{"id":"baf5d06e-8436-494b-9ecb-052f99a6fdd8","arxiv_id":"2501.01616","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid analog-digital semantic communication framework uses a Cramer-Rao-bound loss and parity-bit refinement to improve image delivery for emergency wireless networks.","lead":"This paper combines analog deep-learning image coding with digital parity-bit refinement so low-power emergency transmitters can send usable scene images over bad wireless channels. The proposed scheme is tested on fire images and reports better reconstruction and fire-detection scores than standard separated and deep joint source-channel coding baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CRLB in Eq. (16) is derived for a surrogate distortion and mishandles the fading coefficient, so the claimed design objective is not a valid bound on reconstruction error.","rationale":"The paper's central claim is twofold: (i) the proposed DA-ESemCom framework outperforms SSCC and DL-based JSCC in fidelity and detection, and (ii) the CRLB in Eq. (16) provides a valid design objective for the semantic codec. The reader correctly identifies that the CRLB derivation is invalid as a bound on the actual reconstruction error. My independent reading confirms the same issue and adds that the probability model in Eq. (6) is internally inconsistent with the channel model in Eq. (1): the effective noise variance should depend on |h_i|^2 when conditioned on h_i, and the expectation over h_i in Eq. (11) should involve |h_i|^2, not h_i^2. The derivation also computes a Fisher information for the parameter s_i from a likelihood that is written for the received symbol y_i, but then equates the inverse Fisher information with the mean squared error between the encoder output and the received symbol. These are not the same quantity, and the CRLB is not a bound on the surrogate distortion of Eq. (4). Because this flawed bound is used directly in the optimization objective (18), the theoretical guidance for codec design is unsupported. I agree with the reader's verdict: the empirical section is suggestive but not independently verifiable, and the flawed CRLB is a load-bearing error. I see no reason to change the REJECT judgment. The proposed check, a small analytical/numerical comparison of Eq. (16) against the true end-to-end distortion for a scalar Gaussian source, would settle whether the concern lands definitively. If the CRLB derivation were corrected and the simulations were reproducible, the framework idea would be worth revisiting, but as written the central claim does not hold.","tokens_in":18067,"tokens_out":7294,"duration_ms":67293,"concrete_test":"Take a scalar Gaussian source s ~ CN(0, sigma_s^2) with f_E(s)=s, h ~ CN(0,1), semantic noise variance sigma_eps^2, and channel noise variance sigma_n^2. Compute the true MMSE reconstruction error E[(s-hat(s))^2] numerically and compare it with the right-hand side of Eq. (16). Also compute the Fisher information from the correct conditional density p(y|s,h) = CN(h s, |h|^2 sigma_eps^2 + sigma_n^2) and check whether Eq. (11) is recovered. If Eq. (16) does not lower-bound the true end-to-end distortion, the CRLB-based training objective is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is the CRLB derivation in Section II-C, which underwrites the training objective (18). Eq. (4) defines the semantic transmission distortion as E[(f_E(s_i)-y_i)^2], the error between the encoder output and the received symbol, not the end-to-end reconstruction error E[(s_i-hat(s)_i)^2] that the CRLB in Eq. (5) is claimed to lower-bound. The CRLB applies to the variance of an unbiased estimator of s_i; it cannot bound this surrogate quantity. Additionally, Eq. (6) treats the effective noise as CN(0, sigma_eps^2/(1+sigma_eps^2+sigma_n^2)) and Eq. (9)-(11) drop the fading coefficient by replacing E[(partial(h_i f_E)/partial s)^2] with E[(partial f_E/partial s)^2], even though h_i ~ CN(0,1) gives E[h_i^2]=0 and the conditional noise variance is |h_i|^2 sigma_eps^2 + sigma_n^2. Consequently, Eq. (16) is not a valid lower bound on reconstruction distortion, and the loss built from it is not a theoretically grounded design objective. Since this flawed CRLB is one of the paper's two main contributions, the central claim is not supported even if the empirical comparisons were reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DA-ESemCom, a digital-analog transmission framework for emergency wireless communications. The analog part is a deep-learning-based semantic codec with a performance-constrained model that incorporates semantic noise and channel noise; the digital part is a distributed source coding (DSC) scheme based on LDPC parity bits that refines the analog reconstruction. The authors derive a Cramér–Rao lower bound (CRLB) for the proposed semantic coding model, use it to define a training loss, and present simulations showing that the full framework outperforms classical SSCC (JPEG+LDPC) and DL-based JSCC baselines in PSNR, LPIPS, MS-SSIM, and object-detection mAP over a Rayleigh fading channel. The paper also includes an ablation study on model size.","tokens_in":18399,"tokens_out":7495,"duration_ms":65482,"significance":"The problem addressed is relevant and timely: emergency transmitters with limited hardware, no ARQ, and low SNR can benefit from robust semantic communication. The hybrid analog-digital architecture—using DSC parity to refine a semantic analog reconstruction—is a sensible and potentially useful engineering idea, and the empirical comparisons cover both fidelity and task-oriented detection metrics. The paper also attempts a theoretical justification of the codec design via CRLB, which, if valid, would be a valuable design principle. However, the theoretical derivation contains load-bearing errors: the distortion being bounded is not the end-to-end reconstruction error, the fading coefficient is treated inconsistently with the stated complex Gaussian model, and the final CRLB expression does not follow algebraically from the preceding equations. Because the CRLB is the basis for the training objective and the claimed performance bottleneck, the central theoretical contribution is not supported. The empirical results remain suggestive, but they are not sufficient to salvage the paper's main claims as written.","major_comments":[{"comment":"The distortion D_overall in Eq. (4) is defined as the expected squared error between the encoder output f_E(s_i) and the received symbol y_i, not as the end-to-end reconstruction error E[(s_i - \\hat{s}_i)^2]. The CRLB in Eq. (5) is a lower bound on the variance of an unbiased estimator of a parameter (here s_i) from the observation y_i. It cannot be applied to this surrogate quantity, which is not an estimator variance. Consequently, the bound in Eq. (5) and the training objective (18) built on it do not lower-bound the actual image reconstruction distortion.","section":"Section II-C, Eq. (4)–(5)"},{"comment":"The complex fading coefficient h_i ~ CN(0,1) is mishandled. In Eq. (1), the effective noise is h_i \\epsilon_n + n, whose conditional variance given h_i is |h_i|^2 \\sigma_\\epsilon^2 + \\sigma_n^2, not the value \\sigma_\\epsilon^2/(1+\\sigma_\\epsilon^2+\\sigma_n^2) used in Eq. (6). Moreover, for h_i ~ CN(0,1), E[h_i^2] = 0, so the replacement in Eq. (11) of E[(\\partial(h_i f_E)/\\partial s)^2] by E[(\\partial f_E/\\partial s)^2] is invalid. The Fisher information for this complex channel model should involve |h_i|^2 or an expectation over |h_i|^2; the derivation as written effectively treats h_i as a real unit-power scalar. This error propagates into Eqs. (13) and (16).","section":"Section II-C, Eqs. (1), (6), (9)–(11)"},{"comment":"Equation (16) does not follow algebraically from Eq. (13). From Eq. (13), the first term is \\sigma_\\epsilon^2 / [(1+\\sigma_\\epsilon^2+\\sigma_n^2) \\sum_l E[(\\partial f_E/\\partial s)^2]], whereas Eq. (16) gives a different numerator (\\sigma_\\epsilon^2(1+\\sigma_n^2)+\\sigma_n^2(1+\\sigma_\\epsilon^2)). The second term in Eq. (13) is \\sigma_{t,i}^2 \\sigma_{s,i}^2 \\sigma_\\epsilon^2 / L, which with \\rho_i = \\sigma_{t,i}/\\sigma_{s,i} equals \\rho_i^2 \\sigma_{s,i}^4 \\sigma_\\epsilon^2 / L, not \\sigma_\\epsilon^2/(L \\rho_i^2) as in Eq. (16). The design insight that \\rho_i^2 should be maximized and the subsequent loss (18) are therefore based on an incorrect formula.","section":"Section II-C, Eqs. (13)–(16)"},{"comment":"The conditional density p(t_i|s_i) is derived incorrectly. From s_i = t_i + \\epsilon_n with independent Gaussian t_i and \\epsilon_n, the conditional distribution of t_i given s_i is Gaussian with mean (\\sigma_t^2/(\\sigma_t^2+\\sigma_\\epsilon^2)) s_i and variance \\sigma_t^2 \\sigma_\\epsilon^2/(\\sigma_t^2+\\sigma_\\epsilon^2), not the expression in Eq. (8). This invalidates the Fisher information computation for the synonymous-mapping term and hence the corresponding CRLB term.","section":"Section II-C, Eq. (8)"}],"minor_comments":[{"comment":"The text says 'Estimate the various of the semantic noise'; 'various' should be 'variance'.","section":"Algorithm 1, line 13"},{"comment":"The symbol f_{Ieal}^E should be f_{Ideal}^E; the typo appears in Algorithm 1 and in Section II-D.","section":"Algorithm 1 and text"},{"comment":"The figure captions contain the misspelling 'Porposed' for 