PaperlayerView exact revision on arXiv

Source-linked AI summary

Approximate message-passing decoder and capacity-achieving sparse superposition codes

Jean Barbier, Florent Krzakala

arXiv:1503.08040v4cs.ITmath.IT

TL;DR

Sparse superposition codes can reach Shannon capacity, but iterative AMP decoding encounters a phase transition before capacity. This paper analyzes AMP with power allocation and spatial coupling, finding both can reach capacity in the proper limit and that spatial coupling performs better in finite-size comparisons.

  • Problem

    Although sparse superposition codes can achieve capacity, earlier tractable decoders had poor asymptotic and finite-blocklength performance, and AMP faces a decoding threshold before capacity.

    Source: §I · p. 2, §VII · p. 23

  • Method

    The paper analyzes AMP for sparse superposition codes using replica and state-evolution methods, considering power allocation and spatial coupling.

    Source: §I · p. 2, §VII · p. 23

  • Results

    Both power allocation and spatial coupling allow AMP to achieve capacity in the proper limit, while spatial coupling outperforms optimized power allocation in finite-size reconstruction error and noise robustness.

    Source: §VII · p. 23

  • Takeaways & Limitations

    A spatially coupled fast Hadamard-based solver provides an efficient approach that handles very large instances and performs well at finite lengths.

    Source: §VII · p. 23

    Without a suitable modification, AMP encounters a sharp phase transition that blocks decoding before capacity.

    Source: §VII · p. 23

Abstract

from arXiv · show

We study the approximate message-passing decoder for sparse superposition coding on the additive white Gaussian noise channel and extend our preliminary work [1]. We use heuristic statistical-physics-based tools such as the cavity and the replica methods for the statistical analysis of the scheme. While superposition codes asymptotically reach the Shannon capacity, we show that our iterative decoder is limited by a phase transition similar to the one that happens in Low Density Parity check codes. We consider two solutions to this problem, that both allow to reach the Shannon capacity: i) a power allocation strategy and ii) the use of spatial coupling, a novelty for these codes that appears to be promising. We present in particular simulations suggesting that spatial coupling is more robust and allows for better reconstruction at finite code lengths. Finally, we show empirically that the use of a fast Hadamard-based operator allows for an efficient reconstruction, both in terms of computational time and memory, and the ability to deal with very large messages.

I. INTRODUCTION

The introduction presents AMP as a tractable decoder for sparse superposition codes and motivates strategies to close the gap between its decoding threshold and Shannon capacity. The paper develops the code setting, analyses, and finite-length comparisons for power allocation and spatial coupling.

I. INTRODUCTION

  • Prior sparse superposition code work established capacity achievement with power allocation and MAP decoding, while earlier tractable decoders had poor asymptotic and finite-blocklength performance.The paper extends its earlier AMP presentation and proposes power allocation and spatial coupling to make AMP capacity achieving in the appropriate limit.Source: §I · p. 2

A. Related works

  • The authors relate AMP phase transitions to those in LDPC codes and connect their analysis to compressed sensing, while noting that state-evolution exactness is not yet rigorous for structured sparsity.They also cite rigorous results for power allocation and replica analysis in related settings.Source: A. Related works · pp. 2–3

B. Main contributions of the present study

  • The study gives a self-contained AMP derivation from loopy belief propagation and analyzes performance across power allocation and spatial-coupling settings.It also compares state evolution with replica analysis and studies the large-section limit.Source: B. Main contributions of the present… · p. 3
  • State evolution analysis identifies a sharp AMP performance transition distinct from the code’s optimum transition, while the structured-sparsity setting leaves exactness conjectural.The paper also analyzes optimal performance using the heuristic replica method and conjectures its single-letter MMSE formulation is exact.Source: B. Main contributions of the present… · p. 3
  • Proper power allocation removes AMP’s blocking BP threshold, while spatial coupling offers an alternative route to capacity-achieving performance.The paper also reports better finite-size reconstruction and noise robustness for properly designed spatial coupling than for optimized power allocation.Source: B. Main contributions of the present… · p. 3
  • At finite blocklength, properly designed spatial coupling is reported to improve performance and robustness to noise despite gains from power allocation.The comparison is between spatial coupling and power allocation strategies.Source: B. Main contributions of the present… · p. 3

