{"id":"72293d9b-f955-4bb2-b527-eac6db722c57","arxiv_id":"2504.13601","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper introduces a spatially coupled VAMP decoder and shows its asymptotic state evolution reaches capacity for sparse superposition codes when the design matrix spectrum concentrates at one.","lead":"This paper proposes a spatially coupled VAMP decoder for sparse superposition codes and uses state evolution to argue it can reach Shannon capacity under a condition on the design matrix spectrum. The decoder might matter for communication because it achieves lower section error rates than prior VAMP-based schemes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The capacity claim rests on an unproven state-evolution assumption for the uniform-diagonalized SC-VAMP update; a direct SE-vs-simulation comparison would settle whether it holds.","rationale":"The reader's conditional verdict is appropriate. The paper's central claim is not disproven, but it rests on an unproven SE assertion for the specific uniform-diagonalized SC-VAMP update. Without a proof or a reference establishing SE for this exact concatenated message schedule, ACCEPT would be too strong. At the same time, the gap is plausibly fillable, so REJECT would also be too strong. The sign error in Proposition 1's proof is secondary: it is correctable and does not by itself undermine the intended threshold argument. A direct empirical SE check, comparing predicted and measured per-block MSE across increasing L and several W values, is inexpensive and would largely settle whether the SE assumption holds. Thus the verdict remains CONDITIONAL, matching the reader's assessment.","tokens_in":12396,"tokens_out":14560,"duration_ms":145332,"concrete_test":"Run the Section V experiment with DCT and Gaussian ensembles at L=2^12, 2^14, and 2^16 (B=16, snr=15, \\Gamma=16, W=2), and for each iteration record empirical per-block MSE E_k^c. Compare these curves with the finite-B SE recursion from Section IV. The SE assumption is supported only if the empirical E_k^c approaches the SE prediction as L grows and the empirical algorithmic threshold approaches \\vartheta^{-1}R_{IT}; repeat with W=4 and W=8 to stress the uniform-diagonalization step.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The capacity claim is secured only if the SE recursion in Section IV exactly tracks Algorithm 1. That recursion is asserted via 'it follows' rather than proved for this algorithm. The critical step is the concatenating/uniform-diagonalization update in Algorithm 1, lines 20-21, where per-block precisions \\hat\\eta_{2,c} are collapsed to \\eta_{2,r}=|W_r|/\\sum_{c\\in W_r}\\hat\\eta_{2,c}^{-1}. Because each block c feeds several overlapping matrices A_r, the messages entering a given r are not conditionally independent given x_c; VAMP's scalar SE closure is not automatic here. No theorem in the paper, and no cited result clearly covering this exact concatenated update, establishes that the scalarized messages preserve the per-block MSE asymptotically. Proposition 1 and the capacity conclusion apply only to this assumed SE. A second, correctable gap: the proof of Proposition 1 states \\sum_{r=c}^{W-1}1/r\\le\\ln(W/c), which is false as written; the needed direction is \\ge. The intended argument is apparent, but as printed the proof is not fully rigorous.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spatially coupled VAMP (SC-VAMP) decoder for sparse superposition codes with spatially coupled right-rotationally-invariant design matrices. The algorithm is derived from a factor graph with expectation-consistent message passing and a uniform-diagonalization approximation. The authors write state-evolution recursions for the per-block MSE and prove a threshold-saturation result (Proposition 1) for the limiting SE recursion. A Cauchy-Schwarz bound on the information-theoretic threshold R_IT is then used to conclude that the SC-SS code with SC-VAMP is capacity-achieving over the AWGN channel when the design spectra converge to δ1 as α→0. Numerical experiments compare SC-VAMP with plain VAMP and exponential-power-allocation VAMP, reporting a lower section error rate; the code is publicly available.","tokens_in":1202,"tokens_out":1184,"duration_ms":75236,"significance":"If the SE recursion is valid and the large-section limit is made precise, the paper would give a new capacity-achieving decoding scheme that combines spatial coupling with rotationally invariant designs, extending prior VAMP results