{"id":"9bbe2fd5-55a8-464d-baf4-4f350865b2d3","arxiv_id":"2501.07030","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A diffusion-model-based signal detector is claimed to beat ML detection, but the supporting theory has algebraic errors and the claim contradicts ML optimality.","lead":"This paper proposes using a diffusion model to clean noise from received wireless signals before decoding, claiming it can beat the maximum likelihood detector that is normally considered optimal. The mathematical derivation is flawed, and the claim contradicts a basic result in detection theory, so the main result should not be taken at face value.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's SNR-to-timestep formula is algebraically wrong: at 15 dB Eq. (21) gives t≈0.015, the correct root of Eq. (19) is t≈0.030, and Table II lists 0.0467, so the t/α framework and the distribution-alignment argument built on it collapse.","rationale":"I read the paper as claiming a new theory that lets a diffusion receiver beat ML in AWGN MIMO. The load-bearing condition is that Theorem 1 and Theorem 2 correctly map each SNR to a diffusion timestep and a scaling that makes the input in-distribution. The algebraic check above shows this condition fails internally: the closed-form t does not solve the paper's own matching equation, and Table II is inconsistent with either formula. This is not a disagreement with the community consensus about ML; it is a direct correctness defect in the manuscript. It also explains the reader's distribution-mismatch observation: the mismatch is not a subtle second-order effect but a consequence of using wrong t and α. I do not rest the objection on the general claim that no receiver can beat ML, because the baseline is joint ML and symbol error is not the criterion minimized by that baseline; the internal error is sufficient. Because the reader already rejected the paper and this concern reinforces that rejection, I recommend no change to the verdict. No ad hominem is intended; the findings concern the equations and tables.","tokens_in":20414,"tokens_out":12486,"duration_ms":124930,"concrete_test":"Analytically re-derive Eq. (21) from Eq. (19) for the case γ=Es∥Hs∥². Substitute σ²=10^{-1.5} and report the root t; if the result is not Table II's 0.0467, recompute α from Eq. (22) with the corrected t and verify whether αr and xt have matched per-component noise variance. Then rerun the BPSK 8×8 experiment of Fig. 6 using the corrected (t,α): if the SER advantage over ML disappears, the claim is unsupported; if it survives, the advantage comes from the neural network, not from the stated theory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires three linked steps: (i) Theorem 1 maps channel SNR to a diffusion timestep t; (ii) Theorem 2 scales the received signal so that αr is approximately a diffusion latent; (iii) the trained denoiser then reduces noise below the channel level. Step (i) is already invalid. In the proof of Theorem 1, after setting x0=Hs as in Remark 2, Eq. (19) reduces to 1/σ² = (1−t)²/t. Solving the quadratic t²−(σ^{-2}+2)t+1=0 gives t = [σ^{-2}+2 − sqrt(σ^{-4}+4σ^{-2})]/2. The paper instead states Eq. (21), t = [2σ²+1 − sqrt(1+4σ²)]/(4σ²), which is not this solution. At 15 dB, σ²=10^{-1.5}: Eq. (21) yields t≈0.015, the root of Eq. (19) is t≈0.030, and Table II reports t=0.0467. These three values cannot be reconciled by any unit convention. Moreover, Table II sets α=1−t, which would require α²σ²=t, i.e. (1−t)²σ²=t; at 15 dB this gives 0.0287 instead of 0.0467. The proof of Theorem 2 explicitly concedes that matching mean norms 'cannot directly guarantee that the distribution of r and xt is consistent' and then asserts the missing distributional match without proof. Since the claimed optimality of the timestep in Fig. 6 and the OOD-free scaling in Fig. 7 both depend on these numbers, the theoretical foundation of the SER claim is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a diffusion-model-based MIMO signal detector that claims to achieve lower symbol error rate than maximum-likelihood detection for BPSK and QAM in AWGN channels, while maintaining O(n^2) complexity. The theoretical part develops an SDE description of diffusion, derives an SNR-to-timestep mapping (Theorem 1), and introduces a scaling rule intended to make received signals in-distribution for the diffusion model (Theorem 2). The practical part uses a Diffusion