C. Outline

  • The paper proceeds from the code setting and AMP decoder to state evolution, replica analysis, and numerical comparisons of power allocation and spatial coupling.Its appendices provide derivations of AMP, state evolution, and replica analysis.Source: C. Outline · p. 4

A. Sparse superposition codes: setting

III. THE APPROXIMATE MESSAGE-PASSING DECODER AND SPATIAL COUPLING

Direct belief propagation is intractable on the densely connected factor graph, motivating AMP as a lower-complexity decoder. AMP faces a phase transition that spatial coupling is designed to overcome, though threshold saturation is conjectured rather than proved here.

A. Why belief-propagation is not an option for decoding with sparse superposition codes

  • Loopy BP requires combinatorially many terms and 2ML messages on the densely connected graph, making it intractable even for small signals.The paper contrasts this with BP's efficiency on low-connectivity, tree-like graphs.Source: A. Why belief-propagation is not an… · p. 6

B. The approximate message-passing decoder for sparse superposition codes

  • AMP is derived by expanding dense-graph BP to second order and then reducing the number of messages, yielding a Gaussian approximation.Its estimates track codeword and message quantities alongside posterior estimates and variances of x.Source: B. The approximate message-passing… · p. 7
  • AMP is asymptotically Bayes-optimal only below the BP threshold; power allocation or spatial coupling can overcome the transition.The transition location depends on the noise variance.Source: B. The approximate message-passing… · p. 7

C. Spatial coupling

  • Spatial coupling uses a seeded boundary and finite-window connections across a chain to trigger a reconstruction wave from the easy seed into harder subsystems.The coding operator has a block structure, with the seed block measured at a higher rate than the remaining blocks.Source: C. Spatial coupling · pp. 8–9
  • The paper conjectures that its spatially coupled matrices achieve threshold saturation but leaves the proof for future work.It reports explicit examples approaching the desired threshold; a footnote notes later rigorous proofs for spatially coupled sparse superposition codes.Source: C. Spatial coupling · p. 8

D. The fast Hadamard-based coding operator

The decoder uses structured Hadamard operators to reduce the cost and memory burden of matrix operations. Their performance approaches that of random Gaussian operators as section size increases, while random operators are advantageous at smaller section sizes.

D. The fast Hadamard-based coding operator

  • The operator is block-structured, with nonzero blocks formed either from random Hadamard modes or i.i.d. Gaussian submatrices, and rescaled to give the codeword unit power.The Hadamard construction uses random selections and reordering of modes.Source: D. The fast Hadamard-based coding… · p. 9
  • O(N ln N) replaces O(N^2) matrix-multiplication cost, and the matrix need not be stored, enabling fast decoding of very large messages.The operator applies fast transforms and simple sums rather than costly direct matrix multiplications.Source: D. The fast Hadamard-based coding… · p. 10
  • As section size B increases, Hadamard-based operators quickly approach random i.i.d. Gaussian operators in decoding performance; random operators are advantageous at low B.The comparison fixes snr = 100 and measures distance to the BP threshold.Source: D. The fast Hadamard-based coding… · p. 10, Figure 6 · p. 10

IV. RESULTS OF THE STATE EVOLUTION ANALYSIES

The state evolution analysis tracks AMP’s error-rate dynamics and predicts its behavior with Gaussian operators, while remaining an accurate, though approximate, predictor for Hadamard operators. Results show an error floor and a threshold-dependent decoding transition, with the floor decreasing as section size or SNR increases, or rate decreases.