from exponential power allocation to a spatially coupled construction. The empirical comparison is a genuine strength, and the public code makes the experiments reproducible. However, the central capacity claim is conditional on unproved state evolution for the specific overlapping-matrix, diagonalized update, and the asymptotic order of the coupling parameters is not quantified. The contribution is therefore a plausible and well-supported conjecture with a clear algorithmic proposal, rather than a fully demonstrated theorem.","major_comments":[{"comment":"The SE equations are introduced with 'it follows' and no theorem establishes that the scalarized per-block MSE recursions track Algorithm 1. The concatenating/uniform-diagonalization step collapses the per-block precisions (η̂_{2,c}) into a single scalar η_{2,r} by harmonic mean. Because each block c feeds several overlapping matrices A_r, the messages entering a given r are not conditionally independent given x_c, so the VAMP/EC scalar closure conditions are not automatic. Please either prove SE for this exact update or cite a theorem that covers overlapping spatially coupled blocks with a section-wise denoiser; without this, Proposition 1 and the capacity conclusion apply only to the assumed SE.","section":"§IV (Algorithm 1, lines 20–21)"},{"comment":"The printed inequality ∑_{r=c}^{W-1} 1/r ≤ ln(W/c) is false for 1 ≤ c ≤ W; the correct direction is ≥. The subsequent upper bound on τ^0_c requires a lower bound on the harmonic sum, so the displayed proof of the initial saturation wave is not rigorous as written. This appears easily correctable, but it is load-bearing because it establishes the first set of zero ψ^0_c blocks.","section":"§IV, Proposition 1 proof"},{"comment":"The passage from finite-B SE to the limit equations uses the phase transition of E_2(γ) and assumes α = Θ(log B/B) → 0 without quantifying error terms. The statement that 'for sufficiently large W with Γ > W^2, ϑ → 1' does not specify how W, Γ, B, and L must scale relative to one another and to the target rate gap Δ. Please state explicit asymptotics, or provide a non-asymptotic bound, so that 'capacity-achieving' has a precise meaning and the iteration count K = 1 + ⌈Γ/(2g)⌉ is compatible with the code length.","section":"§IV, large-section limit"},{"comment":"The capacity conclusion relies on equality in the Cauchy-Schwarz bound, which occurs only when ρ0 = δ1, while Proposition 1 requires all F_r to coincide. The paper should clarify whether the claimed capacity result assumes a common limiting spectrum for all r or only that each ρ_supp,r → δ1 as α→0, and, in the latter case, explain how the proof adapts when F_r differs for finite r.","section":"§IV, after Proposition 1"}],"minor_comments":[{"comment":"The phrase 'g is a integer greater than or equal to 1' contains a typo; it should be 'an integer'.","section":"§IV, Proposition 1 proof"},{"comment":"It would be more informative to state the numerical values of the algorithmic and information-theoretic thresholds used in the comparison.","section":"Fig. 3 caption"},{"comment":"The definition N(r) = |W_r| 1{r≤W} + W 1{r>W} is redundant; if N(r) = |W_r|, it may be simpler to say so explicitly.","section":"Algorithm 1, line 14"},{"comment":"The GitHub URL contains a space ('SC-V AMP') due to formatting; ensure it is printed as a single clickable hyperlink.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The main gap is the unproved state evolution for the proposed decoder. If the authors can supply a proof or a precise citation covering the overlapping-block uniform-diagonalized VAMP update, the paper would be a solid contribution. Without that, the capacity claim should be explicitly labeled as conditional on SE, and the paper's value rests on the algorithm and simulations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"X, this paper extends VAMP decoding to spatially coupled sparse superposition codes and claims capacity for AWGN when the design matrices satisfy the spectra criterion. The combination of SC structure with VAMP is new, and the empirical gain over exponential-decay-PA VAMP is real, though shown on a single instance. The proof of Proposition 1 follows the standard threshold-saturation pattern and is mostly correct, but there are inequality typos: the printed sum_{r=c}^{W-1} 1/r <= ln(W/c) should be >= for the deduction to go through, and h is called decreasing when it is increasing. These are easy fixes. The soft spot is the state evolution assumption. Section