{"id":"89c9f261-76ac-4e3c-bdc3-ef100961daa3","arxiv_id":"2507.17196","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HSC transmits a PCA-like complementary projection of the image error and combines it with the decoded semantic image, reducing MSE at the same or modestly increased payload.","lead":"This paper proposes hybrid semantic communication (HSC), which sends a small complementary projection of the original image alongside a neural semantic representation, then merges the two at the receiver to cut reconstruction error. A generalist reader might care because it offers a controllable way to push semantic image transmission toward pixel-accurate fidelity, a known weak point of semantic communication.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Under Rayleigh fading, Alice computes the CR from her clean simulation of Bob's decode, not from Bob's actual noisy decode, so the CR corrects the wrong error; the NCR adapters are a statistical fix, leaving the 74% fading gain ungrounded in the derived theory.","rationale":"The paper proposes supplementing a VAE-based semantic stream with a complementary representation CR = [A | AX], where A is the top-d eigenvector matrix of the per-image error B = (X - X_hat)(X - X_hat)^T, combined at the receiver as X_tilde = A†AX + (I - A†A)X_hat. The central clean-channel mathematics is correct: I re-derived Eqs. (11)-(20); A†A is a symmetric idempotent projection for real full-row-rank A, B is generically full rank (X - X_hat is L×L, not a vector), and the eigenvalue-tail expression MSE = sum_{l>d} lambda_l follows, with d = L giving X_tilde = X exactly. This is real support for the construction and is parameter-light. The single most load-bearing weakness is the synchronization premise of Section III.B. In the clean channel Alice can reproduce Bob's decode exactly, which is why the error-free 91% claim is not attacked here. In the fading channel, Eq. (3) makes Bob's decode random, and nothing in the protocol gives Alice the received SR before she must choose A; few-shot fine-tuning only reduces the mismatch in expectation, and the NCR adapters are trained statistically over the channel distribution, so the CR compensates the wrong residual. The paper's statement that the optimal CR 'remains useful under non-ideal channels, provided that channel perturbations can be separately corrected' is an assertion, with no oracle baseline or released artifacts to verify the 74% fading number. I also flag that the 'error-free' experiments still pass the CR through lossy JPEG2000 and 16QAM/LDPC, so the quantitative match between the measured curves and Eq. (19) is not established; this is secondary to the synchronization issue. Because the clean-channel theory and the general mechanism hold, the verdict should remain CONDITIONAL rather than moving to ACCEPT or REJECT, with release of code/data and the oracle test as the decisive conditions.","tokens_in":8233,"tokens_out":20843,"duration_ms":219275,"concrete_test":"Run the fading experiment with an oracle CR: since the channel model and noise are known in simulation, Alice computes the CR from the true receiver error e_true = X - f^{-1}_phi(hz*sqrt(P)*z + nz), and compare the achieved MSE against the paper's protocol, where the CR uses the clean-simulated e_sim = X - f^{-1}_phi(z), at each d for SNR = 0, 2.5, 5 dB with the same fine-tuned models and NCR adapters. If the oracle-vs-simulated gap is small relative to the reported 0.74 reduction factor, the synchronization concern is benign; if the gap is large, the fading-channel gains are not attributable to the optimal-CR theory, and the claim that the optimal CR 'remains useful under non-ideal channels' is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The clean-channel derivation (Eqs. 8-20) is internally sound: for real full-row-rank A, A†A is a symmetric idempotent projector, and with A = [e1...ed]^T the MSE equals Tr((I-A†A)B) = sum_{l>d} lambda_l. The load-bearing premise for the practical claims is synchronization: Section III.B says 'Alice shares the same decoding function f^{-1}_phi as Bob to synchronize the generation of X_hat.' This is exact only when the SR channel is clean. Under the Rayleigh fading model in Eq. (3), Bob decodes from z_hat = hz*sqrt(P)*z + nz, so X_hat_Bob != f^{-1}_phi(z) = X_hat_Alice. The CR is then built from e = X - X_hat_Alice, which is not the error Bob will actually experience, and the optimality of A in Eq. (20) and the closed form MSE = sum_{l>d} lambda_l no longer apply. The paper does acknowledge this and adds NCR adapters (Eq. 23) trained to minimize ||X - X_tilde|| over the channel distribution; because the adapters see X only at