IV. RESULTS OF THE STATE EVOLUTION ANALYSIES

  • Hadamard-based operators reach a lower final SER and generally converge in fewer iterations than i.i.d. Gaussian operators, while closely following state evolution at large signal sizes.The comparison uses constant power allocation, section size B = 64, and rates on both sides of R_BP.Source: Figure 7 · p. 11

A. State evolution for homogeneous coding operators and constant power allocation

  • State evolution monitors AMP dynamics and performance as signal size tends to infinity; it is derived for i.i.d. Gaussian matrices and predicts Hadamard-operator behavior approximately.The analysis does not extend to ensembles such as row-orthogonal matrices in this paper.Source: §IV · p. 10, A. State evolution for homogeneous… · p. 11
  • Experiments are restricted to relatively low SNR because at high SNR, finite messages and Monte Carlo samples may not reveal the very low error floor.At high SNR, curves can stop without reaching an error floor, making decoding appear perfect.Source: A. State evolution for homogeneous… · p. 12
  • Above R_BP, AMP fails to reconstruct and converges to the high SER predicted by state evolution; below it, reconstruction succeeds up to a parameter-dependent error floor.Predictions match experiments better for larger messages and predict Hadamard-operator behavior especially well at large B.Source: A. State evolution for homogeneous… · p. 12
  • For every finite section size B, an asymptotic error floor remains, but it decreases rapidly as B or SNR increases, or as the rate decreases.The floor is also present for the optimal MMSE estimator and can be made arbitrarily small by increasing B at fixed SNR while retaining low-complexity AMP decoding.Source: A. State evolution for homogeneous… · p. 12
  • With Hadamard operators, AMP complexity scales as O(BL ln(BL)); increasing B can reduce the floor without the exponential complexity barrier described for LDGM decoding.For i.i.d. Gaussian operators, AMP complexity scales as O((BL)^2).Source: A. State evolution for homogeneous… · p. 12

B. State evolution for spatially coupled coding operators and constant power allocation

State evolution and replica analysis characterize decoding with spatial coupling and nonconstant power allocation, as well as the phase transitions of constant-power codes. Spatial coupling lets AMP improve as section size grows, while the replica analysis distinguishes the algorithmic threshold from the optimal threshold.

B. State evolution for spatially coupled coding operators and constant power allocation

  • For Hadamard-based spatially coupled operators, the final error is at least as good as state evolution predicts, and convergence is slightly faster than with Gaussian matrices.The comparison is between the predicted block section error rates and actual reconstruction error.Source: B. State evolution for spatially… · p. 13

C. State evolution for homogeneous coding operators and non constant power allocation

  • Grouping sections by power and scaling corresponding column blocks yields a constant-power structured operator equivalent to the original system from the state-evolution perspective.For G groups, the construction divides the operator into LB/G column blocks and scales each group’s block by its power allocation.Source: C. State evolution for homogeneous… · p. 13

A. The replica symmetric potential of sparse superposition codes with constant power allocation

  • The replica potential’s extrema match state-evolution fixed points, and its competing maxima identify the AMP-blocking hard phase and the optimal threshold.AMP can be interpreted as ascending the potential; a high-error local maximum can block it from reaching the low-error global maximum.Source: A. The replica symmetric potential of… · pp. 14–15

B. Results from the replica analysis for sparse superposition codes with constant power allocation and finite section size

  • Spatial coupling lets AMP improve as section size increases by entering the hard phase between the BP and optimal thresholds.Without spatial coupling, AMP performance monotonously declines in terms of possible communication rate as section size grows.Source: B. Results from the replica analysis… · p. 16
  • The optimal threshold approaches capacity faster with section size than the BP threshold approaches its asymptote; both trends are suggestively power-law, not established fits.The plotted lines are guides, and low scaling exponents may also indicate slow logarithmic convergence.Source: B. Results from the replica analysis… · p. 16, Figure 12 · p. 16
  • For high-noise regimes, AMP attains the plotted optimal SER without spatial coupling; in lower-noise regimes, that SER is optimal only below Ropt or where no transition occurs.The replica analysis also identifies whether a parameter setting is in the easy, hard, or impossible inference phase.Source: B. Results from the replica analysis… · p. 16

C. Large section limit for sparse superposition codes with constant power allocation

A heuristic large-section analysis uses a replica potential to characterize the asymptotic decoding landscape. It finds that the optimal threshold is Shannon capacity, while AMP without spatial coupling has a lower threshold because its initialization can remain trapped at high error.

