Further Results on Finite-Length Analysis of
BATS Codes
Shenghao Yang
∗
and Raymond W. Yeung
†
∗
School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen
†
Institute of Network Coding and Dept. of Information Engineering, The Chinese University of Hong Kong
Abstract—BATS codes were proposed for communication
through networks with packet loss. A BATS code consists of
an outer code and an inner code. The outer code is a matrix
generalization of a fountain code, which works with the inner
code that comprises random linear coding at the intermediate
network nodes. In this paper, we present some new results on
the finite-length performance of BATS codes with respect to
both belief propagation (BP) decoding and inactivation decoding.
Our results reveal explicitly how the number of batches used
for decoding affect the decoding performance for both BP
decoding and inactivation decoding. We further derive i) the
error exponent of BP decoding, ii) a finite summation expression
of the expected number of batches consumed by BP decoding, and
iii) the asymptotic behavior of the number of inactive symbols
required when the number of batches used for decoding goes to
infinity.
I. INTRODUCTION
Proposed for communication through networks with packet
loss, a BATS code consists of an outer code and an inner
code [1], [2]. As a matrix generalization of fountain codes,
the outer code generates potentially unlimited number of
batches, each of which consists of 𝑀 coded symbols. The
inner code comprises (random) linear network coding [3]–
[5] at the intermediate network nodes, performed on symbols
belonging to the same batch. When 𝑀 =1, the outer code
reduces to that of an LT code (or Raptor code if precode
is applied), and network coding of the batches becomes to
forwarding. BATS codes preserve salient features of fountain
codes, in particular, their ratelessness property and low en-
coding/decoding complexity. Compared with ordinary random
linear network coding schemes [6]–[8], BATS codes not only
have lower encoding/decoding complexity, but also smaller
coefficient vector overhead and intermediate node caching
requirement.
The asymptotic performance of BATS codes with belief
propagation (BP) decoding has been analyzed in [2]. A suffi-
cient condition for the BP decoder to recover a given fraction
of the input symbols with high probability was obtained. The
asymptotic analysis helps us to design BATS codes with good
performance for a large number of input symbols (e.g., tens
of thousands). However, for codes with a relatively small
number of input symbols, the error bound obtained in the
asymptotic analysis is rather loose (if valid), and the degree
distribution optimized asymptotically does not always give a
good performance.
The finite-length performance of BATS codes with respect
to both belief propagation (BP) decoding and inactivation
decoding has been studied in [9]. For a fixed number of input
symbols and a fixed number of batches used for decoding,
recursive formulae were obtained to calculate i) the probability
distribution of the BP decoding stopping time, and ii) the ex-
pected number of inactive symbols for inactivation decoding.
In this paper, we demonstrate that finite-length analysis can
also help to derive some explicit analytical results, which
provides insights on the decoding performance of both BP
decoding and inactivation decoding. Specifically, we find that
both the probability that BP decoding stops at time 𝑡 and the
probability that an input symbol is inactivated at time 𝑡 in
inactivation decoding can be expressed exactly as the power-
sum form
∑
2
𝑡
−1
𝑖=0
𝑐
𝑖
𝑒
𝑛
𝑖
, where 𝑛 is the number of batches
used for decoding, 𝑐
𝑖
is a function of the number of input
symbols and 𝑡, and 0 ≤ 𝑒
𝑖
≤ 1 is a function of the number
of input symbols, the degree distribution, the transfer matrix
rank distribution and 𝑡. Note that i) both 𝑒
𝑖
and 𝑐
𝑖
do not
depend on 𝑛, ii) for both decoders 𝑒
𝑖
are the same and
only the coefficients 𝑐
𝑖
are different. This expression reveals
clearly how the decoding failure probability (for BP decoding)
and the expected number of inactivations (for inactivation
decoding) decrease with the number of batches. Based on the
power-sum formulae, we further derive i) the error exponent
of BP decoding, ii) a finite summation expression of the
expected number of batches consumed by BP decoding, and
iii) the asymptotic behavior of the number of inactive symbols
required when the number of batches used for decoding goes
to infinity.
LT codes are BATS codes with unit batch size. Our results
also provide new analytical results for LT codes since, as far
as we know, there are no corresponding results for LT codes
available in the literature.
II. P
RELIMINARIES
In this paper, we use 0 as the origin of the index for vectors
and matrices. For an 𝑚 × 𝑛 matrix A, we denote by rk(A) its
rank and by A[𝑖
1
:𝑖
2
,𝑗
1
:𝑗
2
] (0 ≤ 𝑖
1
≤ 𝑖
2
≤ 𝑚 − 1, 0 ≤ 𝑗
1
≤
𝑗
2
≤ 𝑛−1) the submatrix of A formed by the entries between
the 𝑖
1
-th and 𝑖
2
-th rows and the 𝑗
1
-th to 𝑗
2
-th columns. We
also write A[𝑖, 𝑗
1
:𝑗
2
]=A[𝑖:𝑖, 𝑗
1
:𝑗
2
], A[𝑖, 𝑗:] = A[𝑖:𝑖, 𝑗:𝑛−1],
etc. For real numbers 𝑥 and 𝑦, denote their minimum and
maximum by 𝑥 ∧ 𝑦 and 𝑥 ∨ 𝑦, respectively.
978-1-5090-2482-7/16/$31.00 ©2016 IEEE