Near-Ideal M-ary LDGM Quantization with. Recovery. Qingchuan Wang, Student Member, IEEE, Chen He, Member, IEEE, and Lingge Jiang, Member, IEEE.
1830
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 7, JULY 2011
Near-Ideal ๐ -ary LDGM Quantization with Recovery Qingchuan Wang, Student Member, IEEE, Chen He, Member, IEEE, and Lingge Jiang, Member, IEEE AbstractโFor iterative mean-square error (MSE) quantizers with alphabet size ๐ = 2๐พ using low-density generator-matrix (LDGM) code constructions, an efficient recovery algorithm is proposed, which adjusts the priors used in belief propagation (BP) to limit the impact of previous non-ideal decimation steps. Based on an analysis of the BP process under ideal or nonideal decimation, the algorithm first estimates the conditional probability distributions describing the effect of non-ideal decimation, then adjusts the priors to make the distributions match the ideal situation. As shown in simulation results, the recovery algorithm can improve quantization performance greatly, reducing the shaping loss to as low as 0.012 dB, while the increase in computational complexity is modest thanks to the use of FFT techniques. Index TermsโLow-density generator-matrix, quantization, decimation, recovery.
I. I NTRODUCTION
S
PARSE-GRAPH codes have recently found some use in long-block lossy source coding problems due to their potential to achieve near-ideal rate-distortion performance at a lower computational complexity than traditional methods like trellis-coded quantization (TCQ) [1]. Although structured constructions such as polar codes [2] have been shown to be fast and effective in many channel and source coding problems [3], including those involving side information and binning [4], more randomized ones based on low-density generator matrix (LDGM) codes remain attractive due to their more moderate block length requirements, efficient integration with e.g. superposition coding schemes, as well as the availability of well-established optimization methods from low-density parity-check (LDPC) literature. Indeed, being duals to LDPC codes used in channel coding [5], LDGM codes are known to be able to approach the Shannon limit under optimal encoding for the binary symmetric case [6], and with appropriate modulation mappings, for more general sources possibly requiring non-uniform reconstruction alphabets as well [7]; variants with additional parity [8] or Hamming weight [9] constraints have also been proposed to improve finite-degree performance or to allow channel coding and binning. Naturally, most practical encoding algorithms (or quantizers) for such sparse-graph codes employ some form of message passing. Although more elaborate algorithms like survey propagation exist [6], [7], ordinary belief propagation Paper approved by Z. Xiong, the Editor for Distributed Coding and Processing of the IEEE Communications Society. Manuscript received August 5, 2010; revised December 19, 2010. The authors are with the Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, 200240, China (e-mail: {r6144, chenhe, lgjiang}@sjtu.edu.cn). This paper was supported by the National Natural Science Foundation of China Grants No. 60772100, 60832009, and 60872017, as well as National 863 Program Grant No. 2009AA011505. Digital Object Identifier 10.1109/TCOMM.2011.061511.100462
(BP) appears to be sufficient for quantization when a good degree distribution is used [10]. In any case, decimation steps (i.e. hard decisions) usually need to be carried out according to e.g. the BP marginals (extrinsic probabilities) in order to make BP converge to one codeword among many similarly good ones; other positive feedback mechanisms used in [11] and [12] serve essentially the same purpose. For randomized LDGM constructions, loops in the factor graph and the limited number of iterations causes the extrinsic probabilities from BP to be approximate; this leads to non-ideal decimation choices, which adversely affects future iterations and makes the performance of such BP algorithms difficult to predict theoretically. Our works mainly focus on a specific lossy source coding problem, mean-square error (MSE) quantization of Euclidean space [13, Sec. II-C], which plays an important role in highrate source coding, as well as various channel coding schemes on Gaussian channels, such as the shaping component of dirty paper coding [14]. As shown in previous works [15] and [16], an LDGM-based construction using BP and decimation for encoding can approach the Shannon limit of this problem quite well, and methods for optimizing the degree distribution, pace of decimation, etc. have been proposed. The non-ideal decimation problem above also exists here; in [16, Sec. VI-C], a recovery step run before each BP iteration to adjust the priors has been found to reduce its impact and improve quantization performance significantly. However, the recovery algorithm in [16] is somewhat ad hoc and poorly understood, and it is only applicable to binary constructions, whose alphabet limitation leaves a significant gap to the Shannon limit. In this paper, after giving an overview of the quantization algorithm in Section II, in Section III we will present some theoretical arguments in an attempt to gain a better understanding of non-ideal decimation and how recovery might be performed. Based on this analysis, a recovery algorithm will be designed in Section IV for LDGM-based MSE quantization with ๐ -ary alphabet, whose computational complexity turns out to be quite modest. The simulation results in Section V demonstrate that recovery can improve quantization performance greatly and approach the Shannon limit to within 0.012 dB. Finally, Section VI concludes the paper. A. Notations and Conventions Notations are similar to those in [16]. Bold letters denote sequences, vectors or matrices whose elements are indicated by subscripts, e.g. ๐ = (๐ฆ1 , . . . , ๐ฆ๐ ); conversely, a vector or matrix can also be defined element-wise, e.g. (๐โj2๐๐๐ฃ/๐ )๐๐ฃ is the DFT (discrete Fourier transform) matrix. (โ