training time, this is a statistical repair, not a correction of the instantaneous wrong residual. No oracle experiment (CR computed from the true received SR), no feedback protocol, and no code or data are provided, and the Fig. 4 ablations compare fine-tuning variants rather than isolating the synchronization mismatch. A secondary compounding gap: even the 'error-free' CR experiments route A and AX through lossy JPEG2000 (1/5) and 16QAM/LDPC, so the measured curves mix unmodeled compression distortion with the eigen-truncation sum, leaving quantitative agreement with Eq. (19) unverified. The clean-channel theory stands; the channel-robustness claims do not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Hybrid Semantic-Complementary Transmission (HSC) for image communication. Alice transmits a neural semantic representation (SR) through a VAE-based encoder-decoder and additionally transmits a complementary representation (CR) formed as C=[A|AX], where A is a projection matrix. Bob recomposes the image as X_tilde = A^† A X + (I - A^† A) X_hat. The paper derives that, under an error-free CR channel, the optimal A consists of the d dominant eigenvectors of the error matrix B=(X-X_hat)(X-X_hat)^T, giving closed-form MSE = sum_{l=d+1}^L lambda_l. Fine-tuning and a neural-augmented CR adapter are proposed for fading channels. Simulations on MNIST and Flickr8k with VAE and VQ-VAE report large MSE reductions, including up to 91% reduction at equal payload and 74% reduction under fading.","tokens_in":8627,"tokens_out":7242,"duration_ms":78909,"significance":"The central idea is attractive: it gives semantic communication a controllable knob to drive pixel-level MSE to zero and directly addresses the known saturation of SR-based reconstruction. The range-null space decomposition is a natural and promising mechanism, and the paper explicitly validates the approach on two datasets and two semantic autoencoder architectures. The clean-channel derivation, once the proof is repaired, yields a simple closed-form result, which is a genuine strength. However, the current validation does not cleanly separate the derived theory from unmodeled codec and channel effects, the fading results depend on an unverified synchronization assumption, and the equal-payload comparison is not fully controlled. The paper is worth revising rather than rejecting, but the remaining gaps are load-bearing for the strongest advertised claims.","major_comments":[{"comment":"The step from Eq. (18) to Eq. (19) is not valid for a general full-row-rank matrix A. In Eq. (18), e_l^T (A^† A) e_l equals ||A^† A e_l||^2, not ||A e_l||^2, and the inequality ||A e_l||^2 <= ||e_l||^2 = 1 is false when the rows of A are not orthonormal (for example, A=[2 0 ... 0] gives ||A e_1||^2 = 4). The minimization in Eq. (10) should be formulated over the rank-d orthogonal projector P = A^† A; the final eigenvector solution A=[e_1 ... e_d]^T is correct and can be recovered from the variational bound on sum_l lambda_l ||P e_l||^2, but the proof as written needs a substantive correction, not just a typo fix.","section":"Eq. (19)"},{"comment":"The closed-form optimality of Eq. (20) and the formula MSE = sum_{l>d} lambda_l require Alice to know Bob's decoded image X_hat exactly. Under the fading model of Eq. (3), Bob decodes from z_hat = h_z sqrt(P) z + n_z, so f_phi^{-1}(z_hat) differs from the X_hat Alice simulates from the clean z. The CR is therefore constructed from the wrong residual, and Eq. (20) no longer characterizes the achievable MSE. The NCR adapters in Eq. (23) are trained in expectation over the channel distribution and do not repair the instantaneous wrong residual. The 74% fading improvement reported in Section IV.C is not grounded in the derived theory; an oracle experiment using the true received SR, or a protocol that feeds back or estimates the actual residual, is needed to validate the claimed fading gain.","section":"Section III.B and Eq. (23)"},{"comment":"The experiments do not test the derived theory cleanly. Even for the so-called error-free comparisons, the CR is source-coded with JPEG2000 at 1/5, channel-coded with LDPC rate 1/2, and mapped to 16QAM, whereas Eqs. (8)-(20) assume exact delivery of C=[A|AX]. The measured MSE therefore mixes unmodeled compression distortion and possible residual channel errors with the eigen-truncation term sum_{l>d} lambda_l. As a result, the reported curves cannot quantitatively verify the closed-form expression. To support the theoretical claim, the authors should either simulate the ideal CR channel defined by the model or extend the analysis to include the distortion introduced by the digital source and channel coding