there exists then exactly one
z ∈ F2n/2 such that a(z + ω)i ∈ F2n/2 ; and this sum is null if a ∈ F2n/2
(this can be checked, if a = 0 thanks to the balancedness of g, and if a 6= 0
because y ranges over F2n/2 and a(z + ω)i 6∈ F2n/2 ). This completes the
proof. 2

Lemma 4 Let n be any positive integer. Let a and ω be two elements of

the set F2n \ F2n/2 and let i be co-prime with 2n − 1. There exists a unique
n/2 n/2
element z ∈ F2 such that a(z + ω)i ∈ F2 .

n/2
Proof. Let j be the inverse of i modulo 2n − 1. We have a(z + ω)i ∈ F2 if
n/2 n/2 n/2
and only if z ∈ ω + a−j × F2 . The sets ω + a−j × F2 and F2 are two
n/2 n/2
flats whose directions a−j × F2 and F2 are subspaces whose sum is direct
and equals F2n . Hence, they have a unique vector in their intersection. 2

104
 Relationships between the notion of hyper-bent function and cyclic codes
P every hyper-bent function f : F2 → F2 ,
are studied in [82]. It is proved that n

can be represented as: f (x) = ri=1 trn (ai xti ) ⊕ , where ai ∈ F2n ,  ∈ F2
and w2 (ti ) = n/2. Consequently, all hyper-bent functions have algebraic
degree n/2.
In [102] is proved that, for every even n, every λ ∈ F∗2n/2 and every r ∈
n
]0; 2 [ such that the cyclotomic cosets of 2 modulo 2n/2 +1 containing respec-

r
tively 2r − 1 and 2r + 1 have size n and such that the function tr n2 λx2 +1
  r n/2 r n/2

is balanced on F2n/2 , the function trn λ x(2 −1)(2 −1) + x(2 +1)(2 −1)
r
is bent (i.e. hyper-bent) if and only if the function tr n2 x−1 + λx2 +1 is

also balanced on F2n/2 .

Remark. In [56] have been determined those Boolean functions on Fn2 such
that, for a given even integer k (2 ≤ k ≤ n−2), any of the Boolean functions
on Fn−k
2 , obtained by keeping constant k coordinates among x1 , · · · , xn , is
bent (i.e. those functions which satisfy the propagation criterion of degree
n − k and order k, see Section 8). These are the four symmetric bent
functions (see Section 10). They were called hyper-bent in [56] but we keep
this term for the notion introduced by Youssef and Gong.

6.8 Superclasses: partially-bent functions, partial bent func-

tions and plateaued functions
We have seen that bent functions can never be balanced, which makes them
improper for a direct cryptographic use. This has led to a research of super-
classes of the class of bent functions, whose elements can have high nonlin-
earities, but can also be balanced (and possibly, be m-resilient with large
m or satisfy P C(l) with large l). A first super-class having these proper-
ties has been obtained as the set of those functions which achieve a bound
expressing some trade-off between the number of non-balanced derivatives
(i.e. of nonzero auto-correlation coefficients) of a Boolean function and the
number of nonzero values of its Walsh transform. This bound, given in the
next proposition, had been conjectured in [301] and has been proved later
in [53].
Proposition 26 Let n be any positive integer. Let f be any Boolean func-
on Fn2 . Let us denote
tion n n
o the cardinalities of the sets {b ∈ F2 | F(Db f ) 6= 0}
and b ∈ Fn2 | fbχ (b) 6= 0 by N∆f and Nfb , respectively. Then:
χ

N∆f × Nfb ≥ 2n . (53)

χ

105
Moreover, N∆f × Nfb = 2n if and only if, for every b ∈ Fn2 , the derivative
χ
Db f is either balanced or constant. This property is also equivalent to the
fact that there exist two linear subspaces E (of even dimension) and E 0
of Fn2 , whose direct sum equals Fn2 , and Boolean functions g, bent on E, and
h, affine on E 0 , such that:

∀x ∈ E, ∀y ∈ E 0 , f (x + y) = g(x) ⊕ h(y). (54)

Inequality (53) comes directly from Relation (25): since the value of the
auto-correlation coefficient F(Db f ) lies between −2n and 2n for every b,
2
we have N∆f ≥ 2−n b∈Fn (−1)u·b F(Db f ) = 2−n fbχ (u), for every u ∈ Fn2 ,
P
2
2
P
fχ b 2 (u)
u∈Fn
and thus N∆f ≥ 2−n maxu∈Fn2 fbχ (u). And we have Nfb ≥ 2
2 =
χ maxu∈Fn fbχ (u)
2
22n
2 . This proves Inequality (53). This inequality is an equality
maxu∈Fn fbχ (u)
2
if and only if both inequalities above are equalities, that is, if and only if,
for every b, the auto-correlation coefficient F(Db f ) equals 0 or 2n (−1)u0 ·b ,
2 2
where maxu∈Fn fb (u) = fb (u0 ), and if f is plateaued. The condition that
2 χ χ
Db f is either balanced or constant, for every b, is sufficient to imply that f
has the form (54): E 0 is the linear kernel of f and the restriction of f to E
has balanced derivatives. Conversely, any function of the form (54) is such
that Relation (53) is an equality. 2

These functions such that N∆f × Nfb = 2n are called partially-bent func-
χ
tions. Every quadratic function is partially-bent. Partially-bent functions
share with quadratic functions almost all of their nice properties (Walsh
spectrum easier to calculate, potential good nonlinearity and good resiliency
order), see [53]. In particular, the values of the Walsh transform equal 0 or
0
±2dim(E )+dim(E)/2 .

A generalization of Relation (53) has been obtained in [307]:

Proposition 27 Let ϕ be any nonzero n-variable pseudo-Boolean function.

Let Nϕ = |{x ∈ Fn2 / ϕ(x) 6= 0}| and Nϕb = |{u ∈ Fn2 / ϕ(u)
b 6= 0}|, then
Nϕ × Nϕb ≥ 2n .
Equality occurs if and only if there exists a number λ and a flat F of Fn2
such that ϕ(x) = λ(−1)u·x if x ∈ F and ϕ(x) = 0 otherwise.

Proof. Denoting by 1ϕ the indicator of the support {x ∈ Fn2 / ϕ(x) 6= 0} of ϕ,

b gives, for every u ∈
and replacing ϕ(x) by 1ϕ (x) ϕ(x) in the definition of ϕ,

106
Fn2 : ϕ(u) = x∈Fn 1ϕ (x)ϕ(x)(−1)u·x . Applying then the Cauchy-Schwartz
P
b
2
b2 (u) ≤ Nϕ x∈Fn ϕ2 (x) = 2−n Nϕ v∈Fn ϕ b2 (v) (accord-
P P
inequality gives ϕ
2 2
ing to Parseval’s relation (3)). Hence, ϕ b2 (u) ≤ 2−n Nϕ × Nϕb maxv∈Fn2 ϕb2 (v).
Choosing u such that ϕ 2
b (u) is maximum gives the desired inequality, since,
according to Parseval’s inequality, and ϕ being nonzero, this maximum can-
not be null.
Equality occurs if and only if all of the inequalities above are equalities, that
is, ϕb2 (v) takes only one nonzero value (say µ) and there exists a number λ
such that, for every u such that ϕ b2 (u) = µ, we have ϕ(x) 6= 0 ⇒ ϕ(x) =
u·x
λ(−1) . This is equivalent to the condition stated at the end of Proposi-
tion 27. 2

Partially-bent functions must not be mistaken for partial bent functions,

studied by P. Guillot in [172]. By definition, the Fourier transforms of par-
tial bent functions take exactly two values42 λ and λ + 2n/2 on Fn2 ∗ (n even).
Rothaus’ bound on the degree generalizes to partial bent functions. The
dual fe of f , defined by fe(u) = 0 if fb(u) = λ and fe(u) = 1 if fb(u) = λ + 2n/2 ,
is also partial bent; and its dual is f . Two kinds of partial bent functions
f exist: those such that fb(0) − f (0) = −λ(2n/2 − 1) and those such that
fb(0) − f (0) = (2n/2 − λ)(2n/2 + 1). This can be proved by applying Parse-
val’s Relation (23). The sum of two partial bent functions of the same kind,
whose supports have at most the zero vector in common, is partial bent. A
potential interest of partial bent functions is in the possibility of using them
as building blocks for constructing bent functions.

In spite of their good properties, partially-bent functions, when they are

not bent, have by definition nonzero linear structures and so do not give full
satisfaction. The class of plateaued functions, already encountered above in
Subsection 4.1 (and sometimes called three-valued functions) is a natural
extension of that of partially-bent functions. They have been first studied
by Zheng and Zhang in [363]. A function is called plateaued if its squared
Walsh transform takes at most one nonzero value, that is, if its Walsh trans-
form takes at most three values 0 and ±λ (where λ is some positive integer,
that we call the amplitude of the plateaued function). Bent functions are
plateaued and, according to Parseval’s Relation (23), a plateaued function
is bent if and only if its Walsh transform never takes the value 0.
Note that, according to Parseval’s relation again, denoting as above by Nfb
χ

42
Partial bent functions are the indicators of partial difference sets.

107
the cardinality of the support {a ∈ Fn2 / fbχ (a) 6= 0} of the Walsh transform of
2
a given n-variable Boolean function f , we have N b × maxa∈Fn fb (a) ≥ 22n
fχ 2 χ

and therefore, according to Relation (35) relating

! the nonlinearity to the

Walsh transform: nl(f ) ≤ 2n−1 1− q1 . Equality is achieved if and

Nfc
χ
only if f is plateaued.
Still because of Parseval’s relation, the amplitude λ of any plateaued func-
tion must be of the form 2r where r ≥ n/2 (since Nfb ≤ 2n ). Hence, the
χ

values of the Walsh transform of a plateaued function are divisible by 2n/2

if n is even and by 2(n+1)/2 if n is odd. The class of plateaued functions
contains those functions which achieve the best possible trade-offs between
resiliency, nonlinearity and algebraic degree: the order of resiliency and the
nonlinearity of any Boolean function are bounded by Sarkar et al.’s bound
(see Section 7 below) and the best compromise between those two criteria
is achieved by plateaued functions only; the third criterion – the algebraic
degree – is then also optimum. Other properties of plateaued functions can
be found in [42].
Plateaued functions can be characterized by second-order covering sequences
(see [93]):

Proposition 28 A Boolean function f on Fn2 is plateaued if and only if

there exists λ such that, for every x ∈ Fn2 :
X
(−1)Da Db f (x) = λ2 . (55)
a,b∈Fn
2

The proof is very similar to that of Proposition 52 and λ is necessarily the

amplitude of the plateaued function. Indeed, a function f is plateaued with
2
amplitude λ if and only if, for every u ∈ Fn2 , we have fbχ (u) fbχ (u) − λ2 = 0,
3
that is, fbχ (u) − λ2 fbχ (u) = 0. Applying the Fourier transform to both terms
of this equality and using Relation (20), we see that this is equivalent to the
fact that, for every a ∈ Fn2 , we have:
X
(−1)f (x)⊕f (y)⊕f (x+y+a) = λ2 (−1)f (a) .
x,y∈Fn
2

The fact that quadratic functions are plateaued is a direct consequence of

Proposition 28, since their second-order derivatives are constant; and Propo-
sition 28 gives more insight on the relationship between the nonlinearity of

108
a quadratic function and the number of its nonzero second-order derivatives.

P. Langevin proved in [230] that, if f is a plateaued function, then the coset

f ⊕ R(1, n) of the Reed-Muller code of order 1, is an orphan of R(1, n).
The notion of orphan has been introduced in [178] (with the term “urcoset”
instead of orphan) and studied in [31]. A coset of R(1, n) is an orphan if it is
maximum with respect to the following partial order relation: g ⊕ R(1, n) is
smaller than f ⊕ R(1, n) if there exists in g ⊕ R(1, n) an element g1 of weight
nl(g) (that is, of minimum weight in g ⊕ R(1, n)), and in f ⊕ R(1, n) an
element f1 of weight nl(f ), such that supp(g1 ) ⊆ supp(f1 ). Clearly, if f is a
function of maximum nonlinearity, then f ⊕ R(1, n) is an orphan of R(1, n)
(the converse is false, since plateaued functions with non-optimum nonlin-
earity exist). The notion of orphan can be used in algorithms searching for
functions with high nonlinearities.

6.9 Normal and non-normal bent functions

As observed in [54] (see Theorem 9 above), if a bent function f is normal
(resp. weakly-normal), that is, constant (resp. affine) on an n/2-dimensional
flat b+E (where E is a subspace of Fn2 ), then its dual fe is such that fe(u)⊕b·u
is constant on E ⊥ (resp. on a + E ⊥ , where a is a vector such that f (x) ⊕ a · x
is constant on E). Thus, fe is weakly-normal. Moreover, we have already
seen that f (resp. f (x) ⊕ a · x) is balanced on each of the other cosets of the
flat. H. Dobbertin used this idea to construct balanced functions with high
nonlinearities from normal bent functions (see Subsection 7.5.1).
The existence of non-normal (and even non-weakly-normal) bent functions,
i.e. bent functions which are non-constant (resp. non-affine) on every n/2-
dimensional flat, has been shown, contradicting a conjecture made by several
authors that such bent function did not exist. It is proved  in2k[139] that
 the
k
so-called Kasami function defined over F2n by f (x) = trn ax2 −2 +1 , with
gcd(k, n) = 1, is bent if n is not divisible by 3 and if a ∈ F2n is not a cube.
As shown in [45], if a ∈ F4 \ F2 and k = 3, then for n = 10, the function
f (x) ⊕ trn (b) is non-normal for some b, and for n = 14, the function f is
not weakly normal. Cubic bent functions on 8 variables are all normal, as
shown in [101].
The direct sum (see definition in Subsection 6.4) of two normal functions
is obviously a normal function, while the direct sum of two non-normal
functions can be normal. What about the sum of a normal bent function
and of a non-normal bent function? This question has been studied in [79].

109
To this aim, a notion more general than normality has been introduced as
follows:
Definition 8 Let U ⊆ V be two vector spaces over F2 . Let β : U → F2 and
f : V → F2 be bent functions. Then we say that f is a normal extension of
β, in symbols β  f , if there is a direct decomposition V = U ⊕ W1 ⊕ W2
such that
(i) β(u) = f (u + w1 ) for all u ∈ U, w1 ∈ W1 ,
(ii) dim W1 = dim W2 .
The relation  is transitive and if β  f then the same relation exists
between the duals: βe  fe.
A bent function is normal if and only if   f , where  ∈ F2 is viewed as a
Boolean functions over the vector space F02 = {0}.
Examples of normal extensions are given in [79] (including the construction
of Theorem 10 and its particular cases, the indirect sum and the extension
of Maiorana-McFarland type).
The clarification about the sum of a normal bent function and of a non-
normal bent function comes from the two following propositions:
Proposition 29 Let fi : Vi → F2 , i = 1, 2, be bent functions. The direct
sum f1 ⊕ f2 is normal if and only if bent functions β1 and β2 exist such that
fi is a normal extension of βi (i = 1, 2) and either β1 and β2 or β1 and
β2 ⊕ 1 are linearly equivalent.
Proposition 30 Suppose that β  f for bent functions β and f . If f is
normal, then also β is normal.
Hence, since the direct sum of a bent function β and of a normal bent
function g is a normal extension of β, the direct sum of a normal and a
non-normal bent function is always non-normal.
Normal extension leads to a secondary construction of bent functions:
Proposition 31 Let β be a bent function on U and f a bent function on
V = U × W × W . Assume that β  f . Let
β 0 : U → F2
be any bent function. Modify f by setting for all x ∈ U , y ∈ W

f 0 (x, y, 0) = β 0 (x),
while f 0 (x, y, z) = f (x, y, z) for all x ∈ U , y, z ∈ W , z 6= 0. Then f 0 is bent
and we have β 0  f 0 .

110
6.10 Kerdock codes
For every even n, the Kerdock code Kn [211] is a supercode of R(1, n) (i.e.
contains R(1, n) as a subset) and is a subcode of R(2, n). More precisely
Kn is a union of cosets fu ⊕ R(1, n) of R(1, n), where the functions fu are
quadratic (one of them is null and all the others have algebraic degree 2).
The difference fu ⊕ fv between two distinct functions fu and fv being bent,
Kn has minimum distance 2n−1 − 2n/2−1 (n even), which is the best possible
minimum distance for a code equal to a union of cosets of R(1, n), according
to the covering radius bound. The size of Kn equals 22n . This is the best
possible size for such minimum distance (see [129]). We describe now how
the construction of Kerdock codes can be simply stated.

6.10.1 Construction of the Kerdock code

The function M
f (x) = xi xj (56)
1≤i<j≤n

(which can also be defined as f (x) = wH2(x) [mod 2]) is bent43 because


M
the kernel of it associated symplectic form ϕ(x, y) = xi yj equals
1≤i6=j≤n
{0}. Thus, the linear code R(1, n) ∪ (f ⊕ R(1, n)) has minimum distance
2n−1 − 2n/2−1 .
We want to construct a code of size 22n with this same minimum distance.
We use the structure of field to this aim. We have recalled in Subsection 2.1
some properties of the field F2m (where m is any positive integer). In par-
ticular, we have seen that there exists α ∈ F2m (called a primitive element)
m
such that F2m = {0, α, α2 , · · · , α2 −1 }. Moreover, there exists α, primitive
2 m−1
element, such that (α, α2 , α2 , · · · , α2 ) is a basis of the vectorspace F2m .
Such basis is called a normal basis. If m is odd, then there exists a self-dual
i j
normal basis, that is, a normal basis such that: trm (α2 +2 ) = 1 if i = j and
i j
trm (α2 +2 ) = 0 otherwise, where trm is the trace function over F2m .
m−1
Consequence: for all x = x1 α + · · · + xm α2 in F2m , we have
m m
j +1
M M
trm (x) = xi trm (x2 )= xi xi+j ,
i=1 i=1

(where xi+j is replaced by xi+j−m if i + j > m).

43
We shall see in Section 10 that it is, up to the addition of affine functions, the sole
symmetric bent function.

111
The function f of Relation (56), viewed as a function f (x, xn ) on F2m × F2 ,
where m = n − 1 is odd – say m = 2t + 1 – can now be written as44 :
 
t
j
X
f (x, xn ) = trm  x2 +1  ⊕ xn trm (x) .
j=1

Notice that the associated symplectic form associated to f equals trm (x)trm (y)⊕
trm (xy) ⊕ xn trm (y) ⊕ yn trm (x).

Let us denote f (ux, xn ) by fu (x, xn ) (u ∈ F2m ), then Kn is defined as the

union, when u ranges over F2m , of the cosets fu ⊕ R(1, n).
Kn contains 2n+1 affine functions and 22n − 2n+1 quadratic bent func-
tions. Its minimum distance equals 2n−1 − 2n/2−1 because the sum of
two distinct functions fu and fv is bent. Indeed, the kernel of the asso-
ciated symplectic form equals the set of all ordered pairs (x, xn ) verifying
trm (ux)trm (uy) ⊕ trm (u2 xy) ⊕ xn trm (uy) ⊕ yn trm (ux) = trm (vx)trm (vy) ⊕
trm (v 2 xy) ⊕ xn trm (vy) ⊕ yn trm (vx) for every (y, yn ) ∈ F2m × F2 , that is,
utrm (ux) + u2 x + xn u = vtrm (vx) + v 2 x + xn v and trm (ux) = trm (vx); it is
a simple matter to show that it equals {(0, 0)}.

Open problem: Other examples of codes having the same parameters

exist [205]. All are equal to subcodes of the Reed-Muller code of order 2, up
to affine equivalence. We do not know how to obtain the same parameters
with non-quadratic functions. This would be useful for cryptographic pur-
poses as well as for the design of sequences for code division multiple access
(CDMA) in telecommunications.

Remark.
The Kerdock codes are not linear. However, they share some nice properties
with linear codes: the distance distribution between any codeword and all
the other codewords does not depend on the choice of the codeword (we say
that the Kerdock codes are distance-invariant; this results in the fact that
their distance enumerators are equal to their weight enumerators); and, as
proved by Semakov and Zinoviev [329], the weight enumerators of the Ker-
dock codes satisfy a relation similar to Relation (33), in which C is replaced
by Kn and C ⊥ is replaced by the so-called Preparata code of the same length
(we say that the Kerdock codes and the Preparata codes are formally dual).
An explanation of this astonishing property has been recently obtained [175]:
44
Obviously, this expression can be taken as the definition of f .

112
the Kerdock code is stable under an addition inherited of the addition in
Z4 = Z/4Z (we say it is Z4 -linear) and the Mac Williams identity still holds
in this different framework. Such an explanation had been an open problem
for two decades.

7 Resilient functions
We have seen in Subsection 4.1 that the combining functions in stream
ciphers must be m-resilient with large m. As any cryptographic functions,
they must also have high algebraic degrees and high nonlinearities.
Notation: by an (n, m, d, N )- function, we mean an n-variable, m-resilient
function having algebraic degree at least d and nonlinearity at least N .
There are necessary trade-offs between n, m, d and N .

7.1 Bound on algebraic degree

Siegenthaler’s bound states that any m-resilient function (0 ≤ m < n − 1)
has algebraic degree smaller than or equal to n − m − 1 and that any (n − 1)-
resilient function is affine45 . This can be proved directly by using Rela-
tion (3) and the originalLdefinition of resiliency given by Siegenthaler (Defi-
nition 3), since the bit x∈Fn / supp(x)⊆I f (x) equals the parity of the weight
2
of the restriction of f obtained by setting to 0 the coordinates of x which lie
outside I. Note that instead of using the original Siegenthaler’s definition
in the proof of Siegenthaler’s bound, we can also use the characterization by
Xiao and Massey, recalled in Theorem 3, together with the Poisson summa-
tion formula (18) applied to ϕ = f and with E ⊥ = {x ∈ Fn2 | supp(x) ⊆ I},
where I has size strictly greater than n − m − 1. But this gives a less simple
proof. Siegenthaler’s bound is also a direct consequence of a characterization
of resilient functions46 through their NNFs and of the fact that the algebraic
degrees of Boolean functions are smaller than or equal to their numerical
degrees:

Proposition 32 [88] Let n be any positive integer and m < n a non-

negative integer. A Boolean function f on Fn2 is m-resilient if and only
if the NNF of the function f (x) ⊕ x1 ⊕ · · · ⊕ xn has degree at most n − m − 1.
45
Siegenthaler also proved that any n-variable m-th order correlation-immune function
has degree at most n − m. This can be shown by using similar methods as for resilient
functions. Moreover, if such function has weight divisible by 2m+1 then it satisfies the
same bound as m-resilient functions.
46
A similar characterization of correlation-immune functions can be found in [63].

113
Proof. Let us denote by g(x) the function f (x) ⊕ x1 ⊕ · · · ⊕ xn . For each
vector a ∈ Fn2 , we denote by a the componentwise complement of a equal
to a + (1, · · · , 1). We have fbχ (a) = gc
χ (a). Thus, f is m-resilient if and only
if, for each vector u of weight greater than or equal to n − m, the number
gcχ (u) is null. Consider the NNF of g:

X
g(x) = λI xI .
I∈P(N )

According to Relations (30), (31) and (12) applied to g, we have for nonzero
u: X
wH (u)+1
gc
χ (u) = (−1) 2n−|I|+1 λI ,
I∈P(N ) | supp(u)⊆I

and for nonempty I:

X
λI = 2−n (−2)|I|−1 gc
χ (u).
u∈Fn
2 | I⊆supp(u)

We deduce that gc χ (u) is null for every vector u of weight greater than or
equal to n − m if and only if the NNF of g has degree at most n − m − 1.2

Proposition 32 can also be proved by using the Xiao-Massey characteriza-

tion (again) and Relation (8) relating the values of the coefficients of the
NNF to the values of the function, applied to g(x) = f (x) ⊕ x1 ⊕ · · · ⊕ xn .