)T and (โ
)H denotes matrix transposition and Hermitian transposition, โฅโ
โฅ is the Euclidean norm, 1[โ
] is 1 if the condition is true and 0 otherwise, โ and โ denote addition and
c 2011 IEEE 0090-6778/11$25.00 โ
WANG et al.: NEAR-IDEAL ๐-ARY LDGM QUANTIZATION WITH RECOVERY
subtraction modulo-2 or modulo-๐ (should be clear from context), (โ
)โ is the modulo-๐ operation on a real number into โ โ [โ๐/2, ๐/2), and (โ
)โ ๐ is the element-wise modulooperation into โ ๐ . In the BP algorithm, symbols like ๐b๐ , b ๐bc ๐๐ , ๐๐ are binary or ๐ -ary probability tuples representing the priors, messages and extrinsic probabilities in BP; each ๐ -ary probability tuple ๐ is a tuple of ๐ real numbers representing a probability distribution over {0, . . . , ๐ โ 1}, with each component denoted by ๐(๐ข), ๐ข = 0, . . . , ๐ โ1. For conciseness, all probability tuples are implicitly normalized; that is, when we define an ๐ -ary probability tuple ๐ by writing ๐(๐ข) = ๐๐ข , ๐ข = 0, . . . , ๐ โ 1, it actually means that ๐(๐ข) = ๐๐ข /(๐0 + โ
โ
โ
+ ๐๐โ1 ), and later mentions of ๐(๐ข) refer to these normalized components. For ๐ข = 0, . . . , ๐ โ 1, ๐ข denotes the โsure-๐ขโ probability tuple with ๐ข(๐ข) = 1 and all 1 1 other components being zero, while โ = ( ๐ ,..., ๐ ) denotes the โunknownโ probability tuple. For ๐ -ary probability tuples ๐โฒ and ๐โฒโฒ , ๐โฒ โ ๐โฒโฒ and ๐โฒ โ ๐โฒโฒ are also ๐ -ary probability tuples (๐โฒ โ ๐โฒโฒ )(๐ข) โ ๐โฒ (๐ข)๐โฒโฒ (๐ข) and (๐โฒ โ ๐โฒโฒ )(๐ข) โ โ๐โ1with โฒ โฒ โฒโฒ โฒ ๐ขโฒ =0 ๐ (๐ข )๐ (๐ข โ ๐ข ), ๐ข = 0, . . . , ๐ โ 1, similar to the variable-node and check-node operations in LDPC literature. The informativeness of an ๐ -ary probability tuple ๐ is measured in bits according to ๐ผ(๐) โ log ๐ +
๐โ1 โ
๐(๐ข) log ๐(๐ข);
(1)
๐ข=0
for example, among binary probability tuples, 0 and 1 are the most informative and โ is the least. Often we will have a random ๐ -ary probability tuple ๐ and a random variable ๐ขโ โ {0, . . . , ๐ โ 1}, where ๐(๐ขโ ) is uniform and ๐(๐ โฃ ๐ขโ ) satisfies, for any deterministic ๐โฒ , ๐(๐ = ๐โฒ โฃ ๐ขโ = 0) = ๐(๐ = ๐โฒ โ ๐ข โฃ ๐ขโ = ๐ข), ๐ข = 0, . . . , ๐ โ 1, (2) due to the dithering performed below. ๐(๐ โฃ ๐ขโ ) can then be fully characterized by ๐(๐ โฃ ๐ขโ = 0), and we extend similar notions in LDPC analysis [17] and call the latter the density of ๐ with respect to ๐ขโ , and say ๐ has a symmetric density (or simply is symmetric) w.r.t. ๐ขโ if, for any deterministic ๐โฒ , ๐(๐ขโ = ๐ข โฃ ๐ = ๐โฒ ) = ๐โฒ (๐ข),
๐ข = 0, . . . , ๐ โ 1.
(3)
Eq. (3) implies that the mutual information ] [๐โ1 โ โ โ โ ๐(๐ข โฃ ๐) log ๐(๐ข โฃ ๐) ๐ผ(๐ข ; ๐) = log ๐ + E๐
(4)
๐ขโ =0
is equal to E [๐ผ(๐)]; if E [๐ผ(๐)] > ๐ผ(๐ขโ ; ๐), ๐ is then said to be over-confident w.r.t. ๐ขโ . II. OVERVIEW OF THE Q UANTIZATION P ROBLEM AND A LGORITHM We consider MSE quantization with reconstruction alphabet size ๐ โ 2๐พ , which is equivalent to the follows [16]: given ๐ , rate ๐
and block length ๐, design a codebook ๐ฐ โ {0, . . . , ๐ โ 1}๐ with โฃ๐ฐโฃ = 2๐b and ๐b โ ๐๐
, as well as a quantization algorithm that quantizes a source sequence ๐ ห uniformly distributed in [0, ๐ )๐ into a ๐ โ ๐ฐ, such that the modulo-๐ MSE ๐ 2 โ E [๐(ห ๐ , ๐)] with
1831
ฮผcu sj ฮผbc is
c1
u1
c2 b1
c3
b2
u2
c4
c2nโ1
bnb ฮผcb si
c2n
un
ฮผuc js Fig. 1. The factor graph of the LDGM quantizer when ๐ = 4. Since the scrambling sequence ๐ has already been fixed when the quantization algorithm is run, the corresponding variable nodes can be omitted. 2
๐(ห ๐ , ๐) โ ๐1 โฅ(ห ๐ โ ๐)โ ๐ โฅ is minimized. As ๐ โ โ, ๐โ2 (๐
) โ (2๐๐(2๐
/๐ )2 )โ1 is a lower bound of ๐ 2 that becomes tight as ๐ increases (see [13] and [16, Sec. II]), so we define 10 log10 (๐ 2 /๐โ2 (๐
)) as the shaping loss in decibels. The ๐ -ary LDGM codebook is constructed in a way similar to [16, Sec. VII]. We let ๐ฎ be an ๐b ร ๐c (๐c โ ๐พ๐) binary low-density generator matrix from a suitably optimized LDGM code ensemble, ๐น โ {0, . . . , ๐ โ 1}๐ and ๐ โ {0, 1}๐พ๐ be predetermined i.i.d. uniform dithering and scrambling sequences known at both the encoder and the decoder. For each ๐ โ {0, 1}๐b , we divide ๐ โ ๐(๐, ๐) โ ๐๐ฎ โ ๐ into ๐ sub-sequences ๐ห๐ โ (๐๐1 , . . . , ๐๐๐พ ) with ๐๐ โ ๐พ(๐ โ 1) + ๐, ๐ = 1, . . . , ๐, and map each resulting ๐ห๐ into one ๐ -ary ๐๐ ) with ๐๐ (โ
) โ ๐(โ
) โ ๐ฟ๐ being the symbol ๐ข๐ = ๐๐ (ห dithered version of Gray mapping ๐(โ
