{"id":"9adf7e4a-0c50-40b9-aa27-9c39f27ce5d7","arxiv_id":"2501.11860","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"BD-QMAP, a Bayesian despeckler adapted from Q-MAP, is applied to piecewise constant Markov sources and paired with a theoretical MSE lower bound.","lead":"This paper adapts the Q-MAP Bayesian denoiser to multiplicative speckle noise, yielding a despeckling method called BD-QMAP for structured stationary sources. It simplifies the method for piecewise constant Markov sources and derives a lower bound on the achievable mean squared error.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's lower-bound proof relies on the false premise that the true jump locations minimize MSE; the actual objective is log-likelihood, so the bound does not apply to the implemented BD-QMAP.","rationale":"The reader identified the eta-coefficient discrepancy between Eq. (5) and Lemma 1 and the Corollary 1 sign mismatch, both of which are real. My independent reading found a deeper problem in the proof of Theorem 1 itself: the proof assumes that the true jump locations minimize the squared-error loss, but the estimator defined in Eq. (13) minimizes a negative log-likelihood. These two objectives need not agree, and there are plausible regimes where they disagree sharply (close levels that are better merged, constant segments that are better split). If that is the case, the genie-aided lower bound does not lower-bound the actual BD-QMAP solution, even when k is known. The conditional verdict would require fixing the displayed algebra, but the proof's key inequality is structurally unsupported, not merely misprinted. A small exhaustive enumeration test can settle whether the inequality is false; if it fails, Theorem 1 cannot be used as a performance certificate for BD-QMAP, and the central theoretical claim of the paper collapses, leaving only the empirical results. I therefore recommend REJECT rather than CONDITIONAL, because the main theoretical contribution is not a matter of missing details but of an invalid inference from a genie estimator to the proposed algorithm.","tokens_in":14942,"tokens_out":18846,"duration_ms":203934,"concrete_test":"Construct a small synthetic instance, e.g. n=6, X = [1, 1, 1.1, 1.1, 1.1, 1.1] (one true jump), with W i.i.d. N(0,1). For many noise draws, compute: (a) the MSE of the genie-aided estimator that knows the true jump locations and applies Eq. (8); and (b) the MSE of the segmentation that minimizes Eq. (13) with k=1, found by exhaustive enumeration. If the average MSE of (b) is below the average MSE of (a), the central inequality in the proof of Theorem 1 is false. Repeat this comparison for sources generated from the actual first-order Markov prior to determine whether the theorem can bound the implemented BD-QMAP estimator.","verdict_should_be":"REJECT","load_bearing_attack":"The most load-bearing concern is in the proof of Theorem 1 (§4.2). After assuming the number of jumps k is known, the paper defines the estimator as the minimizer of Eq. (13), a negative log-likelihood over all segmentations with k jumps. The proof then lower-bounds its MSE by the MSE of a genie-aided ML estimator that is given the true jump locations, asserting that 'the expected error is minimized when the locations of the jumps are detected correctly.' This assertion is not justified by the optimization: Eq. (13) minimizes a likelihood, not a squared-error loss. For sources with close adjacent levels, merging two neighboring levels can dramatically reduce MSE while changing the likelihood only slightly; conversely, splitting a constant segment lowers the log-likelihood (because smaller samples make the log-mean smaller) but increases estimator variance. Hence the minimizer of Eq. (13) can have lower expected MSE than the true-location genie estimator, which would reverse the inequality used in Eqs. (16)-(17). Even if the algebraic constant and sign errors in Corollary 1 were repaired, Theorem 1 as stated would still not certify the MSE of the actual BD-QMAP estimator (Eq. 3), which does not know k. The theoretical lower bound therefore targets a different, genie-aided problem than the one solved by the proposed method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BD-QMAP, a Bayesian despeckling method for multiplicative Gaussian speckle, defined in Eq. (3) as the minimization of a negative log-likelihood fidelity term plus a regularizer built from