Page 1

International Symposium on Information Theory and its Applications, ISITA2008

Auckland, New Zealand, 7-10, December, 2008

An Efficient SAT Encoding of Circuit Codes

Yury CHEBIRYAK†and Daniel KROENING‡

†Computer Systems Institute

ETH Zurich

8092 Zurich, Switzerland

E-mail: yury.chebiryak@inf.ethz.ch

‡Computing Laboratory

Oxford University

Oxford, OX1 3QD, UK

E-mail: kroening@comlab.ox.ac.uk

Abstract

Circuit codes in hypercubes are generalized snake-in-

the-box codes and are used in analog-to-digital conver-

sion devices. The construction of the longest known

circuit codes is based on either an exhaustive search or

an algorithm that restricts the search to the codes with

periodic coordinate sequences. In this paper, we de-

scribe an efficient SAT encoding of circuit codes, which

enabled us to obtain new circuit codes.

1. INTRODUCTION

In 1958,

the snake-in-the-box problem—finding a binary code

that has unit distance between adjacent code words

and minimum distance two between all other code

words [14].The search for snakes is motivated by

the theory of error-correcting codes (as the vertices of

a solution to the snake or coil in the box problems

can be used as a Gray code that can detect single-

bit errors), electrical engineering, computer network

topologies [1], Systems Biology [9], etc. Approaches

to find long snakes range from studies of mathematical

constructions (e.g. binary necklaces [17]) and certain

patterns in lower dimensions [20, 19] to genetic algo-

rithms [1, 2, 22].

R. C. Singleton generalized the concept of snake-

in-the-box codes to circuit codes with a parameter

spread [21]. A circuit code of spread δ has unit distance

between adjacent code words, and minimum distance δ

between code words δ apart in the ordered sequence.

For example, the circuit codes with the spread δ = 2

are the coil-in-the-box, and the codes with δ = 1 and

2ndistinct code words are the Hamiltonian cycles of

the n-cube. Singleton then presented constructions for

circuit codes for spreads up to 7.

Circuit codes are useful in correcting and limiting

errors in analog-to-digital conversion (see [16]). The

W. H. Kautz brought attention to

This research is supported in part by an award from IBM

Research and by ETH Research Grant TH-19 06-3.

longer the code, the greater the accuracy of the sys-

tem (while the greater the spread, the greater the

error-detection capability). Therefore, determining the

length of the longest n-dimensional circuit code of

spread δ is of interest [8, 23].

V. Klee showed the construction of a code with even

spread δ by extending a code of spread δ using a code

of spread δ−1 [15]. K. Deimer described a method for

finding a circuit code of spread δ and length L−k in

the n-cube from a circuit code in dimension n + 1 of

spread δ+1 and length L [4]. Here k is the number of

times a certain transition is taken (this transition num-

ber is then removed from the transition sequence). It is

not evident that finding such a code (of higher spread

and length, in higher dimension) is easier than the tar-

get circuit code. Paterson and Tuliani presented a con-

struction method based on binary necklaces [17], gen-

eralizing ideas for obtaining single-track circuit codes

of [7]. In earlier work, we improved lower bounds and

proved optimality of circuit codes for 14 different sets of

parameters (n,δ) [24]. The approach uses a SAT-solver

and is not limited to specific values of a spread.

In this paper, we present new (longer than previ-

ously known) circuit codes that we have obtained using

a novel efficient propositional satisfiability encoding.

2. CIRCUIT CODES

Consider an ordered sequence C of L binary code

words W0,W1,...,WL−1. Let dk,m

ming distance between code words Wkand Wm:

=??{ i | Wk[i] ?= Wm[i]}??,

let dCbe the cyclic distance between code words in the

sequence [12]:

H

denote the Ham-

dk,m

H

where Wk[i] denotes the i-th bit of k-th code word, and

dk,m

C

= min{|k − m|, L − |k − m|} .