stages.","section":"Section IV.A and Figs. 2-3"},{"comment":"The equal-payload comparison is ambiguously described. For the SC baseline in Fig. 2, the SR size k is increased while d=0, whereas for HSC, k is fixed at 128 and d is increased, but the text does not state whether the VAE or VQ-VAE is retrained for each k. The horizontal axis in Fig. 2 is labeled mu, even though mu is used for the fading coefficient variance in Eq. (3) and the compression ratio is already defined as eta in Eq. (21). In addition, no error bars or multiple-seed results are reported, so the crossover in Fig. 3 and the claimed 91% MSE reduction are not statistically supported. Please clarify the experimental protocol, define the plotted quantity unambiguously, and report variability over channel realizations and training runs.","section":"Section IV.B, Figs. 2-3"}],"minor_comments":[{"comment":"The expression for the compression ratio eta is dimensionally unclear: |z| is a number of complex symbols, |X| is a number of real pixels, and R is introduced as a ratio, but the formula mixes these quantities without stating how bits and channel uses are counted. Please define each cardinality and rederive eta consistently.","section":"Eq. (21)"},{"comment":"The paper switches between treating images as matrices X in R^{L x L} and as vectors when forming B=(X-X_hat)(X-X_hat)^T. Please state explicitly that X is flattened before constructing B, and make the dimensions of A, AX, and C consistent throughout.","section":"Throughout"},{"comment":"There are minor typographical errors, including 'complimentary' for 'complementary' in the Section III.B heading and 'technqiues' in the Conclusion. These should be corrected.","section":"Section III.B heading and Conclusion"},{"comment":"The caption of Fig. 4 and the surrounding text do not specify how many channel realizations are averaged when reporting the fading results. Please state the number of Monte Carlo trials and report confidence intervals.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising core idea and the clean-channel result is essentially correct after reformulating the proof over orthogonal projectors. The revision should focus on repairing Eq. (19), adding an oracle or feedback-based experiment for the fading case, and separating the codec effects from the theoretical MSE. I also note that no code or data are provided; given that the simulations are central to the practical claims, the journal may wish to encourage a reproducibility statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. The core idea is simple and the linear algebra is correct: with a clean complementary channel, projecting the original image onto the top-d eigenvectors of the error matrix and combining it with the semantic decoder's output in the null space gives MSE equal to the sum of the remaining eigenvalues. That is a nice closed-form result, and it gives semantic communication a genuine tunable fidelity knob. The authors apply range-null space decomposition to a new problem—complementing a neural semantic stream with a classical digital stream—so this is a legitimate new application even though the individual tools are known.\n\nThe paper does several things well. The derivation in Eqs. (8)-(20) is clean, and I checked the idempotence step. The experiments show large MSE reductions, and the method works with both VAE and VQ-VAE backbones. The high-resolution example is a nice touch. I also appreciate that the compression ratio includes the cost of transmitting the projection matrix itself, which many such schemes forget.\n\nThe soft spots are real but not fatal. Most importantly, the fading-channel analysis assumes Alice can reproduce Bob's decoded image before transmission. That is only true if the SR channel is clean. Under Rayleigh fading, Bob decodes from a noisy received signal, so Alice's simulated X_hat is not the image Bob actually gets, and the optimal projection computed from the clean error matrix corrects the wrong residual. The NCR adapters are a statistical patch, not a correction of the instantaneous mismatch. The paper acknowledges this indirectly, but the 74% fading gain is never tied to the derived theory. A quick oracle experiment—where Alice somehow knows the true received SR—would have settled whether the mismatch is the bottleneck. Second, even the 'error-free' CR path runs through JPEG2000 and 16QAM/LDPC, so the measured curves mix unmodeled compression and modulation distortion with the eigen-truncation sum. That makes it hard to verify Eq. (19) quantitatively. Third, the equal-payload comparison in Figure 2 is ambiguous: how do you increase k for a fixed architecture without retraining? The text does not say. Finally, no code, no data, no error bars.