IV asserts that the scalarized, uniform-diagonalized SC-VAMP update tracks per-block MSE almost surely, but no theorem is supplied for this exact concatenated update. Because blocks overlap, the messages entering a given r are not conditionally independent given the signal, so VAMP's scalar SE closure is not automatic. The EC derivation motivates the algorithm; it does not prove SE. If SE fails, the capacity conclusion collapses. This is an addressable gap, but the paper as written should have either proven it or cited a result that covers it. The Cauchy-Schwarz capacity bound is sound: with unit mean spectrum, F(x) <= 1/(sigma^2+x), so R_IT <= C. The condition Gamma > W^2 to make theta -> 1 is stated without quantitative justification, a minor omission. For whom: information theorists working on AMP/VAMP and sparse regression codes. The algorithm is useful; the capacity claim is conditional. I would send it to peer review and ask a referee to verify whether the SE recursion is covered by existing results, e.g., Takeuchi's SC-OAMP, or needs new proof. If the SE gap is closed, this is a solid incremental contribution.","headline":"A plausible extension of VAMP to spatially coupled superposition codes whose capacity claim rests on an unproven state-evolution closure; worth refereeing if the SE step is addressed.","tokens_in":13107,"tokens_out":5266,"would_cite":true,"duration_ms":44183,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A24"],"pacs":[],"model":"deepseek-v4-flash","headline":"Sparse superposition codes with a spatially coupled VAMP decoder are claimed to be capacity-achieving over the AWGN channel when the design matrices meet the spectra criterion.","keywords":["sparse superposition codes","spatial coupling","vector approximate message passing","state evolution","capacity-achieving codes","AWGN channel","spectra criterion","section error rate"],"falsifier":"Run SC-VAMP on a large but finite instance with a right-rotationally invariant design matrix whose limiting spectrum is the point mass $\\delta_1$, at a rate below $\\frac12\\log(1+\\mathrm{snr})$, and compare the measured per-block mean squared error with the state-evolution prediction; a mismatch that does not vanish as the dimensions grow would falsify the SE tracking and hence the capacity claim.","tokens_in":12242,"feed_emoji":"📶","tokens_out":8082,"duration_ms":71636,"temperature":0.7,"pith_summary":"This paper proposes a spatially coupled variant of the vector approximate message passing (VAMP) decoder for sparse superposition codes, a coding scheme in which each codeword is a sparse vector multiplied by a large random matrix. The authors claim that, when the design matrices satisfy the spectra criterion, this SC-VAMP decoder is capacity-achieving over the additive white Gaussian noise channel. The argument runs through a state evolution analysis showing that decoding errors are pushed from the boundary blocks inward and vanish after finitely many iterations for all rates below an information-theoretic threshold, which is then shown to equal channel capacity. A reader would care because this supplies a polynomial-time decoder that provably reaches the Shannon limit, and simulations show it beats the previous VAMP decoder with exponential power allocation at finite block lengths.","feed_headline":"Spatial coupling makes sparse superposition codes capacity-achieving","feed_subtitle":"For Gaussian-channel codes, a state-evolution analysis shows a VAMP decoder reaches the Shannon limit when spectra concentrate.","key_machinery":"The load-bearing mechanism is the rescaled state-evolution recursion given by equations (1)-(4), with variables $\\sigma_r^k$, $\\tau_c^k$, $\\psi_c^k$, and $\\phi_r^k$, together with the phase-transition behaviour of the Bayes denoiser: $\\lim_{B\\to\\infty}E_2(\\gamma)=\\mathbb{I}\\{(\\lim \\log B/\\gamma)>1/2\\}$. Proposition 1 turns this recursion into a threshold-saturation argument: with coupling width $W$ and block count $\\Gamma$, if $R_{\\mathrm{all}}<\\vartheta^{-1}R_{\\mathrm{IT}}$ and $W>\\max\\{\\lceil1/l^*(\\vartheta)\\rceil,\\lceil1/h^*(\\vartheta,\\Delta)\\rceil\\}$, then $\\psi_c^k=0$ for an expanding set of blocks, so decoding succeeds. The capacity step is the identity $F(x)=\\mathbb{E}_{\\rho_0}[\\lambda/(\\lambda x+\\sigma^2)]$, whose Cauchy-Schwarz bound $F(x)\\le(\\sigma^2+x)^{-1}$ yields $R_{\\mathrm{IT}}\\le C$, with equality exactly when $\\rho_0=\\delta_1$. The decoder itself is derived from the factor graph by