Proposition 32 has been used by X.-D. Hou in [193] for constructing resilient
functions. Siegenthaler’s bound gives an example of the trade-offs which
must be accepted in the design of combiner generators47 . Sarkar and Maitra
showed in [321] that the values of the Walsh Transform of an n-variable,
m-resilient (resp. m-th order correlation-immune) function are divisible
by 2m+2 (resp. 2m+1 ) if m ≤ n − 2 (a proof of a slightly more precise result
is given in the next subsection, at Theorem 13)48 . This Sarkar-Maitra’s
divisibility bound (which implies in particular that the weight of any m-th
order correlation-immune function is divisible by 2m ) allows also to deduce
Siegenthaler’s bound, thanks to Proposition 11 applied with k = m+2 (resp.
k = m + 1).
47
One approach to avoid such trade-off is to allow memory in the nonlinear combination
generator, that is, to replace the combining function by a finite state machine, see [277].
48
More is proved in [63, 94]; in particular: if the weight of an m-th order correlation-
immune is divisible by 2m+1 , then the values of its Walsh Transform are divisible by
2m+2 .

114
7.2 Bounds on the nonlinearity
Sarkar-Maitra’s divisibility bound, recalled at the end of the previous sub-
section, has provided a nontrivial upper bound on the nonlinearity of re-
silient functions, independently obtained by Tarannikov [345] and by Zheng
and Zhang [366]: the nonlinearity of any m-resilient function (m ≤ n − 2)
is bounded above by 2n−1 − 2m+1 . This bound is tight, at least when
m ≥ 0.6 n, see [345, 346]49 . We shall call it Sarkar et al.’s bound . No-
tice that, if an m-resilient function f achieves nonlinearity 2n−1 − 2m+1 ,
then f is plateaued. Indeed, the distances between f and affine functions
lie then between 2n−1 − 2m+1 and 2n−1 + 2m+1 and must be therefore equal
to 2n−1 − 2m+1 , 2n−1 and 2n−1 + 2m+1 because of the divisibility result of
Sarkar and Maitra. Thus, the Walsh transform of f takes three values 0 and
±2m+2 . Moreover, it is proved in [345] that such function f also achieves
Siegenthaler’s bound (and as proved in [261], achieves minimum sum-of-
squares indicator). These last properties can also be deduced from a more
precise divisibility bound shown later in [63]:
Theorem 13 Let f be any n-variable m-resilient function (m ≤ n − 2)
and let d be its algebraic degree. The values of the Walsh transform of f
n−m−2
are divisible by 2m+2+b d c . Hence the nonlinearity of f is divisible
n−m−2
by 2m+1+b d c .
The approach for proving this result was first to use the numerical normal
form (see [63]). Later, a second proof using only the properties of the Fourier
transform was given in [94]:
Proof. The Poisson summation formula (18) applied to ϕ = fχ and to the
vectorspace E = {u ∈ Fn2 / ∀i ∈ N, ui ≤ vi } where v is some vector of Fn2 ,
whose orthogonal equals E ⊥ = {u ∈ Fn2 / ∀i ∈ N, ui ≤ vi ⊕ 1}, gives
wH (v)
P P
u∈E fχ (u) = 2 x∈E ⊥ fχ (x). It is then a simple matter to prove the
b
result by induction on the weight of v, starting with the vectors of weight
m + 1 (since it is obvious for the vectors of weights at most m), and using
McEliece’s divisibility property (see Subsection 3.1). 2
A similar proof shows that the values of the Walsh transform of any m-
n−m−1
th order correlation-immune function are divisible by 2m+1+b d c (and
n−m−2 n−m−2
by 2m+2+b d c if its weight is divisible 2m+1+b d c , see [94]).
49
Also Zheng and Zhang [366], showed that the upper bound on the nonlinearity of
correlation-immune functions of high orders is the same as the upper bound on the non-
linearity of resilient functions of the same orders. The distances between resilient functions
and Reed-Muller codes of orders greater than 1 have also been studied by Kurosawa et al.
and by Borissov at al. [221, 23].

115
Theorem 13 gives directly a more precise upper bound on the nonlinear-
ity of any m-resilient function of degree d: this nonlinearity is bounded
n−m−2
above by 2n−1 − 2m+1+b d c . This gives a simpler proof that it can
be equal to 2n−1 − 2m+1 only if d = n − m − 1, i.e. if Siegenthaler’s
bound is achieved. Moreover, the proof above also shows that the non-
linearity of any m-resilient n-variable Boolean function is bounded above
n−m−2
by 2n−1 − 2m+1+b d c where d is the minimum algebraic degree of the
restrictions of f to the subspaces {u ∈ Fn2 / ∀i ∈ N, ui ≤ vi ⊕ 1} such that v
has weight m + 1 and fbχ (v) 6= 0.

If 2n−1 − 2m+1 is greater than the best possible nonlinearity of all balanced
functions (and in particular if it is greater than the covering radius bound)
then, obviously, a better bound exists. In the case of n even, the best
possible nonlinearity of all balanced functions being strictly smaller than
2n−1 − 2n/2−1 , Sarkar and Maitra deduce that nl(f ) ≤ 2n−1 − 2n/2−1 −
2m+1 for every m-resilient function f with m ≤ n/2 − 2. In the case of n
odd, they state that nl(f ) is smaller than or equal to the highest multiple
of 2m+1 , which is less than or equal to the best possible nonlinearity of
all Boolean functions. But a potentially better upper bound can be given,
whatever is the parity of n. Indeed, Sarkar-Maitra’s divisibility bound shows
that fbχ (a) = ϕ(a) × 2m+2 where ϕ(a) is integer-valued. But Parseval’s
Relation (23) and the fact that fbχ (a) is null for every vector a of weight ≤ m
imply X
ϕ2 (a) = 22n−2m−4
a∈Fn
2 / wH (a)>m

and, thus,
s
22n−2m−4 2n−m−2
maxn |ϕ(a)| ≥ m n

=q
.
2n − i=0 i
P
a∈F2
2n − m n
P
i=0 i
& '
n−m−2
Hence, we have maxa∈Fn2 |ϕ(a)| ≥ q 2 P , and this implies:
n
2n − mi=0 ( i )

 
2n−m−2
nl(f ) ≤ 2n−1 − 2m+1  q
 .
 (57)
 2n − m n
 P
i=0

i

When n is even and m ≤ n/2 − 2, this number is always less than or

equal to the number 2n−1 − 2n/2−1 − 2m+1 (given by Sarkar and Maitra),

116
 n−m−2
because q 2 P is strictly greater than 2n/2−m−2 and 2n/2−m−2 is an
n
2n − m
i=0 ( i )
& '
n−m−2
integer, and, thus, q 2 P is at least 2n/2−m−2 + 1. And when
n
2n − m
i=0 ( i )

n increases, the right hand-side of Relation (57) is strictly smaller than

2n−1 − 2n/2−1 − 2m+1 for an increasing number of values of m ≤ n/2 − 2
(but this improvement does not appear when we compare the values we
obtain with this bound to the values indicated in the table given by Sarkar
and Maitra in [321], because the values of n they consider in this table are
small).
When n is odd, it is difficult to say if Inequality (57) is better than the
bound given by Sarkar and Maitra, because their bound involves a value
which is unknown for n ≥ 9 (the best possible nonlinearity of all balanced
Boolean functions). In any case, this makes (57) better usable.
nH (m/n)
We know (see [258], page 310) that m n √2 2
P

i=0 i ≥ , where H2 (x) =
8m(1−m/n)
−x log2 (x) − (1 − x) log2 (1 − x) is the so-called binary entropy function and
satisfies H2 ( 12 − x) = 1 − 2x2 log2 e + o(x2 ). Thus, we have
 
2n−m−2
nl(f ) ≤ 2n−1 − 2m+1 
 
r . (58)
 2n − √2nH2 (m/n)
 

 8m(1−m/n) 

7.3 Bound on the maximum correlation with subsets of N

An upper bound on the maximum correlation of m-resilient functions with
respect to subsets I of N can be directly deduced from Relation (40) and
from Sarkar et al.’s bound. Note that we get an improvement by using
that the support of fbχ , restricted to the set of vectors u ∈ Fn2 such that
P|I|
ui = 0, ∀i 6∈ I, contains at most i=m+1 |I|


i vectors. In particular, if |I| =
m + 1, the maximum correlation of f with respect to I equals 2−n |fbχ (u)|,
where u is the vector of support I, see [38, 47, 358]. The optimal number
of LFSRs which should be considered together in a correlation attack on a
cryptosystem using an m-resilient combining function is m + 1, see [38].

7.4 Relationship with other criteria

The relationships between resiliency and other criteria have been studied
in [105, 261, 348, 365]. For instance, m-resilient P C(l) functions can exist
only if m + l ≤ n − 1. This is a direct consequence of Relation (27), relating

117
the values of the Walsh transform of a function on a flat a+E to the autocor-
relation coefficients of the function on a flat b + E ⊥ , applied with a = b = 0,
E = {x ∈ Fn2 ; xi = 0, ∀i ∈ I} and E ⊥ = {x ∈ Fn2 ; xi = 0, ∀i 6∈ I}, where I
has size n − m: if l ≥ n − m then the right-hand term of (27) is non-zero
while the left-hand term is null. Equality m + l = n − 1 is possible only if
l = n − 1, n is odd and m = 0 [365, 105]. The known upper bounds on the
nonlinearity (see Section 7) can then be improved with the same argument.

The definition of resiliency has been weakened in [27, 89, 222] in or-
der to relax some of the trade-offs recalled above without weakening the
cryptosystem against the correlation attack.
Resiliency is related to the notion of corrector (useful for the generation
of random sequences having good statistical properties) introduced by P.
Lacharme in [224].

7.5 Constructions
High order resilient functions with high algebraic degrees, high nonlineari-
ties and good immunity to algebraic attacks are needed for applications in
stream ciphers using the combiner model. But designing constructions of
Boolean functions meeting all these cryptographic criteria is still a challenge
nowadays (while we would need numerous such functions in order to be able
choosing among them functions satisfying additional design criteria). The
primary constructions (which allow designing resilient functions without us-
ing known ones) lead potentially to wider classes of functions than secondary
constructions (recall that the number of Boolean functions on n − 1 vari-
ables is only equal to the square root of the number of n-variable Boolean
functions). But the known primary constructions of such Boolean functions
do not lead to very large classes of functions. In fact, only one reasonably
large class of Boolean functions is known, whose elements can be analyzed
with respect to the cryptographic criteria recalled in Subsection 4.1. So we
observe some imbalance in the knowledge on cryptographic functions for
stream ciphers: much is known on the properties of resilient functions, but
little is known on how constructing them. Examples of m-resilient functions
achieving the best possible nonlinearity 2n−1 − 2m+1 (and thus the best al-
gebraic degree) have been obtained for n ≤ 10 in [292, 320, 321] and for
every m ≥ 0.6 n [345, 346] (n being then not limited). But n ≤ 10 is too
small for applications and m ≥ 0.6 n is too large (because of Siegenthaler’s
bound) and almost nothing is known on the immunity of these functions to
algebraic attacks. Moreover, these examples give very limited numbers of

118
functions (they are often defined recursively or obtained after a computer
search) and many of these functions have cryptographic weaknesses such as
linear structures (see [105, 261]). Balanced Boolean functions with high non-
linearities have been obtained by C. Fontaine in [154] and by E. Filiol and
C. Fontaine in [152], who made a computer investigation - but for n = 7, 9
which is too small - on the corpus of idempotent functions. These functions,
whose ANFs are invariant under the cyclic shifts of the coordinates xi , have
been called later rotation symmetric (see Subsection 10.6).

7.5.1 Primary constructions

Maiorana-McFarland’s construction: An extension of the class of bent
functions that we called above the Maiorana-McFarland original class has
been given in [35], based on the same principle of concatenating affine func-
tions50 (we have already met this generalization in Section 6, that we shall
call the Maiorana-McFarland general construction): let r be a positive in-
teger smaller than n; we denote n − r by s; let g be any Boolean function
on Fs2 and let φ be a mapping from Fs2 to Fr2 . Then, we define the function:
r
M
fφ,g (x, y) = x · φ(y) ⊕ g(y) = xi φi (y) ⊕ g(y), x ∈ Fr2 , y ∈ Fs2 (59)
i=1

where φi (y) is the i-th coordinate function of φ(y).

For every a ∈ Fr2 and every b ∈ Fs2 , we have seen in Subsection 6.4 that
X
r
φ,g χ (a, b) = 2
fd (−1)g(y)⊕b·y . (60)
y∈φ−1 (a)

This can be used to design resilient functions: if every element in φ(Fs2 )

has Hamming weight strictly greater than k, then fφ,g is m-resilient with
m ≥ k (in particular, if φ(Fs2 ) does not contain the null vector, then fφ,g
is balanced). Indeed, if wH (a) ≤ k then φ−1 (a) is empty in Relation (60);
hence, if wH (a) + wH (b) ≤ k then fd φ,g χ (a, b) is null. The k-resiliency of fφ,g
under this hypothesis can also be deduced from the facts that any affine
function x ∈ Fr2 7→ a · x ⊕ ε (a ∈ Fr2 nonzero, ε ∈ F2 ) is (wH (a) − 1)-resilient,
and that any Boolean function equal to the concatenation of k-resilient func-
tions is a k-resilient function (see secondary construction 3 below).

50
These functions have also been studied under the name of linear-based functions
in [1, 355].

119
Degree: The algebraic degree of fφ,g is at most s+1 = n−r+1. It equals s+1
if and only if φ has algebraic degree s (i.e. if at least one of its coordinate
functions has algebraic degree s). If we assume that every element in φ(Fs2 )
has Hamming weight strictly greater than k, then φ can have algebraic de-
gree s only if k ≤ r − 2, since if k = r − 1 then φ is constant. Thus, if m = k
then the algebraic degree of fφ,g reaches Siegenthaler’s bound n − k − 1 if
and only if either k = r − 2 and φ has algebraic degree s = n − k − 2 or
k = r − 1 and g has algebraic degree s = n − k − 1. There are cases where
m > k (see [118, 64, 65]). An obvious one is when each set φ−1 (a) has even
size and the restriction of g to this set is balanced: then m ≥ k + 1.

Nonlinearity: Relations (35), relating the nonlinearity to the Walsh trans-

form, and (60) lead straightforwardly to a general lower bound on the non-
linearity of Maiorana-McFarland’s functions (first observed in [327]):

nl(fφ,g ) ≥ 2n−1 − 2r−1 maxr |φ−1 (a)| (61)

a∈F2

(where |φ−1 (a)| denotes the size of φ−1 (a)). A more recent upper bound
& '
r
nl(fφ,g ) ≤ 2n−1 − 2r−1 maxr |φ−1 (a)| (62)
a∈F2

obtained in [64] strengthens a bound previously obtained in [107, 108] which

stated nl(fφ,g ) ≤ 2n−1 − 2r−1 .

Proof of (62): The sum

 2  
X X X X
 (−1)g(y)+b·y  =  (−1)g(y)+g(z)+b·(y+z) 
b∈Fs2 y∈φ−1 (a) b∈Fs2 y,z∈φ−1 (a)

equals 2s |φ−1 (a)| (since the sum b∈Fs (−1)b·(y+z) is null if y 6= z). The
P
2
maximum of a set of values being always greater than or equal to its mean,
we deduce  
 X 
  p
g(y)+b·y  −1
maxs  (−1)  ≥ |φ (a)|

b∈F2 
y∈φ−1 (a) 
and thus, according to Relation (60):
& '
r
max |fbχ φ,g (a, b)| ≥ 2r maxr |φ−1 (a)| .
a∈Fr2 ;b∈Fs2 a∈F2

120
Relation (35) completes the proof. 2

This new bound allowed characterizing the Maiorana-McFarland’s func-

tions fφ,g such that wH (φ(y)) > k for every y and achieving nonlinearity
s −1
r+ 2
2
2n−1 − 2k+1 : the inequality nl(fφ,g ) ≤ 2n−1 − qP implies either that
r
i=k+1 (ri)
r = k + 1 or r = k + 2.
If r = k + 1, then φ is the constant (1, · · · , 1) and n ≤ k + 3. Either s = 1
and g(y) is then any function in one variable, or s = 2 and g is then any
function of the form y1 y2 ⊕ `(y) where ` is affine (thus, f is quadratic).
If r = k + 2, then φ is injective, n ≤ k + 2 + log2 (k + 3), g is any function
on n − k − 2 variables and d◦ fφ,g ≤ 1 + log2 (k + 3).
A simple example of k-resilient Maiorana-McFarland’s functions such
that nl(fφ,g ) = 2n−1 − 2k+1 (and thus achieving Sarkar et al.’s bound) can
be given for any r ≥ 2s − 1 and for k = r − 2 (see [64]). And, for every
even n ≤ 10, Sarkar et al.’s bound with m = n/2 − 2 can be achieved by
Maiorana-McFarland’s functions. Also, functions with high nonlinearities
but achieving not Sarkar et al.’s bound exist in Maiorana-McFarland’s class
(for instance, for every n ≡ 1 [ mod 4], there exist such n−1
4 -resilient functions
n−1
n
on F2 with nonlinearity 2 n−1 − 2 2 ).

Generalizations of Maiorana-McFarland’s construction have been

introduced in [64] and [93]; the latter generalization has been further gen-
eralized into a class introduced in [68]. A motivation for introducing such
generalizations is that Maiorana-McFarland’s functions have the weakness
that x 7→ fφ,g (x, y) is affine for every y ∈ Fs2 and have high divisibilities
of their Fourier spectra (indeed, if we want to ensure that f is m-resilient
with large value of m, then we need to choose r large; then the Walsh spec-
trum of f is divisible by 2r according to Relation (60); there is also a risk
that this property can be used in attacks, as it is used in [48] to attack
block ciphers). The functions constructed in [64, 93] are concatenations of
quadratic functions instead of affine functions. This makes them harder to
study than Maiorana-McFarland’s functions. But they are more numerous
and more general. Two classes of such functions have been studied:
- the functions of the first class are defined as:
t
M
fψ,φ,g (x, y) = x2i−1 x2i ψi (y) ⊕ x · φ(y) ⊕ g(y),
i=1

121
with x ∈ Fr2 , y ∈ Fs2 , where n = r + s, t = 2r , and where ψ : Fs2 → Ft2 ,
 

φ : Fs2 → Fr2 and g : Fs2 → F2 can be chosen arbitrarily;

- the functions of the second class are defined as:

fφ1 ,φ2 ,φ3 ,g (x, y) = (x · φ1 (y)) (x · φ2 (y)) ⊕ x · φ3 (y) ⊕ g(y),

with x ∈ Fr2 , y ∈ Fs2 , where φ1 , φ2 and φ3 are three functions from Fs2 into
Fr2 and g is any Boolean function on Fs2 . The size of this class roughly equals
 r 2s 3 s s
(2 ) × 22 = 2(3r+1)2 (the exact number, which is unknown, is smaller
since a same function can be represented in this form in several ways) and is
s s s
larger than the size of the first class, roughly equal to (2t )2 × (2r )2 × 22 =
s
2(t+r+1)2 .
The second construction has been generalized in [68]. The functions of this
generalized class are the concatenations of functions equal to the sums of r-
variable affine functions and of flat-indicators:
ϕ(y)
Y
∀(x, y) ∈ Fr2 × Fs2 , f (x, y) = (x · φi (y) ⊕ gi (y) ⊕ 1) ⊕ x · φ(y) ⊕ g(y),
i=1

where ϕ is a function from Fs2 into {0, 1, · · · , r}, φ1 , · · · , φr and φ are functions
from Fs2 into Fr2 such that, for every y ∈ Fs2 , the vectors φ1 (y), · · · , φϕ(y) (y)
are linearly independent, and g1 , · · · , gr and g are Boolean functions on Fs2 .
There exist formulae for the Walsh transforms of the functions of these
classes, which result in sufficient conditions for their resiliency and in bounds
on their nonlinearities (see [64, 68]).

Other constructions: We first make a preliminary observation. Let k <

n. For any k-variable function g, any surjective linear mapping L : Fn2 → Fk2
and any element s of Fn2 ; the function f (x) = g◦L(x)⊕s·x is (d−1)-resilient,
where d is the Hamming distance between s and the linear code C whose
generator matrix equals the matrix of L. Indeed, for any vector a ∈ Fn2
of Hamming weight at most d − 1, the vector s + a does not belong to C.
This implies that the Boolean function f (x) ⊕ a · x is linearly equivalent
to the function g(x1 , · · · , xk ) ⊕ xk+1 , since we may assume without loss of
generality that L is systematic (i.e. has the form [Idk |N ]); it is therefore
balanced. But such function f having nonzero linear structures, it does not
give full satisfaction.

A construction derived from PS ap construction is introduced in [58] to obtain

resilient functions: let k and r be positive integers and n ≥ r; we denote

122
n − r by s; the vectorspace Fr2 is identified to the Galois field F2r . Let g be
any Boolean function on F2r and φ an F2 -linear mapping from Fs2 to F2r ; set
a ∈ F2r and b ∈ Fs2 such that, for every y in Fs2 and every z in F2r , a + φ(y)
is nonzero and φ∗ (z) + b has weight greater than k, where φ∗ is the adjoint
of φ (satisfying u · φ(x) = φ∗ (u) · x for every x and u, that is, having for
matrix the transpose of that of φ). Then, the function
 
x
f (x, y) = g ⊕ b · y, where x ∈ F2r , y ∈ Fs2 , (63)
a + φ(y)

is m-resilient with m ≥ k. There exist bounds on the nonlinearities of these

functions (see [65]), similar to those existing for Maiorana-McFarland’s func-
tions. But this class has much fewer elements than Maiorana-McFarland’s
class, because φ is linear.

Dobbertin’s construction: in [141] is given a nice generalization of a method,

introduced by Seberry et al. in [328], for modifying bent functions into
balanced functions with high nonlinearities. He observes that most known
bent functions on Fn2 (n even) are normal (that is, constant on at least one
n/2-dimensional flat). Up to affine equivalence, we can then assume that
n/2 n/2
f (x, y), x ∈ F2 , y ∈ F2 is such that f (x, 0) = ε (ε ∈ F2 ) for every
n/2
x ∈ F2 and that ε = 0 (otherwise, consider f ⊕ 1).
n/2 n/2
Proposition 33 Let f (x, y), x ∈ F2 , y ∈ F2 be any bent function such
n/2
that f (x, 0) = 0 for every x ∈ F2 and let g be any balanced function
n/2
on F2 . Then the Walsh transform of the function h(x, y) = f (x, y) ⊕
δ0 (y)g(x), where δ0 is the Dirac symbol, satisfies:

h
c (u, v) = 0 if u = 0 and h
χ
c (u, v) = fb (u, v) + gc(u) otherwise.
χ χ χ (64)

c (u, v) = fb (u, v)−P n/2 (−1)u·x +P n/2 (−1)g(x)⊕u·x =

Proof. We have hχ χ x∈F x∈F
2 2
fbχ (u, v) − 2n/2 δ0 (u) + gc
χ (u). The function g being balanced, we have gcχ (0) =
n/2 n/2
0. And fb (0, v) equals 2
χ for every v, since f is null on F 2× {0} and
n/2
according to Relation (46) applied to E = {0} × F2 and a = b = 0 (or see
the remark after Theorem 9). 2

We deduce that:

max |h
c (u, v)| ≤ max |fb (u, v)| + max |c
χ χ gχ (u)|,
n/2 n/2 n/2
u,v∈F2 u,v∈F2 u∈F2

123
i.e. that 2n − 2nl(h) ≤ 2n − 2nl(f ) + 2n/2 − 2nl(g), that is:

nl(h) ≥ nl(f ) + nl(g) − 2n/2−1 = 2n−1 − 2n/2 + nl(g).

