Pith. sign in

REVIEW 2 major objections 4 minor 30 references

Semantic Rate-Distortion Theory with Applications

T0 review · 2 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper establishes a rate limit for semantic compression and derives a closed form for binary sources.

desk verdict The central equivalence theorem rests on an invalid max/expectation interchange; the binary example makes the failure concrete. read the letter →

arxiv 2509.10061 v1 pith:FAEEP4FX submitted 2025-09-12 cs.IT eess.SPmath.IT

Yi-Qun Zhao , Zhi-Ming Ma , Geoffrey Ye Li , Shuai Yuan , Tong Ye , Chuan Zhou This is my paper
classification cs.ITeess.SPmath.IT MSC 94A1594A17
keywords semanticcommunicationrate-distortiontheorycompressionrate-distortion-perceptionconditionalprobabilitydistortionbinarysymmetricsourcetotalvariationdistancecodingtheorem
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper develops a rate-distortion theory for semantic communication in which the transmitter and receiver care about the probability distribution of a latent meaning S given the observed symbol X, not just about the symbol itself. It defines an information-semantic rate-distortion function RI(Dp,Do) as the minimum mutual information over channels Y|X that keep both the conditional-semantic distortion and the ordinary symbol distortion within budgets, and proves, under a lower-semicontinuity condition, that this equals the operational minimum achievable rate. For a doubly symmetric binary source with total-variation semantic distortion and Hamming observation distortion, it obtains a closed form: the rate is 1 - h2((1 - sqrt(1 - 2Dp/C))/2) when Dp is small enough, and the classical binary rate-distortion function otherwise. Experiments with a digit-recognition task show that adding a conditional-semantic-distortion penalty improves downstream recognition accuracy and lowers the bit rate needed for a target accuracy. The paper's central claim is that semantic ambiguity and polysemy can be treated as a fidelity constraint on the receiver's inferred meaning distribution.

What carries the argument

The central object is the information semantic rate-distortion function RI(Dp,Do) = min_{p_{Y|X}: Edp(pS|X,pS|Y)<=Dp, Edo(X,Y)<=Do} I(X;Y), defined with a conditional semantic probability distortion dp and an observation distortion do. The proof machinery uses the Poisson representation lemma to construct codes for the achievability direction (Theorem 1), a standard converse with a lower-semicontinuity condition for Theorem 2, and symmetry reduction plus calculus for the binary closed form (Theorem 3). The key conceptual mechanism is that semantic fidelity is measured at the level of the posterior distribution p_{S|X} versus p_{S|Y}, not by a point estimate of meaning, which lets ambiguity a

What would settle it

Set n=2 for a binary source with Hamming observation distortion and Do=0.25. Under the paper's Definition 3, E[max_i d_H(X_i,Y_i)] <= 0.25 forces the two symbols to be reproduced exactly with probability at least 0.75, so the operational rate per symbol must be at least about 0.59 bits; the paper's symbolic-only formula gives 1-h2(0.25) ≈ 0.189 bits per symbol, which is impossible for the operational problem.

Watch

Extended reading notes

Core claim

The central claim is that the minimum rate at which a semantic source can be compressed, subject to a budget on how much the receiver's inferred semantic distribution p_{S|Y} can differ from the transmitter's p_{S|X}, is the solution of a single-letter optimization problem. Theorems 1 and 2 together state RI(Dp,Do) = RO(Dp,Do) when RI is lower semicontinuous, so the information-theoretic quantity is the true operational limit. Theorem 3 evaluates this limit for the binary doubly symmetric case: with C=|1-2q| and a(Do) as defined, R(Dp,Do) equals 1 - h2((1 - sqrt(1 - 2Dp/C))/2) for Dp <= a(Do), and equals the classical 1 - h2(min{Do,1/2}) above that threshold. The idea is that semantic ambigu

Load-bearing premise

The load-bearing assumption is that sequence-level distortion can be controlled component-wise—specifically, that E[max_i dp(p_{S_i|X_i}, p_{S_i|Y_i})] equals max_i E[dp(...)] and likewise for the observation distortion; without this interchange, the operational rate-distortion function may not equal the information-theoretic function computed in the paper.