); this yields a codeword ๐ โ ๐(๐, ๐), and all 2๐b codewords from the given ๐ฎ, ๐น and ๐ form the codebook ๐ฐ. Fig. 1 shows the factor graph [18] describing the code; as in [16], the variable nodes for all ๐๐ โs have the same right-degree, denoted by ๐b . The BP messages used in the algorithm, denoted by e.g. ๐bc ๐๐ , are also depicted in the figure; the two subscripts are the indices of its source and destination nodes. Given parameter ๐ก > 0, the quantization algorithm carries out BP with a priori probabilities ห u (๐ข) = ๐โ๐ก(ห๐ฆ๐ โ๐ข)2โ /๐(ห ๐ ๐ฆ๐ ), ๐ข = 0, . . . , ๐ โ 1 (5) ๐ โ๐โ1 โ๐ก(ห๐ฆ๐ โ๐ข)2 โ is the nor๐ฆ๐ ) โ for each ๐ข๐ , where ๐(ห ๐ข=0 ๐ malization factor that will henceforth be omitted as noted in the conventions, and ๐ ห is an adjusted version of the source sequence ๐ ห using the proposed recovery algorithm. To make BP converge, a number of decimation steps are performed in each iteration, each of which fixes a certain ๐๐โ to a hard decision ๐โ โ {0, 1} by setting its prior ๐b๐โ to the sure message ๐โ . The choice of ๐โ and ๐โ in each decimation step is made according to the extrinsic probabilities ๐๐b of the ๐๐ โs from BP. Specifically, we can use either the greedy decimator (GD), which chooses ๐โ and ๐โ among the undecimated positions such that ๐๐bโ (๐โ ) is maximized, or the typical (probabilistic) decimator (TD), which chooses an undecimated bit index ๐โ
1832
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 7, JULY 2011
randomly and then ๐โ โ {0, 1} with probability proportional to ๐๐bโ (๐โ ); while GD performs better in practice, TD is more amenable to analysis and thus used there. For conciseness, the quantization algorithms using TD/GD with recovery will henceforth be called TD-R and GD-R, respectively, while if recovery is not done (i.e. with ๐ห made the same as ๐ห) they are called plain TD or GD. The number of decimation steps performed in each iteration, โฉ known โช as the pace of decimation, is controlled via ๐ผbc โ ๐ผ(๐bc ) where โจโ
โฉ means averaging ๐๐ over the subscripts ๐ and ๐ . Based on the analysis in [16, Sec. VI-D], in each iteration we simply carry out decimation until ๐ผbc increases by at least ๐b โ2 2(๐b โ 1) ๐๐ผbc = (1 โ ๐ผbc ) 2(๐b โ1) ๐๐ ๐ฟ0 ๐b
(6)
since the last iteration, which ensures that the algorithm finishes within ๐ฟ0 iterations. When the block length ๐ is small, the actual iteration count ๐ฟ may be significantly smaller than ๐ฟ0 ; a โthrottlingโ mechanism has been used in [16] to reduce the shaping loss in this case at the cost of higher ๐ฟ, but since it does not improve the performance-speed tradeoff significantly, it is not adopted here for simplicity. The entire quantization algorithm used in this paper can now be outlined in Fig. 2, which is largely the same as that in [16, Sec. VII]. The main difference is the introduction of the recovery algorithm at the beginning of each iteration, which is the focus of this paper. As the recovery algorithm ห u โs must be updated recomputes ๐ห in every iteration, the ๐ ๐ accordingly, rather than being computed once before BP; to reduce the computational cost, (5) is pre-evaluated for different ๐ฆห๐ at step size 1/32 and then approximated via linear interpolation, which turns out to have negligible performance impact. Another difference is that the ๐cu ๐ ๐ โs are now updated after the decimation steps, unlike in [16] where they were cu updated at the same time as the ๐cb ๐ ๐ โs. This allows the ๐๐ ๐ โs to include the information in the ๐bc ๐๐ โs from the current iteration, thus making the ๐๐u โs used in recovery in the next iteration more up-to-date; the importance of this will be explained in Section III-D. Finally, the introduction of ๐น and ๐, which is necessary in analysis, also leads to some minor changes. III. T HEORETICAL A RGUMENTS FOR R ECOVERY A. Analysis Framework for the Decimation Process Decimation in the LDGM quantization algorithm is both a significant obstacle in the algorithmโs analysis and the main reason for the use of recovery. The analysis in [16] was for TD and under ideal decimation in the sense defined below, which was sufficient for the purpose of degree distribution optimization. On the other hand, when designing a recovery algorithm, a good understanding of the behavior under nonideal decimation is necessary, so we will attempt below to extend the ideas in [16] to model this. To avoid complications, we assume without loss of generality that the bits in ๐ are decimated by TD sequentially rather than in a random order, and the number of decimation steps carried out in each iteration is likewise assumed to be deterministic over the ensemble defined below, unaffected by statistical fluctuations in ๐ผbc .
Input: Quantizer parameters ๐ฎ, ๐น, ๐, ๐ก, source sequence ๐ ห Output: Quantized codeword ๐ = ๐(๐, ๐) labeled by ๐ cu ๐bc ๐๐๐ โ โ, ๐๐๐ ๐ โ โ, ๐ = 1, . . . , ๐b , ๐ = 1, . . . , ๐, ๐ = 1, . . . , ๐พ ๐b๐ โ โ, ๐ = 1, . . . , ๐b โฐ โ {1, 2, . . . , ๐b } {the set of bits not yet decimated} repeat {beliefโpropagation iteration} cu ๐ ), ๐ = 1, . . . , ๐, ๐ข = ๐ (ห ๐๐u (๐ข) โ ๐พ ๐ ๐) = 0, . . . , ๐ โ 1 ๐ ๐=1 ๐๐๐ ๐ (ห Compute ๐ ห from ๐ ห and the ๐๐u โs using the recovery algorithm ห u (๐ข) โ ๐โ๐ก(๐ฆห๐ โ๐ข)2โ , ๐ = 1, . . . , ๐, ๐ข = 0, . . . , ๐ โ 1 ๐ ๐ for ๐ = 1 to ๐ โ and ๐ = 1 to ๐พ do โ หu ๐uc ๐)) ๐โฒ โ=๐ ๐cu ๐๐โฒ ), ๐ = 0, 1 ๐ ห:ห ๐๐ =๐ ๐๐ (๐๐ (ห ๐๐๐ (๐) โ ๐๐โฒ ๐ (ห end for for ๐ = ๐๐ = 1 to ๐c do( ) uc bc cb ๐cb bc โ{๐} ๐๐โฒ ๐ , ๐ โ ๐ฉ๐ โ
๐ ๐ โ (๐๐๐ โ ๐๐ ) โ โ๐โฒ โ๐ฉโ
๐ end for for ๐ = 1 to ๐b (do ) b cb bc ๐bc ๐๐ โ ๐๐ โ โ๐ โฒ โ๐ฉ cb โ{๐ } ๐๐ โฒ ๐ , ๐ โ ๐ฉ๐โ
โ
๐
๐๐b โ โ๐ โฒ โ๐ฉ cb ๐cb โฒ โ
๐ ๐ ๐ end for while โฐ โ= โ
and more decimation is necessary in this iteration do Choose the bit index ๐โ to decimate and its value ๐โ bc โ โ ๐b๐โ โ ๐โ , ๐bc ๐โ ๐ โ ๐ , ๐ โ ๐ฉ๐โ โ
{decimate ๐๐ to ๐ } โฐ โ โฐโ{๐โ } end while ( ) bc ๐cu bc ๐๐โฒ ๐ , ๐ = ๐๐ = 1, . . . , ๐c ๐ ๐ โ ๐๐ โ โ๐โฒ โ๐ฉโ
๐ until โฐ = โ
๐๐ โ 0 (resp. 1) if ๐b๐ = 0 (or 1), ๐ = 1, . . . , ๐b
Fig. 2. A brief outline of the quantization algorithm. ๐ฉ๐โ
bc = ๐ฉโ
๐cb is the cb = ๐ฉ bc is the positions of 1โs in the ๐-th row of ๐ฎ, and similarly ๐ฉ๐ โ
โ
๐ positions of 1โs in the ๐ -th column of ๐ฎ.