Applying recursively this principle (if n/2 is even, g can be constructed in

the same way), we see that if n = 2k n0 (n0 odd), Dobbertin’s method allows
n 0 n0 −1
reaching the nonlinearity 2n−1 − 2n/2−1 − 2 4 −1 − · · · − 2n −1 − 2 2 since
0
we know that, for every odd n0 , the nonlinearity of functions on Fn2 can
0 n0 −1
be as high as 2n −1 − 2 2 , and that balanced (quadratic) functions can
achieve this value. If n0 ≤ 7 then this value is the best possible and 2n−1 −
n 0 n0 −1
2n/2−1 −2 4 −1 −· · ·−2n −1 −2 2 is therefore the best known nonlinearity of
balanced functions in general. For n0 > 7, the best nonlinearity of balanced
0 n0 −1
n0 -variable functions is larger than 2n −1 − 2 2 (see the paragraph devoted
n 0
to nonlinearity in Section 4.1) and 2n−1 − 2n/2−1 − 2 4 −1 − · · · − 22n −1 −
0
2n + nl(g), where g is an n0 -variable balanced function, can therefore reach
higher values.
Unfortunately, according to Relation (64), Dobbertin’s construction can-
not produce m-resilient functions with m > 0 since, g being a function de-
n/2
fined on F2 , there cannot exist more than one vector a such that gc χ (a)

equals ±2n/2 .

7.5.2 Secondary constructions

There exist several simple secondary constructions, which can be combined
to obtain resilient functions achieving the bounds of Sarkar et al. and Siegen-
thaler. We list them below in chronological order. As we shall see in the
end, they all are particular cases of a single general one.

I Direct sum of functions

A. Adding a variable
Let f be an r-variable t-resilient function. The Boolean function on Fr+1
2 :

h(x1 , · · · , xr , xr+1 ) = f (x1 , · · · , xr ) ⊕ xr+1

is (t + 1)-resilient [336]. If f is an (r, t, r − t − 1, 2r−1 − 2t+1 ) function51 ,

then h is an (r + 1, t + 1, r − t − 1, 2r − 2t+2 ) function, and thus achieves
Siegenthaler’s and Sarkar et al.’s bounds. But h has the linear structure
(0, · · · , 0, 1).
51
Recall that, by an (n, m, d, N )- function, we mean an n-variable, m-resilient function
having algebraic degree at least d and nonlinearity at least N .

124
B. Generalization
If f is an r-variable t-resilient function (t ≥ 0) and if g is an s-variable
m-resilient function (m ≥ 0), then the function:
h(x1 , · · · , xr , xr+1 , · · · , xr+s ) = f (x1 , · · · , xr ) ⊕ g(xr+1 , · · · , xr+s )

is (t+m+1)-resilient. This comes from the easily provable relation h

c (a, b) =
χ

fb (a) × gc(b), a ∈ F , b ∈ F . We have also d h = max(d f, d◦ g) and,

χ χ
r s
2
◦
2
◦

thanks to Relation (35) relating the nonlinearity to the Walsh transform,

nl(h) = 2r+s−1 − 21 (2r −2nl(f ))(2s −2nl(g)) = 2r nl(g)+2s nl(f )−2nl(f )nl(g).
Such decomposable function does not give full satisfaction. Moreover, h has
low algebraic degree, in general. And if nl(f ) = 2r−1 − 2t+1 (t ≤ r − 2) and
nl(g) = 2s−1 − 2m+1 (m ≤ s − 2), i.e. if nl(f ) and nl(g) have maximum pos-
sible values, then nl(h) = 2r+s−1 − 2t+m+3 and h does not achieve Sarkar’s
and Maitra’s bound (note that this is not in contradiction with the proper-
ties of the construction recalled in I.A, since the function g(xr+1 ) = xr+1 is
1-variable, 0-resilient and has null nonlinearity).
Function h has no nonzero linear structure if and only if f and g both have
no nonzero linear structure.

II. Siegenthaler’s construction

Let f and g be two Boolean functions on Fr2 . Let us consider the function
h(x1 , · · · , xr , xr+1 ) = (xr+1 ⊕ 1)f (x1 , · · · , xr ) ⊕ xr+1 g(x1 , · · · , xr )
on Fr+1
2 . Note that the truth-table of h can be obtained by concatenating
the truth-tables of f and g. Then:
c (a1 , · · · , ar , ar+1 ) = fb (a1 , · · · , ar ) + (−1)ar+1 gc(a1 , · · · , ar ).
h (65)
χ χ χ

Thus:
1. If f and g are m-resilient, then h is m-resilient [336]; moreover, if
for every a ∈ Fr2 of Hamming weight m + 1, we have fbχ (a) + gc χ (a) = 0,
then h is (m + 1)-resilient. Note that the construction recalled in I.A
corresponds to g = f ⊕ 1 and satisfies this condition. Another possible
choice of a function g satisfying this condition (first pointed out in [35])
is g(x) = f (x1 ⊕ 1, · · · , xr ⊕ 1) ⊕ , where  = m [ mod 2], since gc χ (a) =
f (x)⊕⊕(x⊕(1,···,1))·a +w (a)
P
x∈Fr2 (−1) = (−1) fχ (a). It leads to a function h
H b
having also a nonzero linear structure (namely, the vector (1, · · · , 1));
2. The value max |h
c (a1 , · · · , ar , ar+1 )| is bounded above by the num-
χ
a1 ,···,ar+1 ∈F2
ber max |fbχ (a1 , · · · , ar )| + max gχ (a1 , · · · , ar )|; this implies 2r+1 −
|c
a1 ,···,ar ∈F2 a1 ,···,ar ∈F2

125
2nl(h) ≤ 2r+1 − 2nl(f ) − 2nl(g), that is nl(h) ≥ nl(f ) + nl(g);
a. if f and g achieve maximum possible nonlinearity 2r−1 − 2m+1 and if h is
(m + 1)-resilient, then the nonlinearity 2r − 2m+2 of h is the best possible;
b. if f and g are such that, for every vector a, at least one of the numbers
fbχ (a), gc
χ (a) is null (in other words, if the supports of the Walsh transforms

 g are disjoint), then we have maxa1 ,···,ar+1 ∈F2 |hχ (a1 , · · · , ar , ar+1 )| =
of f and c
max maxa ,···,a ∈F |fb (a1 , · · · , ar )|; maxa ,···,a ∈F |c
1 r 2 χ 1 r 2 g (a1 , · · · , ar )| . Hence we
χ

have 2r+1 − 2nl(h) = 2r− 2 min(nl(f ), nl(g)) and nl(h) equals therefore
2r−1 + min(nl(f ), nl(g)); thus, if f and g achieve best possible nonlinearity
2r−1 − 2m+1 , then h achieves best possible nonlinearity 2r − 2m+1 ;
3. If the monomials of highest degree in the algebraic normal forms of f
and g are not all the same, then d◦ h = 1 + max(d◦ f, d◦ g). Note that this
condition is not satisfied in the two cases indicated above in 1, for which h
is (m + 1)-resilient.
4. For every a = (a1 , · · · , ar ) ∈ Fr2 and every ar+1 ∈ F2 , we have, de-
noting (x1 , · · · , xr ) by x: D(a,ar+1 ) h(x, xr+1 ) = Da f (x) ⊕ ar+1 (f ⊕ g)(x) ⊕
xr+1 Da (f ⊕ g)(x) ⊕ ar+1 Da (f ⊕ g)(x). If d◦ (f ⊕ g) ≥ d◦ f , then D(a,1) h is
non-constant, for every a. And if, additionally, there does not exist a 6= 0
such that Da f and Da g are constant and equal to each other, then h admits
no nonzero linear structure.
This construction allows obtaining:
- from any two m-resilient functions f and g having disjoint Walsh spectra,
achieving nonlinearity 2r−1 − 2m+1 and such that d◦ (f ⊕ g) = r − m − 1, an
m-resilient function h having algebraic degree r − m and having nonlinearity
2r − 2m+1 , that is, achieving Siegenthaler’s and Sarkar et al.’s bounds; note
that this construction increases (by 1) the algebraic degrees of f and g;
- from any m-resilient function f achieving algebraic degree r − m − 1 and
nonlinearity 2r−1 − 2m+1 , a function h having resiliency order m + 1 and
nonlinearity 2r − 2m+2 , that is, achieving Siegenthaler’s and Sarkar et al.’s
bounds and having same algebraic degree as f (but having nonzero linear
structures).
So it allows, when combining these two methods, to keep best tradeoffs be-
tween resiliency order, algebraic degree and nonlinearity, and to increase by
1 the degree and the resiliency order.
Generalization: let (fy )y∈Fs2 be a family of r-variable m-resilient functions;
then the function on Fr+s 2 defined by f (x, y) = fy (x) (x ∈ Fr2 , y ∈ Fs2 ) is
m-resilient. Indeed, we have fbχ (a, b) = y∈Fs (−1)b·y fc
P
y χ (a). The function f
2
corresponds to the concatenation of the functions fy ; hence, this secondary
construction can be viewed as a generalization of Maiorana-McFarland’s

126
construction (in which the functions fy are m-resilient affine functions).

More on the resilient functions, achieving high nonlinearities, and con-

structed by using, among others, the secondary constructions above (as well
as algorithmic methods) can be found in [216, 291].

III. Tarannikov’s elementary construction

Let g be any Boolean function on Fr2 . We define the Boolean function
h on Fr+1 2 by h(x1 , · · · , xr , xr+1 ) = xr+1 ⊕ g(x1 , · · · , xr−1 , xr ⊕ xr+1 ). By
the change of variable xr ← xX r ⊕ xr+1 , we see that the Walsh transform
hχ (a1 , · · · , ar+1 ) is equal to
c (−1)a·x⊕g(x1 ,···,xr )⊕ar xr ⊕(ar ⊕ar+1 ⊕1)xr+1 ,
x1 ,···,xr+1 ∈F2
where a = (a1 , · · · , ar−1 ) and x = (x1 , · · · , xr−1 ); if ar+1 = ar then this value
is null and if ar = ar+1 ⊕ 1 then it equals 2 gc χ (a1 , · · · , ar−1 , ar ). Thus:
1. nl(h) = 2 nl(g);
2. If g is m-resilient, then h is m-resilient. If, additionally, gc χ (a1 , · · · , ar−1 , 1)
is null for every vector (a1 , · · · , ar−1 ) of weight at most m, then for every
such vector gc χ (a1 , · · · , ar−1 , ar ) is null for every ar and h is then (m + 1)-
resilient, since if ar = ar+1 ⊕ 1 then (ar , ar+1 ) has weight 1; note that, in
such case, if g has nonlinearity 2r−1 −2m+1 then the nonlinearity of h, which
equals 2r − 2m+2 achieves then Sarkar et al.’s bound too. The condition that
gcχ (a1 , · · · , ar−1 , 1) is null for every vector (a1 , · · · , ar−1 ) of weight at most m
is achieved if g does not actually depend on its last input bit; but the con-
struction is then a particular case of the construction recalled in I.A. The
condition is also achieved if g is obtained from two m-resilient functions,
by using Siegenthaler’s construction (recalled in II), according to Relation
(65).
3. d◦ f = d◦ g if d◦ g ≥ 1.
4. h has the nonzero linear structure (0, · · · , 0, 1, 1).

Tarannikov combined in [345] this construction with the constructions re-

called in I and II, to build a more complex secondary construction, which
allows to increase in the same time the resiliency order and the algebraic
degree of the functions and which leads to an infinite sequence of functions
achieving Siegenthaler’s and Sarkar et al.’s bounds. Increasing then, by us-
ing the construction recalled in I.A, the set of ordered pairs (r, m) for which
such functions can be constructed, he deduced the existence of r-variable
m-resilient functions achieving Siegenthaler’s and Sarkar et al.’s bounds for
any number of variables r and any resiliency order m such that m ≥ 2r−7 3
and m > 2r − 2 (but these functions have nonzero linear structures). in [292],

127
Pasalic et al. slightly modified this more complex Tarannikov’s construction
into a construction that we shall call Tarannikov et al.’s construction, which
allowed, when iterating it together with the construction recalled in I.A, to
relax slightly the condition on m into m ≥ 2r−10
3 and m > 2r − 2.

IV. Indirect sum of functions

Tarannikov et al.’s construction has been in its turn generalized into the
following construction. All the secondary constructions listed above are
particular cases of it.

Theorem 14 [67] Let r and s be positive integers and let t and m be non-
negative integers such that t < r and m < s. Let f1 and f2 be two r-variable
t-resilient functions. Let g1 and g2 be two s-variable m-resilient functions.
Then the function

h(x, y) = f1 (x) ⊕ g1 (y) ⊕ (f1 ⊕ f2 )(x) (g1 ⊕ g2 )(y); x ∈ Fr2 , y ∈ Fs2

is an (r + s)-variable (t + m + 1)-resilient function. If f1 and f2 are dis-

tinct and if g1 and g2 are distinct, then the algebraic degree of h equals
max(d◦ f1 , d◦ g1 , d◦ (f1 ⊕f2 )+d◦ (g1 ⊕g2 )); otherwise, it equals max(d◦ f1 , d◦ g1 ).
The Walsh transform of h takes value

c (a, b) = 1 fc
h i 1 h i
hχ 1 (a) g 1χ (b) + g 2χ (b) + f
c2 (a) g 1χ (b) − g 2χ (b) . (66)
2 χ 2 χ
d d d d

If the Walsh transforms of f1 and f2 have disjoint supports and if the Walsh
transforms of g1 and g2 have disjoint supports, then

nl(h) = min 2r+s−2 + 2r−1 nl(gj ) + 2s−1 nl(fi ) − nl(fi )nl(gj ) . (67)


i,j∈{1,2}

In particular, if f1 and f2 are two (r, t, −, 2r−1 −2t+1 ) functions with disjoint
Walsh supports, if g1 and g2 are two (s, m, −, 2s−1 − 2m+1 ) functions with
disjoint Walsh supports, and if f1 ⊕ f2 has degree r − t − 1 and g1 ⊕ g2 has
algebraic degree s − m − 1, then h is a (r + s, t + m + 1, r + s − t − m −
2, 2r+s−1 − 2t+m+2 ) function, and thus achieves Siegenthaler’s and Sarkar
et al.’s bounds.
Proof. We have:
 
X X
h
c (a, b) =
χ
 (−1)f1 (x)⊕a·x  (−1)g1 (y)⊕b·y
y∈Fs2 / g1 ⊕g2 (y)=0 x∈Fr2

128
  
X X
+  (−1)f2 (x)⊕a·x  (−1)g1 (y)⊕b·y
y∈Fs2 / g1 ⊕g2 (y)=1 x∈Fr2
X X
= fc
1 χ (a) (−1)g1 (y)⊕b·y + fc
2 χ (a) (−1)g1 (y)⊕b·y
y∈Fs2/ y∈Fs2/
g1 ⊕g2 (y)=0 g1 ⊕g2 (y)=1
!
X
g1 (y)⊕b·y 1 + (−1)(g1 ⊕g2 )(y)
= fc
1 χ (a) (−1)
2
y∈Fs2
!
X
g1 (y)⊕b·y 1 − (−1)(g1 ⊕g2 )(y)
+ fc
2 χ (a) (−1) .
2
y∈Fs2

We deduce Relation (66). If (a, b) has weight at most t + m + 1 then a has

weight at most t or b has weight at most m; hence we have h c (a, b) = 0.
χ
Thus, h is t + m + 1-resilient.
If f1 ⊕ f2 and g1 ⊕ g2 are non-constant, then the algebraic degree of h equals
max(d◦ f1 , d◦ g1 , d◦ (f1 ⊕f2 )+d◦ (g1 ⊕g2 )) because the terms of highest degrees
in (g1 ⊕ g2 )(y) (f1 ⊕ f2 )(x), in f1 (x) and in g1 (y) cannot cancel each others.
We deduce from Relation (66) that if the supports of the Walsh transforms
of f1 and f2 are disjoint, as well as those of g1 and g2 , then:
 
1
max |hχ (a, b)| =
c max max |fi (a)| maxs |gbj (b)|
b
(a,b)∈Fr2 ×Fs2 2 i,j∈{1,2} a∈Fr2 b∈F2

and according to Relation (35) relating the nonlinearity to the Walsh trans-
form, this implies:
1
2r+s − 2nl(h) = max ((2r − 2nl(fi ))(2s − 2nl(gj ))) ,
2 i,j∈{1,2}
which is equivalent to Relation (67). 2
This construction is sometimes called the indirect sum of resilient functions
Note that function h, defined this way, is the concatenation of the four func-
tions f1 , f1 ⊕ 1, f2 and f2 ⊕ 1, in an order controled by g1 (y) and g2 (y).
Examples of pairs (f1 , f2 ) (or (g1 , g2 )) satisfying the hypotheses of Theorem
14 can be found in [67].

V. Constructions without extension of the number of variables

Proposition 22 leads to the following construction:
Proposition 34 [70] Let n be any positive integer and k any non-negative
integer such that k ≤ n. Let f1 , f2 and f3 be three k-th order correlation

129
immune (resp. k-resilient) functions. Then the function s1 = f1 ⊕ f2 ⊕ f3 is
k-th order correlation immune (resp. k-resilient) if and only if the function
s2 = f1 f2 ⊕ f1 f3 ⊕ f2 f3 is k-th order correlation immune (resp. k-resilient).
Moreover:
3
!
1 X
nl(s2 ) ≥ nl(s1 ) + nl(fi ) − 2n−1 (68)
2
i=1

and if the Walsh supports of f1 , f2 and f3 are pairwise disjoint (that is, if
at most one value χc
fi (s), i = 1, 2, 3 is nonzero, for every vector s), then
 
1
nl(s2 ) ≥ nl(s1 ) + min nl(fi ) . (69)
2 1≤i≤3

Proof. Relation (50) and the fact that, for every (nonzero) vector a of weight
at most k, we have fc i χ (a) = 0 for i = 1, 2, 3 imply that sc 1 χ (a) = 0 if and
only if sc
2χ (a) = 0. Relations (68) and (69) are also direct consequences of
Relation (50) and of Relation (35) relating the nonlinearity to the Walsh
transform. 2

Note that this secondary construction is proper to allow achieving high al-
gebraic immunity with s2 , given functions with lower algebraic immunities
f1 , f2 , f3 and s1 , since the support of s2 can be made more complex than
those of these functions. This is done without changing the number of vari-
ables and keeping similar resiliency order and nonlinearity.

Remark. Let g and h be two Boolean functions on Fn2 with disjoint sup-
ports and let f be equal to g ⊕ h = g + h. Then, f is balanced if and
only if wH (g) + wH (h) = 2n−1 . By linearity of the Fourier transform, we
have: fb = gb + b h. Thus, if g and h are m-th order correlation-immune, then
f is m-resilient. For every nonzero a ∈ Fn2 , we have |fbχ (a)| = 2 |fb(a)| ≤
2 |b
g (a)| + 2 |b
h(a)| = |c
gχ (a)| + |h
c (a)|. Thus, assuming that f is balanced, we
χ
have nl(f ) ≥ nl(g) + nl(h) − 2n−1 . The algebraic degree of f is bounded
above by (and can be equal to) the maximum of the algebraic degrees of g
and h.

The most part of the secondary constructions of bent functions described

in Section 6.4 can be altered into constructions of correlation-immune and
resilient functions, see [58].

130
7.6 On the number of resilient functions
It is important to ensure that the selected criteria for the Boolean functions,
supposed to be used in some cryptosystems, do not restrict the choice of the
functions too severely. Hence, the set of functions should be enumerated.
But this enumeration is unknown for most criteria, and the case of resilient
functions is not an exception in this matter. We recall below what is known.
As for bent functions, the class of balanced or resilient functions produced
by Maiorana-McFarland’s construction is far the widest class, compared to
the classes obtained from the other usual constructions, and the number of
provably balanced or resilient Maiorana-McFarland’s functions seems negli-
gible with respect to the total number of functions with the same properties.
For balanced functions, this can be checked: for every positive r, the num-
ber of balanced Maiorana-McFarland’s functions (59) obtained by choosing
s
φ such that φ(y) 6= 0, for every y, equals (2r+1 − 2)2 , and is smaller than
n−1
or equal to 22 (since r ≥ 1). It is quite negligible with respect to the
2n

22n + 12
number 2n−1 ≈ √π2n of all balanced functions on Fn2 . The number of m-
resilient Maiorana-McFarland’s functions obtained by choosing φ such that

2n−r
wH (φ(y)) > m for every y equals 2 ri=m+1 ri
 P
, and is probably also
very small compared to the number of all m-resilient functions. But this
number is unknown.
The exact numbers of m-resilient functions is known for m ≥ n − 3 (see [35],
where (n−3)-resilient functions are characterized) and (n−4)-resilient func-
tions have been characterized [75, 26].
As for bent function, an upper bound comes directly from the Siegen-
thaler bound on the algebraic
Pn−m−1 n
degree: the number of m-resilient functions
is bounded above by 2 i=0 ( i ) . This bound is the so-called naive bound.

In 1990, Yang and Guo published an upper bound on the number of first-
order correlation-immune (and thus on resilient) functions. At the same
time, Denisov obtained a much stronger result (see below) but his result
being published in russian, it was not known internationally. His paper was
translated into english two years later but was not widely known either. This
explains why several papers appeared with weaker results. Park, Lee, Sung
and Kim [294] improved upon Yang-Guo’s bound. Schneider [325] proved
that the number of m-resilient n-variable Boolean functions is less than:
 n−i−1
2i ( m−1 )
n−m
Y 
.
2i−1
i=1

131
but this result was known, see [158]. A general upper bound on the number
of Boolean functions whose distances to affine functions are all divisible
by 2m has been obtained in [90]. It implies an upper bound on the number
of m-resilient functions which improves upon previous bounds for about half
the values of (n, m) (it is better for m large). This bound divides the naive
Pn−m−1 m−1
bound by approximately 2 i=0 ( i )−1 if m ≥ n/2 and by approximately
2m+1 −1
22 if m < n/2.
An upper bound on m-resilient functions (m ≥ n/2 − 1) partially improving
upon this latter bound was obtained for n/2 − 1 ≤ m < n − 2 in [84]: the
number of n-variable m-resilient functions is lower than:
n

n−m 
i
(n−i−1
m−1 )
Pn−m−2 n
( ) n−m−1
Y 2
2 i=0 i +
m+1 i−1
.
2(n−m−1)+1 i=1 2

The expressions of these bounds seem difficult to compare mathematically.

Tables have been computed in [84].
The problem of counting resilient functions is related to counting integer
solutions of a system of linear equations, see [281].
An asymptotic formula for the number of m-resilient (and also for m-th
order correlation-immune functions), where m is very small compared to n -
√
namely m = o( n) - was given by O. Denisov in [131]. This formula was not
correct for m ≥ 2 and a correction was given by the same author in [132] (as
well as a simpler proof): the number of m-resilient functions is equivalent
to
m  
  X !
n − m n n p
exp2 2n − − log2 π/2 .
2 m i
i=0

For large resiliency orders, Y. Tarannikov and D. Kirienko showed in [347]

that, for every positive integer m, there exists a number p(m) such that for
n > p(m), any (n − m)-resilient function f (x1 , · · · , xn ) is equivalent, up to
permutation of its input coordinates, to a function of the form g(x1 , · · · , xp(m) )⊕
xp(m)+1 ⊕ · · · ⊕ xn . It is then a simple matter to deduce that the number
Pp(m)
of (n − m)-resilient functions equals i=0 A(m, i) ni , where A(m, i) is the

number of i-variable (i − m)-resilient functions that depend on all inputs

x1 , x2 , . . . , xi nonlinearly. Hence, it is equivalent to A(m,p(m))
p(m)! n
p(m) for m

constant when n tends to infinity, and it is at most Am np(m) , where Am

depends on m only. It is proved in [348] that 3 · 2m−2 ≤ p(m) ≤ (m − 1)2m−2
and in [347] that p(4) = 10; hence the number of (n − 4)-resilient functions
equals (1/2)n10 + O(n9 ).