A path in a hypercube is a sequence of code words,

in which consecutive elements have unit Hamming dis-

1235978-1-4244-2069-8/08/$25.00 ©2008 IEEE

Page 2

tance. In a cycle, the first and the last code words also

have unit Hamming distance:

∀k ∈ {0,1,...,L} :

dk,k+1 mod L

C

= 1 =⇒ dk,k+1 mod L

H

= 1 .

The definitions of circuit codes by Paterson

andTuliani[17] andby

ergelt [18] differ slightly, but were proven equivalent

by L. Haryanto [12].

Preparata andNiev-

Definition 1 (Circuit code [13]). A length L, spread δ

circuit code in the n-cube (or (n, L, δ)-CC) is a cyclic

path C of L binary n-tuples W0,W1,...,WL−1with the

property that for all k,m ∈ {0,1,...,L − 1},

dk,m

H

< δ =⇒ dk,m

C

< δ .

(1)

3. PROPOSITIONAL SAT ENCODING

A SAT solver determines whether a propositional

formula, given in conjunctive normal form (CNF), is

satisfiable. If it is, the solver provides a satisfying as-

signment to the variables in the formula.

In this section, we describe our encoding of a search

for circuit codes into a propositional SAT formula in

detail. Then, we improve it using an observation about

the circuit codes’ structure and obtain new codes.

3.1. The Satisfiability Problem

Let V = {x0,x1,...,xt−1} be a set of Boolean

variables.A literal is a variable xi or its negation

¬xi. Let φ be a propositional formula over the vari-

ables in V . The propositional satisfiability (SAT) prob-

lem [10] is to determine whether there exists an assign-

ment of truth values to the variables in V such that the

formula φ evaluates to true.

3.2. Encoding of Circuit Codes

Our goal is to construct a formula with satisfying

assignments corresponding to the coordinates of the

nodes forming a spread-δ circuit code of L code words

in the n-cube. For this purpose, we define n·L Boolean

variables denoted by Wi[j], where i ∈ {0,...,L−1} and

j ∈ {0,...,n−1}. The Boolean vector Widenotes the

coordinates of node number i of the code, where Wi[0]

corresponds to the right-most bit of the coordinates of

node Wi.

For a sequence of nodes W0,W1,...,WL−1, we en-

code the following constraints:

1. To form a cycle in an n-cube, the neighbouring

nodes of a sequence must be adjacent. The adja-

cency is expressed using the Hamming distance:

φcycle:=

?L−1

k=0

?dk,k+1 mod L

H

= 1?.

(2)

2. The formula (1) suggests to pick pairs of code-

words Wkand Wm, that are at least δ apart in the

sequence1and require their Hamming distances

to be at least δ:

?

φδ:=

0≤k<m<L

?dk,m

C

≥ δ ⇒ dk,m

H

≥ δ) .

(3)

The propositional formula

φCC:= φcycle∧ φδ

(4)

encodes an (n, L, δ)-CC. A satisfying assignment

of φCCcontains the coordinates of some circuit code

with these parameters.

We encoded formulae (2) and (3) using once-twice

chains and bitonic sorting networks respectively (for

details, see [3]) and obtained new circuit codes in [24].

3.3. A More Efficient Encoding

One can reduce the number of variables and clauses

by a factor of two.

Consider nodes Wk and Wm together with neigh-

bours of Wm: Wm−1 and Wm+1. If Wk is at least δ

apart in the sequence from each of these three nodes,

we have to require that their Hamming distances are

at least δ each (see Figure 1).

Suppose that k and m are of the same parity, i.e. the

distance dk,m

C

is even. Then, by 2-colourability of a hy-

percube [11], the Hamming distance is also even. If

the value of the spread is odd, we can strengthen the

property (3) to dk,m

H

≥ δ + 1, which implies the sepa-

rability condition for Wk and neighbours of Wm (be-

cause Hamming distances for them may decrease only

by one). In Eq. (3) we can therefore reduce the number

of constraints by about one half due to redundancy.