C. Large section limit for sparse superposition codes with constant power allocation

  • The large-section analysis heuristically evaluates the replica potential using a random-energy-model analogy, with an alternative replica derivation offered as corroboration.The authors state that the analysis is not rigorous.Source: C. Large section limit for sparse… · p. 17
  • At large section size, the potential has possible maxima at E=0 and E=1, representing zero-error and high-error states.The high-error maximum at E=1 is stable when the rate exceeds the asymptotic BP threshold.Source: C. Large section limit for sparse… · p. 19
  • AMP initialized at E=1 is limited by the asymptotic BP threshold below capacity; above it, spatial coupling or power allocation is required for optimal decoding.The high-error local maximum explains the computational gap between message-passing and optimal performance.Source: C. Large section limit for sparse… · p. 19
  • The two potential maxima have equal value at R = log2(1 + snr)/2, the AWGN Shannon capacity, so the optimal threshold reaches capacity as section size grows.The analysis concludes that the optimal section error rate vanishes in this limit.Source: C. Large section limit for sparse… · p. 19

D. Optimality of the AMP decoder with a proper power allocation

The section constructs a blockwise power allocation and shows that, in the stated asymptotic regime, it removes the AMP decoding barrier below Shannon capacity. With this allocation, AMP achieves perfect reconstruction for rates R < C.

D. Optimality of the AMP decoder with a proper power allocation

  • The construction divides the signal into G blocks, assigning all nonzero sections in block g the same value c_g.The analysis assumes 1 ≪ B ≪ LG and 1 ≪ G, where LG = L/G is the number of sections per group.Source: D. Optimality of the AMP decoder with… · p. 19
  • Decoding proceeds block by block: if sections through g̃ − 1 are decoded, the next block is decodable when its effective noise variance satisfies Σ² < 1/2.When this condition holds for every block, AMP advances through the signal without a BP threshold blocking reconstruction.Source: D. Optimality of the AMP decoder with… · p. 20
  • For rates R < C, the proper power allocation eliminates spurious potential maxima, so AMP reaches perfect reconstruction with SER = 0.The result is stated for the allocation in (44) and the large-section analysis.Source: D. Optimality of the AMP decoder with… · pp. 19–20

VI. NUMERICAL EXPERIMENTS FOR FINITE SIZE SIGNALS

Finite-size experiments compare spatial coupling with power allocation and other AMP configurations, while the conclusion summarizes the resulting performance and theoretical analysis. The experiments report that spatial coupling outperforms per-rate optimized power allocation in the tested comparison, with a high-rate boundary to that advantage.

VI. NUMERICAL EXPERIMENTS FOR FINITE SIZE SIGNALS

  • At B = 256, the spatially coupled AMP transition is sharp and sharpens as L increases; larger signals also decode closer to asymptotic transitions with lower error floors.For small B, the gap between the BP threshold and capacity is too small for spatial coupling to yield real improvement over full operators.Source: §VI · p. 21
  • AMP with spatially coupled Hadamard operators outperforms the iterative successive decoder for any B in the reported comparison, and the homogeneous or coupled operators perform well for M < 64000.The conclusion reports that AMP without spatial coupling also beats the successive decoder for manageable sizes.Source: §VI · p. 22
  • The red on-line AMP curve with optimized power allocation outperforms the blue off-line implementation using the same parameters.The authors attribute the difference to decoder implementation: the off-line method requires state-evolution preprocessing, while their on-line decoder does not.Source: §VI · p. 22, Figure 18 · p. 23
  • Spatial coupling with Hadamard operators greatly outperforms per-rate optimized power allocation, except at very high rates where spatial coupling cannot decode.The comparison is independent of AMP implementation and operator choice; the mixed strategy performs poorly because power allocation worsens the spatial-coupling design.Source: §VI · p. 22, Figure 18 · p. 23