132
8 Functions satisfying the strict avalanche and prop-
agation criteria
In this section, we are interested in the functions (and more particularly, in
the balanced functions) which achieve P C(l) for some l < n (the functions
achieving P C(n) are the bent functions and they cannot be balanced).

8.1 P C(l) criterion

It is shown in [180, 60, 61] that, if n is even, then P C(n−2) implies P C(n); so
we can find balanced n-variable P C(l) functions for n even only if l ≤ n − 3.
For odd n ≥ 3, it is also known that the functions which satisfy P C(n − 1)
are those functions of the form g(x1 ⊕ xn , · · · , xn−1 ⊕ xn ) ⊕ `(x), where g is
bent and ` is affine, and that the P C(n − 2) functions are those functions
of a similar form, but where, for at most one index i, the term xi ⊕ xn may
be replaced by xi or by xn (other equivalent characterizations exist [61]).
The only known upper bound on the algebraic degrees of P C(l) functions
is n − 1. A lower bound on the nonlinearity of functions satisfying the
propagation criterion exists [360] and can be very easily proved: if there
exists an l-dimensional subspace F such that, for every nonzero a ∈ F , the
1
derivative Da f is balanced, then nl(f ) ≥ 2n−1 − 2n− 2 l−1 ; Relation (27),
relating the values of the Walsh transform of a function on a flat a + E to
the autocorrelation coefficients of the function on a flat b + E ⊥ , applied to
2
any a ∈ Fn2 , with b = 0 and E = F ⊥ , shows indeed that every value fbχ (u) is
bounded above by 22n−l ; it implies that P C(l) functions have nonlinearities
1
bounded below by 2n−1 − 2n− 2 l−1 . Equality can occur only if l = n − 1 (n
odd) and l = n (n even).
The maximum correlation of Boolean functions satisfying P C(l) (and in
particular, of bent functions) can be directly deduced from Relations (40)
and (27), see [38].

8.1.1 Characterizations
There exist characterizations of the propagation criterion. A first obvious
one is that, according to Relation (24), i.e. to the Wiener-Khintchine The-
2
orem, f satisfies P C(l) if and only if u∈Fn (−1)a·u fbχ (u) = 0 for every
P
2
nonzero vector a of weight at most l. A second one is:

Proposition 35 [61] Any n-variable Boolean function f satisfies P C(l) if

133
and only if, for every vector u of weight at least n − l, and every vector v:
X 2
fbχ (w + v) = 2n+wH (u) .
w u

This is a direct consequence of Relation (27). A third characterization is

given in Subsection 8.2 below (apply it to k = 0).

8.1.2 Constructions
Maiorana-McFarland’s construction can be used to produce functions satis-
fying the propagation criterion: the derivative D(a,b) (x, y) of a function of
the form (59) being equal to x · Db φ(y) ⊕ a · φ(y + b) ⊕ Db g(y), the function
satisfies P C(l) under the sufficient condition that:
1. for every nonzero b ∈ Fs2 of weight smaller than or equal to l, and ev-
ery vector y ∈ Fs2 , the vector Db φ(y) is nonzero (or equivalently every set
φ−1 (u), u ∈ Fr2 , either is empty or is a singleton or has minimum distance
strictly greater than l);
2. every linear combination of at least one and at most l coordinate func-
tions of φ is balanced (this condition corresponds to the case b = 0).
Constructions of such functions have been given in [60, 61, 223].
According to Proposition 35, Dobbertin’s construction cannot produce
functions satisfying P C(l) with l ≥ n/2. Indeed, if u is for instance the
vector with n/2 first coordinates equal to 0, and with n/2 last coordinates
equal to 1, we have, according to Relation (64): h c 2 (w) = 0 for every w  u.
χ

8.2 P C(l) of order k and EP C(l) of order k criteria

According to the characterization of resilient functions and to the definitions
of P C and EP C criteria, we have:

Proposition 36 [302] A function f satisfies EP C(l) (resp. P C(l)) of or-

der k if and only if, for any vector a of Hamming weight smaller than or
equal to l and any vector b of Hamming weight smaller than or equal to k, if
(a, b) 6= (0, 0) (resp. if (a, b) 6= (0, 0) and if a and b have disjoint supports)
then: X
(−1)f (x)⊕f (x+a)⊕b·x = 0.
x∈Fn
2

A recent paper [308] gives the following characterization:

134
Proposition 37 Any n-variable Boolean function f satisfies EP C(l) (resp.
P C(l)) of order k if and only if, for every vector u of weight at least n − l,
and every vector v of weight at least n − k (resp. of weight at least n − k
and such that v and u have disjoint supports):
X
gχ (w) = 2wH (u)+wH (v) ,
fbχ (w)c
w u

where g is the restriction of f to the vectorspace {x ∈ Fn2 / x  v}.

This can be proved by applying Poisson summation formula (17) to the

function (a, b) 7→ D
[ a fχ (b).
Preneel showed in [300] that SAC(k) functions have algebraic degrees at
most n − k − 1 (indeed, all of their restrictions obtained by fixing k input
coordinates have algebraic degrees at most n − k − 1). In [253], the criterion
SAC(n − 3) was characterized through the ANF of the function, and its
properties were further studied. A construction of P C(l) of order k functions
based on Maiorana-McFarland’s method is given in [223] (the mapping φ
being linear and constructed from linear codes) and generalized in [60, 61]
(the mapping φ being not linear and constructed from nonlinear codes). A
construction of n-variable balanced functions satisfying SAC(k) and having
algebraic degree n−k −1 is given, for n−k −1 odd, in [223] and, for n−k −1
even, in [320] (where balancedness and nonlinearity are also considered).
It is shown in [61] that, for every positive even l ≤ n − 4 (with n ≥ 6) and
every odd l such that 5 ≤ l ≤ n − 5 (with n ≥ 10), the functions which
satisfy P C(l) of order n − l − 2 are the functions of the form:
M
xi xj ⊕ h(x1 , · · · , xn )
1≤i<j≤n

where h is affine.

9 Algebraic immune functions

We have recalled in Section 4.1 the different algebraic attacks on stream
ciphers and the related criteria of resistance for the Boolean functions used
in their pseudo-random generators. We study now these criteria more in
details and we describe the known functions satisfying them.

135
9.1 General properties of the algebraic immunity and its re-
lationship with some other criteria
We have seen that the algebraic immunity of any n-variable Boolean function
is bounded above by dn/2e and that the functions used in stream ciphers
must have an algebraic immunity close to this maximum. Note also that for
any functions f and g depending on distinct variables, we have AI(f ⊕ g) ≤
AI(f ) + AI(g). Indeed, for some , η ∈ F2 , let h be an annihilator of degree
AI(f ) of f ⊕  and k an annihilator of degree AI(g) of g ⊕ η, then the
product of h and k is a nonzero annihilator of degree at most AI(f ) + AI(g)
of f ⊕ g ⊕  ⊕ η.

9.1.1 Algebraic immunity of random functions

Random functions behave well with respect to algebraic immunity52 : it has
been proved in [133] that, for all
qa < 1, when n tends to infinity, AI(f ) is
n n n


almost surely greater than 2 − 2 ln 2a ln 2 .

9.1.2 Algebraic immunity of monomial functions

It has been shown in [285] that, if the number of runs r(d) of 1’s in the binary
expansion of the exponent d of a power function trn (axd ) (that is, the num-
√
ber of full subsequences of consecutive 1’s) is smallerlthanm n/2, then the
√
algebraic immunity is bounded above by r(d)b nc + b√nnc − 1. Note that
this bound is better than the general bound dn/2e for only a negligible part
of power mappings, but it concerns however all of those whose exponents
have a constant 2-weight or a constant number of runs - the power functions
studied as potential S-boxes in block ciphers enter in this framework (see
the chapter “Vectorial Boolean Functions for Cryptography”). Moreover,
the bound is further improved when n is odd and the function is almost
bent (see this same chapter for a definition): the algebraic immunity of such
√
functions is bounded above by 2 b nc.

9.1.3 Functions in odd numbers of variables with optimal alge-

braic immunity
In [39], A. Canteaut has observed the following property:
52
No result is known on the behavior of random functions against fast algebraic attacks.

136
Proposition 38 If an n-variable balanced function f , with n odd, admits
no non-zero annihilator of algebraic degree at most n−1
2 , then it has optimum
algebraic immunity n+1
2 .

This means that we do not need to check also that f ⊕ 1 has no non-zero
annihilator of algebraic degree at most n−1 2 for showing that f has optimum
algebraic immunity. Indeed, consider the Reed-Muller code of length 2n
and of order n−1 2 . This code is self-dual (i.e. is its own dual), according to
Theorem 2. Let G be a generator matrix of this code. Each column of G
is labeled by the vector of Fn2 obtained by keeping its coordinates of indices
2, · · · , n+1. Saying that f has no non-zero annihilator of algebraic degree at
most n−1 2 is equivalent to saying that the matrix obtained by selecting those
columns of G corresponding to the elements of the support of f has full rank
P n−1 2 n

n−1 . By hypothesis, f has weight 2n−1 . Since the order of the
i=0 i = 2
columns in G can be freely chosen, we shall assume for simplicity that the
columns corresponding to the support of f are the 2n−1 first ones. Then
we have G = (A | B) where A is an invertible 2n−1 × 2n−1 matrix (and the
matrix G0 = A−1 × G = (Id | A−1 × B) is also a generator matrix). In terms
of coding theory, the support of the function is an information set. Then
the complement of the support of f is also an information set (i.e. B is also
invertible): otherwise, there would exist a vector (z | 0), z 6= 0, in the code
and this is clearly impossible since G and G0 are also parity-check matrices
of the code.

9.1.4 Relationship between normality and algebraic immunity

If an n-variable function f is k-normal then its algebraic immunity is at
most n − k, since the fact that f (x) =  ∈ F2 for every x ∈ A, where A
is a k-dimensional flat, implies that the indicator of A is an annihilator of
f + . This bound is tight since, being symmetric the majority function
is bn/2c-normal for every n (see below) and has algebraic immunity dn/2e.
Obviously, AI(f ) ≤ ` does not imply conversely that f is (n − `)-normal,
since when n tends to infinity, for every a > 1, n-variable Boolean functions
are almost surely non-(a log2 n)-normal [66] (note that k ∼ a log2 n implies
that n − k ∼ n) and the algebraic immunity is always bounded above by
n/2.

137
9.1.5 Relationship between algebraic immunity, weight and non-
linearity
PAI(f )−1 n
Pn−AI(f ) n

It can be easily shown that i=0 i ≤ wH (f ) ≤ i=0 i : the left-
hand side inequality must be satisfied since, otherwise, the number wH (f ) of
equations in the linear system expressing that a function of algebraic degree
at most AI(f ) − 1 is an annihilator of f would have a number of equations
smaller than its number of unknowns (i.e. the number of coefficients in its
algebraic normal form) and it would therefore have non-trivial solutions, a
contradiction. The right-hand side inequality is obtained from the other
one by replacing f by f ⊕ 1. This implies that a function f such that
AI(f ) = n+12 (n odd) must be balanced.
It has been shown in [121] and [76] that low nonlinearity implies low algebraic
immunity (but high algebraic immunity does not imply high nonlinearity):
it can be easily proved that, for every function h of algebraic degree r, we
have AI(f ) − r ≤ AI(f ⊕ h) ≤ AI(f ) + r, and this implies:
AI(f )−2  
X n
nl(f ) ≥
i
i=0

and more generally:

AI(f )−r−1  
X n
nlr (f ) ≥ .
i
i=0

These bounds have been improved in all cases for the first order nonlinearity
into
AI(f )−2  
X n−1
nl(f ) ≥ 2
i
i=0

by Lobanov [256] and in most cases for the r-th order nonlinearity into
AI(f )−r−1  
X n−r
nlr (f ) ≥ 2
i
i=0

(in fact, the improvement was slightly stronger than this, but more complex)
in [71]. Another improvement:
AI(f )−r−1   AI(f )−r−1  
X n X n−r
nlr (f ) ≥ +
i i
i=0 i=AI(f )−2r

138
(which always improves upon the bound of [76] and improves upon the bound
of [71] for low values of r) has been subsequently obtained by Mesnager in
[280].

9.2 The problem of finding functions achieving high alge-

braic immunity and high nonlinearity
We know that functions achieving optimal or suboptimal algebraic immu-
nity and in the same time high algebraic degree and high nonlinearity must
exist thanks to the results of [133, 311]. But knowing that almost all func-
tions have high algebraic immunity does not mean that constructing such
functions is easy.
The bounds of [71] and [280] seen above are weak53 and Lobanov’s bound,
which is tight, does not assure that the nonlinearity is high
enough:
n−1 n
• For n even and AI(f ) = n2 , it gives nl(f ) ≥ 2n−1 −2 n/2−1 = 2n−1 − n/2

which is much smaller than the best possible nonlinearity 2n−1 −2n/2−1 and,
more problematically, much smaller than the asymptotic almost sure non-
linearity of Boolean functions, which
√ is, when n tends to ∞, located in the
neighbourhood of 2n−1 − 2n/2−1 2n ln 2 as we saw. Until recently, the best
nonlinearity reached by the known functions with optimum AI was that of
the majority function and of the iterative

construction (see

more details be-
n−1 n−1 n−1 1 n
low on these functions): 2 − n/2 = 2 − 2 n/2 [124]. This was a
little better than what gives Lobanov’s bound but insufficient.
• For n odd and AI(f ) = n+1 2 , Lobanov’s bound gives nl(f ) ≥ 2
n−1 −
n−1 n−1 − 1 n



(n−1)/2 ' 2 2 (n−1)/2 which is a little better than in the n even
case, but still far from the average nonlinearity of Boolean functions. Until
recently, the best known nonlinearity was that of the majority function and
matched this bound.
Efficient algorithms have been given in [5, 134] for computing the algebraic
immunity and tables are given in [5].

9.3 The functions with high algebraic immunity found so far

and their parameters
Sporadic functions Balanced highly nonlinear functions in up to 20 vari-
ables (derived from power functions) with high algebraic immunities have
been exhibited in [83] and [5].
53
Their interest is to be valid for every function with given algebraic immunity.

139
Infinite classes of functions The majority function (first proposed by
J.D. Key, T.P. McDonough and V.C. Mavron in the context of the erasure
channel [213] - rediscovered by Dalai et al. in the context of algebraic im-
munity [124]), defined as f (x) = 1 if wH (x) ≥ n/2 and f (x) = 0 otherwise,
has optimum algebraic immunity54 . It is a symmetric function (which can
represent a weakness) and its nonlinearity is insufficient. Some variants have
also optimum algebraic immunity.
A nice iterative construction of an infinite class of functions with optimum
algebraic immunity has been given in [122] and further studied in [76]; how-
ever, the functions it produces are neither balanced nor highly nonlinear.
All of these functions are weak against fast algebraic attacks, as shown in
[5].
More numerous functions with optimum algebraic immunity were given in
[72]. Among them are functions with better nonlinearities. However, the
method of [72] did not allow to reach high nonlinearities (see [96]) and some
functions constructed in [246, 247] seem still worse from this viewpoint.
Hence, the question of designing infinite classes of functions achieving all
the necessary criteria remained open after these papers.
A function with optimum algebraic immunity, apparently (according to com-
puter investigations) good immunity to fast algebraic attacks, provably much
better nonlinearity than the functions mentioned above and in fact, accord-
ing to computer investigations, quite sufficient nonlinearity has been exhib-
ited very recently in [151, 81]:

Theorem 15 Let n be any positive integer and α a primitive element of the

field F2n . Let f be the balanced Boolean function on F2n whose support equals
n−1
{0, 1, α, · · · , α2 −2 }. Then f has optimum algebraic immunity dn/2e.

Proof.
Let g bePany Boolean function of algebraic degree at most dn/2e − 1. Let
n
g(x) = 2i=0−1 gi xi be its univariate representation in the field F2n , where
gi ∈ F2n is null if the 2-weight w2 (i) of i is at least dn/2e (which implies in
particular that g2n −1 = 0).
If g is an annihilator of f , then we have g(αi ) = 0 for every i = 0, · · · , 2n−1 −
2, that is, the vector (g0 , · · · , g2n −2 ) belongs to the Reed-Solomon code over
n−1
F2n of zeroes 1, α, · · · , α2 −2 (see [258]). According to the BCH bound,
if g is non-zero, then this vector has Hamming weight at least 2n−1 . We
54
Changing wH (x) ≥ n/2 into wH (x) > n/2 or wH (x) ≤ n/2 or wH (x) < n/2 changes
the function into an affinely equivalent one, up to addition of the constant 1, and therefore
does not change the AI.

140
briefly recall how this lower bound can be simply proved in our framework.
By definition, we have:
     
g(1) 1 1 1 ··· 1 g0
n
 g(α)   1 α α2 ··· α2 −2   g1 
   n −2)   
 g(α2 )   1 α 2 α 4 · · · α 2(2   g2 
 = × 
 ..   .. .. .. ..   .. 
 .   . . . ··· .   . 
2 n −2 n −2 n −2) n −2)(2n −2)
g(α ) 1 α 2 α 2(2 ··· α (2 g2n −2
P2n −2
which implies (since k=0 α(i−j)k equals 1 if i = j and 0 otherwise):

···
     
g0 1 1 1 1 g(1)
n
 g1   1 α−1 α−2 ··· α−(2 −2)   g(α) 
n
     

 g2 =
  1 α−2 α−4 ··· α−2(2 −2) ×
  g(α2 ) 

 ..   .. .. .. ..   .. 
 .   . . . ··· .   . 
n −2) n −2) n n n −2
g2n −2 1 α−(2 α−2(2 · · · α−(2 −2)(2 −2) g(α2 )
n−1
··· g(α2 −1 )
   
1 1 1
n−1 −1) n−1 n −2)
α−(2 α−2 α−(2
n−1
 ···   g(α2 ) 
= ×
   
.. .. .. .. 
 . . ··· .   . 
n−1 n n−1 n n n n −2
α−(2 −1)(2 −2) α−2 (2 −2) · · · α−(2 −2)(2 −2) g(α2 )
n−1 n
Suppose that at least 2n−1 of the gi ’s are null. Then, g(α2 −1 ), · · · , g(α2 −2 )
satisfy a homogeneous system whose matrix is obtained from the latter
matrix above by erasing 2n−1 − 1 rows. This is a 2n−1 × 2n−1 Vander-
monde matrix and its determinant is therefore non-null. This implies that
n−1 n
g(α2 −1 ), · · · , g(α2 −2 ) and therefore g must then be null. Hence the vec-
tor (g0 , · · · , g2n −2 ) has weight at least 2n−1 .
Moreover, suppose that the vector (g0 , · · X · , g2n −2 ) has Hamming weight 2n−1
exactly. Then n is odd and g(x) = xi ; but this contradicts the
0≤i≤2n −2
w2 (i)≤(n−1)/2

fact that g(0) = 0. We deduce that the vector (g0 , · · · , g2n −2 ) has Hamming
weight strictly greater than 2n−1 , leading to a contradiction with the fact
that g has algebraic degree at most dn/2e − 1, since the number of integers
of 2-weight at most dn/2e − 1 is not strictly greater than 2n−1 .
Let g be now a non-zero annihilator of f ⊕ 1. The vector (g0 , · · · , g2n −2 ) be-
n−1 n
longs then to the Reed-Solomon code over F2n of zeroes α2 −1 , · · · , α2 −2 .
According to the BCH bound (which can be proven similarly as above), this
vector has then Hamming weight strictly greater than 2n−1 . We arrive to

141
the same contradiction. Hence, there does not exist a non-zero annihilator
of f or f ⊕1 of algebraic degree at most dn/2e−1 and f has then (optimum)
algebraic immunity dn/2e. 2
It is shown in [81] that the univariate representation of f equals
n −2
2X
αi
1+ xi (70)
i=1
(1 + αi )1/2
n−1
where u1/2 = u2 , which shows that f has algebraic degree n − 1 (which
is optimum for a balanced function), and that we have:
n
nl(f ) ≥ 2n−1 − n · ln 2 · 2 2 − 1.

It could be checked, for small values of n, that the exact value of nl(f ) is
much better than what gives this lower bound and seems quite sufficient for
resisting fast correlation attacks (for these small values of n, it behaves as
2n−1 − 2n/2 ). Finally, the function seems to show good immunity against
fast algebraic attacks: the computer investigations made using Algorithm 2
of [5] suggest the following properties:
• No nonzero function g of algebraic degree at most e and no function
h of algebraic degree at most d exist such that f g = h, when (e, d) =
(1, n − 2) for n odd and (e, d) = (1, n − 3) for n even. This has been
checked for n ≤ 12 and we conjecture it for every n.

• For e > 1, pairs (g, h) of algebraic degrees (e, d) such that e+d < n−1
were never observed. Precisely, the non-existence of such pairs could
be checked exhaustively for n ≤ 9 and e < n/2, for n = 10 and e ≤ 3
and for n = 11 and e ≤ 2. This suggests that this class of functions,
even if not always optimal against fast algebraic attacks, has a very
good behavior.
Hence, the functions of this class gather all the properties needed for al-
lowing the stream ciphers using them as filtering functions to resist all the
main attacks (the Berlekamp-Massey and Rønjom-Helleseth attacks, fast
correlation attacks, standard and fast algebraic attacks). They are the only
functions found so far for which such properties could be shown. There
remains at least one attack against which the resistance of the functions
should be evaluated: the algebraic attack on the augmented function (this
obliges to consider all possible update functions of the linear part of the
pseudo-random generator).

142
 The construction of Proposition 22 allows increasing the complexity of
Boolean functions while keeping their high nonlinearities and may allow
increasing their algebraic immunity as well.

10 Symmetric functions
A Boolean function is called a symmetric function if it is invariant under
the action of the symmetric group (i.e. if its output is invariant under per-
mutation of its input bits). Its output depends then only on the Hamming
weight of the input. So, in other words, f is symmetric if and only if there
exists a function f # from {0, 1, · · · , n} to F2 such that f (x) = f # (wH (x)).
Such functions are of some interest to cryptography, as they allow to imple-
ment in an efficient way nonlinear functions on large numbers of variables.
Let us consider for example an LFSR filtered by a 63 variable symmetric
function f , whose input is the content of an interval of 63 consecutive flip-
flops of the LFSR. This device may be implemented with a cost similar to
that of a 6 variable Boolean function, thanks to a 6 bit counter calculating
the weight of the input to f (this counter is incremented if a 1 is shifted in
the interval and decremented if a 1 is shifted out). However, the pseudo-
random sequence obtained this way has correlation with transitions (sums
of consecutive bits), and a symmetric function should not take all its in-
puts in a full interval. In fact, it is not yet completely clarified whether
the advantage of allowing much more variables and the cryptographic weak-
nesses these symmetric functions may introduce result in an advantage for
the designer or for the attacker, in more sophisticated devices.