quantized empirical statistics. For piecewise-constant first-order Markov sources, the paper claims in Lemma 1 that the method reduces to a jump-penalized segmentation problem, and it derives a lower bound on reconstruction MSE in Theorem 1 and Corollary 1. Experiments on synthetic piecewise-constant sources compare BD-QMAP with classical filters and with a genie-aided ML estimator, reporting improved PSNR. The practical algorithm is clearly implementable via Viterbi decoding, but the theoretical claims contain several unresolved inconsistencies.","tokens_in":15243,"tokens_out":11764,"duration_ms":115394,"significance":"If the theoretical results were valid, the paper would make a useful contribution: a tractable Bayesian treatment of multiplicative noise for structured sources, a concrete Viterbi implementation, and a certified performance benchmark. The experimental comparison is informative and includes a parameter study of the regularization weight and quantization level. However, the advertised lower bound and the claimed equivalence between BD-QMAP and a jump-penalized segmentation are not established: the regularizer coefficient in Eq. (5) is inconsistent with Lemma 1, the proof of Theorem 1 uses a false genie-aided optimality assertion, and the theorem and corollary disagree in signs and constants. These issues are load-bearing for the central theoretical claims.","major_comments":[{"comment":"The jump penalty in Eq. (5) has coefficient λ + η/b, with η = -λ log q0 - log(1 - q0 + q0 2^{-b}), while Lemma 1 in Eq. (7) and the proof in Eq. (10) use the coefficient λ + 1/b with no factor η. Since η is not identically equal to 1 (for example, with q0 = 0.1, λ = 1, b = 2, one obtains η ≈ 0.31), Lemma 1 does not describe the optimization in Eq. (5). The simplification of the BD-QMAP regularizer to a jump count is therefore unsupported as stated, and any theorem built on that simplification does not apply to BD-QMAP as defined.","section":"Section 3.2, Lemma 1, Eq. (5) and Eq. (10)"},{"comment":"The proof of Theorem 1 asserts that 'the expected error is minimized when the locations of the jumps are detected correctly.' This is not a property of the objective in Eq. (13), which minimizes negative log-likelihood, not squared-error risk. A segmentation that merges two close adjacent levels, or splits a long constant block, can have lower MSE than the true-location genie estimator while having a similar or even larger likelihood. Consequently, the inequality E[||X^n - hat X^n||^2] ≥ E[||X^n - hat X^{n,ML}||^2] used in Eqs. (16)-(17) is not justified. Without this inequality, Theorem 1 only computes the MSE of a genie-aided estimator with known jump locations; it does not lower-bound the MSE of the fixed-k estimator in Eq. (13), let alone the actual BD-QMAP optimization in Eq. (3).","section":"Section 4.2, Eq. (13) and following paragraph"},{"comment":"There are two independent discrepancies in the constant terms. First, Theorem 1 defines c1 = 3 - sqrt(2/π) - 2√π, but the proof in Eq. (23) evaluates the T = 1 and T = 2 contributions as q0(3 - sqrt(2/π) - √π - q0(2 - √π)), which requires c1 = 3 - sqrt(2/π) - √π rather than the printed value. Second, Corollary 1 changes the + q0^2 c2 term of Theorem 1 into - q0^2 c2. These are not cosmetic differences, since the resulting bound can change sign for small q0; as printed, the theorem, its proof, and the corollary cannot all be correct.","section":"Theorem 1, Corollary 1, and Eq. (23)"},{"comment":"The simulations use a uniform source distribution on [0,1], but Theorem 1 and the BD-QMAP objective in Eq. (3) are stated under the assumption x_m > 0. Moreover, for the b-bit quantized search space X_b^n, the value 0 is a valid grid point, and at u_i = 0 the fidelity term log u_i^2 is undefined (it diverges to -∞). The paper neither imposes x_m > 0 in the experiments nor specifies how the zero quantization bin is handled, so the comparison in Figure 2 and Table 1 between the empirical PSNR and the Theorem 1 bound is not covered by the stated assumptions.","section":"Section 5.3 and Theorem 1 assumptions"}],"minor_comments":[{"comment":"The abstract contains a typo: 'depseckler' should be 'despeckler'.","section":"Abstract"},{"comment":"In Eq. (10), the sum is written with index j but the summand uses i; this should be indexed consistently.","section":"Section 4.1, Eq. (10)"},{"comment":"The derivation of the regularizer simplification