Transformer backbone and reports simulations for Nr = 4, 8, 16 and modulations up to 256-QAM, including ablations over timestep, scaling, and network size.","tokens_in":20917,"tokens_out":9859,"duration_ms":90115,"significance":"If the claims were correct, the paper would overturn the long-standing optimality of ML detection for equally likely symbols and would provide a principled way to generalize a learned detector across SNRs without retraining. The paper also proposes a computationally attractive one-step denoising receiver and addresses a real problem, namely the poor SNR generalization of neural detectors. These are meaningful contributions. However, the derivation contains elementary algebraic and probabilistic errors: the SDE reduction is internally inconsistent, the closed-form solution in Theorem 1 does not solve its own equation, the reported timesteps in Table II are not reproducible from the stated formulas, and the distribution-alignment argument in Theorem 2 is only a norm-matching condition, not a distributional match. The central claim of outperforming ML is therefore not supported by the presented theory or by reproducible experiments.","major_comments":[{"comment":"The SDE derivation is internally inconsistent. Setting g_t = 1 after Eq. (10) discards the variance-preserving structure of the DDPM forward process; the subsequent constraint Eq. (12) (∫_0^t h_s ds + x_0 = 0) combined with Eq. (11) implies that x_t is pure Brownian noise independent of x_0. This directly contradicts Eq. (14), which keeps the x_0 term. Eq. (16) is then asserted without derivation, and its variance Δt(t−Δt)/t I does not follow from the reverse SDE in Eq. (15). Since the denoising update in Eq. (17) and the one-step assumption in Remark 1 rest on these equations, the theoretical foundation of the method is not established.","section":"Section IV-A, Eqs. (10)-(14)"},{"comment":"The closed-form solution in Eq. (18) does not solve Eq. (19). With X = E[||Hs||^2]/(σ^2γ), Eq. (19) reduces to X = (1−t)^2/t, whose root in (0,1] is t = (2+X − sqrt(X^2+4X))/2; Eq. (18) divides by 4 instead of 2, giving a factor-of-two error. The same factor-of-two error appears in Remark 2, Eq. (21): with γ = E[||Hs||^2], the correct root is (2σ^2+1 − sqrt(1+4σ^2))/(2σ^2), not the expression with 4σ^2 in the denominator. The proof also sets E[||ϵ||^2] = 1, which is only valid for scalar noise; for a d-dimensional standard Gaussian vector, E[||ϵ||^2] = d, so the derived timestep depends on the signal dimension in an unacknowledged way.","section":"Theorem 1, Eqs. (18)-(21)"},{"comment":"Table II is not reproducible from Eq. (21) or from Eq. (19). At 15 dB SNR, σ^2 = 10^{-1.5}: Eq. (21) gives t ≈ 0.014, the correct root of Eq. (19) gives t ≈ 0.030, and Table II lists t = 0.0467. The table also sets α = 1 − t, which is compatible with the norm-matching equation only if (1−t)^2σ^2 = t; at 15 dB, (1−t)^2σ^2 ≈ 0.029 while t = 0.0467. Thus the tabulated values cannot have been computed from Theorems 1 and 2, and the validation of the 'optimal' timestep in Fig. 6 is not supported by the reported data.","section":"Table II and Section VI-D"},{"comment":"Matching the expected squared norm of αr to that of x_t does not yield a distributional match; the proof itself concedes that the scaling 'cannot directly guarantee that the distribution of r and xt is consistent.' The subsequent assertion that r and x_t share the same distribution form because both are linear combinations of a signal and zero-mean Gaussian noise is not a proof, and the per-component noise variances differ: Var(αr | s) = α^2σ^2, whereas Var(x_t | x_0) = t under the paper's convention. With α = 1−t and the tabulated t values, α^2σ^2 ≠ t (e.g., ≈0.029 vs 0.047 at 15 dB), so the denoiser is fed out-of-distribution inputs. The OOD robustness claimed in Fig. 7 therefore does not follow from the theory.","section":"Theorem 2 and Section IV-B"},{"comment":"The estimation step is not derived. Under the forward model x_t = (1−t)x_0 + √t ϵ, the natural reconstruction would be Hs = (αr − √t ϵ)/(1−t) or the appropriate posterior mean, but Eq. (24) instead uses Hs = αr − t h_t − ϵ_t with no normalization and no derivation. Combined with the absence of a proof that this receiver can strictly outperform the ML decision rule for equally likely symbols in AWGN, the central performance claim in the abstract is not backed by a valid theoretical