expectation-consistent message passing with uniform diagonalization, which approximates per-block messages by Gaussians with a common precision.","core_discovery":"On its own terms, the paper establishes that the spatially coupled sparse superposition (SC-SS) code, decoded by the proposed SC-VAMP decoder, is capacity-achieving over the additive white Gaussian noise channel when every design matrix satisfies the spectra criterion: the limiting spectral density $\\rho_{\\mathrm{supp},r}$ of $B^{-1}A_r^T A_r$ converges to the point mass $\\delta_1$ as the section size $B\\to\\infty$ with aspect ratio $\\alpha\\to 0$. The proof works through state evolution: for rates $R_{\\mathrm{all}}$ below the information-theoretic threshold $R_{\\mathrm{IT}}=\\frac12\\int_0^1 F(x)\\,dx$, Proposition 1 shows the per-block error indicator $\\psi_c^k$ is driven to zero in a wave from the outermost blocks inward, so after $K=1+\\lceil\\Gamma/(2g)\\rceil$ iterations every block is recovered. The Cauchy-Schwarz bound $F(x)\\le(\\sigma^2+x)^{-1}$ gives $R_{\\mathrm{IT}}\\le\\frac12\\log(1+\\mathrm{snr})=C$, and equality is attained in the point-mass limit, so every rate $R_{\\mathrm{all}}<C$ is decodable. This confirms the VAMP-based capacity conjecture for rotational invariant designs.","pith_inferences":["If the state-evolution tracking can be proven non-asymptotically, the threshold-saturation argument should produce finite-length section error bounds that decay exponentially in the block length; the paper does not prove this step.","The spectra criterion is a universality condition, so the theorem should be read as covering any right-rotationally invariant design whose spectrum concentrates well, not just the Gaussian and DCT matrices simulated.","Because the capacity argument rests only on the Cauchy-Schwarz bound for $F(x)$, the same structure should extend to memoryless channels whenever the denoiser has the required phase transition; the paper mentions this only informally.","The condition $\\Gamma>W^2$ with $W\\to\\infty$ means the coupling overhead $\\vartheta$ tends to one only asymptotically; quantifying the required coupling width for a fixed gap to capacity is a natural next step that the paper leaves open."],"forward_implications":["Under the paper's state-evolution analysis, any rate $R_{\\mathrm{all}}<C$ is decodable with SC-VAMP provided the design matrices satisfy the spectra criterion and the coupling parameters are chosen sufficiently wide.","Decoding error disappears in an inward-moving wave: the outermost blocks are recovered first and the central blocks last, confirming threshold saturation for VAMP-based spatially coupled decoding.","The spectra criterion makes the result universal across right-rotationally invariant designs: Gaussian matrices and structured DCT/Hadamard matrices both admit capacity-achieving decoding, with fast transforms reducing per-iteration cost to $O(BL\\log(BL))$.","In the large-system limit, SC-VAMP reaches the same capacity threshold as VAMP with exponential power allocation, while the simulations here show SC-VAMP attains a lower section error rate at practical finite block lengths.","If state-evolution tracking remains valid at finite sizes, the same threshold-saturation structure should yield non-asymptotic section error bounds, the extension the paper lists as future work."],"supporting_citations":[{"why":"Introduces the VAMP decoder for sparse superposition codes and conjectures that it reaches capacity under the spectra criterion.","marker":"[1]"},{"why":"Proves capacity for VAMP with exponential decay power allocation under the spectra criterion and supplies the baseline for the empirical comparison.","marker":"[2]"},{"why":"Supplies the spatial-coupling construction for sparse superposition codes with AMP decoding that the SC-VAMP design adapts.","marker":"[12]"},{"why":"Provides the phase transition of the denoiser error function $E_2(\\gamma)$ used in the rescaled limit state evolution.","marker":"[13]"},{"why":"Rigorously establishes capacity-achieving spatially coupled sparse superposition codes with AMP decoding, giving the coupling structure and proof style extended here.","marker":"[16]"},{"why":"Defines the vector approximate message passing algorithm whose message-passing rules the SC-VAMP decoder extends.","marker":"[22]"},{"why":"Develops spatially coupled orthogonal AMP and the uniform diagonalization step used to concatenate block messages.","marker":"[28]"},{"why":"Provides the expectation-consistent inference