[Paper Review] Pseudo-codeword Landscape

Michael Chertkov, Mikhail Stepanov|ArXiv.org|Jan 12, 2007

Error Correcting Code Techniques10 references4 citations

TL;DR

This paper introduces a 'dendro' transformation that replaces high-degree parity-check nodes in LDPC codes with degree-three trees, enabling efficient LP decoding while preserving the original code's pseudo-codeword spectrum. The method reveals two distinct error-floor regimes: early error floors from low-effective-distance non-codeword pseudo-codewords, and transient asymptotes from gaps in the spectrum, explaining FER behavior at moderate-to-high SNR.

ABSTRACT

We discuss the performance of Low-Density-Parity-Check (LDPC) codes decoded by means of Linear Programming (LP) at moderate and large Signal-to-Noise-Ratios (SNR). Utilizing a combination of the previously introduced pseudo-codeword-search method and a new "dendro" trick, which allows us to reduce the complexity of the LP decoding, we analyze the dependence of the Frame-Error-Rate (FER) on the SNR. Under Maximum-A-Posteriori (MAP) decoding the dendro-code, having only checks with connectivity degree three, performs identically to its original code with high-connectivity checks. For a number of popular LDPC codes performing over the Additive-White-Gaussian-Noise (AWGN) channel we found that either an error-floor sets at a relatively low SNR, or otherwise a transient asymptote, characterized by a faster decay of FER with the SNR increase, precedes the error-floor asymptote. We explain these regimes in terms of the pseudo-codeword spectra of the codes.

Motivation & Objective

• To analyze the Frame-Error-Rate (FER) performance of LDPC codes under Linear Programming (LP) decoding at moderate and high SNR.
• To explain the observed discrepancy between early error floors and asymptotic FER behavior in terms of pseudo-codeword spectra.
• To develop a complexity-reduction technique for LP decoding that preserves the original code’s pseudo-codeword structure.
• To investigate whether the pseudo-codeword spectrum can predict the transition from transient to asymptotic FER decay.
• To validate the dendro-construction method on real-world LDPC codes, including those with high-check-degree structures.

Proposed method

• Introduce the 'dendro' construction: replace high-degree checks with trees of degree-three checks and punctured degree-two bits, preserving the code’s codewords and pseudo-codewords.
• Prove that the dendro-code and original code have identical sets of pseudo-codewords and identical MAP decoding results for any channel output.
• Apply the pseudo-codeword-search method to compute the effective distance spectra of both original and dendro-codes, enabling analysis of previously intractable codes.
• Use Monte Carlo simulations to validate the predicted FER behavior and compare with theoretical spectra.
• Analyze the frequency spectra of effective distances to distinguish between codes with early error floors and those with transient asymptotes.
• Leverage the large-SNR limit of BP as equivalent to LP decoding, enabling connection to Bethe free energy and instanton analysis.

Experimental results

Research questions

• RQ1 Why do some LDPC codes exhibit early error floors in FER performance, while others show a transient asymptote before the high-SNR limit?
• RQ2 Can the pseudo-codeword spectrum explain the transition from moderate-SNR to high-SNR FER behavior?
• RQ3 Does the dendro-construction preserve the pseudo-codeword structure of the original code while reducing LP decoding complexity?
• RQ4 What role does the gap in the pseudo-codeword spectrum play in determining the FER decay rate at moderate SNR?
• RQ5 How do the frequencies of low-effective-distance pseudo-codewords correlate with the observed error-floor onset?

Key findings

• The $[155,64,20]$ code exhibits a continuous pseudo-codeword spectrum starting at $d_{\text{min}} \approx 16.404$, indicating a strong early error floor consistent with $\sim\exp(-d_{\text{min}}s^2/2)$ decay.
• The $[672,336,16]$ code shows a gap in the spectrum above $d \approx 27.33$, leading to a transient asymptote at moderate SNR with $\sim\exp(-27.33s^2/2)$ decay before transitioning to the $d_{\text{ML}}=16$ asymptote.
• The $[648,324,15]$ and $[648,432,12]$ codes also display a two-stage behavior with gaps, though smaller than in the $[672,336,16]$ code, indicating intermediate asymptotes.
• The $[273,191,18]$ projective geometry code shows a 'staircase' of low-distance codewords with significant frequencies, leading to multiple error-floor features.
• The $[1296,648]$ code has a non-codeword pseudo-codeword with $d_{\text{min}} \approx 19.6$ and a noticeable gap to the next at $d \approx 21.75$, suggesting a one-stage error floor dominated by the lowest effective distance.
• The dendro-construction successfully preserves the pseudo-codeword spectrum, as confirmed by near-identical spectra between dendro-codes and original codes across all tested codes.