Editorial extensions

If this is right

  • If RI=RO, then the operational semantic rate-distortion region is characterized by a single-letter optimization, and any code sequence achieving RI is optimal up to vanishing rate loss.
  • For the binary doubly symmetric case, the closed form shows that for small semantic distortion Dp the rate is set entirely by Dp, while for larger Dp it degenerates to the classical Hamming rate-distortion function; the threshold a(Do) governs the switch.
  • The reported experiments imply that adding a conditional-semantic-distortion penalty can yield large accuracy gains at the same bit rate, e.g., from about 62% to about 94% at 12 bits in the paper's digit-recognition setup.
  • When both semantic and symbolic fidelity are required, the two constraints can be tuned independently, enabling systems that preserve task meaning while limiting observable distortion.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The paper's sequence-level distortion is defined as the maximum per-component distortion, which is stronger than the usual average distortion; under an average-distortion definition the binary closed form would likely change, perhaps to a convex combination of the semantic and symbolic constraints.
  • Editorial inference: The equality RI=RO is stated as conditional on lower semicontinuity, and the appendices interchange expectation with a maximum over components; if that interchange is not valid, the operational interpretation of the closed form requires an alternative proof or a modified distortion definition.
  • Editorial inference: A natural testable extension is to replace total variation with KL divergence in the same binary setup and compare the resulting threshold and rate curves against end-task accuracy, which would indicate which divergence measure better predicts practical semantic fidelity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper develops a semantic rate-distortion framework with dual constraints: a semantic probability distortion constraint (based on conditional distributions p_{S|X}, p_{S|Y}) and a traditional symbolic distortion constraint. It defines an operational semantic rate-distortion function using sequence-level max distortion (Definition 3) and an information semantic rate-distortion function using per-letter constraints (Definition 8). Theorems 1 and 2 claim that these functions coincide under a lower-semicontinuity condition, and Theorem 3 derives a closed-form expression for the doubly symmetric binary source with TV semantic distortion and Hamming symbolic distortion. The paper also presents MNIST experiments indicating that adding a KL semantic term improves recognition accuracy.

Significance. The problem addressed is relevant, and the modeling choice of conditional semantic distributions rather than point estimates is well motivated. The paper correctly leverages known rate-distortion-perception techniques, and the closed-form computation in Theorem 3 is an explicit, non-circular minimization. However, the central claim—that the operational rate-distortion function equals the information rate-distortion function—rests on an invalid interchange of maximum and expectation in the achievability proof. For the paper's own binary Hamming example, the claimed rates are incompatible with Fano's inequality, so the main theorem is false as stated. The experimental section is heuristic and illustrative, not a substitute for a correct coding theorem.

major comments (2)
  1. [Appendix A, Eqs. (32)–(33)] The achievability proof replaces E[max_i d_o(X_i,Y_i)] with max_i E[d_o(X_i,Y_i)] (and similarly for d_p). This identity is false: for nonnegative random variables, E[max] ≥ max E, generally strictly. Under Definition 3, sequence distortion is the maximum, so the constructed channel satisfying per-letter constraints does not meet the operational constraints (8)–(9). This is load-bearing: for binary Hamming distortion, E[max_i d_H] = P(X^n≠Y^n). Fano's inequality gives R ≥ 1 − D_o − h_2(D_o)/n for any code with E[max] ≤ D_o. Taking D_o=0.4 and D_p>a(D_o), Theorem 3 predicts R ≈ 0.029, whereas the Fano bound requires R ≥ 0.6 in the limit. Thus the stated operational rate-distortion function is not achievable; Theorem 1 and the operational reading of Theorem 3 collapse.
  2. [Appendix D, Proposition 5] The proof of Proposition 5 asserts without demonstration that \tilde I(t) is nonincreasing on [ (1−2A)/(1−A), 1 ]. This monotonicity is essential for the claimed minimizer (76). The text says "one can verify" but no derivative or convexity/majorization argument is given. This is a separate gap in the closed-form derivation, though secondary to the achievability error.
minor comments (4)
  1. [Definition 7] The name "Operation Semantic Rate Distortion Function" appears to be a typo for "Operational Semantic Rate Distortion Function"; the term "operational" is used elsewhere. Please harmonize.
  2. [Section 5.1] The experiments treat the pre-trained classifier output as the true semantic conditional distribution p_{S|y}. This is a strong modeling assumption; the experiments should explicitly acknowledge that they illustrate a heuristic objective rather than directly validate the rate-distortion bounds.
  3. [Equation (19)] The two cases of the piecewise expression use inconsistent spacing ("D p" vs "D_p") and the boundary D_p = a(D_o) is only included in the first interval. Please make the notation uniform and state the convention for the boundary.
  4. [Throughout] The manuscript contains several typographical errors (e.g., "efficiency" in the abstract, "Operation" for "Operational", and irregular whitespace in matrix expressions). A careful proofread is needed.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the coding theorem and binary closed form are obtained by explicit optimization; the only author-overlap citation is background, and the main defect is a non-circular mathematical error in Appendix A.