\n\nVerdict: worth a serious referee. The theoretical contribution is sound and the application is plausible, but the empirical claims need to be backed by artifacts and the fading analysis needs either a stronger synchronization argument or an explicit acknowledgment that the optimality is an approximation. I would not cite it in my own work, but I would send it to someone working on semantic communication.","headline":"Clean closed-form MSE and a sensible hybrid architecture, but the fading-channel story rests on a synchronization assumption the paper never actually tests.","tokens_in":9183,"tokens_out":3172,"would_cite":false,"duration_ms":35800,"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":"Hybrid semantic communication sends a small residual image alongside the semantic code, letting receivers drive pixel error toward zero while keeping the payload fixed.","keywords":["hybrid semantic communication","complementary representation","range-null space decomposition","image reconstruction fidelity","mean squared error","VQ-VAE","fading channel denoising","semantic communication"],"falsifier":"Run HSC over a Rayleigh fading channel where Alice computes the CR from a clean semantic decode while Bob's received semantic code is corrupted, then compare the recomposed image's MSE with the predicted tail-eigenvalue sum $\\sum_{\\ell=d+1}^{L}\\lambda_\\ell$; if the measured MSE systematically exceeds the prediction even after NCR fine-tuning, the closed-form optimality claim fails.","tokens_in":8029,"feed_emoji":"📡","tokens_out":7377,"duration_ms":67774,"temperature":0.7,"pith_summary":"Semantic communication can convey the gist of an image but saturates in pixel-level fidelity. This paper proposes hybrid semantic communication (HSC), which transmits a complementary representation alongside the neural semantic representation, carrying the residuals the decoder fails to reproduce. The receiver combines the two through a range-null space projection, so the final image is the semantic reconstruction corrected by the transmitted residuals. The authors derive a closed-form MSE, show the optimal projection is built from the eigenvectors of the error matrix, and demonstrate up to a 91 percent MSE reduction at the same total payload.","feed_headline":"Hybrid semantic link cuts image-reconstruction error by up to 91%","feed_subtitle":"A small residual image alongside the semantic code lets receivers refine details until pixel error reaches zero.","key_machinery":"The central object is the range-null space decomposition induced by a full-rank projection matrix $A\\in\\mathbb{R}^{d\\times L}$ and its Moore-Penrose pseudoinverse $A^\\dagger$. The receiver forms $\\tilde{X}=A^\\dagger A X+(I-A^\\dagger A)\\hat{X}$, using the transmitted projected image $AX$ for the range component and the semantic decoder output $\\hat{X}$ for the null component; because the original image obeys the same decomposition, the reconstruction error collapses to $(I-A^\\dagger A)(X-\\hat{X})$. Minimizing its squared norm gives the eigenvector solution and the closed-form error $\\sum_{\\ell=d+1}^{L}\\lambda_\\ell$. Under channel impairments, a shallow autoencoder-based NCR adapter is trained, with the semantic model fixed, to make the recomposed image robust to corrupted CR inputs.","core_discovery":"The central claim is that sending a complementary representation (CR) in addition to a semantic representation (SR) removes the fidelity ceiling of semantic image transmission. With the optimal projection $A$, the recomposed image is $\\tilde{X}=A^\\dagger A X+(I-A^\\dagger A)\\hat{X}$, where $\\hat{X}$ is the receiver's decoded image and $A$ consists of the $d$ largest eigenvectors of the error matrix $B=(X-\\hat{X})(X-\\hat{X})^T$. The resulting MSE equals $\\sum_{\\ell=d+1}^{L}\\lambda_\\ell$, so the transmitter can trade payload for fidelity by choosing $d$, and exact reconstruction is reached when $d=L$. Simulations on MNIST and Flickr8k show HSC reduces MSE by up to 0.91 times compared with SC at the same total payload, and the proposed neural-augmented CR fine-tuning preserves most of the gain under Rayleigh fading.","pith_inferences":["A testable extension: the same residual-correction argument should apply to diffusion- or GAN-based decoders by computing the projection from the specific image pair rather than a fixed basis.","Because the optimal $A$ depends only on the error matrix, one could try to predict or compress $A$ itself, shrinking the $2dL$ overhead and improving the rate-fidelity tradeoff.","The closed-form MSE defines a rate-distortion operating curve: plotting $\\sum_{\\ell=d+1}^{L}\\lambda_\\ell$ against the payload formula gives a computable comparison point with classical transform coding."],"forward_implications":["With an error-free complementary channel, HSC gives a hard fidelity guarantee: transmitting $d$ eigen-directions of the error matrix leaves MSE $\\sum_{\\ell=d+1}^{L}\\lambda_\\ell$, and $d=L$ gives exact reconstruction.","At a fixed total payload, shifting budget from a larger semantic code to a complementary code lowers MSE more than enlarging the SR, so the fidelity floor of SC is not fundamental.","The recomposition rule applies to any neural image decoder whose output the transmitter can reproduce, including VAE- and VQ-VAE-based semantic transceivers.","For fading channels, few-shot fine-tuning of the semantic model plus NCR adaptation restores most of the ideal-channel MSE reduction."],"supporting_citations":[{"why":"supplies the range-null decomposition used to combine the semantic and complementary images.","marker":"[12]"},{"why":"provides the null-space decomposition concept for filling in missing information.","marker":"[13]"},{"why":"contributes the null-space model used for zero-shot restoration.","marker":"[14]"},{"why":"defines the deep joint source-channel coding architecture that HSC extends with a CR.","marker":"[15]"},{"why":"provides the variational autoencoder formulation behind the semantic encoder and decoder.","marker":"[16]"},{"why":"MNIST is the main dataset for the reconstruction-error simulations.","marker":"[22]"},{"why":"the masked VQ-VAE semantic communication framework is HSC's second evaluated architecture.","marker":"[23]"},{"why":"Flickr8k supplies the high-resolution images used for visual examples.","marker":"[24]"}],"fun_headline_variants":["Semantic plus residual stream hits 91% MSE drop in image relay","Hybrid semantic-code link erases image fidelity ceiling","Add residual bits to semantic code for near-perfect image recovery","Image relay: semantic code plus small residual cuts error 91%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes Alice can reproduce Bob's decoded image before transmitting, so the complementary representation targets the exact error Bob will see; under fading and noise that copy is only approximate, and the NCR adapters repair the mismatch imperfectly.","fun_headline_variants_meta":{"raw":{"variants":["Semantic plus residual stream hits 91% MSE drop in image relay","Hybrid semantic-code link erases image fidelity ceiling","Add residual bits to semantic code for near-perfect image recovery","Image relay: semantic code plus small residual cuts error 91%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000847,"raw_usage":{"total_tokens":3663,"prompt_tokens":901,"completion_tokens":2762,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":2690}},"tokens_in":517,"tokens_out":2762,"duration_ms":20003,"temperature":1.0,"reasoning_tokens":2690,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:55:45.345583+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HSC over a Rayleigh fading channel where Alice computes the CR from a clean semantic decode while Bob's received semantic code is corrupted, then compare the recomposed image's MSE with the predicted tail-eigenvalue sum $\\sum_{\\ell=d+1}^{L}\\lambda_\\ell$; if the measured MSE systematically exceeds the prediction even after NCR fine-tuning, the closed-form optimality claim fails.","supporting_citations":[{"cited_title":"Null-space diffusion sampling for zero-shot point cloud completion,","cited_arxiv_id":null,"evidence_quote":"provides the null-space decomposition concept for filling in missing information."},{"cited_title":"Zero-shot image restoration using denoising diffusion null-space model,","cited_arxiv_id":null,"evidence_quote":"contributes the null-space model used for zero-shot restoration."},{"cited_title":"Deep joint source- channel coding for wireless image transmission,","cited_arxiv_id":null,"evidence_quote":"defines the deep joint source-channel coding architecture that HSC extends with a CR."},{"cited_title":"Auto-Encoding Variational Bayes,","cited_arxiv_id":null,"evidence_quote":"provides the variational autoencoder formulation behind the semantic encoder and decoder."},{"cited_title":"Robust semantic communications with masked vq-vae enabled codebook,","cited_arxiv_id":null,"evidence_quote":"the masked VQ-VAE semantic communication framework is HSC's second evaluated architecture."},{"cited_title":"Framing image description as a ranking task: data, models and evaluation metrics,","cited_arxiv_id":null,"evidence_quote":"Flickr8k supplies the high-resolution images used for visual examples."}],"review_version":1}