VII. CONCLUSION AND FUTURE WORKS

  • Replica analysis finds a sharp message-passing phase transition before capacity, while the Bayesian-optimal limit approaches Shannon capacity as section size increases.The replica and state-evolution analyses predict the phase transitions equivalently; the reported error-floor decay and transition convergence empirically follow power laws.Source: §VII · p. 23
  • The authors present spatially coupled fast Hadamard operators as an efficient, capacity-achieving solver that handles very large instances and performs as well as random coding operators.Their experiments report better reconstruction error and noise robustness than optimized power allocation, including good spatial-coupling performance for small signals.Source: §VII · p. 23

APPENDIX A DERIVATION OF THE APPROXIMATE MESSAGE-PASSING DECODER FROM BELIEF-PROPAGATION

The appendix derives the decoder from loopy belief propagation by approximating messages with Gaussian distributions and then expressing them through moments. Under the dense i.i.d.-matrix limit, the resulting approximations are asymptotically equivalent to belief propagation.

APPENDIX A DERIVATION OF THE APPROXIMATE MESSAGE-PASSING DECODER FROM BELIEF-PROPAGATION

  • The derivation starts from loopy BP with continuous density messages, which are expanded to obtain tractable equations for the discrete signal variables.The section frames this derivation for section-level variables with a known prior.Source: A. Gaussian approximation of… · p. 24, Appendix A · p. 24

A. Gaussian approximation of belief-propagation for dense linear estimation: relaxed belief-propagation

  • The Gaussian approximation yields diagonal-covariance messages, allowing each message to be represented by its first and second moments rather than full distributions.The within-section component independence arises asymptotically for independently drawn matrix entries; the prior retains the section-level correlations.Source: A. Gaussian approximation of… · pp. 24–25
  • The denoisers combine AMP fields with the section prior to estimate posterior marginal means and variances.The section describes the mean denoiser as an MMSE estimator for an effective AWGN channel.Source: A. Gaussian approximation of… · p. 25
  • For dense i.i.d. coding matrices, the Gaussian approximation is asymptotically exact as L → ∞, yielding the same signal estimates as BP up to vanishing corrections.The text attributes this equivalence to the dense matrix and central-limit-like arguments.Source: A. Gaussian approximation of… · p. 25

B. Reducing the number of messages: the approximate message-passing algorithm

The derivation reduces relaxed belief propagation to AMP, incorporates the sparse-superposition prior through denoisers, and further simplifies the decoder for zero-mean matrices with homogeneous variance per block. The simplified decoder is asymptotically equivalent to AMP under those matrix conditions.

B. Reducing the number of messages: the approximate message-passing algorithm

  • TAP corrections convert the relaxed-BP updates from 2MN edge messages to AMP with M + N messages, preserving asymptotic equivalence as L → ∞.The Onsager reaction terms retain the first-order corrections needed to distinguish AMP from a naive mean-field approach.Source: B. Reducing the number of messages:… · pp. 25–26

C. Taking into account the prior for sparse superposition codes

  • For sparse superposition codes, the problem-dependent denoisers are derived from the section prior for a general power allocation.These denoisers complete the decoder derivation for the sparse-superposition prior.Source: C. Taking into account the prior for… · pp. 26–27

D. Further simplifications for random matrices with zero mean and homogeneous variance

  • Replacing squared matrix elements by their variance gives a simpler decoder that is asymptotically equivalent to AMP for zero-mean matrices with homogeneous variance per block.The approximation is justified because the relevant ensemble averages are O(1) while their variances are O(1/N).Source: D. Further simplifications for random… · p. 27, Figure 19 · p. 28

APPENDIX B STATE EVOLUTION ANALYSIS

Under the Bayesian optimal setting, the Nishimori identity relates averages over the signal and posterior samples. With concentration in the large-system limit, the section derives equivalent expressions for the MMSE and posterior expected variance.

