![](bg2.png)
The paper hence focus on nonlinear feedback
functions whose Algebraic Normal Form (ANF) is
given by:
f (x
n−1
,...,x
0
) =
n−1
∑
i=0
c
i
.x
i
+ x
n
+ x
j
.x
k
(3)
for all possible pair j, k such that 0 < j, k < n and
where c
i
are binary coefficients describing whether
the register cell is considered (c
i
= 1) or not (c
i
= 0).
We have conducted an exhaustive exploratory anal-
ysis to find all feedback functions up to the degree
n = 28. To reduce the computing time it has been nec-
essary to find a new way of NLFSR modelling. For
that purpose, we represent NLFSR as directed graphs
whose incidence matrix exhibits specific properties to
express the maximal period property. It is worth men-
tioning that the present study can be easily applied to
any other forms of feedback polynomials (for instance
more quadratic terms).
It is the first exhaustive search to date whereas pre-
vious works only published a very few number of re-
sults due to the search complexity. From the results
obtained, we have identified a number of news re-
sults that can be of high interest to explore further for
n > 28.
In the rest of this article we will consider all oper-
ations in the finite field (F
2
,+,.).
The paper is organized as follows. Section 2 anal-
yses the overall complexity of searching of NLF-
SRs producing m-sequence and presents the related
works. Section 3 introduces our combinatorial model
for NLFSRs and formalizes the maximal period prop-
erty in terms of algebraic equations. Then Section 4
details the particular implementation aspects that have
been used to perform an effective computation. Fi-
nally Section 5 presents the detailed results of our
exhaustive search and identify a few new interesting
properties before concluding in Section 6.
2 PRELIMINARIES
2.1 Complexity Analysis of NLFSR
Search
To date, there are no theoretical results that allow to
easily find maximum period NLFSRs as is the case
for LFSRs (Golomb, 1981). In this section we look at
the approach currently being favoured in recent years.
Let us rewrite (3) in a more simple way:
f (x
n−1
,...,x
0
) = l(x
n−1
,...,x
0
) + x
j
.x
k
(4)
where l(x) = l(x
n−1
,...,x
0
) =
∑
n−1
i=0
c
i
.x
i
+ x
n
is the
linear part of f .
The search for such NLFSRs of length n can be
formalized according the two following steps:
1. Among the 2
n
possible candidates corresponding
l(x), for a given pair i, j fixing the degree-2 mono-
mial, we retain those which validate a certain
number of algebraic or combinatorial properties
I
1
,I
2
,...,I
k
. If these properties are independent,
with respective probabilities of being realized by
a good candidate P(I
i
) = p
i
, then at the end of this
stage we retain N = 2
n
.
∏
k
i=1
p
i
. This step has an
incompressible complexity of 2
n−1
(a symmetry
property presented in Section 2.3 enables to cut
search work in half).
2. For each valid candidate for l(x), we check
whether it is in the maximum period by calcu-
lating the cycle. Complexity is in 2
n
in the
worst case. The average complexity is 2
n−1
from
(Golomb, 1981, Corollary 11, p. 183). No result
is known for this step that would allow us to re-
duce this complexity (except in certain very spe-
cific cases, see (Golomb, 1981, Chapter VII)).
The overall worst-case complexity is therefore 2
2n−1
and the average-case complexity is 2
2n−2
. To reduce
this complexity, the focus must be on the first stage
to reduce the number of candidates to be tested in the
second stage. Some of these properties are already
known (see Section 2.3). In this paper we are going
to add many others thanks to results from graph the-
ory and matrix calculus on the associated incidence
matrix. This significantly reduces the overall com-
plexity of the exhaustive search. This has enabled us
to perform an exploratory analysis up to n = 28.
2.2 Related and Previous Work
Since the Jansen’s seminal thesis in 1989 (Jansen,
1989), the main approach for searching for NLFSRs
of the simplest form considers more or less sophisti-
cated exhaustive search. In (Dabrowski et al., 2014),
this approach has been initiated with parallel comput-
ing.
Later on in (Poluyanenko, 2017), the author stud-
ies NLFSRs implementation on FPGAs and discusses
issues of their optimization. Search method of NLF-
SRs generating M-sequence was given. It was based
on a practical synthesis and explores the possibility of
NLFSR implementation of FPGA.
In (Augustynowicz, 2018), the authors consider
a multi-stages hybrid algorithm which uses Graph-
ics Processor Units (GPU) and developed for process-
ing data-parallel throughput computation. They focus
and give results for feedback polynomials of the form
l(x) + x
i
.x
k
+ x
j
.x
l
(two quadratic monomials). Later
Graph-Based Modelling of Maximum Period Property for Nonlinear Feedback Shift Registers
833