The analysis uses a fully-adjusted sequence ๐, obtained from an idealized recovery algorithm described in Secห tion III-C, in place of ๐ ห in (5) to define the priors ๐u๐ โs, while ๐ u ห and the ๐๐ โs used in practice are regarded as approximations. The ๐๐b used in each TD decimation step then approximates the corresponding โtrueโ extrinsic probabilities โ โ ๐u๐ (๐ข๐ (๐โฒ , ๐โฒ )) ๐๐bโ (๐) โ (๐โฒ ,๐โฒ )โ๐ ๐ ๐โฒ๐ =๐
=
โ
โฒ
โฒ
๐โ๐๐ก๐(๐,๐(๐ ,๐ )) ,
(7)
(๐โฒ ,๐โฒ )โ๐ ๐โฒ๐ =๐
where ๐ is defined as the set of (๐โฒ , ๐โฒ ) consistent with previous decimation choices, meaning here that ๐โฒ is equal to the scrambling sequence ๐, while each ๐โฒ๐โฒ is equal to ๐โ โ {0, 1} if ๐๐โฒ has previously been decimated to ๐โ (i.e. ๐b๐โฒ = ๐โ ), but can be either 0 or 1 if ๐๐โฒ has not been decimated yet (i.e. ๐b๐โฒ = โ). The i.i.d. choice of each ๐๐ in the scrambling sequence ๐, which determines the codebook ๐ฐ, can likewise be viewed as a TD decimation step using โ as the extrinsic probabilities. Assuming for convenience that ๐1 , ๐2 , . . . , ๐๐c are chosen sequentially, the corresponding true extrinsic probabilities for the choice of ๐๐ can similarly be given by โ โ ๐u๐ (๐ข๐ (๐โฒ , ๐โฒ )), (8) ๐๐ aโ (๐) โ (๐โฒ ,๐โฒ )โ๐ ๐ ๐โฒ๐ =๐
where ๐ is still the set of (๐โฒ , ๐โฒ ) consistent with previous decimation choices, i.e. ๐โฒ๐ โฒ = ๐๐ โฒ for ๐ โฒ < ๐ , while the other bits in ๐โฒ as well as the entire ๐โฒ are arbitrary. We can thus define the true typical decimator (TTD) as an idealized version
WANG et al.: NEAR-IDEAL ๐-ARY LDGM QUANTIZATION WITH RECOVERY
1833
of TD that uses respectively ๐๐ aโ and ๐๐bโ in the choice of each ๐๐ and ๐๐ , and say ideal decimation has been performed if all the bits in ๐ and ๐ decimated so far have used the TTD.1 Using TTD in all decimation steps yields each ๐ โ {0, . . . , ๐ โ 1}๐ with probability proportional to ๐โ๐๐ก๐(๐,๐) , giving
with recovery, and say a certain ๐(โ
) is ideal if it matches the corresponding ๐ยฏ(โ
). Now the reference codeword is simply the result of TTD after all decimation steps, so
โซ
๐โ1 โ
โ๐ก(๐ฆโ๐ข)2โ
๐ 1 ๐ E [๐(๐, ๐)] = ๐๐ก2 โ ๐๐ฆ (๐ฆ โ ๐ข)2โ โ
๐ 0 ๐(๐ฆ) ๐ข=0 โซ ๐/2 2 ๐โ๐ก๐ง ๐๐ง = ๐ง2 โ
๐(๐ง) โ๐/2
โ 2
๐ยฏ(๐ขโ๐ โฃ ๐ฆ๐ ) = ๐ยฏ(๐ขโ๐ โฃ ๐) = ๐โ๐ก(๐ฆ๐ โ๐ข๐ )โ /๐(๐ฆ๐ ),
which implies ๐ยฏ(๐ขโ๐ โฃ ๐u๐ ) = ๐u๐ (๐ขโ๐ ), i.e. ๐u๐ has a symmetric density w.r.t. ๐ขโ๐ . As ๐ยฏ(๐ฆ๐ ) is uniform, this density can be computed from โ 2
๐ยฏ(๐ฆ๐ โฃ ๐ขโ๐ ) = ๐โ๐ก(๐ฆ๐ โ๐ข๐ )โ /๐(๐ฆ๐ ).
(9)
(denoted by ๐๐ก in [16]); for example, if ๐ = ๐ห, then the achieved MSE is simply ๐๐ก2 . TD can yield the same result if each ๐๐ aโ can be made equal to โ and each ๐๐bโ equal to the ๐๐b from BP. Similar approaches have been used in [19] and [3] to analyze the decimation process in satisfiability problems and quantization with polar codes. We now consider an ensemble of quantizer instances with a probability measure over it, so that probabilistic analytical approaches including density evolution (DE) can be carried out. Recall that in the DE analysis of the LDPC decoder [17], the probabilities are defined over the possible channel realizations and the LDPC code ensemble with a specific degree distribution, and DE is performed in reference to the transmitted codeword (usually assumed to be all-zero). In contrast, when analyzing the LDGM quantizer with TD or TTD, each quantizer instance in the ensemble has not only a specific ๐ห, a ๐ฎ in the LDGM code ensemble, a dither ๐น but also a specific sequence of values used as the random source for decimation of both ๐ and ๐; for example, for each ๐๐ the random source can be denoted by an i.i.d. uniform variable ๐๐b over [0, 1), such that ๐๐ is decimated to 1[๐๐b โฅ ๐ห๐b (0)], ๐ห๐b being ๐๐b for TD and ๐๐bโ for TTD. Over this ensemble, ๐ as well as all the priors, BP messages, etc. are now random variables whose distributions depend on whether TD or TTD (ideal decimation) is used, and whether recovery is performed. The reference codeword used by DE, denoted by (๐โ , ๐โ ) or the corresponding ๐โ โ ๐(๐โ , ๐โ ), is defined as the quantization result that would be obtained if previous decimation choices were kept and the remaining decimation steps used TTD. Therefore, (๐โ , ๐โ ) is random (specific to each quantizer instance); given ๐ and ๐, it is any (๐โฒ , ๐โฒ ) โ ๐ โฒ โฒ with probability proportional to ๐โ๐๐ก๐(๐,๐(๐ ,๐ )) . Over the entire ensemble, by definition it remains unchanged after a decimation step using TTD, but is otherwise also specific to each decimation step. B. Analysis of Ideal Decimation The analysis of a decimation step under ideal decimation (in previous steps), as was done in the DE analysis for degree distribution optimization [16, Sec. V], can be expressed in the above framework as follows. As recovery is not necessary in this case, ๐ is identical to ๐ ห. For clarity, we use ๐ยฏ(โ
) to denote a pdf under ideal decimation, ๐(โ
) when TD is used 1 Note that the definition of TTD here also encompasses the concept of true typical source generator (TTSG) introduced in [16, Sec. V-A], by viewing the choice of bits in ๐ as decimation steps as well.