full rationale

The paper's central claim is a coding theorem equating an operational semantic rate-distortion function (Definition 7) with an information-theoretic one (Definition 8). These are distinct objects: one is an infimum over code sequences satisfying expected max-distortion constraints, the other is a minimum over conditional distributions satisfying per-symbol expected distortion constraints. The equivalence is not assumed; Theorem 1 is proved via a Poisson-representation code construction and Theorem 2 via a standard converse argument. The binary closed form in Theorem 3 is derived from an explicit constrained minimization of I(w,z) under the constraints Lambda(w,z) <= Dp and Gamma(w,z) <= Do. No parameter is fitted to the target rate, and no load-bearing conclusion is imported from a self-citation. The only author-overlapping reference is [5], which is background on deep-learning semantic communication and is not used to justify the theorems. The main substantive concern is mathematical rather than circular: Appendix A's equations (32)-(33) assert E[max_i d(X_i,Y_i)] = max_i E[d(X_i,Y_i)] for i.i.d. components, which is false in general and invalidates the achievability proof as written. This is an error in the derivation chain, not a definitional reduction of the conclusion to the assumptions. Accordingly, no circular step is identified; the score reflects only the minor background self-citation and the paper's close reliance on known rate-distortion-perception methods, which it explicitly acknowledges.

Assumptions & free parameters 1 free parameters · 6 assumptions · 1 invented entities

The central theorem depends on the hidden-source model, the Markov structure S - X - Y, perfect channels, lower semicontinuity, and the Poisson representation lemma. No free parameters enter the theoretical derivation. The experimental part introduces a hand-tuned gamma and an unverified classifier-proxy assumption for the semantic distribution.

free parameters (1)
  • gamma = 0, 0.01, 0.1, 0.5, 100
    Tunable weight in the experimental loss (Equation 21). It controls the tradeoff between symbolic MSE and semantic KL distortion and is chosen by hand; it is not used in the theoretical rate-distortion derivation.
assumptions (6)
  • domain assumption The pair (S,X) has a known joint distribution p_{S,X}, and the sequence (S_i,X_i) is i.i.d.
    Section 2.3 introduces the source model. The semantics S are hidden and only X can be coded.
  • domain assumption Markov chain S - X - Y, equivalently p_{S,X,Y} = p_{Y|X} p_{S,X}.
    Definition 8 states Y and S are conditionally independent given X. This is essential to the rate-distortion formulation but is not derived from a physical encoder model.
  • domain assumption The channel is perfect; semantic compression is the only source of loss.
    Section 2.1 states 'we assume perfect channel and focus on semantic compression.'
  • domain assumption The information semantic rate-distortion function R_I(D_p,D_o) is lower semicontinuous, or the finite-alphabet sufficient condition (16) holds.
    Theorem 2 requires lower semicontinuity to equate operational and information definitions. Proposition 2 gives a sufficient condition but it is not proven for all distortion measures used.
  • standard math Poisson representation lemma of [25].
    Appendix A uses this lemma to construct the achievability code for Theorem 1.
  • ad hoc to paper The pre-trained classifier output is treated as the true semantic conditional distribution p_{S|y}, and the original label is treated as p_{S|x}.
    Section 5.1 uses a one-hot label for p_{S|x_tr} and a classifier soft output for p_{S|y_tr}. This contradicts the paper's stated motivation that p_{S|x} is ambiguous and polysemous, and it is an unvalidated proxy for the semantic distributions in the theory.
invented entities (1)
  • Latent semantic variable S
    purpose: Models the intrinsic meaning behind the observable symbol X, which the communication system tries to preserve.
    S is a latent construct inherited from semantic communication models; the paper provides no direct measurement of S. In the experiments it is approximated by classifier outputs and one-hot labels, not by an independent handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic Rate-Distortion Theory with Applications." pith.science (2026). https://pith.science/paper/FAEEP4FX