A. The Bayesian optimal setting and the Nishimori identity

  • The Bayesian optimal setting assumes a known signal prior and channel, with estimates drawn from the resulting posterior distribution.Under these conditions, averaging a function of the signal and one posterior replica permits replacing the signal with an independent posterior replica.Source: A. The Bayesian optimal setting and… · p. 29
  • Concentration of macroscopic observables and overlaps lets the analysis use their expectations in the asymptotic limit.The section combines this assumption with the Nishimori identity to derive the needed overlap relations.Source: A. The Bayesian optimal setting and… · p. 29
  • As L →∞, the MMSE satisfies E = q − 2m + Q = Q − m = 1 − m.The displayed equivalences follow from the preceding identities.Source: A. The Bayesian optimal setting and… · p. 29
  • The posterior expected variance V and MMSE are asymptotically equal.The section links the posterior expected variance to the MMSE using concentration and the Nishimori identity.Source: A. The Bayesian optimal setting and… · p. 29

B. Coding matrices with homogeneous variance

For homogeneous-variance coding matrices with constant power allocation, the analysis derives an AMP state evolution recursion from the decoder's field fluctuations. A Nishimori-based equivalent form has computational advantages but can be numerically cumbersome.

B. Coding matrices with homogeneous variance

  • The analysis assumes homogeneous entry variance and constant power allocation, which makes the sections symmetric.Nonconstant power allocation is treated as a special case of the structured-matrix analysis.Source: B. Coding matrices with homogeneous… · p. 30
  • The decoder-field fluctuations are analyzed through their Gaussian moments, retaining leading O(1) terms under the stated large-size concentration assumption.The derivation notes that an alternative relaxed-BP route would not require the same assumption.Source: B. Coding matrices with homogeneous… · p. 30
  • The derivation yields the state evolution recursion for AMP's MSE per component under the constant-power sparse superposition prior.The section first defines the per-component MSE and then obtains the final recursion using the prior.Source: B. Coding matrices with homogeneous… · p. 32
  • The Nishimori identity gives an equivalent state evolution in which the updated MSE is the averaged MMSE of an effective B-dimensional AWGN channel.The denoiser is the Bayes-optimal MMSE estimator for that channel.Source: B. Coding matrices with homogeneous… · p. 32
  • The equivalent state evolution is faster to compute, but finite numerical precision can make it negative when a difference is very small.In such cases, it can be more cumbersome to use than the original recursion.Source: B. Coding matrices with homogeneous… · p. 32

C. Coding matrices with a block structure

For block-structured coding matrices, the state evolution generalizes to track error by block, with variances determined by block indices. The derivation retains the large-system regime in which AMP is valid for structured matrices.

C. Coding matrices with a block structure

  • The structured-matrix derivation assumes L ≫ Lc, Lr and allows matrix-entry variances to differ across blocks.It follows the homogeneous-variance derivation, with the variance now depending on block indices.Source: C. Coding matrices with a block… · pp. 32–33
  • Cross terms cancel at dominant order, and the resulting block-structured state evolution is given by equations (22) and (23).The block-indexed variance is the key change from the homogeneous case.Source: C. Coding matrices with a block… · pp. 33–34
  • The analysis defines the asymptotic MSE separately for each signal block induced by the spatially coupled operator.It also defines the average variance of estimates within each block and uses the Nishimori identities to equate it with the block MSE.Source: C. Coding matrices with a block… · pp. 33–34

APPENDIX C REPLICA ANALYSIS

The replica analysis derives a replica-symmetric free entropy for constant power allocation using the replica trick and an overlap-based saddle-point calculation. Its derivation relies on heuristic assumptions, including an interchange of limits that is not rigorous.