(10)
(11)
For asymptotically large ๐, DE can then be carried out to obtain the densities of the ๐๐ a โs and ๐๐b โs after any given number of BP iterations with different initial symmetric densities of BP messages. The corresponding ๐๐ aโ and ๐๐bโ are bound by these densities in terms of physical degradation, making it possible to evaluate how close each ๐๐ aโ matches โ and each ๐๐bโ matches ๐๐b . For given degree distributions, this DE analysis yields a monotonicity threshold ๐กthr such that, as long as ๐ก โค ๐กthr , when the iteration count ๐ฟ goes to infinity along with ๐ (i.e. decimation is performed slowly), the mean-square difference between each ๐๐ aโ , ๐๐bโ and respectively โ, ๐๐b goes to zero. In other words, assuming ideal decimation in previous steps, when the current decimation step is carried out using TD after a sufficiently large number of BP iterations, it will yield the same result as the TTD with high probability and thus be ideal as well. This MSE ๐๐ก2thr potentially achievable by TD, as defined in (9), can be very close to the theoretical limit ๐โ2 (๐
) when using degree distributions optimized for a high ๐กthr . The extrinsic probabilities ๐๐u of each ๐ข๐ , given by ๐๐u (๐ข) โ
๐พ โ
๐cu ๐๐ ), ๐๐ ๐ (ห
๐ข = ๐๐ (ห ๐) = 0, . . . , ๐ โ 1, (12)
๐=1
are used by the proposed recovery algorithm as a summarization of the BP messages. The corresponding true extrinsic probabilities can be defined as โ โ ๐u๐ โฒ (๐ข๐ โฒ (๐โฒ , ๐โฒ )). (13) ๐๐uโ (๐ข) โ (๐โฒ ,๐โฒ )โ๐ ๐ โฒ โ=๐ ๐ข๐ (๐โฒ ,๐โฒ )=๐ข
As the ๐u๐ โฒ โs have symmetric densities, it is easy to prove that each ๐๐uโ is symmetric w.r.t. ๐ขโ๐ as well, i.e. ๐ยฏ(๐ขโ๐ โฃ ๐๐uโ ) = ๐๐uโ (๐ขโ๐ ); ๐ฆ๐ (or ๐u๐ ) โ ๐ขโ๐ โ ๐๐uโ can also be shown to form a Markov chain, as ๐๐uโ only depends on the ๐ฆ๐ โฒ โs with ๐ โฒ โ= ๐. Using (11), we thus have ๐ยฏ(๐ฆ๐ โฃ ๐๐uโ ) = =
๐โ1 โ ๐ข=0 ๐โ1 โ
๐ยฏ(๐ขโ๐ = ๐ข โฃ ๐๐uโ ) โ
๐ยฏ(๐ฆ๐ โฃ ๐ขโ๐ = ๐ข) (14) ๐๐uโ (๐ข)
โ
๐ยฏ(๐ฆ๐ โฃ ๐ขโ๐
= ๐ข).
๐ข=0
Since the neighborhood in the factor graph involved in computing ๐๐u is loop-free with high probability as ๐ โ โ, for asymptotically large ๐, ๐๐u can be shown to be symmetric w.r.t. ๐ขโ๐ and ๐ฆ๐ โ ๐ขโ๐ โ ๐๐u is a Markov chain as well, so (14) remains true with ๐๐uโ replaced by ๐๐u .
1834
C. Decimation Errors and Idealized Recovery Plain TD gives poor quantization performance in practice, and its reason can be understood as follows. The finite ๐ and ๐ฟ used in practice leaves a small but finite mean-square difference between ๐๐bโ and ๐๐ aโ and respectively ๐๐b and โ, so each TD decimation step has a finite probability to give a result different from that of TTD, which we call decimation errors for convenience. Without recovery (๐ = ๐ ห), such erroneous decimations will usually not favor codewords close to ๐ห as much as the TTD, and after ๐ shrinks due to decimation, the subsequent reference codeword will, on average, have a larger distance to ๐ ห. The quantization error (ห ๐ โ ๐โ )โ ๐ is analogous to the noise in the LDPC decoder; its increase in magnitude will change the density of the ๐u๐ โs compared to ideal decimation, usually making them over-confident, which will slow down BP convergence and reduce the informativeness of the ๐๐b โs in future BP iterations, thus worsening the quality of future decimation choices. The recovery algorithm intends to allow BP to recover from decimation errors. It is inspired by the analysis of the similar issue in binary erasure quantization (BEQ) [16, Sec. VI-B], where erroneous decimations are evident as contradictions among the BP messages and the source sequence ๐ห, and can similarly result in less informative ๐๐b โs in the future. In that case, a solution is to flip the contradictory bits in the source sequence ๐ห and use the resulting flipped sequence ๐ ห to generate the ๐u๐ โs for subsequent iterations (also reminiscent of the approach in [12]). These flipped bits intuitively represent errors already made which we do not attempt to fix, and flipping them can be shown to make future ๐๐b โs as informative as if no decimation errors occurred, so fewer decimation errors will be made in the future. The proposed recovery algorithm for ๐ -ary MSE quantization uses a similar idea, which can be better explained by first considering an idealized recovery algorithm that defines the ๐ above and is approximated by the actual one. Whenever ๐ and thus the reference codeword ๐โ changes due to decimation errors, the algorithm adjusts ๐ accordingly to make each ๐(๐ฆ๐ โฃ ๐ขโ๐ ) over the ensemble identical to its ideal version (11). This keeps the density of each ๐u๐ generated from this ๐ฆ๐ , and thus that of each ๐๐b computed from these ๐u๐ โs in subsequent iterations, identical to their counterparts under ideal decimation (which are computable via DE for asymptotically large ๐), so decimation errors no longer degrade the quality of future decimations. When the quantization algorithm finishes after decimating all bits in ๐, ๐ will contain only the resulting (๐, ๐), which will also be the reference codeword then; since ๐(๐ฆ๐ โฃ ๐ขโ๐ ) is ideal, E [๐(๐, ๐)] = E [๐(๐, ๐โ )] is also equal to the ideal ๐๐ก2 , while the difference between ๐ห and ๐ correspond to decimation errors that cause the actual MSE ๐ , ๐)] to be higher. ๐ 2 = E [๐(ห While adjustment of each ๐ฆ๐ causes ๐ขโ๐ to change as well, the desired ๐(๐ฆ๐ โฃ ๐ขโ๐ ) can still be achieved in idealized recovery using ๐๐uโ from (13), which may be approximated by ๐๐u in practice. To see this, recall that given ๐ and ๐, the reference codeword is any (๐โฒ , ๐โฒ ) โ ๐ with probability โฒ โฒ proportional to ๐โ๐๐ก๐(๐,๐(๐ ,๐ )) , so ignoring normalization,