'Proposed'.","section":"Figures 3–6"},{"comment":"The acronym 'AWGM' should be 'AWGN'.","section":"Eq. (1)"},{"comment":"Reference [23] is Slepian and Wolf's noiseless coding of correlated sources, which is not the standard citation for the Cramér–Rao lower bound; a proper CRLB reference is needed. Reference [25] on texture stationarity is not the usual basis for the sample-variance formula in Eq. (7).","section":"References"},{"comment":"The baseline description does not specify the JPEG quality factor, the LDPC code rate, or the symbol-to-bandwidth mapping for the classical SSCC scheme, which are needed to reproduce the comparison.","section":"Section V-C"},{"comment":"The 'ideal semantic transmission' baseline is vaguely defined as 'full, noiseless, and high-dimensional semantic feature vectors with full-resolution constellation'; a quantitative definition of bandwidth, power, and constellation resolution is needed to interpret why the proposed DA-ESemCom can surpass it.","section":"Section V-C"},{"comment":"The notation \\hat{S} is used in Eq. (3) for the decoder output, but the final output of the DA-ESemCom framework is also denoted S in Section IV; please disambiguate these.","section":"Section III and Eq. (3)"}],"recommendation":"reject","confidential_remarks":"The paper's theoretical contribution is the CRLB-based design objective in Section II, and this derivation has fundamental errors that are not fixable by local corrections: the definition of distortion is a surrogate, the fading coefficient is treated inconsistently, and the algebra leading to Eq. (16) is wrong. The empirical DA-ESemCom framework may still be viable as a heuristic, but the paper's central claims about the CRLB and the performance bottleneck would require a substantially rewritten manuscript. The relationship to the authors' prior work in IEEE Network (reference [26]) should also be clarified in any revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — here's the short version: the DA-ESemCom framework is a sensible extension of the known hybrid digital-analog idea to emergency image transmission, and the simulations suggest it works. But the CRLB derivation in Section II-C, which is one of the paper's two main contributions, does not hold up. The distortion in Eq. (4) is defined as E[(f_E(s_i)-y_i)^2] — the error between the encoder output and the received symbol — not the end-to-end reconstruction error E[(s_i - \\hat{s}_i)^2]. The CRLB is a bound on the variance of an unbiased estimator of s_i; it cannot bound that surrogate quantity. The fading coefficient is also mishandled: with h_i ~ CN(0,1), E[h_i^2]=0, so replacing E[(∂h_i f_E/∂s)^2] with E[(∂f_E/∂s)^2] in Eq. (11) is wrong. And Eq. (16) does not follow arithmetically from Eq. (13). Because the training loss in Eq. (18) includes the CRLB-derived term, the 'DL-based JSCC (Proposed loss)' baseline is partly testing the loss that is supposed to be justified by the theory. That is circular.\n\nWhat is genuinely new: the emergency-communication framing, the synonymous-mapping model for dropped patches, and the lightweight power/bandwidth allocation that prioritizes the digital parity part. The paper does a decent job on the empirical side — PSNR, LPIPS, mAP, visual examples, and an ablation over model sizes. The authors also state the inherent fidelity bottleneck honestly. But no code or data are released, and the 'ideal semantic transmission' upper bound is not specified tightly enough to be a fair comparison.\n\nProportionate bottom line: the system idea may be sound and the empirical gains plausible, but the theoretical contribution as written is not valid. This is a paper for the semantic-communications and emergency-wireless crowd; they will find the architecture and the resource allocation interesting, but they should not cite the CRLB result.\n\nMy recommendation: don't desk-reject — the empirical work deserves scrutiny and the CRLB can be fixed or dropped. But I would send it back for major revision, and before I'd trust the numbers I'd want code/data and a fairer baseline.","headline":"Hybrid digital-analog semantic transmission is a plausible engineering idea, but the CRLB that anchors the theory is invalid, so the paper needs major revision.","tokens_in":18899,"tokens_out":3482,"would_cite":false,"duration_ms":31961,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid analog-digital semantic codec with a CRLB-trained encoder and distributed-source-coding parity refinement outperforms classical separated and purely neural joint source-channel coding for emergency image transmission.","keywords":["emergency wireless communication","semantic