3.4. Evaluation

1With such a formulation, for codes with higher spreads there

are fewer of pairs to consider (given that dimension and length

are fixed), hence an encoding of these codes requires fewer vari-

ables and clauses. This is advantageous as performance of exist-

ing approaches decays with increasing spread (e.g., the construc-

tion in [17] uses special kinds of binary necklaces and finding

them for higher spreads is hard).

1236

Page 3

Wm+1

Wm−1

Wm

Wk

dH≥ δdH≥ δ

dH≥ δ

dH= 1dH= 1

Figure 1: Constraints on code words to form a circuit

code.

We generalize this observation, by modifying the

Eq. (3) as follows:

?

∧ dk,m

The modified encoding of a (10,84,3)-CC takes

49.9% less variables and clauses. The runtime2for this

instance is decreased by 51.5%.

Using the efficient encoding we obtained 11 new cir-

cuit codes (see Table 1).

φδ?:=

0≤k<m<L

≥ δ?⇒ dk,m

??(dk,m

H

C

mod 2 ?= δ mod 2)

≥ δ + 1

C

?

.

(5)

References

[1] D. Casella and W. Potter, “Using evolutionary

techniques to hunt for the snakes and coils,” The

2005 IEEE Congress on Evolutionary Computa-

tion, vol. 3, pp. 2499–2505, 2005.

[2] D. A. Casella, “New lower bounds for the snake-

in-the-box and the coil-in-the-box problems: Us-

ing evolutionary techniques to hunt for snakes and

coils,” Master’s thesis, The University of Georgia,

2005.

[3] Y. Chebiryak and D. Kroening, “Towards a clas-

sification of Hamiltonian cycles in the 6-cube,”

Journal on Satisfiability, Boolean Modeling and

Computation (JSAT), vol. 4, pp. 57–74, 2008.

[4] K. Deimer, “Some new bounds for the maximum

length of circuit codes.” IEEE Transactions on In-

formation Theory (IT), vol. 30, no. 5, pp. 754–756,

1984.

2For our experiments, we use the MiniSat, written by E´ en and

S¨ orensson [6]. MiniSat provides interfaces for incremental solving

and All-SAT. The current version uses preprocessing techniques

from QBF-Solvers [5]. All experiments were carried out on a PC

with an Intel Xeon (3.0-GHz, 4-GB RAM, running Linux) with

a timeout of 24 h.

[5] N. E´ en and A. Biere, “Effective preprocessing in

SAT through variable and clause elimination,” in

Theory and Applications of Satisfiability Testing

(SAT), ser. LNCS, vol. 3569.

61–75.

Springer, 2005, pp.

[6] N. E´ en and N. S¨ orensson, “An extensible SAT-

solver,” in Theory and Applications of Satisfiabil-

ity Testing (SAT), ser. LNCS.

vol. 2919, pp. 502–518.

Springer, 2004,

[7] T. Etzion and K. G. Paterson, “Near optimal

single-track gray codes.” IEEE Transactions on

Information Theory (IT), vol. 42, no. 3, pp. 779–

789, 1996.

[8] A. A. Evdokimov, “Maximal length of circuit in a

unitary n-dimensional cube,” Mathematical Notes,

vol. 6, no. 3, pp. 642–648, 1969.

[9] L. Glass, “Combinatorial aspects of dynamics

in biological systems,” in Statistical mechanics

and statistical methods in theory and applications,

U. Landman, Ed.Plenum, 1977, pp. 585–611.

[10] J. Gu, P. Purdom, J. Franco, and B. Wah, “Al-

gorithms for the satisfiability (SAT) problem: a

survey,” DIMACS, pp. 19–152, 1997.

[11] F. Harary, J. P. Hayes, and H.-J. Wu, “ A survey of

the theory of hypercube graphs,” Comput. Math.

Appl., vol. 15, pp. 277–289, 1988.

[12] L.Haryanto,

codes and families of such codes covering the hy-

percube,” Ph.D. dissertation, Delft University of

Technology, January 2007.