@misc{pith2026250910061,
  author       = {Pith},
  title        = {Pith review of: Semantic Rate-Distortion Theory with Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FAEEP4FX}},
  note         = {Machine review of arXiv:2509.10061}
}
read the original abstract

Artificial intelligence (AI) is ushering in a new era for communication. As a result, the establishment of a semantic communication framework is putting on the agenda. Based on a realistic semantic communication model, this paper develops a rate-distortion framework for semantic compression. Different from the existing works primarily focusing on decoder-side estimation of intrinsic meaning and ignoring its inherent issues, such as ambiguity and polysemy, we exploit a constraint of conditional semantic probability distortion to effectively capture the essential features of practical semantic exchanges in an AI-assisted communication system. With the help of the methods in rate-distortion-perception theory, we establish a theorem specifying the minimum achievable rate under this semantic constraint and a traditional symbolic constraint and obtain its closed-form limit for a particular semantic scenario. From the experiments in this paper, bounding conditional semantic probability distortion can effectively improve both semantic transmission accuracy and bit-rate efficiency. Our framework bridges information theory and AI, enabling potential applications in bandwidth-efficient semantic-aware networks, enhanced transceiver understanding, and optimized semantic transmission for AI-driven systems.

Figures

Figures reproduced from arXiv: 2509.10061 by the authors.

Figure 1
Figure 1. Semantic Communication Model. methods are also sufficient for transmitting the extracted semantics (if estimated version of the semantics sˆ = f1(x) and codeword ws = f2(ˆs), then just let semantic encoding function in our framework f = f2 ◦ f1), thereby yielding enhanced generality. Starting from this perspective, the semantic communication procedure can be modeled as fol￾lows, as depicted in [PITH_FULL_IMAGE:figu… view at source ↗
Figure 2
Figure 2. The remaining part of this subsection will establish the essential definitions required for the subsequent analysis. Let us start from the definition of the semantic probability distortion for given observation — the central concern in semantic communication. Definition 1 Semantic Probability Distortion Measure (Based on the Observation): A semantic 7 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 2
Figure 2. Semantic Communication Model with Dual Fidelity Metrics. probability distortion measure (based on the observation) is defined as a function dp : P(S) × P(S) → R +, (1) where P(S) denotes the set of all possible probability distributions over semantics set S. Distortion dp(pS|x, pS|y) quantifies the discrepancy between probability distributions of the intrinsic meaning based on extrinsic observations x and y. Definit… view at source ↗
Figures from the paper (7 more)
Figure 3
Figure 3. Figure 3: The Closed-Form Expression of R(Dp, Do) for the Binary Case. rate determination. 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.2 0.4 0.6 0.8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 [PITH_FULL_IMAGE:figures/full_fig_p015_3.png]
Figure 4
Figure 4. Figure 4: R(Dp, Do) for ρ = 0.5 and q = 0.9 with One Variable Fixed in the Binary Case. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Training Workflow. The Kullback-Leibler (KL) Divergence is computed between the image’s original digit label, which is represented as a single-point distribution, i.e. the true semantic distribution given original image xtr, namely pS|xtr , and distribution pS|ytr to c…
Figure 6
Figure 6. Figure 6: Evaluating Workflow. We compare all recognition results against ground-truth digit labels to calculate the recog￾nition accuracy. Crucially, the accuracy reflects traditional communication performance for γ = 0 while the accuracy quantifies semantic communication effic…
Figure 7
Figure 7. Figure 7: Accuracy under Different Rates and γ Values. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]
Figure 8
Figure 8. Figure 8: The Original and Recovered Images for Traditional and Semantic Communication at 10 Bits. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_8.png]
Figure 9
Figure 9. Figure 9: The Original and Recovered Images for Traditional and Semantic Communication at 8 Bits [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 4 linked inside Pith

  1. [1]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,” The Bell System Technical Jour- nal, vol. 27, no. 3, pp. 379–423, 1948