communication","digital-analog transmission","joint source-channel coding","distributed source coding","Cramér-Rao lower bound","image reconstruction","object detection"],"falsifier":"Train the same semantic codec with the true end-to-end mean squared error between source and final reconstruction as the loss and compare it against the proposed CRLB-based loss on the paper's fading-channel setup; if the CRLB-based loss does not improve PSNR or mAP, or if Eq. (11)'s gradient identity fails when complex Rayleigh fading is inserted directly, the claimed role of Eq. (16) as a design objective is not supported.","tokens_in":17868,"feed_emoji":"📡","tokens_out":10597,"duration_ms":96060,"temperature":0.7,"pith_summary":"The paper proposes a hybrid transmission scheme for emergency wireless image delivery, where the transmitter has limited neural-network capacity and UDP provides no retransmissions. The scheme pairs an analog deep-learning semantic encoder with a digital refinement path: analog symbols carry a coarse semantic estimate, and low-density parity-check digits from distributed source coding correct that estimate at the receiver. To train the analog codec, the paper derives a Cramér-Rao lower bound for the combined semantic-noise and channel-noise distortion and uses that bound, together with reconstruction error, as the optimization objective. In simulations over fading channels, the hybrid scheme achieves better reconstruction fidelity (PSNR, LPIPS, MS-SSIM) and better fire-detection scores (mAP50, mAP50:95) than classical separated source-channel coding and than DL-based joint source-channel coding at all tested SNR values. The claim is that this removes the fidelity ceiling caused by constrained neural-network capacity without requiring retransmission or a large model.","feed_headline":"Analog semantic coding plus digital parity beats either alone","feed_subtitle":"Fading-channel tests show better reconstruction and fire-detection than digital-only or semantic-only coding.","key_machinery":"The load-bearing identity is the Cramér-Rao lower bound of the overall distortion, expressed in Eq. (16) as a sum over transmitted patches of a term involving the expected squared gradient of the semantic encoder mapping and a sum over synonymously mapped patches of a term governed by the fitting coefficient $\\rho_i^2$. The paper uses this expression as the optimization objective for training the semantic codec. The second mechanism is the DA-ESemCom architecture: an analog semantic encoder selects $Q$ image patches with the lowest fitting coefficients and maps them to channel symbols, a powerful decoder forms a coarse estimate, and a digital DSC path transmits LDPC parity bits derived from DCT-quantized sources, treating the estimate as side information. A lightweight resource-allocation rule gives digital transmission the power and bandwidth needed for a target error probability and gives the remaining bandwidth to the analog path.","core_discovery":"The paper's central claim is that a small emergency transmitter can exceed the reconstruction ceiling of a purely analog neural semantic codec by adding a low-rate digital correction path. The analog path transmits semantically encoded image patches and yields a coarse estimate at the receiver; the digital path, built on distributed source coding, sends only LDPC parity bits, and the receiver uses the coarse estimate as side information to refine the final reconstruction. The paper further claims that the Cramér-Rao lower bound derived for the overall distortion, Eq. (16), is a valid design objective: training the semantic encoder to minimize that bound along with reconstruction MSE improves adaptivity to both semantic noise and channel noise. In the reported simulations, the proposed DA-ESemCom framework outperforms classical SSCC and two DL-based JSCC baselines in reconstruction quality and object-detection performance, and it even exceeds the PSNR of an ideal noiseless semantic-transmission baseline once the SNR passes about 2 dB.","pith_inferences":["A testable extension the paper does not report: an ablation of DA-ESemCom trained with plain reconstruction MSE versus the proposed CRLB-based loss would isolate how much of the gain comes from the digital parity path and how much from the new training objective.","The same analog-plus-digital-parity idea could transfer to other task-oriented emergency sensing modalities, such as compressed video or point-cloud frames, wherever a coarse semantic estimate can serve as side information for distributed source coding.","The paper's resource-allocation rule is heuristic; a joint optimization over the number of analog patches $Q$, the