“Constructing Snake-in-the-box

[13] A. P. Hiltgen and K. G. Paterson, “Single-track

circuit codes,” IEEE Transactions on Information

Theory (IT), vol. 47, no. 6, pp. 2587–2595, 2001.

[14] W. H. Kautz, “Unit distance error checking

codes,” IRE Trans. on Electronic Computers,

vol. 7, pp. 179–180, 1958.

[15] V. Klee, “A method for constructing circuit

codes,” J. ACM, vol. 14, no. 3, pp. 520–528, 1967.

[16] ——, “The use of circuit codes in analog-to-digital

conversion,” Graph Theory and its Applications,

pp. 121–132, 1970.

[17] K. G. Paterson and J. Tuliani, “Some new circuit

codes.” IEEE Transactions on Information The-

ory (IT), vol. 44, no. 3, pp. 1305–1309, 1998.

1237

Page 4

Table 1: New Circuit codes

n

10

11

12

L

86

68

98

δ

3

4

4

Transition sequence

461328167a3869a3495769a4629832173497531a5761547852793487591a754a615a8624973526758a1687

546138b5497b6521b79a5b14985b279543268597b213879245a836b29a6872a4b629

18523b95a641b89a175321bc4238b5928cba91c83b59c16b748391ca596342b53a6174a381ba58cb23891a

b237c4a82bca

8ca57968a473b91a547218ca5b143a79b463ca54732c8ab31279a342

1b264a98b56da39715bc49a18b246ad5b1c4a987d3cba86d5c

85d473c2b517e94362d795384c719da46731b542961d8349e157c843b615294bac

d4c61b9de3f7bad436785ab23e18579a4d8f295e48c19da68eb94236df1e9267ba5d293f4bde8369a2

gdbc146g7dec234gdfebc157af48637ce519b7dcgf94a52b713f92db63afc827b1ef8379b6f48dgbe6c9854

be1c7ag452b368f5492

b951g74e2b81d5c4ef238a147d92561efb48276gf31d47a26935d7efac4b9de8a2713d6e

472h8g1f37e2dg5h94c762eaf8cg697bh3fgdc4273659e1fdag259b73fha5e968b274ce9d561fc7ae981gc5b69

fc4g7e813f9h5gca7bfehd4g56738eh9fcg4213e5fb94dg7ehf38bac9ge5682b

12

13

14

15

16

56

50

66

82

106

5

6

6

6

6

16

17

17

72

90

64

7

7

8

[18] F. Preparata and J. Nievergelt,

preserving codes,” IEEE Transactions on Infor-

mation Theory (IT), vol. 20, no. 5, pp. 643–649,

September 1974.

“Difference-

[19] D. Rajan and A. Shende, “Maximal and re-

versible snakes in hypercubes,” in 24th Annual

Australasian Conference on Combinatorial Math-

ematics and Combinatorial Computing, 1999.

[20] K. G. Scheidt, “Searching for patterns of snakes

in hypercubes,” Proceedings of the second annual

CCSC on Computing in Small Colleges Northwest-

ern conference, pp. 168–176, 2000.

[21] R. C. Singleton, “Generalized snake-in-the-box

codes,” IEEE Transactions on Electronic Comput-

ers, vol. EC-15, no. 4, pp. 596–602, 1966.

[22] D. Tuohy, W. Potter, and D. Casella, “Searching

for Snake-in-the-Box codes with evolved pruning

models,” in Int. Conf. on Genetic and Evolution-

ary Methods, 2007, pp. 3–9.

[23] A. D. Wyner, “Note on circuits and chains of

spread k in the n-cube,” IEEE Trans. Comput.,

vol. C-20, no. 4, pp. 474–474, April 1971.

[24] I. Zinovik, D. Kroening, and Y. Chebiryak, “Com-

puting binary combinatorial Gray codes via ex-

haustive search with SAT-solvers,” IEEE Trans-

actions on Information Theory, vol. 54, no. 4, pp.

1819–1823, April 2008.

1238