argument, and the simulation results cannot compensate for the inconsistencies in the supporting formulas.","section":"Section V, Eqs. (23)-(25)"}],"minor_comments":[{"comment":"The channel matrix H is defined as N_r × N_r, but a MIMO system normally has N_t transmit and N_r receive dimensions; the model and complexity statements should specify the full dimensions.","section":"Section III"},{"comment":"The notation in the variance computation, 'D[ϵ] − E[ϵ]^2', is unclear for vector-valued ϵ and should be written as E[||ϵ||^2]; as written it obscures the dimension dependence discussed above.","section":"Theorem 1 proof"},{"comment":"In the computation of E[||x_t||^2], the text writes E_{x0}[||x0||] instead of E[||x0||^2]; the missing square propagates into the final formula for α.","section":"Theorem 2 proof"},{"comment":"Equation (13) places a stochastic integral in the mean of a Gaussian distribution, which is not meaningful; the conditional distribution should be stated after conditioning on the drift path or with the drift treated accordingly.","section":"Eq. (13)"},{"comment":"Several figure captions and axis labels are garbled or incomplete (e.g., the axis text in Fig. 3 and Fig. 8), and Fig. 8's caption is partially unreadable; these should be cleaned for publication.","section":"Figures 3-10"},{"comment":"The sentence 'we propose a DM-based signal detection method is proposed as follows' is ungrammatical and should be rewritten.","section":"Section V, sentence after Eq. (23)"},{"comment":"The reverse SDE in Eq. (15) is attributed to [18], a preprint on decoupled diffusion models; the paper should instead cite the standard score-based SDE literature (e.g., Song et al., 2021) for this equation.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript claims a fundamental result, that diffusion-model denoising can beat ML detection, but the supporting derivation contains elementary algebraic errors and an internally inconsistent SDE treatment. The tabulated timesteps do not match the stated formulas, and no code is provided to verify the simulations. I see no way to repair the central theoretical claims within a normal revision cycle; a future submission might be viable if the theoretical claims were substantially weakened and the empirical study were presented honestly as a heuristic detector comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a genuinely new combination — diffusion-model denoising as a MIMO receiver with an SNR-dependent starting timestep and scaling to avoid out-of-distribution inputs — and the authors deserve credit for seeing the OOD problem explicitly. But the theory is not coherent, and the headline claim about beating ML is a load-bearing result that the math does not carry.\n\nThe SDE derivation in Section IV-A is shaky: setting g_t=1 and Eq. (12) turns x_t into pure noise, and Eq. (16) is asserted without derivation. More concretely, Theorem 1's closed-form solution does not solve its own equation. With x0=Hs, Eq. (19) reduces to 1/σ²=(1−t)²/t, whose root at 15 dB is about 0.030. The paper's Eq. (21) gives about 0.015, and Table II reports 0.0467. Three different answers. Table II's α=1−t also cannot be reconciled with Theorem 2's formula or with the noise-level alignment: the scaled received noise variance α²σ² equals about 0.029 at 15 dB while the diffusion latent's noise variance t is 0.047. The proof of Theorem 2 admits the distribution match is not guaranteed and then asserts it. Given that the claimed SER advantage and the optimality plots in Figs. 6-7 rest on these numbers, the central claim is unsupported.\n\nOne nuance the reader's report got wrong: beating joint ML on per-symbol SER in MIMO is not impossible, because joint vector ML is not the symbol-wise optimal rule. So the paper would not \"overturn a foundational result\" even if its numbers were right. But the paper's own \"beyond ML\" framing still overstates the case, and the theoretical machinery does not do the work.\n\nFurther softness: no code, no error bars, and Fig. 8 scores high-SNR performance with a separately retrained model, which undercuts the \"no fine-tuning\" claim. The ablation study (depth and width) is a good instinct, but it cannot rescue the foundation.