DSC coding rate, and the power split would likely improve on the priority rule, but the paper does not claim such optimality.","A useful diagnostic: comparing the CRLB-based loss against the true end-to-end reconstruction error would show whether Eq. (16) mainly regularizes the encoder or directly drives the reported fidelity gains."],"forward_implications":["Constrained transmitters can gain most of the benefit of a large neural codec: in the ablation, the proposed framework with the small base model improves MS-SSIM over DL-based JSCC with the large model by roughly 0.3 dB.","Reconstruction fidelity stops saturating at high SNR: the hybrid scheme exceeds the paper's ideal noiseless semantic-transmission baseline in PSNR once the SNR exceeds about 2 dB.","AI-driven fire detection benefits across the whole SNR range: both mAP50 and mAP50:95 scores improve over the baselines and approach the scores of original images at high SNR.","The classical cliff effect of separated source-channel coding can be avoided while retaining digital fidelity, because the analog path provides graceful degradation at low SNR and the DSC path adds precision at high SNR.","The framework works without retransmission, making it compatible with the UDP transport used in emergency wireless networks."],"supporting_citations":[{"why":"Provides the DL-based JSCC baseline whose analog semantic coding the proposed framework extends and compares against.","marker":"[10]"},{"why":"Cited as the source of the Cramér-Rao bound used to derive the overall distortion bound in Eq. (16).","marker":"[23]"},{"why":"Supplies the precedent that a Cramér-Rao bound can guide the design of good encoder mappings.","marker":"[24]"},{"why":"The authors' earlier digital-analog transmission framework that this work adapts to the emergency semantic communication setting.","marker":"[26]"},{"why":"Establishes hybrid digital-analog transmission as a way to outperform purely digital transmission concepts.","marker":"[27]"},{"why":"The LDPC code used as the distributed source coding scheme in the digital refinement path.","marker":"[30]"},{"why":"The object detector used to measure detection performance on reconstructed images.","marker":"[31]"},{"why":"The COCO fire-image dataset used to evaluate reconstruction and detection performance.","marker":"[37]"},{"why":"The JPEG source codec used in the classical separated source-channel coding baseline.","marker":"[38]"}],"fun_headline_variants":["Hybrid digital-analog semantic codec beats pure analog for emergencies","Emergency semantic com: add digital parity to boost reconstruction","Digital-analog hybrid lifts semantic video beyond analog ceiling","Semantic coding with digital side info wins in poor channels","Analog semantic core plus digital parity: emergency link gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation's load-bearing premise is that the distortion in Eq. (4), and hence the Cramér-Rao lower bound and the training loss built from it, is computed for the error between the encoder output and the received symbol, with the fading coefficient treated as a real unit-power scalar, rather than for the true end-to-end reconstruction error under complex Rayleigh fading.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid digital-analog semantic codec beats pure analog for emergencies","Emergency semantic com: add digital parity to boost reconstruction","Digital-analog hybrid lifts semantic video beyond analog ceiling","Semantic coding with digital side info wins in poor channels","Analog semantic core plus digital parity: emergency link gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1406,"prompt_tokens":1034,"completion_tokens":372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":292}},"tokens_in":650,"tokens_out":372,"duration_ms":4246,"temperature":1.0,"reasoning_tokens":292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:24:32.945379+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same semantic codec with the true end-to-end mean squared error between source and final reconstruction as the loss and compare it against the proposed CRLB-based loss on the paper's fading-channel setup; if the CRLB-based loss does not improve PSNR or mAP, or if Eq. (11)'s gradient identity fails when complex Rayleigh fading is inserted directly, the claimed role of Eq. (16) as a design objective is not supported.","supporting_citations":[{"cited_title":"This diversity is convenient to evaluate the performance of the system under different fire scenarios","cited_arxiv_id":null,"evidence_quote":"The COCO fire-image dataset used to evaluate reconstruction and detection performance."}],"review_version":1}