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 7, JULY 2011
๐(๐ขโ๐ = ๐ข โฃ ๐, ๐) = =
โ โฒ
โฒ
(๐ ,๐ )โ๐ ๐ข๐ (๐โฒ ,๐โฒ )=๐ข
โ
โ
โฒ
โฒ
๐โ๐๐ก๐(๐,๐(๐ ,๐ ))
(๐โฒ ,๐โฒ )โ๐ ๐ข๐ (๐โฒ ,๐โฒ )=๐ข
๐u๐ โฒ (๐ข๐ โฒ (๐โฒ , ๐โฒ )) = (๐u๐ โ ๐๐uโ )(๐ข);
(15)
๐โฒ
that is, the dependence on ๐ and ๐ can be subsumed into ๐u๐ (a function of ๐ฆ๐ ) and ๐๐uโ (a function of ๐ and the other elements of ๐). ๐(๐ขโ๐ โฃ ๐ฆ๐ , ๐๐uโ ) is thus also given by (15) and identical to the ideal ๐ยฏ(๐ขโ๐ โฃ ๐ฆ๐ , ๐๐uโ ). Using this fact, we let ๐นยฏ๐๐uโ (โ
) and ๐น๐๐uโ (โ
) be the cdfs corresponding to respectively ๐ยฏ(๐ฆ๐ โฃ ๐๐uโ ) and ๐(ห ๐ฆ๐ โฃ ๐๐uโ ), and define ๐ฆ๐ โ ๐นยฏ๐โ1 ๐ฆ๐ )) uโ (๐น๐ uโ (ห ๐ ๐
(16)
as the fully-adjusted version of ๐ฆห๐ . Conditioned on ๐๐uโ , ๐ฆ๐ ) is uniformly distributed over [0, 1], so the appli๐น๐๐uโ (ห uโ cation of ๐นยฏ๐โ1 ยฏ(๐ฆ๐ โฃ ๐๐uโ ). We then uโ (โ
) makes ๐(๐ฆ๐ โฃ ๐๐ ) = ๐ ๐ have ๐(๐ฆ๐ , ๐ขโ๐ โฃ ๐๐uโ ) = ๐ยฏ(๐ฆ๐ , ๐ขโ๐ โฃ ๐๐uโ ) and ๐(๐ฆ๐ โฃ ๐ขโ๐ , ๐๐uโ ) = ๐ยฏ(๐ฆ๐ โฃ ๐ขโ๐ , ๐๐uโ ) = ๐ยฏ(๐ฆ๐ โฃ ๐ขโ๐ ), the final equality implied by the Markov chain property in Section III-B, so ๐(๐ฆ๐ โฃ ๐ขโ๐ ) now matches ๐ยฏ(๐ฆ๐ โฃ ๐ขโ๐ ), as desired. It should be noted that updating a certain ๐ฆ๐ in this manner changes the corresponding ๐u๐ and thus the ๐๐uโโฒ โs for ๐ โฒ โ= ๐, so ๐ฆ๐ โฒ has to be updated again according to the new ๐(ห ๐ฆ๐ โฒ โฃ ๐๐uโโฒ ) in โ order to maintain the desired ๐(๐ฆ๐ โฒ โฃ ๐ข๐ โฒ ). A rigorous definition of ๐ will thus require an iterative sequential update process, with each step updating one ๐ฆ๐ ; examination of the BEQ case leads us to conjecture that this process will always converge to a fixed point. In any case, such details are no longer relevant when ๐๐uโ is approximated by ๐๐u below. D. Principles of the Proposed Recovery Algorithm The proposed recovery algorithm approximates the idealized version by using ๐๐u instead of ๐๐uโ : let ๐นยฏ๐๐u (โ
) and ๐น๐๐u (โ
) be the cdfs corresponding to respectively ๐ยฏ(๐ฆ๐ โฃ ๐๐u ) and ๐(ห ๐ฆ๐ โฃ ๐๐u ), the algorithm yields, for each ๐, ๐ฆห๐ โ ๐นยฏ๐โ1 ๐ฆ๐ )). u (๐น๐ u (ห ๐ ๐
(17)
While ๐ยฏ(๐ฆ๐ โฃ ๐๐u ) is well approximated by the asymptotic result (14) (with ๐๐u in place of ๐๐uโ ), ๐(ห ๐ฆ๐ โฃ ๐๐u ) must, in practice, be u estimated from the (ห ๐ฆ๐ , ๐๐ ) samples in the actual quantizer instance. Approximating ๐๐uโ with ๐๐u causes the following inaccuracies: u โ ๐๐ is computed with a finite number of BP iterations and ห uโฒ โs from thus fails to include the information in the ๐ ๐ faraway nodes in the factor graph. หuโฒ โs from the ๐ห given by โ The priors used are the ๐ ๐ the actual recovery algorithm, which can be inaccurate compared to the ๐u๐ โฒ โs from idealized recovery due to both estimation errors in ๐(ห ๐ฆ๐ โฃ ๐๐u ) and the incomplete recovery issue discussed below. ห uโฒ โs involved in the computation of the current ๐ u โ The ๐ ๐ ๐ come from recovery steps in earlier iterations, which do not take more recent decimation errors into account. As the result, while (17) makes ๐(ห ๐ฆ๐ โฃ ๐๐u ) identical to ๐ยฏ(๐ฆ๐ โฃ ๐๐u ) in the absence of estimation errors, this does not imply that
WANG et al.: NEAR-IDEAL ๐-ARY LDGM QUANTIZATION WITH RECOVERY
๐(ห ๐ฆ๐ โฃ ๐ขโ๐ ) will match ๐ยฏ(๐ฆ๐ โฃ ๐ขโ๐ ), as is the case for idealized recovery. In general, the effect is that ๐ ห is insufficiently adjusted compared to ๐; for example, when ๐๐u = โ, ๐ยฏ(๐ฆ๐ โฃ ๐๐u ) and ๐(ห ๐ฆ๐ โฃ ๐๐u ) will likely both be uniform distributions, so ๐ฆห๐ will be identical to ๐ฆห๐ even if ๐๐uโ indicates that ๐ฆ๐ should be adjusted to some other value. However, such incomplete recovery does not cause much performance degradation in practice (actually none in case of BEQ). Roughly speaking, if ๐๐u is uninformative at some iteration in an quantizer instance, most of the corresponding cb outgoing BP messages ๐uc ๐๐ and ๐๐ ๐ will be uninformative as well and not much affected by the difference between ๐ฆห๐ and ๐ฆ๐ due to incomplete recovery, while the affected messages are also prevented from propagating far with the use of recovery in future iterations. In effect, recovery is being carried out incrementally, adjusting ๐ฆห๐ toward ๐ฆ๐ as the corresponding ๐๐u becomes informative. Due to the reliance on future recovery to stop the propagation of BP messages affected by incomplete recovery, it is important that recovery be carried out in every iteration, and u the ๐cu ๐ ๐ โs be updated late to make the ๐๐ โs more up-to-date (see Section II). Indeed, the algorithm often exhibits oscillatory behavior otherwise, which can affect ๐ผbc and thus disrupt the pace of decimation so much as to increase the shaping loss above 1dB. IV. T HE R ECOVERY A LGORITHM Based on the above principles, we will now explain in detail how to estimate ๐(ห ๐ฆ๐ โฃ ๐๐u ) and perform the recovery using (17) with reasonable computational complexity. A. Modeling ๐(ห ๐ฆ๐ โฃ ๐๐u ) Since the LDGM code ensemble is symmetric with respect to symbol permutation, ๐(ห ๐ฆ๐ โฃ ๐๐u ) is the same for any ๐. Its estimation must still rely on the ๐ (ห ๐ฆ๐ , ๐๐u ) samples from the current quantizer instance, so some assumptions must be made on it in order to reduce the number of parameters to estimate and minimize random estimation errors. Recall that ๐๐u is asymptotically symmetric w.r.t. ๐ขโ๐ as ๐ โ โ under ideal decimation; this asymptotic symmetry should remain true under non-ideal decimation but idealized recovery, so it is reasonable to assume approximately that it holds under the actual recovery algorithm as well. Analogous to (14), we can then express ๐(ห ๐ฆ๐ โฃ ๐๐u ) as ๐(ห ๐ฆ๐ โฃ ๐๐u ) = =
๐โ1 โ ๐ข=0 ๐โ1 โ
๐(๐ขโ๐ = ๐ข โฃ ๐๐u )๐(ห ๐ฆ๐ โฃ ๐ขโ๐ = ๐ข, ๐๐u ) (18) ๐๐u (๐ข)๐(ห ๐ฆ๐ โฃ ๐ขโ๐ = ๐ข, ๐๐u ).