machinery, including uniform diagonalization, used in the factor-graph derivation.","marker":"[30]"},{"why":"States the spectra criterion and the informal expectation that VAMP-based sparse superposition codes reach capacity for memoryless channels.","marker":"[32]"}],"fun_headline_variants":["SC-VAMP decoder hits Shannon limit for sparse codes","Spatially coupled VAMP achieves capacity in Gaussian channel","Capacity-achieving sparse superposition codes via SC-VAMP","SC-VAMP reaches channel capacity with state evolution proof","Sparse codes hit capacity with spatially coupled VAMP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the decoder's average error in each block is exactly what the state-evolution recursion predicts, and that the coupling can be made arbitrarily wide so the rate loss vanishes; if either fails for the concatenated SC-VAMP update, the capacity claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["SC-VAMP decoder hits Shannon limit for sparse codes","Spatially coupled VAMP achieves capacity in Gaussian channel","Capacity-achieving sparse superposition codes via SC-VAMP","SC-VAMP reaches channel capacity with state evolution proof","Sparse codes hit capacity with spatially coupled VAMP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000624,"raw_usage":{"total_tokens":2901,"prompt_tokens":965,"completion_tokens":1936,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":1858}},"tokens_in":581,"tokens_out":1936,"duration_ms":12487,"temperature":1.0,"reasoning_tokens":1858,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:03:31.747899+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SC-VAMP on a large but finite instance with a right-rotationally invariant design matrix whose limiting spectrum is the point mass $\\delta_1$, at a rate below $\\frac12\\log(1+\\mathrm{snr})$, and compare the measured per-block mean squared error with the state-evolution prediction; a mismatch that does not vanish as the dimensions grow would falsify the SE tracking and hence the capacity claim.","supporting_citations":[{"cited_title":"Orthogonal approximate message-passing for spatially coupled linear models,","cited_arxiv_id":null,"evidence_quote":"Develops spatially coupled orthogonal AMP and the uniform diagonalization step used to concatenate block messages."},{"cited_title":"Sparse superposition codes under V AMP decoding with generic rotational invariant coding matrices,","cited_arxiv_id":null,"evidence_quote":"Introduces the VAMP decoder for sparse superposition codes and conjectures that it reaches capacity under the spectra criterion."},{"cited_title":"Capacity- achieving sparse regression codes via vector approximate message pass- ing,","cited_arxiv_id":null,"evidence_quote":"Proves capacity for VAMP with exponential decay power allocation under the spectra criterion and supplies the baseline for the empirical comparison."},{"cited_title":"Approximate message-passing decoder and capacity achieving sparse superposition codes,","cited_arxiv_id":null,"evidence_quote":"Supplies the spatial-coupling construction for sparse superposition codes with AMP decoding that the SC-VAMP design adapts."},{"cited_title":"Proof of threshold saturation for spatially coupled sparse superposition codes,","cited_arxiv_id":null,"evidence_quote":"Provides the phase transition of the denoiser error function $E_2(\\gamma)$ used in the rescaled limit state evolution."},{"cited_title":"Capacity-achieving spatially coupled sparse superposition codes with AMP decoding,","cited_arxiv_id":null,"evidence_quote":"Rigorously establishes capacity-achieving spatially coupled sparse superposition codes with AMP decoding, giving the coupling structure and proof style extended here."},{"cited_title":"Vector approximate message passing,","cited_arxiv_id":null,"evidence_quote":"Defines the vector approximate message passing algorithm whose message-passing rules the SC-VAMP decoder extends."},{"cited_title":"Expecta- tion consistent approximate inference: Generalizations and convergence,","cited_arxiv_id":null,"evidence_quote":"Provides the expectation-consistent inference machinery, including uniform diagonalization, used in the factor-graph derivation."},{"cited_title":"Sparse superposition codes with rotational invariant coding matrices for memoryless channels,","cited_arxiv_id":null,"evidence_quote":"States the spectra criterion and the informal expectation that VAMP-based sparse superposition codes reach capacity for memoryless channels."}],"review_version":1}