is imported from [27, Sec. 3.2] without being reproduced. Given that the coefficient mismatch in Lemma 1 depends on that derivation, the relevant argument from [27] should be stated or at least sketched.","section":"Section 3.2, Eq. (5) and Section 5.1"},{"comment":"Table 1 reports mean PSNR values without standard deviations or per-trial statistics; with 100 test signals, reporting only single numbers makes it difficult to judge whether the gains are stable across realizations.","section":"Table 1"}],"recommendation":"reject","confidential_remarks":"The experimental part is potentially salvageable, but the theoretical core is not. The inconsistency between Eq. (5) and Lemma 1 could be a fixable typo, but the proof of Theorem 1 relies on an optimality claim that is false for the given likelihood objective, and the theorem/corollary constants do not match the proof. Because the lower bound is a stated central contribution, this is not a local repair within the scope of a revision. I would not recommend acceptance until the theoretical claims are either replaced by a proof that genuinely bounds the implemented estimator or removed from the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful part of this paper is the algorithm: BD-QMAP adapts the Q-MAP regularizer to multiplicative noise, gives a tractable Viterbi implementation on quantized sequences, and the 1D piecewise-constant experiments show real gains over Lee, Kuan, Frost, and TV baselines. That part is solid and worth building on.\n\nThe theoretical part is not in shape. Theorem 1's proof hinges on the claim that, when the number of jumps is known, the expected squared error is minimized when the jump locations are detected correctly. That is not justified and is likely false: the estimator in Eq. (13) minimizes a negative log-likelihood, not a squared-error loss. For close adjacent levels or short segments, a segmentation that merges two true segments and splits a long constant run can lower expected MSE while changing the likelihood only slightly. Even if the algebra were fixed, the bound would target a genie-aided estimator that knows k and the true locations, not the BD-QMAP estimator from Eq. (3), which knows neither. The theorem as stated does not certify the proposed method.\n\nThere are also unresolved internal inconsistencies. Eq. (5) writes the jump penalty coefficient as λ + η/b, with η a nontrivial function of q0 and b, but Lemma 1 and the proof in Eq. (10) use λ + 1/b; no derivation reconciles them. Corollary 1 flips the sign of the q0²c2 term relative to Theorem 1, which looks like a typo but is unacknowledged. These are the kind of errors a referee should catch before publication.\n\nThe experiments are confined to 1D synthetic piecewise-constant signals; there are no SAR or optical images. The regularizer weights are estimated from 10^7 samples of the same source model used for testing—standard in this line of work, but worth stating as a limitation.\n\nWho is this for? People working on Bayesian despeckling or Q-MAP-style estimators. The empirical method is worth trying, and the simplified objective in Lemma 1 (modulo the coefficient issue) is a useful observation. But the theoretical claim of an MSE lower bound should not be taken as established.\n\nMy recommendation: send it to peer review, but flag that the theory section needs major repair. A serious referee should demand either a corrected proof of Theorem 1 or an explicit retreat to a conjecture, plus reconciliation of the penalty coefficient between Eq. (5) and Lemma 1.","headline":"A clean empirical extension of Q-MAP to speckle noise, but the theoretical lower bound is not proved and the paper's own equations don't agree.","tokens_in":15731,"tokens_out":3993,"would_cite":false,"duration_ms":46357,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","62F15","62M05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes BD-QMAP, a Bayesian despeckler for multiplicative Gaussian speckle, and proves that for piecewise-constant first-order Markov sources it reduces to jump-penalized segmentation with a guaranteed MSE lower bound.","keywords":["speckle noise","despeckling","Bayesian estimation","multiplicative noise","piecewise constant Markov source","quantized maximum a posteriori","MSE lower bound"],"falsifier":"Evaluate the full BD-QMAP objective of Eq. (5) and the simplified objective of Eq. (7) for a fixed piecewise-constant sequence