๐ข=0
In general, the deviation of ๐(ห ๐ฆ๐ โฃ ๐ขโ๐ , ๐๐u ) from the ideal โ ๐ยฏ(๐ฆ๐ โฃ ๐ข๐ ) is related to the pace of decimation and can vary with ๐๐u ; analysis of BEQ suggests that there is usually more deviation when ๐๐u is informative. However, this effect is difficult to estimate because only the (ห ๐ฆ๐ , ๐๐u ) samples with u informative ๐๐ โs are useful in the estimation. In practice, attempts to model this dependence do not improve the results, so we ignore it and model ๐(ห ๐ฆ๐ โฃ ๐ขโ๐ , ๐๐u ) as ๐(ห ๐ฆ๐ โฃ ๐ขโ๐ , ๐๐u ) = ๐z ((ห ๐ฆ๐ โ ๐ขโ๐ )โ ),
(19)
1835
such that ๐(ห ๐ฆ๐ โฃ ๐๐u ) =
๐โ1 โ
๐๐u (๐ข)๐z ((ห ๐ฆ๐ โ ๐ข)โ ).
(20)
๐ข=0
๐z (โ
) is further made to satisfy the constraints ๐z (๐ง) = ๐z (โ๐ง), 0 โค ๐ง < ๐/2, ๐โ1 โ
๐z ((๐ฆ โ ๐ข)โ ) = 1, ๐ฆ โ [0, ๐ ),
(21) (22)
๐ข=0
where (22) corresponds to theโซuniformity of ๐(ห ๐ฆ๐ ) and implies the normalization condition โ ๐z (๐ง) ๐๐ง = 1. Note that the ideal ๐ยฏ(๐ฆ๐ โฃ ๐ขโ๐ ) in (11) also has the form of (19) with the 2 corresponding ๐ยฏz (๐ง) โ ๐โ๐ก๐ง /๐(๐ง) (๐ง โ โ) satisfying (21) and (22). The interval [0, ๐ ) of possible ๐ฆห๐ values is now equally divided into ๐ ๐ subintervals ๐ด๐ ๐ฃ , ๐ = 0, . . . , ๐ โ 1, ๐ฃ = 0, . . . , ๐ โ 1, with each ๐ด๐ ๐ฃ โ โซ [๐ฃ + ๐ /๐, ๐ฃ + (๐ + 1)/๐). This discretizes ๐z (โ
) into ๐๐ ๐ฃ โ ๐ด๐ ๐ฃ ๐z ((๐ฆ)โ ) ๐๐ฆ, which can be further grouped into column vectors ๐๐ โ (๐๐ ๐ฃ )๐โ1 ๐ฃ=0 , and the constraints above become, for all ๐ and ๐ฃ, ๐๐ ๐ฃ = ๐๐โ1โ๐ ,๐โ1โ๐ฃ , or ๐๐ = R๐๐โ1โ๐ , ๐โ1 โ
๐๐ ๐ฃ = 1/๐, or 1T ๐๐ = 1/๐,
(23) (24)
๐ฃ=0
where R = RT = Rโ1 is such a matrix that R๐ is ๐ with its elements in reverse order, and 1 is the all-one vector. By (23), it is sufficient to estimate those ๐๐ with ๐ = 0, . . . , ๐/2 โ 1, each satisfying (24), from the ๐ (ห ๐ฆ๐ , ๐๐u ) samples according to the discretized version of (20), โ ] ๐โ1 [ ๐๐u (๐ข)๐๐ ,๐ฃโ๐ข , Pr ๐ฆห๐ โ ๐ด๐ ๐ฃ โฃ ๐๐u =
โ๐, ๐ , ๐ฃ.
(25)
๐ข=0
B. Computing the Raw Estimate of ๐๐ ๐ฃ Maximum-likelihood estimation using (25) turns out to be computationally intractable, so we adopt a regularized leastsquares method instead. For each[๐, ๐ and ๐ฃ, the left-hand side ] of (25) is just the expectation E 1[ห ๐ฆ๐ โ ๐ด๐ ๐ฃ โฃ ๐๐u ] , so we can approximate it with its sample 1[ห ๐ฆ๐ โ ๐ด๐ ๐ฃ ] obtained from the actual ๐ฆห๐ , yielding 1[ห ๐ฆ๐ โ ๐ด๐ ๐ฃ ] =
๐โ1 โ
๐๐u (๐ข)๐๐ ,๐ฃโ๐ข + ๐ค๐ ๐๐ฃ ,
โ๐, ๐ , ๐ฃ,
(26)
๐ข=0
๐ค๐ ๐๐ฃ being the sampling error. These equations can be grouped by ๐ and re-expressed as ๐๐ ๐๐ฃ =
๐โ1 โ
โฒ
๐ด๐ฃ๐๐ฃ ๐๐ ๐ฃโฒ + ๐ค๐ ๐๐ฃ , or ๐
๐ = ๐จ๐๐ + ๐๐ ,
(27)
๐ฃ โฒ =0 โฒ
where ๐ด๐ฃ๐๐ฃ โ ๐๐u (๐ฃ โ ๐ฃ โฒ ), ๐๐ ๐๐ฃ โ 1[ห ๐ฆ๐ โ ๐ด๐ ๐ฃ ], and they form respectively an ๐ ๐ร๐ matrix ๐จ and ๐ ๐ร1 vectors ๐
๐ . For each ๐, the corresponding ๐ rows of ๐จ and ๐
๐ are denoted ๐จ๐ and ๐
๐ ๐ , respectively; note that each ๐จ๐ is a circulant matrix that does not vary with ๐ .