10.1 Representation
Let r = 0, · · · , n and let ϕr be the Boolean function whose support is the
set of all vectors of weight r in Fn2 . Then, according to Relation (8) relating
the values of the coefficients of the NNF to the values of the function,   the
|I|
coefficient of xI , I ∈ P(N ), in the NNF of ϕr is: λI = (−1)|I|−r .
r
Mn
Any symmetric function f being equal to f # (r) ϕr , it is therefore equal
r=0
n
X
to f # (r) ϕr , since the functions ϕr have disjoint supports. The coefficient
r=0

143
 n  
X |I|
of xI in its NNF equals then f # (r)(−1)|I|−r and depends only
r
r=0
on the size of I. The NNF of f is then
n n  
X X
# i−r i
f (x) = ci Si (x), where ci = f (r)(−1) (71)
r
i=0 r=0

and where Si (x) is P the i-th elementary symmetric pseudo-Boolean func-

tion whose NNF is I∈P(N )/ |I|=i xI . The degree of the NNF of f equals
max{i/ ci 6= 0}.
We have clearly Si (x) = wHi(x) = wH (x) (wH (x)−1)···(w H (x)−i+1)


i! . According
#
to Relation (71), we see that the univariate function f (z) admits the poly-
nomial representation ni=0 ci zi = ni=0 ci z (z−1)···(z−i+1)
P
P
i! in one variable z,
whose degree equals the degree of the NNF of f . Since this degree is at
most n, and the values taken by this polynomial at n + 1 points are set, this
polynomial representation is unique.
Denoting by σi (x)  the reduction of Si (x) modulo 2, σi (x) equals 1 if and
wH (x)
only if is odd, that is, according to Lucas’ theorem [258], if and
i
only if the binary expansion of i is covered by that of wH (x). Reducing
Relation (71) modulo 2 and writing that j  i when the binary expansion
Plog2 n l−1 , i =
Plog2 n l−1 , j ≤ i ,
of i covers that of j (i.e. j = l=1 jl 2 l=1 il 2 l l
∀l = 1, · · · , log2 n), we deduce from Luca’s theorem again that the ANF of
f is
Mn M
f (x) = λi σi (x), where λi = f # (j). (72)
i=0 ji

Conversely
L (since the Möbius transform is involutive as we saw) f # (i) =
ji λj .
Note that a symmetric
Ln Boolean
Ln function f has algebraic degree 1 if and
only if it equals x
i=1 i or i=1 xi ⊕ 1, that is, if the binary function
#
f (r) equals r [mod L 2] or r + 1 [mod 2], and that it is quadratic if and
only if it equals 1≤i<j≤n xi xj (introduced to generate the Kerdock code)
plus a symmetric function
of algebraicr
degree at most 1, rthat is, if the
# r


function

f (r) equals 2 [mod 2] or 2 + r [mod 2] or 2 + 1 [mod 2]
or 2r + r + 1 [mod 2]. Hence, f has algebraic degree 1 if and only if f #
satisfies f # (r + 1) = f # (r) ⊕ 1 and it has degree 2 if and only if f # satisfies
f # (r + 2) = f # (r) ⊕ 1.
As observed in [49], the algebraic degree of a symmetric function f is at most
2t − 1, for some positive integer t, if and only if the sequence (f # (r))r≥0 is

144
periodic with period 2t . This is a direct consequence of (72). Here again,
it is not clear whether this is a greater advantage for the designer of a
cryptosystem using such symmetric function f (since, to compute the image
of a vector x by f , it is enough to compute the number of nonzero coordinates
x1 , · · · , xt only) or for the attacker.

10.2 Fourier and Walsh transforms

n
X
By linearity, the Fourier transform of any symmetric function f # (r) ϕr
r=0
n
X
equals f # (r) ϕ
cr .
r=0
For every vector a ∈ Fn2 , denoting by ` the Hamming weight of a, we have
n   
X X ` n−`
ϕ
cr (a) = (−1)a·x = (−1)j , denoting by j the
n
j r−j
x∈F2 | wH (x)=r j=0
size of supp(a)∩supp(x). The polynomials Kn,r (X) = nj=0 (−1)j Xj n−X
P


r−j
are called Krawtchouk polynomials. They are caracterized by their generat-
ing series:
Xn
Kn,r (`)z r = (1 − z)` (1 + z)n−`
r=0
and have nice resulting properties (see e.g. [258, 96]).
From the Fourier transform, we can deduce the Walsh transform thanks
to Relation (12).

10.3 Nonlinearity
If n is even, then the restriction of every symmetric function f on Fn2 to the
n/2-dimensional flat:

A = {(x1 , . . . , xn ) ∈ Fn2 ; xi+n/2 = xi ⊕ 1, ∀i ≤ n/2}

is constant, since all the elements of A have the same weight n/2. Thus, f is
n/2-normal55 (see Definition 4). But Relation (42) gives nothing more than
the covering radius bound (36). The symmetric functions which achieve
this bound, i.e. which are bent, have been first characterized by P. Savicky
in [324]: the bent symmetric functions are the four L symmetric functions of
algebraic degree 2 already described above: f1 (x) = 1≤i<j≤n xi xj , f2 (x) =
55
Obviously, this is more generally valid for every function which is constant on the set
{x ∈ Fn
2 ; wH (x) = n/2}.

145
f1 (x) ⊕ 1, f3 (x) = f1 (x) ⊕ x1 ⊕ · · · ⊕ xn and f4 (x) = f3 (x) ⊕ 1. A stronger
result can be proved in a very simple way [169]:

Theorem 16 For every positive even n, the P C(2) n-variable symmetric

functions are the functions f1 , f2 , f3 and f4 above.

Proof. Let f be any P C(2) n-variable symmetric function and let i < j
be two indices in the range [1; n]. Let us denote by x0 the following vec-
tor: x0 = (x1 , · · · , xi−1 , xi+1 , · · · , xj−1 , xj+1 , · · · , xn ). Since f (x) is symmet-
ric, it has the form xi xj g(x0 ) ⊕ (xi ⊕ xj ) h(x0 ) ⊕ k(x0 ). Let us denote by ei,j
the vector of weight 2 whose nonzero coordinates stand at positions i and
j. The derivative Dei,j f of f with respect to ei,j equals (xi ⊕ xj ⊕ 1)g(x0 ).
Since this derivative is balanced, by hypothesis, then g must be equal to the
constant function 1 (indeed if g(x0 ) = 1 then (xi ⊕ xj ⊕ 1)g(x0 ) equals 1 for
half of the inputs and otherwise L it equals 1 for none). Hence, the degree-2-
part of the ANF of f equals 1≤i<j≤n xi xj . 2
Some more results on the propagation criterion for symmetric functions can
be found in [49].

If n is odd, then the restriction of any symmetric function f to the

n+1
2 -dimensional flat

A = {(x1 , . . . , xn ) ∈ Fn2 ; xi+ n−1 = xi ⊕ 1, ∀i ≤ n/2}

2

is affine, since the weight function wH is constant on the hyperplane of A of

equation xn = 0 and on its complement. Thus, f is n+1 2 -weakly-normal. Ac-
cording to Relation (42), this implies that its nonlinearity is upper bounded
n−1
by 2n−1 − 2 2 . It also allows showing that the only symmetric functions
achieving this bound are the same as the 4 functions f1 , f2 , f3 and f4 above,
but with n odd (this has been first proved by Maitra and Sarkar [265], in a
more complex way). Indeed, Relation (42) implies the following result:

Theorem 17 [66] Let n be any positive integer and let f be any symmetric
function on Fn2 . Let l be any integer satisfying 0 < l ≤ n/2. Denote by hl the
symmetric Boolean function on n−2l variables defined by hl (y1 , · · · , yn−2l ) =
f (x1 , · · · , xl , x1 ⊕ 1, · · · , xl ⊕ 1, y1 , · · · , yn−2l ), where the values of x1 , · · · , xl
are arbitrary (equivalently, hl can be defined by h# #
l (r) = f (r + l), for every
0 ≤ r ≤ n − 2l). Then nl(f ) ≤ 2 n−1 −2 n−l−1 l
+ 2 nl(hl ).

Proof: Let A = {(x1 , . . . , xn ) ∈ Fn2 | xi+l = xi ⊕1, ∀i ≤ l}. For every element
x of A, we have f (x) = hl (x2l+1 , · · · , xn ). Let us consider the restriction g

146
of f to A as a Boolean function on Fn−l 2 , say g(x1 , · · · , xl , x2l+1 , · · · , xn ).
Then, since g(x1 , · · · , xl , x2l+1 , · · · , xn ) = hl (x2l+1 , · · · , xn ), g has nonlinear-
ity 2l nl(hl ). According to Relation (42) applied with ha = g, we have
nl(f ) ≤ 2n−1 − 2n−l−1 + 2l nl(hl ). 2

Then, the characterizations recalled above of the symmetric functions

achieving best possible nonlinearity can be straightforwardly deduced. More-
over:
- if, for some integer l such that 0 ≤ l < n−1
 
2 , the nonlinearity of an
n-variable
 symmetric function
 f is strictly greater than 2n−1 − 2n−l−1 +
n−2l−1 n−1
2l 2n−2l−1 − 2b 2 c − 1 = 2n−1 − 2b 2 c − 2l , then, thanks to these
characterizations and to Theorem 17, the function hl must be quadratic,
and f # satisfies f # (r + 2) = f # (r) ⊕ 1, for all l ≤ r ≤ n − 2 − l (this prop-
erty has been observed in [49, Theorem 6], but proved slightly differently);
n−1
- if the nonlinearity of f is strictly greater than 2n−1 − 2b 2 c − 2l+1 ,
then hl either is quadratic or has odd weight, that is, either f # satisfies
f # (r + 2) = f # (r) ⊕ 1 for all l ≤ r ≤ n − 2 − l, or hl has odd weight.

Further properties of the nonlinearities of symmetric functions can be

found in [49, 66].

10.4 Resiliency
There exists a conjecture on symmetric Boolean functions and, equivalently,
on functions defined over {0, 1, · · · , n} and valued in F2 : if f is a non-constant
symmetric Boolean function, then the numerical degree of f (that is, the
degree of the polynomial representation in one variable of f # ) is greater
than or equal to n − 3. It is a simple matter to show that this numerical
2
degree is greater than or equal to n/2 (otherwise, the polynomial f # − f #
would have degree at most n, and being null at n + 1 points, it would equal
the null polynomial, a contradiction with the fact that f is assumed not to
be constant), but the gap between n/2 + 1 and n − 3 is open. According
to Proposition 32, the conjecture is equivalent to saying that there does
not exist any symmetric 3-resilient function. And proving this conjecture
is also a problem on binomial coefficients since the numerical degree of f is
bounded above by d if and only if, for every k such that d < k ≤ n:
k  
r k
X
(−1) f # (r) = 0. (73)
r
r=0

147
Hence, the conjecture is equivalent to saying that Relation (73), with d =
n − 4, has no binary solution f # (0), · · · , f # (n).
J. von zur Gathen and J. R. Roche [161] observed that all symmetric n-
variable Boolean functions have numerical degrees greater than or equal to
n − 3, for any n ≤ 128 (they exhibited Boolean functions with numerical
degree n − 3; see also [166]).
The same authors observed also that, if the number m = n + 1 is a prime,
then all non-constant n-variable symmetric Boolean functions have numer-
ical degree n (and therefore, considering the function g(x) = f (x) ⊕ x1 ⊕
· · · ⊕ xn and applying Proposition 32, all non-affine n-variable symmetric
Boolean functions are not 0-resilient, that is, are unbalanced): indeed, the
n
congruent with (−1)(−2)···(−r) = (−1)r , modulo


binomial coefficient
Pn r being 1·2···r
n
m, the sum r=0 (−1)r nr f # (r) is congruent with r=0 f # (r), modulo m;

P
and Relation (73) with k = n implies then that f # must be constant.
Notice that, applying Relation (73) with k = p − 1, where p is the largest
prime less than or equal to n + 1, shows that the numerical degree of any
symmetric non-constant Boolean function is greater than or equal to p − 1
(or equivalently that no symmetric non-affine Boolean function is (n−p+1)-
resilient): otherwise, reducing (73) modulo p, we would have that the string
f # (0), · · · , f # (k) is constant, and f # having univariate degree less than or
equal to k, the function f # , and thus f itself, would be constant.
More results on the balancedness and resiliency/correlation immunity of
symmetric functions can be found in [21, 283, 354] and more recent ones in
[49, 323].

10.5 Algebraic immunity

We have seen in Section 4.1 that, for every n-variable Boolean function f ,
there exist g 6= 0 and h of algebraic degrees at most dn/2e such that f g = h
(and equivalently, there exists a nonzero annihilator of degree at most dn/2e
of f or of f ⊕1). The same property can be proven when restricting ourselves
to symmetric functions: the elementary symmetric functions of degrees at
most dn/2e and their products with f give a family of 2 (dn/2e + 1) > n + 1
symmetric functions, which must be linearly dependent since they live in
a vectorspace of dimension n + 1. However, given an n-variable symmetric
function f , there do not necessarily exist symmetric functions g 6= 0 and h
of algebraic degrees at most AI(f ) such that f g = h.
We have seen that the majority function, which is symmetric, has optimum
algebraic immunity. In the case n is odd, it is the only symmetric function
having such property, up to the addition of a constant (see [304] which

148
completed a partial result of [245]). In the case n is even, other symmetric
functions exist (up to the addition of a constant and to the transformation
x → x = (x1 ⊕ 1, · · · , xn ⊕ 1)) with this property; more precisions and more
results on the algebraic immunity of symmetric functions can be found in
[28, 252, 303, 304, 305] and the references therein.

10.6 The super-classes of rotation symmetric and Matriochka

symmetric functions
A super-class of symmetric functions, called idempotent or rotation sym-
metric functions (see Subsection 7.5 above), has been investigated from the
viewpoints of bentness and correlation immunity (see e.g. [152, 338]). Re-
cently, it could be proved in [210], thanks to a further investigation on
these functions, that the best nonlinearity of Boolean functions in odd num-
bers of variables is strictly greater than the quadratic bound if and only if
n > 7. Indeed, a function of nonlinearity 241 could be found (while the
quadratic bound gives 240, and the covering radius bound 244), and us-
ing direct sum with quadratic functions, it gave then 11-variable functions
of nonlinearity 994 (while the quadratic bound gives 992 and the covering
radius bound 1000), and 13-variable functions of nonlinearity 4036 (while
the quadratic bound gives 4032 and the covering radius bound 4050). Still
more recently, it was checked that 241 is the best nonlinearity of 9-variable
rotation symmetric functions, but that 9-variable functions whose truth-
tables (or equivalently ANFs) are invariant under cyclic shifts by 3 steps
and under inversion of the order of the input bits can reach nonlinearity
242, which led to 11-variables functions of nonlinearity 996 and 13-variable
functions of nonlinearity 4040. Balanced functions in 13 variables beating
the quadratic bound could also be found. However, this construction gives
worse nonlinearity than the Patterson-Widemann functions for 15 variables
(whose nonlinearity is 16276).
In [238] is introduced the class of Matriochka symmetric functions, which
are the sums of symmetric functions whose sets of variables are different and
nested. Contrary to symmetric functions, they do not depend on the single
weight of the input but on the sequence of the weights of the corresponding
subinputs, and contrary to rotation symmetric functions, they are not in-
variant under cyclic shifts of the input coordinates. They can be almost as
fastly computable as symmetric functions. Their cryptographic parameters
will have to be further studied.

149
 Acknowledgement

We thank Caroline Fontaine for her careful reading of a previous draft of

this chapter.

References
[1] C.M. Adams and S.E. Tavares. Generating and Counting Binary Bent
Sequences, IEEE Trans. Inf. Theory, vol 36, no. 5, pp. 1170-1173, 1990.

[2] N. Alon, O. Goldreich, J. Hastad and R. Peralta. Simple constructions

of almost k-wise independent random variables. Random Stuctures and
Algorithms, Vol 3, No 3, pp 289-304, 1992.

[3] A.S. Ambrosimov. Properties of bent functions of q-valued logic over

finite fields. Discrete Mathematics Appl. vol 4, no. 4, pp. 341-350, 1994.

[4] F. Armknecht. Improving fast algebraic attacks. Proceedings of Fast

Software Encryption 2004, Lecture Notes in Computer Science 3017,
pp. 65-82, 2004.

[5] F. Armknecht, C. Carlet, P. Gaborit, S. Künzli, W. Meier and O. Ru-

atta. Efficient computation of algebraic immunity for algebraic and fast
algebraic attacks. Proceedings of EUROCRYPT 2006, Lecture Notes in
Computer Science 4004 , pp. 147-164, 2006.

[6] F. Armknecht and M. Krause. Algebraic Attacks on combiners with

memory. Proceedings of CRYPTO 2003, Lecture Notes in Computer
Science 2729, pp. 162–175, 2003.

[7] F. Armknecht and M. Krause. Constructing single- and multi-output

boolean functions with maximal immunity. Proceedings of ICALP 2006,
Lecture Notes of Computer Science 4052, pp. 180-191, 2006.

[8] F. Arnault and T. P. Berger. Design and properties of a new pseudo-

random generator based on a filtered FCSR automaton. IEEE Trans.
Computers 54 (11), pp. 1374-1383, 2005.

[9] G. Ars and J.-C. Faugère. Algebraic immunities of functions over fi-
nite fields. Proceedings of the conference BFCA 2005, Publications des
universités de Rouen et du Havre, pp. 21-38, 2005.

150
[10] E.F. Assmus. On the Reed-Muller codes. Discrete Mathematics
106/107, pp. 25-33, 1992.
[11] E.F. Assmus and J. D. Key. Designs and their Codes, Cambridge Univ.
Press., Cambridge, 1992.
[12] J. Ax. Zeroes of polynomials over finite fields. American Journal on
Mathematics no. 86, pp. 255-261, 1964.
[13] T. Baignères, P. Junod and S. Vaudenay. How far can we go beyond
linear cryptanalysis? Proceedings of ASIACRYPT 2004, Lecture Notes
in Computer Science 3329, pp. 432-450, 2004.
[14] B. Barak, G. Kindler, R. Shaltiel, B. Sudakov and A. Wigder-
son. Simulating Independence: New Constructions of Con-
densers, Ramsey Graphs, Dispersers, and Extractors. Pro-
ceedings of the 37-th ACM STOC, 2005. Preprint available at
http://www.math.ias.edu/ boaz/Papers/BKSSW.html
[15] E. Berlekamp, Algebraic Coding Theory, McGraw-Hill, New York, 1968.
[16] E.R. Berlekamp and N.J.A. Sloane. Restrictions on the weight distri-
butions of the Reed-Muller codes. Information and Control 14, pp. 442-
446, 1969.
[17] E.R. Berlekamp and L.R. Welch. Weight distributions of the cosets of
the (32,6) Reed-Muller code. IEEE Transactions on Information The-
ory, 18(1), pp. 203-207, 1972.
[18] A. Bernasconi and B. Codenotti. Spectral analysis of Boolean functions
as a graph eigenvalue problem. IEEE Transactions on computers 48 (3),
pp. 345-351, 1999.
[19] A. Bernasconi and I. Shparlinski. Circuit complexity of testing square-
free numbers. Proceedings of STACS 99, 16th Annual Symposium on
Theoretical Aspects of Computer Science, Lecture Notes in Computer
Science 1563, pp. 47-56, 1999.
[20] A. Bhattacharyya, S. Kopparty, G. Shoenebeck, M. Sudan and D. Zuck-
erman. Optimal testing of Reed-Muller codes. Electronic Colloquium on
Computational Complexity, report no. 86, 2009.
[21] J. Bierbrauer, K. Gopalakrishnan and D.R. Stinson. Bounds for resilient
functions and orthogonal arrays. Proceedings of CRYPTO’94, Lecture
Notes in Computer Science 839, pp. 247-256, 1994.

151
[22] Y. Borissov, N. Manev and S. Nikova. On the non-minimal codewords
of weight 2dmin in the binary Reed-Muller code. Proceedings of the
Workshop on Coding and Cryptography 2001, published by Electronic
Notes in Discrete Mathematics, Elsevier, vo. 6, pp. 103-110, 2001. A
revised version has been published in Discrete Applied Mathematics 128
(Special Issue “International Workshop on Coding and Cryptography
(2001)”), pp. 65-74, 2003.

[23] Y. Borissov, A. Braeken, S. Nikova, B. Preneel. On the Covering Radii

of Binary Reed-Muller Codes in the Set of Resilient Boolean Functions.
IEEE Transactions on Information Theory, vol. 51, No.3, pp. 1182-
1189, 2005.

[24] J. Bourgain. On the construction of affine extractors. Geometric &

Functional Analysis GAFA, Vol. 17, No. 1, pp. 33-57, 2007.

[25] A. Braeken, Y. Borisov, S. Nikova and B. Preneel. Classification of

Boolean functions of 6 variables or less with respect to cryptographic
properties. Proceedings of ICALP 2005, Lecture Notes in Computer
Science 3580, pp. 324-334, 2005.

[26] A. Braeken, Y. Borisov, S. Nikova and B. Preneel. Classification of cubic

(n − 4)-resilient boolean functions. IEEE transactions on information
theory 52, no. 4, pp. 1670-1676, 2006.

[27] A. Braeken, V. Nikov, S. Nikova and B. Preneel. On Boolean func-

tions with generalized cryptographic properties. Proceedings of IN-
DOCRYPT’2004, Lecture Notes in Computer Science 3348, pp. 120-
135, 2004.

[28] A. Braeken and B. Preneel. On the Algebraic Immunity of Symmet-

ric Boolean Functions. Proceedings of Indocrypt 2005, Lecture Notes
in Computer Science 3797, pp. 35–48, 2005. Some false results of this
reference have been corrected in Braeken’s PhD thesis entitled “Cryp-
tographic properties of Boolean functions and S-boxes” and available
at URL http://homes.esat.kuleuven.be/ abraeken/thesisAn.pdf.

[29] J. Bringer, V. Gillot and P. Langevin. Exponential sums and Boolean

functions. Proceedings of the conference BFCA 2005, Publications des
universités de Rouen et du Havre, pp. 177-185, 2005.

152
[30] E. Brier and P. Langevin. Classification of cubic Boolean functions of
9 variables. Proceedings of 2003 IEEE Information Theory Workshop,
Paris, France, 2003.

[31] R. A. Brualdi, N. Cai and V. S. Pless. Orphans of the first order Reed-
Muller codes. IEEE Transactions on Information Theory 36, pp. 399-
401, 1990.

[32] P. Camion and A. Canteaut. Construction of t-resilient functions over

a finite alphabet, Proceedings of EUROCRYPT’96, Lecture Notes in
Computer Sciences 1070, pp. 283-293, 1996.

[33] P. Camion and A. Canteaut. Generalization of Siegenthaler in-

equality and Schnorr-Vaudenay multipermutations. Proceedings of
CRYPTO’96, Lecture Notes in Computer Science 1109, pp. 372–386,
1996.

[34] P. Camion and A. Canteaut. Correlation-immune and resilient functions

over finite alphabets and their applications in cryptography. Designs,
Codes and Cryptography 16, 1999.

[35] P. Camion, C. Carlet, P. Charpin, N. Sendrier. On correlation-immune

functions, Proceedings of CRYPTO’91, Lecture Notes in Computer Sci-
ence, vol. 576, pp. 86-100, 1991.

[36] A. Canteaut. On the weight distributions of optimal cosets of the first-

order Reed-Muller code. IEEE Transactions on Information Theory,
47(1), pp. 407-413, 2001.

[37] A. Canteaut. Cryptographic functions and design criteria for block ci-
phers. Proceedings of INDOCRYPT 2001, Lecture Notes in Computer
Science 2247, pp. 1-16, 2001.

[38] A. Canteaut. On the correlations between a combining function and

functions of fewer variables. Proceedings of the Information Theory
Workshop’02, Bangalore, 2002.

[39] A. Canteaut. Open problems related to algebraic attacks

on stream ciphers. Proceedings of Workshop on Coding and
Cryptography WCC 2005, pp. 1-10, 2005. See also a re-
vised version in Lecture Notes in Computer Science 3969,
pp. 120134, 2006. Paper available on the web http://www-
rocq.inria.fr/codes/Anne.Canteaut/Publications/canteaut06a.pdf

153
[40] A. Canteaut. Analysis and design of symmetric ciphers. Habilitation for
directing Theses, University of Paris 6, 2006.