with chosen $q_0$, $b$, and $\\lambda$; if their minimizers differ for any such instance, the claimed reduction in Lemma 1 is false and Theorem 1 does not bound BD-QMAP as defined.","tokens_in":14742,"feed_emoji":"📡","tokens_out":8271,"duration_ms":76728,"temperature":0.7,"pith_summary":"Speckle noise corrupts images in radar, holography, and optical coherence tomography, and many existing despecklers are heuristics without performance guarantees. This paper proposes BD-QMAP, a Bayesian estimator built from the negative log-likelihood of the multiplicative Gaussian noise model plus a regularizer that penalizes disagreement between the quantized empirical distribution of a candidate reconstruction and the known source statistics. For a piecewise-constant first-order Markov source, the regularizer becomes a penalty on the number of jumps, so the optimization reduces to finding a segmentation and averaging $Y_i^2$ over each segment. The paper then proves a lower bound on the mean squared error of such an estimator, giving a theoretical performance certificate that most existing despecklers lack.","feed_headline":"Despeckler for structured sources gets an MSE proof","feed_subtitle":"A Bayesian estimator reduces to jump-penalized segmentation; Theorem 1 certifies its error floor.","key_machinery":"The driving object is the regularizer $c_w(u^n)=\\sum_{a^k} w_{a^k}\\hat{p}_k(a^k|u^n)$, with weights $w_{a^k}=-\\log P([X^k]_b=a^k)$; it summarizes the source's $n$-dimensional distribution in $|X_b|^k$ quantized weights. In BD-QMAP this regularizer is added to the negative log-likelihood of the multiplicative Gaussian model $Y_i=X_iW_i$, $W_i\\sim\\mathcal{N}(0,1)$. The reduction to Eq. (7) is carried by the observation that, for the piecewise-constant first-order Markov source with $k=2$, the regularizer is proportional to the number of jumps $N_J([u^n]_b)$. The proof of Theorem 1 then combines a concentration bound for sums of geometric run lengths (Lemma 2) with a Gamma-ratio estimate to control within-segment estimation error.","core_discovery":"The paper's central claim is that BD-QMAP, defined in Eq. (3) as the minimizer of $\\frac{1}{n}\\sum_{i=1}^n \\left(\\log u_i^2 + \\frac{Y_i^2}{u_i^2}\\right) + \\frac{\\lambda}{b} c_w(u^n)$, is a theoretically grounded despeckler for structured stationary sources. For a piecewise-constant first-order Markov source with jump probability $q_0$ and $k=2$, it simplifies to the jump-penalized segmentation of Eq. (7): choose segment lengths $n_1, \\ldots, n_{k+1}$, minimize $\\sum_j n_j \\log\\left(\\frac{1}{n_j}\\sum_{l\\in I_j} Y_l^2\\right) + \\frac{n}{n-1}(\\lambda+\\frac{1}{b})k$, and reconstruct each segment as the square root of the average of the squared observations. Theorem 1 lower-bounds the expected per-sample MSE of this estimator, given the number of jumps $k=k(n)$, by an expression involving $q_0$, the second moment $\\eta_2=E[X_i^2]$, and a geometric run length $T$; Corollary 1 shows the remainder term is $O(n^{-1/4})$.","pith_inferences":["The same regularizer-to-sparsity reduction might be attempted for higher-order Markov or tree-structured sources; whether the quantized weights collapse to a jump count in those settings is a direct testable extension.","The bound's dependence on geometric run lengths suggests that any source with identically distributed segment lengths will exhibit the same qualitative floor, roughly $1/4$ per segment as $T$ grows.","The two-stage pipeline---coarse quantized jump detection followed by continuous ML refinement---suggests a transferable recipe for coherent imaging: discrete structure search first, then continuous parameter estimation.","A numerical check of the $\\eta$ coefficient in Eq. (5) for small $b$ would either close the gap between the stated objective and the proof or show that the theorem applies to a slightly different estimator."],"forward_implications":["BD-QMAP turns despeckling of piecewise-constant sources into a tractable jump-detection problem solvable by the Viterbi algorithm, followed by per-segment averaging of squared observations.","The lower bound of Theorem 1 gives a benchmark against which practical despecklers can be measured, including the genie-aided ML estimator used in the experiments.","The method does not require an upper bound on speckle power or local stationarity assumptions, so it can operate in high-noise regimes where linear adaptive filters