1836
IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 59, NO. 7, JULY 2011
Since ๐๐ = R๐๐โ1โ๐ , for each ๐ โ {0, . . . , ๐/2 โ 1}, (27) for ๐ and ๐ โ 1 โ ๐ can be combined to yield [ ] [ ] ] [ ๐
๐ ๐จ ๐๐ = ๐ + , (28) ๐จR ๐ ๐
๐โ1โ๐ ๐๐โ1โ๐ ห ๐ +๐ or simply ๐
ห๐ = ๐จ๐ ห๐ . The estimation of ๐๐ is then formulated as the following regularized least-squares problem 2 1 1 ห 2 ห ๐ ๐
minimize โ ๐จ๐ + 2 โฅ๐๐ โ ๐ยฏ๐ โฅ ๐ 2 ๐๐ค ๐๐ (29) T subject to 1 ๐๐ = 1/๐, โซ where we have let ๐ยฏ๐ ๐ฃ โ ๐ด๐ ๐ฃ ๐ยฏz (๐ง) ๐๐ง, or ๐ยฏ๐ in vector form, be the โidealโ ๐๐ (which also satisfies 1T ๐ยฏ๐ = 1/๐), and 2 and ๐๐2 can respectively be the regularization parameters ๐๐ค understood as the variance of each ๐ค๐ ๐๐ฃ and the mean-square deviation of each ๐๐ from ๐ยฏ๐ . Solving (29), we obtain the estimate ห + ๐I)โ1 (๐จ หT ๐จ หT ๐
ห๐ + ๐๐ยฏ๐ โ ๐1) ๐๐ = (๐จ = (๐จT ๐จ + R๐จT ๐จR + ๐I)โ1 T
(30)
T
โ
(๐จ ๐
๐ + R๐จ ๐
๐โ1โ๐ + ๐๐ยฏ๐ โ ๐1), 2 where ๐ โ ๐๐ค /๐๐2 and ๐ is the Lagrange dual variable determined by the constraint 1T ๐๐ = 1/๐. T There remains the computation of the ๐ ร ๐ matrix ๐จ โ ๐จ T T T and the ๐ ร 1 vectors ๐จ ๐
๐ . Note that ๐จ ๐จ = ๐ ๐จ๐ ๐จ๐ ; T since each ๐จ๐ is circulant, so is ๐จT ๐ ๐จ๐ and thus ๐จ ๐จ, and the matrix inversion in (30) can be computed efficiently using FFT. Specifically, we first use the circulant property of ๐จT ๐จ to see that (๐จT ๐จ + R๐จT ๐จR + ๐I)1 = (2๐ผ + ๐)1, where ๐ผ is the sum of any row in ๐จT ๐จ; this allows the term with ๐ in (30) to be separated out, yielding
๐๐ = ๐ห๐ โ ๐โฒ 1,
(31)
where ๐ห๐ is (30) sans the ๐-term, and ๐โฒ โ ๐/(2๐ผ + ๐) can be computed from the 1T ๐๐ = 1/๐ constraint as ๐โฒ = (1T ๐ห๐ โ 1/๐)/๐ . Let F โ (๐โj2๐๐๐ฃ/๐ )๐๐ฃ be the DFT matrix with Fโ1 = FH /๐ (both the row index ๐ and the column index ๐ฃ start at 0 for convenience), then (30) can be transformed into F๐ห๐ = (F๐จT ๐จFโ1 + FR๐จT ๐จRFโ1 + ๐I)โ1 โ
F(๐จT ๐
๐ + R๐จT ๐
๐โ1โ๐ + ๐๐ยฏ๐ ).
(32)
โ1 Now F๐จ๐ Fโ1 = diag(A๐ ) and F๐จT = diag(A๐ )โ , ๐ F โ where (โ
) denotes complex conjugation and A๐ (not to be confused with ๐จ๐ ) is an ๐ ร 1 vector representing the DFT of ๐๐u , i.e. ๐โ1 โ ๐โj2๐๐๐ข/๐ ๐๐u (๐ข). (33) (A๐ )๐ โ ๐ข=0 โ
Moreover, FR = ฮฆF where ฮฆ โ diag((๐j2๐๐/๐ )๐โ1 ๐=0 ). Therefore, โ โ1 F๐จT ๐จFโ1 = F๐จT = diag(Q), (34) ๐ ๐จ๐ F ๐ T
โ1
FR๐จ ๐จRF = ฮฆ โ
diag(Q)โ ฮฆโ1 = diag(Q), (35) โ 2 where Q โ ๐ โฃA๐ โฃ is an ๐ ร 1 vector and โฃโ
โฃ takes the complex magnitude element-wise. On the other hand, ๐จT ๐
๐ =
โ
T ๐จT ๐ ๐
๐ ๐ , with each ๐จ๐ ๐
๐ ๐ given by { โ โฒ ๐๐u (๐ฃ โ ๐ฃ โฒ ), if ๐ฆห๐ โ ๐ด๐ ๐ฃ , (๐จT ๐ด๐ฃ๐๐ฃ ๐๐ ๐๐ฃ = ๐ ๐
๐ ๐ )๐ฃ โฒ = / โช๐ฃ ๐ด๐ ๐ฃ , 0, if ๐ฆห๐ โ ๐ฃ (36) and ๐จT ๐
๐โ1โ๐ can be computed analogously. ๐ห๐ can thus be obtained from (32), with the left-multiplication by F and Fโ1 carried out using FFT and inverse FFT respectively, and ๐๐ follows from (31). ๐
C. Further Regularization of the Estimate and Computation of the Adjusted ๐ห As the final step of the recovery algorithm, we regularize the estimated ๐๐ ๐ฃ โs, compute the corresponding ๐น๐๐u (โ
)โs, and use (17) to obtain the adjusted ๐ ห. By (20), each ๐น๐๐u (โ
) is a linear combination ๐น๐๐u (๐ฆ) =
๐โ1 โ
๐๐u (๐ข)๐น๐ข (๐ฆ)
(37)
๐ข=0
of ๐น๐ข (โ
), the cdfs corresponding to ๐z ((โ
โ ๐ข)โ ). For convenience, we define the โperiodically extendedโ cdf ๐น (๐ฆ) โ โซ๐ฆ โฒ ๐ ((๐ฆ )โ ) ๐๐ฆ โฒ for ๐z (โ
), so that ๐น๐ข (๐ฆ) becomes simply z 0 ๐น (๐ฆ โ ๐ข) โ ๐น (โ๐ข). After discretization, for ๐ฆ โ [0, ๐ ] that are integer multiples of 1/๐, โ ๐น (๐ฆ) = ๐๐ ๐ฃ , (38) (๐ ,๐ฃ):๐ฃ+๐ /๐