\n\nWho should read it: someone working on generative-model receivers might skim it for the architecture idea; nobody should rely on the theorems. I would not send the current version to review. The authors should fix the quadratic, correct the scaling argument, and release code; then the idea may deserve a fair hearing.","headline":"A well-intentioned but mathematically broken attempt to turn diffusion denoising into a MIMO receiver; the SNR-to-timestep theorem does not solve its own equation, so the headline result is unsupported.","tokens_in":21394,"tokens_out":4409,"would_cite":false,"duration_ms":44448,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a diffusion-model receiver exceeds maximum-likelihood detection for BPSK and QAM in AWGN MIMO channels.","keywords":["diffusion model","signal detection","maximum likelihood","MIMO","AWGN","symbol error rate","denoising diffusion","stochastic differential equation"],"falsifier":"At 15 dB SNR the paper's own Table II gives $t=0.0467$ but $\\alpha^2\\sigma^2\\approx0.0287$; feed the scaled received signal $\\alpha r$ into the trained network and measure the variance of $\\hat{Hs}-Hs$. If that residual variance is not smaller than the channel noise variance, the distribution-alignment step and the claimed SER advantage fail.","tokens_in":20212,"feed_emoji":"📡","tokens_out":9666,"duration_ms":84639,"temperature":0.7,"pith_summary":"The paper tries to establish that a denoising diffusion model can serve as a receiver front end that removes additive white Gaussian noise from received MIMO signals, so that symbol detection can beat maximum-likelihood (ML) estimation, long regarded as optimal, at lower computational cost. It develops a continuous-time SDE view of diffusion, proves a formula mapping any channel SNR to a diffusion timestep, and proves a scaling rule that maps the received signal into the diffusion latent distribution. With a diffusion transformer (DiT) backbone, the resulting receiver is reported to achieve lower symbol error rate than ML for BPSK and QAM in simulations while keeping complexity at $\\mathcal{O}(n^2)$. If correct, this would mean the noise at the detector is no longer treated as irreducible, and neural receivers could exceed the classical optimal detector.","feed_headline":"Diffusion receiver beats ML detection in simulations","feed_subtitle":"A diffusion-model receiver claims lower symbol error than maximum likelihood at quadratic cost in MIMO AWGN channels.","key_machinery":"The load-bearing object is the continuous-time diffusion process written as the SDE $dx_t=f_t x_t dt+g_t dw_t$, specialised here to latents $x_t=(1-t)x_0+\\sqrt{t}\\epsilon$ with $x_0=Hs$. Theorem 1 solves for $t$ by equating the expected signal-to-noise ratio of $x_t$ with that of the received signal, and Theorem 2 chooses $\\alpha$ so that $E[\\|\\alpha r\\|^2]=E[\\|x_t\\|^2]$. A diffusion transformer (DiT), a Transformer whose self-attention blocks are conditioned on the timestep through scale-and-shift, tokenizes $\\alpha r$ and predicts $h_t$ and $\\epsilon_t$; one denoising step ($\\Delta t=t$) then gives an estimate of $Hs$, and the final symbol estimate is $H^{-1}\\hat{Hs}$.","core_discovery":"On the paper's own terms, the central discovery is that additive Gaussian noise in a received signal can be erased by treating the received signal as a diffusion latent at a suitably chosen timestep. Theorem 1 gives the timestep $t$ that matches the expected signal-to-noise ratio of the diffusion latent $x_t=(1-t)Hs+\\sqrt{t}\\epsilon$ to that of the received signal, and Theorem 2 gives a scaling factor $\\alpha$ that matches the expected squared norm of $\\alpha r$ to that of $x_t$. In this setup $x_0=Hs$, so the trained diffusion transformer predicts the latent components $h_t$ and $\\epsilon_t$, the signal component is recovered as $\\hat{Hs}=\\alpha r-t\\hat{h}_t-\\hat{\\epsilon}_t$, and channel inversion yields the symbol estimate. The simulations report symbol error rates below the ML baseline for BPSK, 4QAM, 16QAM, and 64QAM over $4\\times4$, $8\\times8$, and $16\\times16$ MIMO AWGN channels with one-step denoising.","pith_inferences":["If the advantage is real, the classical statement that ML is the optimal detector is conditional: it holds for receivers that treat channel noise as irreducible, and a denoising preprocessor can move the effective operating point beyond it.","The paper's own Table II shows $\\alpha^2\\sigma^2\\neq t$ (for example, at 15 dB SNR, $\\alpha^2\\sigma^2\\approx0.0287$ while $t=0.0467$), so the claimed distribution alignment is only approximate; a clean follow-up would isolate