saturate or bypass the input.","Because the regularizer is learned from quantized source statistics rather than a parametric prior, the approach is stated for general structured stationary sources, not only the piecewise-constant case tested here."],"supporting_citations":[{"why":"Supplies the Quantized MAP construction and the simplification results for structured sources that BD-QMAP extends to multiplicative noise.","marker":"[27]"},{"why":"Defines information dimension, the criterion used to characterize structured sources and justify the sparsity of the quantized weights.","marker":"[28]"},{"why":"Introduces the structured-source model and uses $q_0$ as the degree-of-structure parameter that drives the piecewise-constant experiments.","marker":"[29]"},{"why":"Provides the tail bound for sums of geometric random variables used as Lemma 2 in the proof of Theorem 1.","marker":"[30]"},{"why":"Supplies the logarithmic inequality used in proving Corollary 1's $O(n^{-1/4})$ remainder.","marker":"[31]"},{"why":"Provides the Viterbi algorithm that makes the BD-QMAPb optimization computationally efficient.","marker":"[32]"}],"fun_headline_variants":["Bayesian despeckling proves a floor for structured noise","Jump-penalized segmentation: proven error floor for speckle","MSE proof for Bayesian despeckler on structured sources","Despeckling via jump segmentation now has a theoretical guarantee","Bayesian despeckler for 1-Markov sources gets an error bound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theoretical results assume that the BD-QMAP regularizer for a piecewise-constant first-order Markov source is exactly equivalent to the simple jump-count penalty in Eq. (7); the proof uses the coefficient $\\lambda+1/b$, while the stated simplification in Eq. (5) has $\\lambda+\\eta/b$, so if that equivalence is not exact, Theorem 1 bounds a different optimization than the one BD-QMAP defines.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian despeckling proves a floor for structured noise","Jump-penalized segmentation: proven error floor for speckle","MSE proof for Bayesian despeckler on structured sources","Despeckling via jump segmentation now has a theoretical guarantee","Bayesian despeckler for 1-Markov sources gets an error bound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000528,"raw_usage":{"total_tokens":2540,"prompt_tokens":934,"completion_tokens":1606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":1515}},"tokens_in":550,"tokens_out":1606,"duration_ms":11656,"temperature":1.0,"reasoning_tokens":1515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:47:14.459932+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the full BD-QMAP objective of Eq. (5) and the simplified objective of Eq. (7) for a fixed piecewise-constant sequence with chosen $q_0$, $b$, and $\\lambda$; if their minimizers differ for any such instance, the claimed reduction in Lemma 1 is false and Theorem 1 does not bound BD-QMAP as defined.","supporting_citations":[{"cited_title":"Bayesian denoising of structured sources and its implications on learning-based denoising,","cited_arxiv_id":null,"evidence_quote":"Supplies the Quantized MAP construction and the simplification results for structured sources that BD-QMAP extends to multiplicative noise."},{"cited_title":"Rényi information dimension: Fundamental limits of almost lossless analog compression,","cited_arxiv_id":null,"evidence_quote":"Defines information dimension, the criterion used to characterize structured sources and justify the sparsity of the quantized weights."},{"cited_title":"Universal compressed sensing for almost lossless recovery,","cited_arxiv_id":null,"evidence_quote":"Introduces the structured-source model and uses $q_0$ as the degree-of-structure parameter that drives the piecewise-constant experiments."},{"cited_title":"Tailboundsforsumsofgeometricandexponentialvariables,","cited_arxiv_id":null,"evidence_quote":"Provides the tail bound for sums of geometric random variables used as Lemma 2 in the proof of Theorem 1."},{"cited_title":"Topsøe,Some Bounds for the Logarithmic Function","cited_arxiv_id":null,"evidence_quote":"Supplies the logarithmic inequality used in proving Corollary 1's $O(n^{-1/4})$ remainder."},{"cited_title":"The viterbi algorithm,","cited_arxiv_id":null,"evidence_quote":"Provides the Viterbi algorithm that makes the BD-QMAPb optimization computationally efficient."}],"review_version":1}