    1948

  2. [2]

    Recent contributions to the mathematical theory of communication,

    W. Weaver, “Recent contributions to the mathematical theory of communication,” ETC: A Review of General Semantics , vol. 10, no. 4, pp. 261–281, 1953. [Online]. A vailable: http://www.jstor.org/stable/42581364 31

    arXiv 1953

  3. [3]

    An outline of a theory of semantic information,

    R. Carnap and Y. Bar-Hillel, “An outline of a theory of semantic information,” Journal of Symbolic Logic, vol. 19, no. 3, pp. 230–232, 1952

    1952

  4. [4]

    Towards a theory of semantic communication,

    J. Bao, P. Basu, M. Dean, C. Partridge, A. Swami, W. Leland, and J. A. Hendler, “Towards a theory of semantic communication,” in 2011 IEEE Netw. Sci. Workshop (NSW) , 2011, pp. 110–117

    2011

  5. [5]

    Deep learning enabled semantic communication systems,

    H. Xie, Z. Qin, G. Y. Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Trans. Signal Process. , vol. 69, pp. 2663–2675, 2021

    2021

  6. [6]

    Semantic communication: From philosophical conceptions towards a mathematical framework,

    J. Gholipour, R. F. Schaefer, and G. P. Fettweis, “Semantic communication: From philosophical conceptions towards a mathematical framework,” 2025. [Online]. A vailable: https://arxiv.org/abs/2505.01342

    arXiv 2025

  7. [7]

    A rate-distortion framework for characterizing semantic information,

    J. Liu, W. Zhang, and H. V. Poor, “A rate-distortion framework for characterizing semantic information,” in 2021 IEEE Int. Symp. Inf. Theory (ISIT) , 2021, pp. 2894–2899

    2021

  8. [8]

    An indirect rate-distortion characterization for semantic sources: General model and the case of gaussian observation,

    J. Liu, S. Shao, W. Zhang, and H. V. Poor, “An indirect rate-distortion characterization for semantic sources: General model and the case of gaussian observation,” IEEE Trans. Commun., vol. 70, no. 9, pp. 5946–5959, 2022

    2022

Show all 30 references
  1. [9]

    Semantic compression with side information: A rate-distortion perspective,

    T. Guo, Y. Wang, J. Han, H. Wu, B. Bai, and W. Han, “Semantic compression with side information: A rate-distortion perspective,” ArXiv, vol. abs/2208.06094, 2022. [Online]. A vailable:https://api.semanticscholar.org/CorpusID:251554744

    2022 arXiv

  2. [10]

    Protecting semantic information using an efficient secret key,

    T. Guo, J. Han, H. Wu, Y. Wang, B. Bai, and W. Han, “Protecting semantic information using an efficient secret key,” in 2022 IEEE Int. Symp. Inf. Theory (ISIT) , 2022, pp. 2660–2665

    2022

  3. [11]

    Rate-distortion theory for strategic semantic communication,

    Y. Xiao, X. Zhang, Y. Li, G. Shi, and T. Başar, “Rate-distortion theory for strategic semantic communication,” in 2022 IEEE Inf. Theory Workshop (ITW) , 2022, pp. 279–284

    2022

  4. [12]

    Semantic communication of learnable concepts,

    F. Pase, S. Kobus, D. Gündüz, and M. Zorzi, “Semantic communication of learnable concepts,” in 2023 IEEE Int. Symp. Inf. Theory (ISIT) , 2023, pp. 731–736

    2023

  5. [13]

    The perception-distortion tradeoff,

    Y. Blau and T. Michaeli, “The perception-distortion tradeoff,” in 2018 IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) , 2018, pp. 6228–6237. 32

    2018

  6. [14]

    Rethinking lossy compression: The rate-distortion-perception tradeoff,

    ——, “Rethinking lossy compression: The rate-distortion-perception tradeoff,” in Proc. 36th Int. Conf. Mach. Learn., ICML 2019, 9-15 June 2019, Long Beach, California, USA , ser. Proceedings of Machine Learning Research, K. Chaudhuri and R. Salakhutdinov, Eds., vol. 97. PMLR, 2...