[41] A. Canteaut, C. Carlet, P. Charpin and C. Fontaine. Propagation char-

acteristics and correlation-immunity of highly nonlinear Boolean func-
tions. Proceedings of EUROCRYPT’2000, Lecture Notes in Computer
Science 187, pp. 507-522 (2000)

[42] A. Canteaut, C. Carlet, P. Charpin and C. Fontaine. On cryptographic

properties of the cosets of R(1, m). IEEE Transactions on Information
Theory vol. 47, no 4, pp. 1494-1513, 2001.

[43] A. Canteaut and P. Charpin. Decomposing bent functions. IEEE Trans-

actions on Information Theory 49, pp. 2004-2019, 2003.

[44] A. Canteaut and P. Charpin and G. Kyureghyan. A new class of mono-

mial bent functions. Finite Fields and Application 14(1), pp.221-241,
2008.

[45] A. Canteaut, M. Daum, H. Dobbertin and G. Leander. Normal and

Non-Normal Bent Functions. Proceedings of the Workshop on Coding
and Cryptography 2003, pp. 91-100, 2003.

[46] A. Canteaut and E. Filiol. Ciphertext only reconstruction of stream

ciphers based on combination generators. Proceedings of Fast Software
Encryption 2000, Lecture Notes in Computer Science 1978, pp. 165-180,
2001.

[47] A. Canteaut and M. Trabbia. Improved fast correlation attacks us-

ing parity-check equations of weight 4 and 5, Advanced in Cryptology-
EUROCRYPT 2000. Lecture notes in computer science 1807, pp. 573-
588, 2000.

[48] A. Canteaut and M. Videau. Degree of Composition of Highly Non-

linear Functions and Applications to Higher Order Differential Crypt-
analysis, Advances in Cryptology, EUROCRYPT2002, Lecture Notes in
Computer Science 2332, pp. 518-533, 2002.

[49] A. Canteaut and M. Videau. Symmetric Boolean functions. IEEE

Transactions on Information Theory 51(8), pp. 2791-2811, 2005.

[50] Jean Robert Du Carlet. La Cryptographie, contenant une très subtile

manière descrire secrètement, composée par Maistre Jean Robert Du

154
 Carlet, 1644. A manuscript exists at the Bibliothèque Nationale (Très
Grande Bibliothèque), Paris, France.

[51] C. Carlet. Codes de Reed-Muller, codes de Kerdock et de Preparata. PhD

thesis. Publication of LITP, Institut Blaise Pascal, Université Paris 6,
90.59, 1990.

[52] C. Carlet. A transformation on Boolean functions, its consequences

on some problems related to Reed-Muller codes, Proceedings of EU-
ROCODE’90, Lecture Notes in Computer Science 514, pp. 42-50, 1991.

[53] C. Carlet. Partially-bent functions, Designs Codes and Cryptography,

3, pp. 135-145, 1993, and Proceedings of CRYPTO’ 92, Lecture Notes
in Computer Science 740, pp. 280-291, 1993.

[54] C. Carlet. Two new classes of bent functions. In Proceedings of EURO-

CRYPT’93, Lecture Notes in Computer Science 765, pp. 77-101, 1994.

[55] C. Carlet. Generalized Partial Spreads, IEEE Transactions on Infor-

mation Theory, vol. 41, no. 5, pp. 1482-1487, 1995.

[56] C. Carlet. Hyper-bent functions. PRAGOCRYPT’96, Czech Technical

University Publishing House, pp. 145-155, 1996.

[57] C. Carlet. A construction of bent functions. Finite Fields and Appli-

cations, London Mathematical Society, Lecture Series 233, Cambridge
University Press, pp. 47-58, 1996.

[58] C. Carlet. More correlation-immune and resilient functions over Galois

fields and Galois rings. Advances in Cryptology, EUROCRYPT’ 97,
Lecture Notes in Computer Science 1233, pp. 422-433, 1997.

[59] C. Carlet. On Kerdock codes, American Mathematical Society. Proceed-

ings of the conference Finite Fields and Applications Fq4, Contempo-
rary Mathematics 225, pp. 155-163, 1999.

[60] C. Carlet. On the propagation criterion of degree ` and order k. Pro-

ceedings of EUROCRYPT’98, Lecture Notes in Computer Science 1403,
pp. 462-474, 1998.

[61] C. Carlet. On cryptographic propagation criteria for Boolean functions.

Information and Computation , vol. 151, Academic Press pp. 32-56,
1999.

155
[62] C. Carlet. Recent results on binary bent functions. Proceedings of the
International Conference on Combinatorics, Information Theory and
Statistics; Journal of Combinatorics, Information and System Sciences,
Vol. 25, Nos. 1-4, pp. 133-149, 2000.

[63] C. Carlet. On the coset weight divisibility and nonlinearity of resilient

and correlation-immune functions, Proceedings of SETA’01 (Sequences
and their Applications 2001), Discrete Mathematics and Theoretical
Computer Science, pp. 131-144, 2001.

[64] C. Carlet. A larger Class of Cryptographic Boolean Functions via

a Study of the Maiorana-McFarland Construction. Proceedings of
CRYPT0 2002, Lecture Notes in Computer Science 2442, pp. 549-564,
2002.

[65] C. Carlet. On the confusion and diffusion properties of Maiorana-

McFarland’s and extended Maiorana-McFarland’s functions. Special Is-
sue “Complexity Issues in Coding and Cryptography”, dedicated to Prof.
Harald Niederreiter on the occasion of his 60th birthday, Journal of
Complexity 20, pp. 182-204, 2004.

[66] C. Carlet. On the degree, nonlinearity, algebraic thickness and non-

normality of Boolean functions, with developments on symmetric func-
tions. IEEE Transactions on Information Theory, vol. 50, pp. 2178-
2185, 2004.

[67] C. Carlet. On the secondary constructions of resilient and bent func-

tions. Proceedings of the Workshop on Coding, Cryptography and Com-
binatorics 2003, published by Birkhäuser Verlag, pp. 3-28, 2004.

[68] C. Carlet. Concatenating indicators of flats for designing cryptographic

functions. Design, Codes and Cryptography volume 36, Number 2,
pp.189 - 202, 2005.

[69] C. Carlet. Partial covering sequences: a method for designing classes

of cryptographic functions. Proceedings of the conference “The first
Symposium on Algebraic Geometry and its Applications” (SAGA’07),
Tahiti, 2007, World Scientific, Series on Number Theory and its Ap-
plications, Vol. 5, pp. 366-387, 2008.

[70] C. Carlet. On bent and highly nonlinear balanced/resilient functions

and their algebraic immunities. Proceedings of AAECC 16, Lecture

156
 Notes in Computer Science 3857, pp. 1-28, 2006. This paper is an ex-
tended version of the paper entitled “Improving the algebraic immunity of re-
silient and nonlinear functions and constructing bent function”, IACR ePrint
Archive http://eprint.iacr.org/ 2004/276.

[71] C. Carlet. On the higher order nonlinearities of algebraic immune func-

tions. Proceedings of CRYPTO 2006, Lecture Notes in Computer Sci-
ence 4117, pp. 584-601, 2006.

[72] C. Carlet. A method of construction of balanced functions with opti-

mum algebraic immunity. To appear in the Proceedings of the Interna-
tional Workshop on Coding and Cryptography, The Wuyi Mountain,
Fujiang, China, June 11-15, 2007, World Scientific Publishing Co. in its
series of Coding and Cryptology, 2008. A preliminary version is available
on IACR ePrint Archive, http://eprint.iacr.org/, 2006/149.

[73] C. Carlet. Recursive lower bounds on the nonlinearity profile of Boolean

functions and their applications. IEEE Transactions on Information
Theory, vol.54, No. 3, pp. 1262-1272, 2008.

[74] C. Carlet. On the higher order nonlinearities of Boolean functions and

S-boxes, and their generalizations. Proceedings of SETA 2008, Lecture
Notes in Computer Science 5203, pp. 345-367, 2008.

[75] C. Carlet and P. Charpin. Cubic Boolean functions with highest re-
siliency. IEEE Transactions on Information Theory, vol. 51, no. 2, pp.
562-571, 2005.

[76] C. Carlet, D. Dalai, K. Gupta and S. Maitra. Algebraic Immunity for

Cryptographically Significant Boolean Functions: Analysis and Con-
struction. IEEE Transactions on Information Theory, vol. 52, no. 7,
pp. 3105-3121, 2006.

[77] C. Carlet, L.E. Danielsen, M.G. Parker and P. Solé Self dual bent func-
tions. Proceedings of the conference BFCA 2008, Copenhagen, to ap-
pear in Lecture Notes in Computer Science.

[78] C. Carlet and C. Ding. Highly Nonlinear Mappings. Special Issue “Com-
plexity Issues in Coding and Cryptography”, dedicated to Prof. Harald
Niederreiter on the occasion of his 60th birthday, Journal of Complexity
20, pp. 205-244, 2004.

157
[79] C. Carlet, H. Dobbertin and G. Leander. Normal extensions of bent
functions. IEEE Transactions on Information Theory, vol. 50, no. 11,
pp. 2880-2885, 2004.

[80] C. Carlet and S. Dubuc. On generalized bent and q-ary perfect nonlinear
functions. Proceedings of Finite Fields and Applications Fq5, Augsburg,
Germany, Springer, pp. 81-94, 2000.

[81] C. Carlet and K. Feng. An infinite class of balanced functions with

optimum algebraic immunity, good immunity to fast algebraic attacks
and good nonlinearity. Proceedings of ASIACRYPT 2008, Lecture Notes
in Computer Science 5350, pp. 425-440, 2008.

[82] C. Carlet and P. Gaborit. Hyper-bent functions and cyclic codes. Jour-
nal of Combinatorial Theory, Series A, 113, no. 3, 466-482, 2006.

[83] C. Carlet and P. Gaborit. On the construction of balanced Boolean

functions with a good algebraic immunity. Proceedings of International
Symposium on Information Theory, ISIT, Adelaide, Australia, 2005. A
longer version of this paper has been published in the Proceedings of
BFCA 2005, Publications des universités de Rouen et du Havre, pp.
1-20, 2005.

[84] C. Carlet and A. Gouget. An upper bound on the number of m-resilient

Boolean functions. Proceedings of ASIACRYPT 2002, Lecture Notes in
Computer Science 2501, pp. 484-496, 2002.

[85] C. Carlet and P. Guillot. A characterization of binary bent functions,

Journal of Combinatorial Theory, Series A, vol. 76, No. 2, pp. 328-335,
1996.

[86] C. Carlet and P. Guillot. An alternate characterization of the bentness

of binary functions, with uniqueness. Designs, Codes and Cryptography,
14, pp. 133-140, 1998.

[87] C. Carlet and P. Guillot. A new representation of Boolean functions,

Proceedings of AAECC’13, Lecture Notes in Computer Science 1719,
pp. 94-103, 1999.

[88] C. Carlet and P. Guillot. Bent, resilient functions and the Numerical
Normal Form. DIMACS Series in Discrete Mathematics and Theoretical
Computer Science, 56, pp. 87-96, 2001.

158
[89] C. Carlet, P. Guillot and S. Mesnager. On immunity profile of Boolean
functions. Proceedings of SETA 2006 (International Conference on Se-
quences and their Applications), Lecture Notes in Computer Science
4086, pp. 364-375, 2006.

[90] C. Carlet and A. Klapper. Upper bounds on the numbers of resilient

functions and of bent functions. This paper was meant to appear in an issue
of Lecture Notes in Computer Sciences dedicated to Philippe Delsarte, Edi-
tor Jean-Jacques Quisquater. But this issue finally never appeared. A shorter
version has appeared in the Proceedings of the 23rd Symposium on Infor-
mation Theory in the Benelux, Louvain-La-Neuve, Belgium, 2002.

[91] C. Carlet and S. Mesnager. On the supports of the Walsh transforms

of Boolean functions. Proceedings of the conference BFCA 2005, Pub-
lications des universités de Rouen et du Havre, pp. 65-82, 2005.

[92] C. Carlet and S. Mesnager. Improving the upper bounds on the covering
radii of binary Reed-Muller codes. IEEE Transactions on Information
Theory 53, pp. 162-173, 2007.

[93] C. Carlet and E. Prouff. On plateaued functions and their construc-

tions. Proceedings of Fast Software Encryption 2003, Lecture notes in
computer science 2887, pp. 54-73, 2003.

[94] C. Carlet and P. Sarkar. Spectral Domain Analysis of Correlation Im-

mune and Resilient Boolean Functions. Finite fields and Applications
8, pp. 120-130, 2002.

[95] C. Carlet and Y. V. Tarannikov. Covering sequences of Boolean func-

tions and their cryptographic significance. Designs, Codes and Cryp-
tography, 25, pp. 263-279, 2002.

[96] C. Carlet, X. Zeng, C. Lei and L. Hu. Further properties of several

classes of Boolean functions with optimum algebraic immunity. Pro-
ceedings of the First International Conference on Symbolic Computa-
tion and Cryptography SCC 2008, LMIB, pp. 42-54, 2008

[97] L. Carlitz and S. Uchiyama. Bounds for exponential sums. Duke Math.
Journal 1, pp. 37-41, 1957.

[98] A.H. Chan and R.A. Games. On the quadratic spans of De Bruijn
sequences. IEEE Transactions on Information Theory, vol. 36, no. 4,
pp. 822-829, 1990.

159
[99] C. Charnes, M. Rötteler and T. Beth. Homogeneous bent functions,
invariants, and designs. Designs, Codes and Cryptography, 26, pp. 139-
154, 2002.

[100] P. Charpin. Open problems on cyclic codes, In “Handbook of Coding

Theory”, Part 1, chapter 11, V. S. Pless, W. C. Huffman, Eds, R. A.
Brualdi, assistant editor, Amsterdam, the Netherlands: Elsevier, 1998.

[101] P. Charpin. Normal Boolean functions. Special Issue “Complexity Is-

sues in Coding and Cryptography”, dedicated to Prof. Harald Niederre-
iter on the occasion of his 60th birthday, Journal of Complexity 20, pp.
245-265, 2004.

[102] P. Charpin and G. Gong. Hyperbent functions, Kloosterman sums and

Dickson polynomials. IEEE Transactions on Information Theory, vol.
54, no. 9, pp. 4230-4238, 2008.

[103] P. Charpin, T. Helleseth, V. Zinoviev. Propagation characteristics of

x → 1/x and Kloosterman sums. Finite Fields and Applications 13
Issue 2, 366-381, 2007.

[104] P. Charpin and G. Kyureghyan. Cubic monomial bent functions: A

subclass of M. SIAM, J. Discr. Math., Vol.22, no.2, pp. 650–665, 2008.

[105] P. Charpin and E. Pasalic. On propagation characteristics of resilient

functions. Proceedings of SAC 2002, Lecture Notes in Computer Science
2595, pp. 356–365, 2002.

[106] P. Charpin, E. Pasalic and C. Tavernier. On bent and semi-bent

quadratic Boolean functions. IEEE Transactions on Information The-
ory, vol. 51, no. 12, pp. 4286-4298, 2005.

[107] S. Chee, S. Lee, K. Kim and D. Kim. Correlation immune functions

with controlable nonlinearity. ETRI Journal, vol 19, no 4, pp. 389-401,
1997.

[108] S. Chee, S. Lee, D. Lee and S. H. Sung. On the correlation immune

functions and their nonlinearity. Proceedings of Asiacrypt’96, Lecture
Notes in Computer Science 1163, pp. 232-243.

[109] V. Chepyzhov and B. Smeets. On a fast correlation attack on cer-

tain stream ciphers. Proceedings of EUROCRYPT’91, Lecture Notes in
Computer Science 547, pp. 176-185, 1992.

160
[110] B. Chor, O. Goldreich, J. Hastad, J. Freidmann, S. Rudich and R.
Smolensky. The bit extraction problem or t-resilient functions. Proceed-
ings of the 26th IEEE Symposium on Foundations of Computer Science,
pp. 396-407, 1985.

[111] G. Cohen, I. Honkala, S. Litsyn and A. Lobstein. Covering codes.

North-Holland, 1997.

[112] Higher Order Correlation Attacks, XL algorithm and Cryptanalysis

of Toyocrypt. Proceedings of ICISC 2002, Lecture notes in computer
science 2587, pp. 182-199, 2003.

[113] N. Courtois. Fast Algebraic Attacks on Stream Ciphers with Linear

Feedback. Proceedings of CRYPTO 2003, Lecture Notes in Computer
Science 2729, pp. 177-194, 2003.

[114] N. Courtois. Algebraic attacks on combiners with memory and several

outputs. Proceedings of ICISC 2004, Lecture notes in computer science
3506, pp. 3-20, 2005.

[115] N. Courtois. Cryptanalysis of SFINKS. Proceedings of ICISC 2005.

Also available at IACR ePrint Archive http://eprint.iacr.org/, Report
2005/243, 2005.

[116] N. Courtois. General Principles of Algebraic Attacks and New Design

Criteria for Components of Symmetric Ciphers. AES 4 Conference, Lec-
ture Notes in Computer Science 3373, 2004.

[117] N. Courtois and W. Meier. Algebraic Attacks on Stream Ciphers with

Linear Feedback. Proceedings of EUROCRYPT 2003, Lecture Notes in
Computer Science 2656, pp. 346-359.

[118] T. W. Cusick. On constructing balanced correlation immune functions.

Proceedings of SETA’98, (Sequences and their Applications 1998), Dis-
crete Mathematics and Theoretical Computer Science, pp. 184-190,
1999.

[119] T. W. Cusick, C. Ding and A. Renvall, Stream Ciphers and Number

Theory, North-Holland Mathematical Library 55. Amsterdam: North-
Holland/Elsevier, 1998.

[120] D.M. Cvetkovic, M. Doob and H. Sachs. Spectra of graphs. Academic

Press, 1979.

161
[121] D. K. Dalai, K. C. Gupta and S. Maitra. Results on Algebraic Immu-
nity for Cryptographically Significant Boolean Functions. Proceedings
of Indocrypt 2004, Lecture Notes in Computer Science 3348, pp. 92-106,
2004

[122] D. K. Dalai, K. C. Gupta and S. Maitra. Cryptographically Significant

Boolean functions: Construction and Analysis in terms of Algebraic
Immunity. Fast Software Encryption 2005, Lecture Notes in Computer
Science 3557, pp. 98-111, 2005.

[123] D. K. Dalai, K. C. Gupta and S. Maitra. Notion of algebraic immunity

and its evaluation related to fast algebraic attacks. Proceedings of the
conference BFCA 2006, Publications des universités de Rouen et du
Havre, pp. 107-124, 2006.

[124] D. K. Dalai, S. Maitra and S. Sarkar. Basic Theory in Construction

of Boolean Functions with Maximum Possible Annihilator Immunity.
Designs, Codes and Cryptography, Volume 40, Number 1, Pages 41–58,
July 2006. IACR ePrint Archive http://eprint.iacr.org/, No. 2005/229,
15 July, 2005.

[125] D. Dalai and S. Maitra. Balanced Boolean functions with (more than)
maximum algebraic immunity. Proceedings of the Workshop on Coding
and Cryptography (in the memory of Hans Dobbertin) WCC 2007, pp.
99-108, 2007.

[126] M. Daum, H. Dobbertin and G. Leander. An algorithm for checking

normality of Boolean functions. Proceedings of the Workshop on Coding
and Cryptography 2003, pp. 133-142, 2003.

[127] M. Daum, H. Dobbertin and G. Leander. Short description of an al-

gorithm to create bent functions. Private communication.

[128] E. Dawson and C.-K. Wu. Construction of correlation immune Boolean

functions. Proceedings of ICICS 1997, pp. 170-180, 1997.

[129] P. Delsarte. An algebraic approach to the association schemes of coding

theory. PhD thesis. Université Catholique de Louvain (1973)

[130] P. Delsarte. Four fundamental parameters of a code and their combi-

natorial significance. Information and Control, vol. 23 (5), pp. 407-438,
1973.

162
[131] O. Denisov. An asymptotic formula for the number of correlation-
immune of order k Boolean functions. Discrete Mathematics Appl., Vol.
2, No. 4; pp. 407-426, 1992. Translation of a russian article in Diskret-
naya Matematika 3, pp. 25-46, 1990.
[132] O. Denisov. A local limit theorem for the distribution of a part of
the spectrum of a random binary function. Discrete Mathematics and
Applications, V. 10, No 1, pp. 87-102, 2000.
[133] F. Didier. A new upper bound on the block error probability after
decoding over the erasure channel. IEEE Transactions on Information
Theory, vol. 52, pp. 4496- 4503, 2006.
[134] F. Didier. Using Wiedemann’s algorithm to compute the immunity
against algebraic and fast algebraic attacks. Proceedings of Indocrypt
2006, Lecture Notes in Computer Science 4329, pp. 236-250.
[135] J. Dillon. A survey of bent functions. NSA Technical Journal Special
Issue, pp. 191-215, 1972.
[136] J. F. Dillon. Elementary Hadamard Difference sets. Ph. D. Thesis,
Univ. of Maryland, 1974.
[137] J. F. Dillon. Elementary Hadamard Difference sets, Proceedings of the
Sixth S-E Conf. Comb. Graph Theory and Comp., Winnipeg Utilitas
Math, pp. 237-249, 1975.
[138] J. Dillon. More DD difference sets. To appear in Designs, Codes and
Cryptography (on line), 2008.
[139] J. F. Dillon and H. Dobbertin. New cyclic difference sets with Singer
parameters. Finite Fields and Their Applications 10, pp. 342-389, 2004.
[140] J. F. Dillon and G. McGuire. Near bent functions on a hyperplane.
Finite Fields and Their Applications Vol. 14, Issue 3, pp. 715-720, 2008.
[141] H. Dobbertin. Construction of bent functions and balanced Boolean
functions with high nonlinearity. Proceedings of Fast Software Encryp-
tion, Second International Workshop, Lecture Notes in Computer Sci-
ence 1008, pp. 61-74, 1995.
[142] H. Dobbertin, P. Felke, T. Helleseth and P. Rosenthal. Niho type cross-
correlation functions via Dickson polynomials and Kloosterman sums.
IEEE Transactions on Information Theory, vol. 52, no. 2, pp. 613-627,
2006.

163
[143] H. Dobbertin and G. Leander. Bent functions embedded into the recur-
sive framework of Z-bent functions. Designs, Codes and Cryptography,
Vol. 49, no. 1-3, pp. 3-22, 2008.

[144] H. Dobbertin, G. Leander. A survey of some recent results on bent

functions. Proceeding of SETA 2004, Lecture Notes in Computer Sci-
ence 3486, pp. 1-29, 2005.

[145] H. Dobbertin, G. Leander, A. Canteaut, C. Carlet, P. Felke and P.

Gaborit. Construction of Bent Functions via Niho Power Functions.
Journal of Combinatorial Theory, Series A, Volume 113, Issue 5, pp.
779-798, 2006.

[146] S. Dubuc. Characterization of linear structures. Designs, Codes and

Cryptography vol. 22, pp. 33-45, 2001.

[147] I. Dumer, G. Kabatiansky and C. Tavernier. List decoding of Reed-

Muller codes up to the Johnson bound with almost linear complexity.
Proceedings of ISIT 2006. Seattle, USA.

[148] J. H. Evertse. Linear structures in block ciphers. Proceedings of EU-

ROCRYPT’87, Lecture Notes in Computer Science 304, pp. 249-266,
1988.

[149] J. C. Faugère. “Fast Gröbner. Algebraic cryptanalysis of HFE and

filter generators”.Proceedings of the Workshop on Coding and Cryptog-
raphy 2003, pp. 175-176, 2003.

[150] J.-C. Faugère and G. Ars. An Algebraic Cryptanalysis of Nonlinear

Filter Generators using Gröbner bases. Rapport de Recherche INRIA
4739, 2003.