how much of the gain survives when the per-component noise variance is deliberately mismatched.","The same ‘match average power by scaling’ recipe could transfer to other denoising-inverse problems with known Gaussian noise levels, but that transfer is not established by the paper and would need separate tests.","A stress test that fixes the total norm while changing the per-component noise variance would reveal whether the reported SER advantage comes from the diffusion denoiser itself or from the subsequent zero-forcing/channel-inversion step."],"forward_implications":["A single trained diffusion receiver can operate across a wide range of SNRs without fine-tuning, because $t$ and $\\alpha$ are computed in closed form from the channel noise power.","The method gives a concrete $\\mathcal{O}(n^2)$ inference path where ML enumeration is exponential, so high-order QAM and large MIMO become feasible in the reported setting.","Because denoising happens before the symbol decision, the effective noise seen by the detector is smaller than the channel noise, which is the stated mechanism for exceeding ML.","One-step denoising keeps inference latency at roughly a single neural-network forward pass rather than a long iterative diffusion sampling chain.","The ablation results indicate the gain comes from the diffusion-based formulation rather than from network capacity, since changing DiT depth and width barely changes SER."],"supporting_citations":[{"why":"Defines the DDPM forward and reverse Markov chains and the training objective the method is built on.","marker":"[13]"},{"why":"Supplies the SDE and reverse-SDE formulation used in the paper's theoretical proof.","marker":"[17]"},{"why":"Supplies the reverse update equations and the $h_t$/$\\epsilon_t$ prediction loss used by the DiT.","marker":"[18]"},{"why":"Provides the self-attention mechanism and the quadratic complexity analysis for the DiT backbone.","marker":"[20]"},{"why":"The DnCNN-based detection baseline that the proposed method compares against and extends.","marker":"[6]"},{"why":"The Transformer-based detection baseline the proposed method must beat in the simulations.","marker":"[9]"}],"fun_headline_variants":["Diffusion model erases noise, beats ML detection","Diffusion receiver outperforms ML in MIMO","Erase noise with diffusion, beat ML","Diffusion detection beats ML at lower cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the assumption that scaling the received signal to the same average power as a diffusion training latent is enough to make the network's input match its training distribution, including the noise level per entry; if that fails, the denoiser is out of distribution and the claimed SER gain collapses.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model erases noise, beats ML detection","Diffusion receiver outperforms ML in MIMO","Erase noise with diffusion, beat ML","Diffusion detection beats ML at lower cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000404,"raw_usage":{"total_tokens":2133,"prompt_tokens":1001,"completion_tokens":1132,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":1074}},"tokens_in":617,"tokens_out":1132,"duration_ms":8507,"temperature":1.0,"reasoning_tokens":1074,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:49.847945+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At 15 dB SNR the paper's own Table II gives $t=0.0467$ but $\\alpha^2\\sigma^2\\approx0.0287$; feed the scaled received signal $\\alpha r$ into the trained network and measure the variance of $\\hat{Hs}-Hs$. If that residual variance is not smaller than the channel noise variance, the distribution-alignment step and the claimed SER advantage fail.","supporting_citations":[{"cited_title":"Decoupled diffusion models: Simultaneous image to zero and zero to noise,","cited_arxiv_id":null,"evidence_quote":"Supplies the reverse update equations and the $h_t$/$\\epsilon_t$ prediction loss used by the DiT."},{"cited_title":"Symbol denoising in high order m-qam using residual learning of deep cnn,","cited_arxiv_id":null,"evidence_quote":"The DnCNN-based detection baseline that the proposed method compares against and extends."},{"cited_title":"Sigt: An efficient end-to-end mimo-ofdm receiver framework based on transformer,","cited_arxiv_id":null,"evidence_quote":"The Transformer-based detection baseline the proposed method must beat in the simulations."}],"review_version":1}