A. Derivation of the replica symmetric free entropy for constant power allocation by the replica method

  • The derivation expresses the replica free entropy through the replica trick, using the replicated partition function and averages over disorder and posterior replicas.It begins with the i.i.d. Gaussian matrix ensemble and constant power allocation.Source: Appendix C · p. 34, A. Derivation of the replica symmetric… · p. 34
  • The replica-symmetric ansatz assumes equivalent replicas and overlaps independent of replica indices in the Bayes-optimal setting.The passage states the ansatz applies when the factor graph is locally tree-like or dense, as in this setting.Source: A. Derivation of the replica symmetric… · p. 35
  • The overlaps reparameterize the replicated partition function, and the Nishimori identity reduces the resulting potential to one free parameter.The derivation then obtains the final replica-symmetric free entropy for the sparse superposition prior.Source: A. Derivation of the replica symmetric… · pp. 35–37
  • The replica-trick derivation assumes the limits n →0 and L →∞ commute, although this interchange is not rigorous.The passage notes that the interchange is heuristically verified in many models, including inference.Source: A. Derivation of the replica symmetric… · p. 36

B. The link between replica and state evolution analysies

The analysis connects extrema of the potential to stationary points of state evolution. It also shows that the potential’s fixed-point conditions reproduce the state-evolution recursion, making the two approaches equivalent for computing typical fixed-point mean-square error.

B. The link between replica and state evolution analysies

  • Extrema of the potential correspond to stationary points of state evolution.The derivation is restricted to constant power allocation; the authors say the generic-allocation derivation is similar.Source: B. The link between replica and state… · p. 37
  • When the stated condition holds, the replica-potential fixed-point conditions recover the stationary state-evolution recursion.The recursion is written with the time index dropped.Source: B. The link between replica and state… · p. 37
  • State evolution and the potential give equivalent estimates of the typical mean-square error of AMP fixed points.The typical error is averaged over the quenched disorder.Source: B. The link between replica and state… · p. 37
  • The fixed-point conditions of the replica potential recover state evolution at its stationary point; this strengthens the claim that replica analysis computes the problem’s potential exactly, though not rigorously.Source: B. The link between replica and state… · p. 37

C. Alternative derivation of the large section limit of the potential via the replica method

The replica method re-derives the large-section potential by averaging the logarithm of a section partition function over disorder. In the large-section limit, its result is consistent with the earlier potential expression.

C. Alternative derivation of the large section limit of the potential via the replica method

  • At fixed Σ, treated as a temperature, the calculation averages the logarithm of a section partition function over disorder using the replica method.The section partition function KB is distinguished from Z, the partition function of the overall signal.Source: C. Alternative derivation of the large… · pp. 37–38
  • The replica-symmetric ansatz treats replicas as equivalent, and the order parameters q and m track replica agreement and membership in the favored transmitted state.The associated entropy counts replica-state configurations compatible with q and m.Source: C. Alternative derivation of the large… · p. 38
  • Of four possible order-parameter states, q = 0, m = 1 is impossible, while q = 1, m = 0 has lower free entropy and is discarded.The remaining ansatz are q = m = 0 and q = m = 1.Source: C. Alternative derivation of the large… · p. 39
  • In the B → ∞ limit, a saddle-point calculation selects the larger of the two remaining ansatz, assuming the limits commute.Their values are evaluated as n → 0.Source: C. Alternative derivation of the large… · p. 39
  • The replica calculation yields the same potential φ(E) as the earlier derivation, consistent with the appendix result.The passage identifies its expression with the second term on the right-hand side of (38).Source: C. Alternative derivation of the large… · p. 39

Related papers

Loading 1503.08040v4…