[151] K. Feng, Q. Liao and J. Yang. Maximal values of generalized algebraic

immunity. Designs, Codes and Cryptography 50, pp. 243-252, 2009.

[152] E. Filiol and C. Fontaine. Highly nonlinear balanced Boolean functions

with a good correlation-immunity. Proceedings of EUROCRYPT’98,
Lecture Notes in Computer Science 1403, pp. 475-488, 1998.

[153] S. Fischer and W. Meier. Algebraic Immunity of S-boxes and Aug-

mented Functions. Proceedings of Fast Software Encryption 2007. Lec-
ture Notes in Computer Science 4593, pp. 366-381, 2007.

164
[154] C. Fontaine. On some cosets of the First-Order Reed-Muller code with
high minimum weight. IEEE Transactions on Information Theory, vol.
45, no. 4, pp. 1237-1243, 1999.

[155] R. Forré. The strict avalanche criterion: spectral properties of Boolean

functions and an extended definition. Proceedings of CRYPTO’88, Lec-
ture Notes in Computer Science 403, pp. 450-468, 1989.

[156] R. Forré. A fast correlation attack on nonlinearly feedforward fil-

tered shift register sequences. Proceedings of EUROCRYPT ’89, Lecture
Notes in Computer Science 434, pp. 586-595, 1990.

[157] R. Fourquet and C. Tavernier. List decoding of second order Reed-

Muller and its covering radius implications. Proceedings of the Work-
shop on Coding and Cryptography 2007 WCC, pp. 147-156, 2007.

[158] J. Friedman. On the bit extraction problem. Proceedings of the 33rd

IEEE Symposium on Foundations of Computer Science, pp. 314-319,
1992.

[159] THE BOOLEAN PLANET. Webpage on the equivalence classes of

Boolean functions in at most 6 variables, maintained by J. Fuller at
URL http://www.booleanfunction.com/

[160] R. G. Gallager. Low density parity check codes. Cambridge, MA: MIT
Press, 1963.

[161] J. von zur Gathen and J. R. Roche. Polynomials with two values.
Combinatorica 17(3), pp. 345-362, 1997.

[162] J. Golić. Fast low order approximation of cryptographic functions.

Proceedings of EUROCRYPT’96, Lecture Notes in Computer Science
1070, pp. 268-282, 1996.

[163] J. Golić. On the security of nonlinear filter generators. Proceedings of

Fast Software Encryption’96, Lecture Notes in Computer Science 1039,
pp. 173-188, 1996.

[164] F. Göloglu and A. Pott. Results on the crosscorrelation and autocor-

relation of sequences. Proceedings of Sequences and their Applications
- SETA 2008 - Lecture Notes in Computer Science 5203, pp. 95-105,
2008.

[165] S.W. Golomb. Shift Register Sequences. Aegean Park Press, 1982.

165
[166] K. Gopalakrishnan, D. G. Hoffman and D. R. Stinson. A Note on
a Conjecture Concerning Symmetric Resilient Functions. Information
Processing Letters 47 (3), pp. 139-143, 1993.

[167] M. Goresky and A. Klapper. Fibonacci and Galois representation of

feedback with carry shift registers. IEEE Transactions on Information
Theory, vol. 48, pp. 2826-2836, 2002.

[168] M. Goresky and A. Klapper. Periodicity and distribution properties

of combined FCSR sequences. Proceedings of Sequences and their Ap-
plications - SETA 2006 - Lecture Notes in Computer Science 4086, pp.
334-341, 2006.

[169] A. Gouget. On the propagation criterion of Boolean functions. Pro-

ceedings of the Workshop on Coding, Cryptography and Combinatorics
2003, Birkhäuser Verlag, pp. 153-168, 2004.

[170] A. Gouget and H. Sibert. Revisiting correlation-immunity in filter gen-

erators. Proceedings of SAC 2007, Lecture Notes in Computer Science
4876, pp. 378-395, 2007.

[171] P. Guillot. Completed GPS Covers All Bent Functions. Journal of

Combinatorial Theory, Series A 93, pp. 242-260, 2001.

[172] P. Guillot. Partial bent functions. Proceedings of the World Multicon-

ference on Systemics, Cybernetics and Informatics, SCI 2000, 2000.

[173] Xiao Guo-Zhen, C. Ding and W. Shan. The stability theory of stream
ciphers, Lecture Notes in Computer Science 561, 1991.

[174] X. Guo-Zhen and J. L. Massey. A Spectral Characterization of

Correlation-Immune Combining Functions. IEEE Trans. Inf. Theory,
vol. 34, no. 3, pp. 569-571, 1988.

[175] A. R. Hammons Jr., P. V. Kumar, A. R. Calderbank, N. J. A. Sloane

and P. Solé. The Z4 -linearity of Kerdock, Preparata, Goethals and re-
lated codes. IEEE Transactions on Information Theory, vol. 40, pp.
301-319, 1994.

[176] P. Hawkes and G. Rose. Rewriting Variables: The Complexity of Fast

Algebraic Attacks on Stream Ciphers. Proceedings of CRYPTO 2004,
Lecture Notes in Computer Science 3152, pp. 390–406, 2004.

166
[177] T. Helleseth, T. Kløve, and J. Mykkelveit. On the covering radius of
binary codes. IEEE Transactions on Information Theory, vol. 24, no.
5, pp. 627-628, 1978.

[178] T. Helleseth and H.F. Mattson Jr. On the cosets of the simplex code.
Discrete Mathematics 56, pp. 169-189, 1985.

[179] S. Hirose and K. Ikeda. Nonlinearity criteria of Boolean functions.

KUIS Technical Report, KUIS-94-0002, 1994.

[180] S. Hirose and K. Ikeda. Complexity of Boolean functions satisfying the

propagation criterion. Proceedings of the 1995 Symposium on Cryptog-
raphy and Information Security, SCIS95-B3.3, 1995.

[181] I. Honkala and A. Klapper. Bounds for the multicovering radii of Reed-
Muller codes with applications to stream ciphers. Designs, Codes and
Cryptography 23, pp. 131-145, 2001.

[182] X.-D. Hou. Some results on the covering radii of Reed-Muller codes.
IEEE Transactions on Information Theory, vol. 39, no. 2, pp. 366-378,
1993.

[183] X.-D. Hou. Classification of cosets of the Reed-Muller code R(m −

3, m). Discrete Mathematics, 128, pp. 203-224, 1994.

[184] X.-D. Hou. The covering radius of R(1, 9) in R(4, 9). Designs, Codes
and Cryptography 8 (3), pp. 285-292, 1995.

[185] X.-D. Hou. AGL(m, 2) acting on R(r, m)/R(s, m). Journal of Algebra
171, pp. 921-938, 1995.

[186] X.-D. Hou. Covering radius of the Reed-Muller code R(1, 7) - a simpler
proof. J. Combin. Theory, Series A 74, pp. 337-341, 1996.

[187] X.-D. Hou. GL(m, 2) acting on R(r, m)/R(r − 1, m). Discrete Mathe-
matics 149, pp. 99-122, 1996.

[188] X.-D. Hou. On the covering radius of R(1, m) in R(3, m). IEEE Trans-
actions on Information Theory,vol. 42, no. 3, pp. 1035-1037, 1996.

[189] X.-D. Hou. The Reed-Muller code R(1, 7) is normal. Designs, Codes
and Cryptography 12, pp. 75-82, 1997.

[190] X.-D. Hou. Cubic bent functions. Discrete Mathematics vol. 189, pp.
149-161, 1998.

167
[191] X.D. Hou. On the coefficients of binary bent functions. Proceedings of
the American American Society, Vol. 128, No. 4, pp. 987-996, 2000.

[192] X.-D. Hou. New Constructions of Bent Functions. Proceedings of the

International Conference on Combinatorics, Information Theory and
Statistics; Journal of Combinatorics, Information and System Sciences,
Vol. 25, Nos. 1-4, pp. 173-189, 2000.

[193] X.-D. Hou. On Binary Resilient Functions. Des. Codes Cryptography

28(1), pp. 93-112, 2003.

[194] X.-D. Hou. Group Actions on Binary Resilient Functions. Appl. Alge-
bra Eng. Commun. Comput. 14(2), pp. 97-115, 2003.

[195] X.-D. Hou. A note on the proof of a theorem of Katz. Finite Fields
and their Applications Volume 11, pp. 316-319, 2005.

[196] X.-D. Hou and P. Langevin. Results on bent functions, Journal of

Combinatorial Theory, Series A, 80, pp. 232-246, 1997.

[197] H. Hu and D. Feng. On quadratic bent functions in polynomial forms.

IEEE Trans. Info. Theory, vol. 53, pp. 2610-2615, 2007.

[198] T. Iwata and K. Kurosawa. Probabilistic higher order differential at-

tack and higher order bent functions. Proceedings of ASIACRYPT’99,
Lecture Notes in Computer Science 1716, pp. 62-74, 1999.

[199] C.J.A. Jansen and D.E. Boekee. The shortest feedback shift register
that can generate a given sequence. Proceedings of CRYPTO’89, Lec-
ture Notes in Computer Science 435, pp. 90-99,1990 (this paper refers
to the classified PhD thesis of C.J.A. Jansen entitled “Investigations
on nonlinear streamcipher systems: construction and evaluation meth-
ods”, Philips).

[200] T. Johansson and F. Jönsson. Improved fast correlation attack

on stream ciphers via convolutional codes. Proceedings of EURO-
CRYPT’99, Lecture Notes in Computer Science 1592, pp. 347-362,
1999.

[201] T. Johansson and F. Jönsson. Fast correlation attacks based on turbo

code techniques. Advances in Cryptology - CRYPTO’99, no. 1666 in
Lecture Notes in Computer Science, pp. 181-197, 1999.

168
[202] T. Johansson and F. Jönsson. Fast correlation attacks through re-
construction of linear polynomials. Advances in Cryptology - CRYPTO
2000, no. 1880 in Lecture Notes in Computer Science, pp. 300-315,
2000.

[203] F. Jönsson. Some results on fast correlation attacks. PhD thesis. Lund
University. 2002.

[204] D. Jungnickel. Difference sets. Contemporary Design Theory: A Col-

lection of Surveys, J. Dinitz and D. R. Stinson eds. John Wiley & Sons,
1992.

[205] W. Kantor. An Exponential Number of Generalized Kerdock Codes.

Inf. and Contr. 53, pp. 74-80, 1982.

[206] T. Kasami. The weight enumerators for several classes of subcodes of

the second order binary Reed-Muller codes. Information and Control,
18, pp. 369-394, 1971.

[207] T. Kasami and N. Tokura. On the weight structure of the Reed Muller
codes, IEEE Trans. Info. Theory, vol. 16, pp. 752-759, 1970.

[208] T. Kasami, N. Tokura, and S. Azumi. On the Weight Enumration of

Weights Less than 2.5d of Reed-Muller Codes. Information and Control,
30:380–395, 1976.

[209] N. Katz. On a theorem of Ax. American Journal of Mathematics 93,

pp. 485-499, 1971.

[210] S. Kavut, S. Maitra and M. D. Yücel. Search for Boolean functions

with excellent profiles in the rotation symmetric class. IEEE Transac-
tions on Information Theory, vol. 53, no. 5, pp. 1743-1751, 2007.

[211] A. M. Kerdock. A class of low-rate non linear codes. Information and

Control 20, pp. 182-187, 1972.

[212] A. Kerckhoffs. La Cryptographie Militaire. Journal des Sciences Mil-

itaires, 1883.

[213] J.D. Key, T.P. McDonough and V.C. Mavron. Information sets and
partial permutation decoding for codes from finite geometries. Finite
Fields and their Applications Volume 12, Issue 2, pp. 232-247, 2006.

169
[214] J. Khan, G. Kalai and N. Linial. The influence of variables on Boolean
functions. IEEE 29th Symp. on foundations of Computer Science, pp.
68-80, 1988.

[215] L.R. Knudsen. Truncated and higher order differentials. Proceedings

of Fast Software Encryption, Second International Workshop, Lecture
Notes in Computer Science 1008, pp. 196-211, 1995.

[216] K. Khoo and G. Gong. New constructions for resilient and highly non-
linear Boolean functions. Proceedings of 8th Australasian Conference,
ACISP 2003, Wollongong, Austrialia, Lecture Notes in Computer Sci-
ence 2727, pp. 498-509, 2003.

[217] K. Khoo, G. Gong and D. Stinson. A New Characterization of Semi-

bent and Bent Functions on Finite Fields. Designs, Codes and Cryp-
tography, Volume 38, Number 2, pp. 279-295, 2006.

[218] A. Klapper and M. Goresky. Feedback Shift Registers, 2-Adic Span,

and Combiners with Memory. Journal of Cryptology, vol. 10, pp. 111-
147. 1997.

[219] L.R. Knudsen and M.P.J. Robshaw. Non-linear approximations in lin-

ear cryptanalysis. Proceedings of EUROCRYPT’96, Lecture Notes in
Computer Science 1070, pp. 224-236, 1996.

[220] P.V. Kumar, R.A. Scholtz and L.R. Welch. Generalized bent functions
and their properties, Journal of Combinatorial Theory, Series A 40, pp.
90-107, 1985.

[221] K. Kurosawa, T. Iwata and T. Yoshiwara. New covering radius of

Reed-Muller codes for t-resilient functions. Proceedings of Selected Areas
in Cryptography, 8th Annual International Workshop, Lecture Notes in
Computer Science 2259, pp. 75 ff, 2001.

[222] K. Kurosawa and R. Matsumoto. Almost security of cryptographic

Boolean functions. IEEE Transactions on Information Theory, vol.50,
no. 11, pp. 2752-2761, 2004.

[223] K. Kurosawa and T. Satoh. Design of SAC/P C(`) of order k Boolean

functions and three other cryptographic criteria. Advances in Cryptol-
ogy, EUROCRYPT’ 97, Lecture Notes in Computer Science 1233, pp.
434-449, 1997.

170
[224] P. Lacharme. Post processing functions for a physical random number
generator. Proceedings of Fast Software Encryption 2008, Lecture Notes
in Computer Science 5086, p 334-342, 2008.

[225] G. Lachaud and J. Wolfmann. The Weights of the Orthogonals of

the Extended Quadratic Binary Goppa Codes. IEEE Trans. Inform.
Theory, vol. 36, pp. 686-692, 1990.

[226] J. Lahtonen, G. McGuire and H. Ward. Gold and Kasami-Welch func-

tions, quadratic forms and bent functions. Advances of Mathematics of
Communication, vol. 1, pp. 243-250, 2007.

[227] X. Lai. Higher order derivatives and differential cryptanalysis. Pro-

ceedings of the ”Symposium on Communication, Coding and Cryptog-
raphy”, in honor of J. L. Massey on the occasion of his 60’th birthday.
1994.

[228] X. Lai. Additive and linear structures of cryptographic functions. Pro-

ceedings of Fast Software Encryption, Second International Workshop,
Lecture Notes in Computer Science 1008, pp. 75-85, 1995.

[229] P. Langevin. Covering radius of RM (1, 9) in RM (3, 9). Eurocode’90,

Lecture Notes in Computer Science 514, pp. 51-59, 1991.

[230] P. Langevin. On the orphans and covering radius of the Reed-Muller

codes. Proceedings of AAECC 9, Lecture Notes in Computer Science
539, pp. 234-240, 1991.

[231] P. Langevin. On generalized bent functions. CISM Courses and Lec-

tures 339 (Eurocode), pp. 147-157, 1992.

[232] P. Langevin and G. Leander. Monomial Bent Functions and Stickel-

berger’s Theorem. TFinite Fields and theirs Applications vol. 14, no.
3, pp. 727-742, 2008.

[233] P. Langevin, G. Leander, P. Rabizzoni, P. Veron and J.-P. Zanotti.

Web page http://langevin.univ-tln.fr/project/quartics/

[234] P. Langevin, P. Rabizzoni, P. Veron, J.-P. Zanotti. On the number of

bent functions with 8 variables. Proceedings of the conference BFCA
2006, Publications des universités de Rouen et du Havre, pp. 125-136,
2007.

171
[235] P. Langevin and P. Solé. Kernels and defaults. American Mathematical
Society (Proceedings of the conference Finite Fields and Applications
Fq4) Contemporary Mathematics 225, pp. 77-85, 1999.
[236] P. Langevin and P. Véron. On the nonlinearity of power functions.
Designs, Codes and Cryptography 37, pp. 31 - 43, 2005.
[237] P. Langevin and J.-P. Zanotti. Nonlinearity of some invariant Boolean
functions. Designs, Codes and Cryptography 36, pp. 131 - 146, 2005.
[238] C. Lauradoux and M. Videau. Matriochka symmetric Boolean func-
tions. Proceedings of International Symposium on Information Theory,
ISIT 2008.
[239] G. Leander. Bent functions with 2r Niho exponents. Proceedings of the
Workshop on Coding and Cryptography 2005, pp. 454-461, 2005.
[240] G. Leander. Monomial bent functions. Proceedings of the Workshop on
Coding and Cryptography 2005, Bergen, pp. 462-470, 2005. And IEEE
Transactions on Information Theory, vol. 52, no. 2, pp. 738-743, 2006.
[241] G. Leander. Another class of non-normal bent functions. Proceedings
of the conference BFCA 2006, Publications des universités de Rouen et
du Havre, pp. 87-98, 2006.
[242] G. Leander and A. Kholosha. Bent functions with 2r Niho exponents.
IEEE Transactions on Information Theory, vol. 52, no. 12, pp. 5529-
5532, 2006.
[243] R. J. Lechner. Harmonic analysis of switching functions. In Recent
Developments in Switching Theory, Academic Press, New York, 1971.
[244] S. Leveiller, G. Zemor, P. Guillot and J. Boutros. A new cryptanalytic
attack for PN-generators filtered by a Boolean function. Proceedings
of Selected Areas of Cryptography 2002, Lecture Notes in Computer
Science 2595, pp. 232 - 249 (2003).
[245] N.Li and W.Qi. Symmetric Boolean functions depending on an odd
number of variables with maximum algebraic immunity. IEEE Trans-
actions on Information Theory, vol. 52, no. 5, pp. 2271-2273, 2006.
[246] N. Li and W.-Q. Qi. Construction and analysis of Boolean functions
of 2t + 1 variables with maximum algebraic immunity. Proceedings of
Asiacrypt 2006, Lecture Notes in Computer Science 4284, pp. 84-98,
2006.

172
[247] N. Li, L. Qu, W.-F. Qi, G. Feng, C. Li and D. Xie. On the construction
of Boolean functions with optimal algebraic immunity. IEEE Transac-
tions on Information Theory, vol. 54, no. 3, pp. 1330-1334, 2008.

[248] R. Lidl and H. Niederreiter, Finite Fields, Encyclopedia of Mathemat-

ics and its Applications, vol. 20, Addison-Wesley, Reading, Massachus-
setts (1983)

[249] S. Ling and C. Xing, Coding Theory, Cambridge: Cambridge Univer-

sity Press, 2004.

[250] N. Linial, Y. Mansour and N. Nisan. Constant depth circuits, Fourier

transform, and learnability. Journal of the Association for Computing
Machinery, vol. 40 (3), pp. 607-620, 1993.

[251] J. H. van Lint. Introduction to coding theory, Springer, New York,

1982.

[252] F. Liu and K. Feng. On the 2m -variable symmetric Boolean functions

with maximum algebraic immunity 2m−1 . Proceedings of the Workshop
on Coding and Cryptography 2007 WCC, pp. 225-232, 2007.

[253] S. Lloyd. Properties of binary functions. Proceedings of EURO-

CRYPT’90, Lecture Notes in Computer Science 473, pp. 124-139, 1991.

[254] S. Lloyd. Counting binary functions with certain cryptographic prop-

erties. Journal of Cryptology 5, pp. 107-131; 1992.

[255] S. Lloyd. Balance, uncorrelatedness and the strict avalanche criterion.

Discrete Applied Mathematics, 41, pp. 223-233, 1993.

[256] M. Lobanov. Tight bound between nonlinearity and algebraic immu-

nity. IACR ePrint Archive http://eprint.iacr.org/ 2005/441.

[257] O.A. Logachev, A.A. Salnikov and V.V. Yashchenko. Bent functions
on a finite Abelian group. Discrete Mathematics Appl. vol 7, N◦ 6, pp.
547-564, 1997.

[258] F. J. MacWilliams and N. J. Sloane. The theory of error-correcting

codes, Amsterdam, North Holland. 1977.

[259] J. A. Maiorana. A classification of the cosets of the Reed-Muller code

R(1, 6). Mathematics of Computation. vol. 57, No. 195, pp. 403-414,
1991.

173
[260] S. Maitra. Highly nonlinear balanced Boolean functions with very good
autocorrelation property. Proceedings of the Workshop on Coding and
Cryptography 2001 published by Electronic Notes in Discrete Mathe-
matics, Elsevier, vo. 6, pp. 355-364, 2001.

[261] S. Maitra. Autocorrelation properties of correlation immune Boolean

functions. Proceedings of INDOCRYPT 2001, Lecture Notes in Com-
puter Science 2247, pp. 242-253, 2001.

[262] S. Maitra, S. Kavut, M. Yucel. Balanced Boolean Function on 13-

variables having Nonlinearity greater than the Bent Concatenation
Bound. Proceedings of the conference BFCA 2008, Copenhagen, to ap-
pear.

[263] S. Maitra and E. Pasalic. Further constructions of resilient Boolean

functions with very high nonlinearity. IEEE Transactions on Informa-
tion Theory, vol.48, no. 7, pp. 1825-1834, 2002.

[264] S. Maitra and P. Sarkar. Enumeration of correlation-immune Boolean

functions. Proceedings of ACISP 1999, pp. 12-25, 1999.

[265] S. Maitra and P. Sarkar. Maximum nonlinearity of symmetric Boolean

functions on odd number of variables. IEEE Transactions on Informa-
tion Theory, vol. 48, pp. 2626-2630, 2002.

[266] S. Maitra and P. Sarkar. Highly nonlinear resilient functions opti-

mizing Siegenthaler’s inequality. Proceedings of CRYPTO’99, Lecture
Notes in Computer Science 1666, pp. 198-215, 1999.

[267] S. Maitra and P. Sarkar. Modifications of Patterson-Wiedemann func-

tions for cryptographic applications. IEEE Transactions on Information
Theory, vol. 48, pp. 278-284, 2002.

[268] S. Maity and S. Maitra. Minimum distance between bent and 1-

resilient Boolean functions. Proceedings of Fast Software Encryption
2004, Lecture Notes in Computer Science 3017, pp. 143-160, 2004.

[269] J. L. Massey. Shift-register analysis and BCH decoding. IEEE Trans-

actions on Information Theory, vol. 15, pp. 122-127, 1969.

[270] J. L. Massey. Randomness, arrays, differences and duality. IEEE

Transactions on Information Theory, vol. 48, pp. 1698-1703, 2002.

174
[271] M. Matsui. Linear cryptanalysis method for DES cipher. Proceedings
of EUROCRYPT’93, Lecture Notes in Computer Science 765, pp. 386-
397, 1994.

[272] R.J. McEliece. Weight congruence for p-ary cyclic codes. Discrete
Mathematics, 3, pp. 177-192, 1972.

[273] R. L. McFarland. A family of noncyclic difference sets, Journal of

Comb. Theory, Series A, no. 15, pp. 1-10, 1973.

[274] W. Meier, E. Pasalic and C. Carlet. Algebraic attacks and decom-

position of Boolean functions. Advances in Cryptology, EUROCRYPT
2004, Lecture Notes in Computer Science 3027, pp. 474-491, 2004.

[275] W. Meier and O. Staffelbach. Fast correlation attacks on stream ci-

phers. Advances in Cryptology, EUROCRYPT’88, Lecture Notes in
Computer Science 330, pp. 301-314, 1988.