    2019

  7. [15]

    A coding theorem for the rate-distortion-perception function,

    L. Theis and A. B. Wagner, “A coding theorem for the rate-distortion-perception function,” in Neural Compression: From Information Theory to Applications – Workshop @ ICLR 2021 ,

    2021

  8. [16]

    On the rate-distortion-perception function,

    J. Chen, L. Yu, J. Wang, W. Shi, Y. Ge, and W. Tong, “On the rate-distortion-perception function,” IEEE J. Sel. Area Inf. Theory , vol. 3, no. 4, pp. 664–673, 2022

    2022

  9. [17]

    The rate-distortion-perception trade-off with side information,

    Y. Hamdi and D. Gündüz, “The rate-distortion-perception trade-off with side information,” in 2023 IEEE Int. Symp. Inf. Theory (ISIT) , 2023, pp. 1056–1061

    2023

  10. [18]

    Conditional rate-distortion-perception trade-off,

    X. Niu, D. Gündüz, B. Bai, and W. Han, “Conditional rate-distortion-perception trade-off,” in 2023 IEEE Int. Symp. Inf. Theory (ISIT) , 2023, pp. 1068–1073

    2023

  11. [19]

    Rate-distortion-perception tradeoff based on the conditional-distribution perception measure,

    S. Salehkalaibar, J. Chen, A. Khisti, and W. Yu, “Rate-distortion-perception tradeoff based on the conditional-distribution perception measure,” IEEE Trans. Inf. Theory , vol. 70, no. 12, pp. 8432–8454, 2024

    2024

  12. [20]

    The rate-distortion-perception trade-off: the role of private randomness,

    Y. Hamdi, A. B. Wagner, and D. Gündüz, “The rate-distortion-perception trade-off: the role of private randomness,” in 2024 IEEE Int. Symp. Inf. Theory (ISIT) , 2024, pp. 1083–1088

    2024

  13. [21]

    Optimal neural compressors for the rate-distortion- perception tradeoff,

    E. Lei, H. Hassani, and S. S. Bidokhti, “Optimal neural compressors for the rate-distortion- perception tradeoff,” 2025. [Online]. A vailable: https://arxiv.org/abs/2503.17558

    2025 arXiv

  14. [22]

    Rate-distortion-perception theory for semantic com- munication,

    J. Chai, Y. Xiao, G. Shi, and W. Saad, “Rate-distortion-perception theory for semantic com- munication,” in 2023 IEEE 31st Int. Conf. Netw. Protocols (ICNP) , 2023, pp. 1–6

    2023

  15. [23]

    Semantic communication via rate distortion perception bottleneck,

    Z. Zhao and C. Wang, “Semantic communication via rate distortion perception bottleneck,”

  16. [24]

    Rate–distortion–perception trade-off in information theory, generative models, and intelligent communications,

    X. Niu, B. Bai, N. Guo, W. Zhang, and W. Han, “Rate–distortion–perception trade-off in information theory, generative models, and intelligent communications,” Entropy, vol. 27, no. 4, 2025. 33

    2025

  17. [25]

    Strong functional representation lemma and applications to coding theorems,

    C. T. Li and A. E. Gamal, “Strong functional representation lemma and applications to coding theorems,” IEEE Trans. Inf. Theory , vol. 64, no. 11, pp. 6967–6978, 2018

    2018

  18. [26]

    T. M. Cover, Elements of Information Theory , 2nd ed. John Wiley & Sons, 2006

    2006

  19. [27]

    Conditional probability models for deep image compression,

    F. Mentzer, E. Agustsson, M. Tschannen, R. Timofte, and L. V. Gool, “Conditional probability models for deep image compression,” in 2018 IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 4394–4402

    2018

  20. [28]

    Generative adversarial networks for extreme learned image compression,

    E. Agustsson, M. Tschannen, F. Mentzer, R. Timofte, and L. Van Gool, “Generative adversarial networks for extreme learned image compression,” in 2019 IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2019, pp. 221–231. 34

    2019

  21. [2021]

    A vailable: https://openreview.net/forum?id=BzUaLGtKecs

    [Online]. A vailable: https://openreview.net/forum?id=BzUaLGtKecs

  22. [2024]

    A vailable: https://arxiv.org/abs/2405.09995

    [Online]. A vailable: https://arxiv.org/abs/2405.09995

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.