[276] W. Meier and O. Staffelbach. Nonlinearity Criteria for Cryptographic

Functions. Advances in Cryptology, EUROCRYPT’ 89, Lecture Notes
in Computer Science 434, pp. 549-562, 1990.

[277] W. Meier and O. Staffelbach. Correlation properties of combiners with

memory in stream ciphers. Advances in Cryptology, EUROCRYPT’90,
Lecture Notes in Computer Science 473, pp. 204-213, 1990.

[278] A. Menezes, P. van Oorschot and S. Vanstone. Handbook of Applied

Cryptography. CRC Press Series on Discrete Mathematics and Its Ap-
plications, 1996.

[279] Q. Meng, H. Zhang, M. Yang and J. Cui. On the degree of homoge-

neous bent functions. Discrete Applied Mathematics Volume 155, Issue
5, pp. 665-669, 2007.

[280] S. Mesnager. Improving the lower bound on the higher order non-
linearity of Boolean functions with prescribed algebraic immunity.
IEEE Transactions on Information Theory, vol. 54, no. 8, pp. 3656
- 3662, 2008. Preliminary version available at IACR ePrint Archive
http://eprint.iacr.org/, 2007/117.

[281] S. Mesnager. On the number of resilient Boolean functions. Proceedings

of the conference “The first Symposium on Algebraic Geometry and its
Applications” (SAGA’07), Tahiti, 2007, published by World Scientific,

175
 Series on Number Theory and its Applications, Vol. 5, pp. 419-433,
2008.

[282] W. Millan, A. Clark and E. Dawson. Heuristic design of cryptograph-

ically strong balanced Boolean functions. EUROCRYPT’98, Advances
in Cryptology, Lecture Notes in Computer Science 1403, 1998.

[283] C. J. Mitchell. Enumerating Boolean functions of cryptographic sig-

nifiance. Journal of Cryptology 2 (3), pp. 155-170, 1990.

[284] J. Mykkelveit. The covering radius of the [128,8] Reed-Muller code

is 56. IEEE Transactions on Information Theory, vol. 26, no. 3, pp.
359-362, 1980.

[285] Y. Nawaz, G. Gong, and K. Gupta. Upper Bounds on Algebraic Immu-

nity of Power Functions. Proceeding of Fast Software Encryption 2006,
Lecture Notes in Computer Science 4047, pp. 375-389, 2006.

[286] N. Nisan and M. Szegedy. On the degree of Boolean functions as real

polynomials. Comput. Complexity 4, pp. 301-313, 1994.

[287] K. Nyberg. Constructions of bent functions and difference sets, EU-

ROCRYPT’90, Advances in Cryptology, Lecture Notes in Computer
Science 473, pp. 151-160, 1991.

[288] L. O’Connor and A. Klapper. Algebraic nonlinearity and its applica-

tions to cryptography. Journal of Cryptology 7, pp. 213-227, 1994.

[289] D. Olejár and M. Stanek. On cryptographic properties of random

Boolean functions. Journal of Universal Computer Science, vol. 4, No.8,
pp. 705-717, 1998.

[290] J. D. Olsen, R. A. Scholtz and L. R. Welch. Bent function sequences,

IEEE Trans. on Inf. Theory, Vol. 28, no. 6, 1982.

[291] E. Pasalic. On Boolean functions in symmetric-key ciphers. Ph.D. The-

sis, 2003.

[292] E. Pasalic, T. Johansson, S. Maitra and P. Sarkar. New constructions

of resilient and correlation immune Boolean functions achieving upper
bounds on nonlinearity. Proceedings of the Workshop on Coding and
Cryptography 2001, published by Electronic Notes in Discrete Mathe-
matics, Elsevier, vo. 6, pp. 425-434, 2001.

176
[293] E. Pasalic and S. Maitra. A Maiorana-McFarland type construction
for resilient Boolean functions on n variables (n even) with nonlinearity
> 2n−1 − 2n/2 + 2n/2−2 . Proceedings of the Workshop on Coding and
Cryptography 2003, pp. 365-374, 2003.

[294] S. M. Park, S. Lee, S. H. Sung, K. Kim. Improving bounds for the num-
ber of correlation-immune Boolean functions. Information Processing
Letters 61, pp. 209-212, 1997.

[295] N.J. Patterson and D.H. Wiedemann. The covering radius of the
[215 , 16] Reed-Muller code is at least 16276. IEEE Transactions on In-
formation Theory, vol. 29, pp. 354-356, 1983.

[296] N.J. Patterson and D.H. Wiedemann. Correction to [295]. IEEE

Transactions on Information Theory, vol. 36, no. 2, pp. 443, 1990.

[297] J. Pieprzyk and X.-M. Zhang. Computing Möbius Transforms of

Boolean Functions and Characterizing Coincident Boolean functions.
Proceedings of the conference BFCA 2007, Publications des universités
de Rouen et du Havre, 2007.

[298] V. S. Pless, W. C. Huffman, Eds, R. A. Brualdi, assistant editor.

Handbook of Coding Theory, Amsterdam, the Netherlands: Elsevier,
1998.

[299] A. Pott. Finite Geometry and Character Theory. Lecture Notes in

Mathematics, vol. 1601, Berlin, Springer Verlag, 1995.

[300] B. Preneel, W. Van Leekwijck, L. Van Linden, R. Govaerts and J.

Vandevalle. Propagation characteristics of Boolean functions, Proceed-
ings of EUROCRYPT’90, Lecture Notes in Computer Sciences 473, pp.
161-173, 1991.

[301] B. Preneel, R. Govaerts and J. Vandevalle. Boolean functions satisfy-

ing higher order propagation criteria, Proceedings of EUROCRYPT’91,
Lecture Notes in Computer Sciences 547, pp. 141-152, 1991.

[302] B. Preneel. Analysis and Design of Cryptographic Hash Functions,

Ph. D. Thesis, Katholieke Universiteit Leuven, K. Mercierlaan 94, 3001
Leuven, Belgium, U.D.C. 621.391.7, 1993.

[303] L. Qu and C. Li. Weight support technique and the symmetric Boolean
functions with maximum algebraic immunity on even number of vari-

177
 ables. Proceedings of INSCRYPT 2007, Lecture Note in Computer Sci-
ence 4990, pp. 271-282.
[304] L. Qu, C. Li and K. Feng. A note on symmetric Boolean functions
with maximum algebraic immunity in odd number of variables. IEEE
Trans. on Inf. Theory, vol. 53, pp. 2908-2910, 2007.
[305] L. Qu, K. Feng, L. Feng and L. Wang. Constructing symmetric
Boolean functions with maximum algebraic immunity. IEEE Trans. on
Inf. Theory, vol. 55, pp. 2406-2412, 2009.
[306] M. Quisquater. Applications of character theory and the Möbius in-
version principle to the study of cryptographic properties of Boolean
functions. PhD thesis, 2004.
[307] M. Quisquater, B. Preneel and J. Vandewalle. A new inequality in
discrete Fourier theory. IEEE Trans. on Inf. Theory, vol. 49, pp. 2038-
2040, 2003.
[308] M. Quisquater, B. Preneel and J. Vandewalle. Spectral characteriza-
tion of cryptographic Boolean functions satisfying the (extended) prop-
agation criterion of degree l and order k. Inf. Process. Lett. 93(1), pp.
25-28, 2005.
[309] C. Riera and M. G. Parker. Generalised Bent Criteria for Boolean
Functions (I). IEEE Transactions on Information theory, vol. 52, no.
9, pp. 4142-4159, 2006.
[310] C. R. Rao. Factorial experiments derived from combinatorial arrange-
ments of arrays. J. Roy. Statist. 9, pp. 128-139, 1947.
[311] F. Rodier. Asymptotic nonlinearity of Boolean functions. Designs,
Codes and Cryptography, no 40:1 2006, pp 59-70.
[312] S. Ronjom, M. Abdelraheem and L. E. Danielsen. Online
database of Boolean Functions. http://www.ii.uib.no/ mo-
hamedaa/odbf/index.html
[313] S. Rønjom and T. Helleseth. A new attack on the filter generator. IEEE
Transactions on Information theory, vol. 53, no. 5, pp. 1752-1758, 2007.
[314] S. Rønjom and T. Helleseth. Attacking the filter generator over
GF (2m ). Proceedings of the International Workshop on the Arithmetic
of Finite Fields, WAIFI 2007, Lecture Notes in Computer Science 4547,
pp. 264-275, June 2007.

178
[315] O. S. Rothaus. On “bent” functions. J. Comb. Theory, 20A, pp. 300-
305, 1976.

[316] B.V. Ryazanov. On the distribution of the spectral complexity of

Boolean functions. Discrete Mathematics Appl., vol. 4, No. 3, pp. 279-
288, 1994.

[317] R. A. Rueppel Analysis and design of stream ciphers Com. and Contr.
Eng. Series, Berlin, Heidelberg, NY, London, Paris, Tokyo 1986

[318] R. A. Rueppel and O. J. Staffelbach. Products of linear recurring se-

quences with maximum complexity. IEEE Transactions on Information
theory, vol. 33, no. 1, 1987.

[319] P. Sarkar. The Filter-Combiner Model for Memoryless Synchronous

Stream Ciphers. Proceedings of CRYPTO 2002, Lecture Notes in Com-
puter Science 2442, pp. 533-548, 2002.

[320] P. Sarkar and S. Maitra. Construction of nonlinear Boolean functions

with important cryptographic properties. Proceedings of EUROCRYPT
2000, Lecture Notes in Computer Science 1807, pp. 485-506, 2000.

[321] P. Sarkar and S. Maitra. Nonlinearity Bounds and Constructions of

Resilient Boolean Functions. Proceedings of CRYPTO 2000, Lecture
Notes in Computer Science 1880, pp. 515-532, 2000.

[322] P. Sarkar and S. Maitra. Construction of nonlinear resilient Boolean

functions using “small” affine functions. IEEE Transactions on Infor-
mation theory, vol. 50, no 9, pp. 2185-2193, 2004.

[323] P. Sarkar and S. Maitra. Balancedness and correlation immunity of

symmetric Boolean functions. Discrete Mathematics 307, pp. 2351-2358,
2007.

[324] P. Savicky. On the bent Boolean functions that are symmetric. Eur.
J. Combinatorics 15, pp. 407-410, 1994.

[325] M. Schneider. A note on the construction and upper bounds of

correlation-immune functions. Proceedings of the 6th IMA Conference,
Lecture Notes In Computer Science 1355, pp. 295-306, 1997. An ex-
tended version appeared under the title “On the construction and upper
bounds of balanced and correlation-immune functions”, Selected Areas
Crypt., pp. 73-87, 1997.

179
[326] J. Seberry and X-.M. Zhang. Constructions of bent functions from two
known bent functions. Australasian Journal of Combinatorics no. 9, pp.
21-35, 1994.

[327] J. Seberry, X-.M. Zhang and Y. Zheng. On constructions and nonlin-

earity of correlation immune Boolean functions. Advances in Cryptology
- EUROCRYPT’93, Lecture Notes in Computer Science 765, pp. 181-
199, 1994.

[328] J. Seberry, X-.M. Zhang and Y. Zheng. Nonlinearly balanced Boolean

functions and their propagation characteristics. Advances in Cryptology
- CRYPTO’93, pp. 49-60, 1994.

[329] N. V. Semakov and V. A. Zinoviev. Balanced codes and tactical con-

figurations. Problems of Info. Trans. 5(3), pp. 22-28 (1969)

[330] A. Shanbhag, V. Kumar and T. Helleseth. An upper bound for the

extended Kloosterman sums over Galois rings. Finite Fields and their
Applications 4, pp. 218-238, 1998.

[331] C. E. Shannon. A mathematical theory of communication. Bell system

technical journal, 27, pp. 379-423, 1948.

[332] C.E. Shannon. Communication theory of secrecy systems. Bell system

technical journal, 28, pp. 656-715, 1949.

[333] C. E. Shannon. The synthesis of two-terminal switching circuits. Bell

system technical journal, 28, pp. 59–98, 1949.

[334] I. Shparlinski. Bounds on the Fourier coefficients of the weighted sum

function. Inf. Process. Lett. 103(3), pp. 83-87, 2007.

[335] I. Shparlinski and A. Winterhof. On the nonlinearity of linear recur-

rence sequences. Applied Mathematics Letters 19, pp. 340-344, 2006.

[336] T. Siegenthaler. Correlation-immunity of nonlinear combining func-

tions for cryptographic applications. IEEE Transactions on Informa-
tion theory, vol. 30, no 5, pp. 776-780, 1984.

[337] T. Siegenthaler. Decrypting a Class of Stream Ciphers Using Cipher-

text Only. IEEE Transactions on Computer, vol. C-34, No 1, pp. 81-85,
1985.

180
[338] P. Stanica, S. Maitra and J. Clark. Results on rotation symmetric
bent and correlation immune Boolean functions. Proceedings of Fast
Software Encryption 2004, Lecture Notes in Computer Science 3017,
pp. 161-177, 2004.

[339] P. Stanica and S. H. Sung. Boolean functions with five controllable

cryptographic properties. Designs, Codes and Cryptography 31, pp. 147-
157, 2004.

[340] V. Strassen. Gaussian elimination is not optimal. Numerische Math.

13, pp. 354-356, 1969.

[341] I. Strazdins. Universal affine classification of Boolean functions. Acta

Applicandae Mathematicae 46, pp. 147-167, 1997.

[342] T. Sugita, T. Kasami and T. Fujiwara. Weight distributions of the

third and fifth order Reed-Muller codes of length 512. Nara Inst. Sci.
Tech. Report, 1996.

[343] S. H. Sung, S. Chee and C. Park. Global avalanche characteristics

and propagation criterion of balanced Boolean functions. Information
Processing Letters 69, pp. 21-24, 1999.

[344] H. Tapia-Recillas and G. Vega. An upper bound on the number of

iterations for transforming a Boolean function of degree greater than
or equal than 4 to as function of degree 3. Designs, Codes and Cryp-
tography 24, pp. 305-312, 2001.

[345] Y. V. Tarannikov. On resilient Boolean functions with maximum pos-

sible nonlinearity. Proceedings of INDOCRYPT 2000, Lecture Notes in
Computer Science 1977, pp. 19-30, 2000.

[346] Y. V. Tarannikov. New constructions of resilient Boolean functions

with maximum nonlinearity. Proceedings of FSE 2001, 8th International
Workshop, FSE 2001, Lecture Notes in Computer Science 2355, pp. 66-
77, 2001.

[347] Y. V. Tarannikov and D. Kirienko. Spectral analysis of high order

correlation immune functions. Proceedings of 2001 IEEE International
Symposium on Information Theory, p. 69, 2001 (full preliminary version
at IACR ePrint Archive http://eprint.iacr.org/).

181
[348] Y. V. Tarannikov, P. Korolev and A. Botev. Autocorrelation coeffi-
cients and correlation immunity of Boolean functions. Proceedings of
Asiacrypt 2001, Lecture Notes in Computer Science 2248, pp. 460-479,
2001

[349] F.Didier and J.Tillich: Computing the Algebraic Immunity Efficiently.

Proceedings of Fast Software Encryption 2006, Lecture Notes in Com-
puter Science 4047, pp. 359-374, 2006.

[350] S. Tsai. Lower bounds on representing Boolean functions as polyno-

mials in Z?m . SIAM J. Discrete Mathematics, vol. 9 (1), pp. 55-62, 1996.

[351] S. F. Vinokurov and N. A. Peryazev. An expansion of Boolean function

into a sum of products of subfunctions. Discrete Mathematics Appl., vol.
3 (5), pp. 531-533, 1993.

[352] A.F. Webster and S.E. Tavares. On the design of S-boxes. In Pro-
ceedings of CRYPTO’85, Lecture Notes in Computer Science 219, pp.
523-534, 1985.

[353] J. Wolfmann. Bent functions and coding theory. Difference Sets, Se-
quences and their Correlation Properties, A. Pott, P. V. Kumar, T.
Helleseth and D. Jungnickel, eds., pp. 393–417. Amsterdam: Kluwer,
1999.

[354] Y. X. Yang and B. Guo. Further enumerating Boolean functions of

cryptographic signifiance. Journal of Cryptology 8 (3), pp. 115-122,
1995.

[355] R. Yarlagadda and J.E. Hershey. Analysis and synthesis of bent se-
quences, IEE proceedings. Part E. Computers and digital techniques,
vol. 136, pp. 112-123, 1989.

[356] A.M. Youssef and G. Gong. Hyper-bent functions. Proceedings of EU-

ROCRYPT 2001, Lecture Notes in Computer Science 2045, Berlin, pp.
406-419, 2001.

[357] N. Y. Yu and G. Gong. Constructions of quadratic bent functions in

polynomial forms. IEEE Transactions on Information Theory, vol. 52,
no. 7, pp. 3291-3299, 2006.

[358] M. Zhang. Maximum correlation analysis of nonlinear combining func-

tions in stream ciphers. Journal of Cryptology 13 (3), pp. 301-313, 2000.

182
[359] X.-M. Zhang and Y. Zheng. GAC - the criterion for global avalanche
characteristics of cryptographic functions. Journal of Universal Com-
puter Science, 1(5), pp. 320-337, 1995.

[360] X.-M. Zhang and Y. Zheng. Auto-correlations and new bounds on

the nonlinearity of Boolean functions. Proceedings of EUROCRYPT’96,
Lecture Notes in Computer Science 1070, pp. 294-306, 1996.

[361] X.-M. Zhang and Y. Zheng. Characterizing the structures of cryp-

tographic functions satisfying the propagation criterion for almost all
vectors. Designs, Codes and Cryptography, 7(1), pp. 11-134, 1996.

[362] X.-M. Zhang and Y. Zheng. The nonhomomorphicity of Boolean func-

tions. Proceedings of SAC 1998, Lecture Notes in Computer Science
1556, pp. 280-295, 1999.

[363] Y. Zheng and X. M. Zhang. Plateaued functions. Proceedings of

ICICS’99, Lecture Notes in Computer Science 1726, pp. 284-300, 1999.

[364] Y. Zheng, X.-M. Zhang, and H. Imai. Restriction, terms and nonlin-
earity of Boolean functions. Theoretical Computer Science, 226(1-2),
pp. 207-223, 1999.

[365] Y. Zheng and X.-M. Zhang. On relationships among avalanche, nonlin-

earity and correlation immunity. Proceedings of Asiacrypt 2000, Lecture
Notes in Computer Science 1976, pp. 470-483, 2000.

[366] Y. Zheng and X.-M. Zhang. Improving upper bound on the nonlinear-
ity of high order correlation immune functions. Proceedings of Selected
Areas in Cryptography 2000, Lecture Notes in Computer Science 2012,
pp. 262-274, 2001.

[367] N. Zierler and W. H. Mills. Products of linear recurring sequences.

Journal of Algebra 27, pp. 147-157, 1973.

183
Index
[N, k, d]-code, 34 complexity criteria, 68
m-sequences, 45 concatenating affine functions, 86
2-weight, 17 confusion, 47
conventional cryptography, 5
absolute indicator, 65 correlation attack, 57
absolute trace representation, 15 correlation-immune function, 56
adjoint operator, 29 coset leader, 16
affine functions, 14 covered, 10
affine invariant, 12 covering radius, 34
affinely equivalent, 13 covering radius bound, 51
algebraic attacks, 61 covering sequence, 74
algebraic degree, 12 cryptanalysis, 5
algebraic immunity, 63 cryptography, 5
Algebraic Normal Form, 9 cyclic code, 35
algebraic thickness, 68 cyclotomic class, 15
amplitude, 107
annihilator, 62 decomposable functions, 92
atomic functions, 9 decryption, 5
auto-correlation function, 27 defining set, 35
Ax’s theorem, 40 derivative, 27
difference set, 79
balanced functions, 56 diffusion, 47
BCH bound, 35 Dillon’s functions, 88
bent functions, 52 Dirac symbol, 23
bent-concatenation bound, 52 direct sum, 92
Berlekamp-Massey algorithm, 44 discrete Fourier transform, 21
binary entropy function, 117 distance enumerator, 41
binary expansion, 17 distance to linear structures, 61
binary Möbius transform, 11 distinguishing attacks, 56
Boolean functions, 6 dual code, 34
dual distance, 41
Cayley graph, 32
dual function, 80
ciphertext, 5
code, 6 encryption, 5
codewords, 6 equivalent codes, 34
coincident functions, 13 error correcting codes, 6
combiner model, 44 eSTREAM Project, 45
complete class of functions, 78 extended propagation criterion, 59

184
extension of Maiorana-McFarland type,linear complexity, 44
95 Linear Feedback Shift Registers, 43
linear kernel, 59
fast correlation attacks, 50 linear structure, 59
Fast Fourier Transform, 21 linearly equivalent, 13
Fast Möbius Transform, 11
feedback coefficients, 44 Möbius transform over integers, 19
feedback polynomial, 43 MacWilliams’ identity, 40
Feedback Shift Register, 46 Maiorana-McFarland general construc-
filter model, 45 tion, 119
flat, 30 Maiorana-McFarland original class, 85
Mattson-Solomon polynomial, 16
generalized degree, 18 maximal odd weighting, 37
Generalized Partial Spread, 102 maximum correlation, 67
generator matrix, 34 maximum length sequences, 45
generator polynomial, 35 McEliece’s theorem, 40
global avalanche criterion, 65 minimum distance, 34
Gold function, 82 monomial functions, 89
Hamming code, 34 naive bound, 100
Hamming distance, 8 non-trivial covering sequence, 74
Hamming weight, 8 nonhomomorphicity, 68
higer order nonlinearity, 54 nonlinearity, 51
hyper-bent functions, 103 nonlinearity profile, 54
normal basis, 111
idempotent functions, 118
normal extension, 109
indicator, 13
normal function, 68
indirect sum of bent functions, 95
numerical degree, 18
indirect sum of resilient functions, 129
Numerical Normal Form, 18
information set, 137
inner product, 14 one time pad, 42
orphan of R(1, n), 108
Kasami function, 83
orthogonal, 24
Kerdock code, 110
keystream, 42 parity check polynomial, 50
Kloosterman sums, 77 parity-check matrix, 34
Krawtchouk polynomials, 144 Parseval’s relation, 27
partial bent functions, 107
level of a covering sequence, 74
Partial Spreads class, 87
LFSR, 43
partially defined, 13
linear code, 34
partially-bent functions, 106

185
perfect nonlinear functions, 79 synchronous, 42
plaintext, 5
plateaued functions, 67 Tarannikov et al.’s construction, 127
Poisson summation formula, 25 three-valued functions, 107
power functions, 89 trace function, 15
primary constructions, 85 trace representation, 16
primitive element, 16 transmission rate, 6
private key cryptography, 5
univariate representation, 15
Propagation Criterion, 59
pseudo-Boolean functions, 18 Vernam cipher, 42
pseudo-random sequences, 42
public key cryptography, 5 Walsh transform, 23
weakly-normal function, 68
quadratic bound, 52 weight distribution, 40
quadratic functions, 69 weight enumerator, 40
rank of ϕf , 70 Weil’s bound, 77
redundancy, 6 Wiener-Khintchine Theorem, 27
Reed-Muller codes, 36
Reed-Solomon code, 35
resiliency order, 57
resilient function, 56
rotation symmetric, 118
Rothaus construction, 92
Rothaus’ bound, 84

Sarkar et al.’s bound, 114

Sarkar-Maitra’s divisibility, 114
second order covering sequence, 102
second order derivative, 66
secondary constructions, 85
semi-bent functions, 71
sign function, 22
Stickelberger theorem, 40
Stream ciphers, 42
Strict Avalanche Criterion, 59
sum-of-squares indicator, 65
support of the codeword, 8
support of the function, 8
symmetric cryptography, 5
symmetric function, 142

186