Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

A single fixed codebook can quantize Gaussian weights within 0.11 bit per dimension of the best covariance-adapted benchmark, for every input covariance at once.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A universal vector-quantization codebook exists that is within 0.11 bit/coordinate of covariance-adaptive waterfilling simultaneously for all input covariances, for Gaussian weights.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Main theorem is solid and worth citing; the 0.11-bit headline is real but conditional on an unreproduced numerical claim, so the abstract overstates it. the 3 major comments →

arxiv 2602.05790 v2 pith:5FXYJD7P submitted 2026-02-05 cs.IT cs.LGmath.ITstat.ML

Price of metric universality in vector quantization is at most 0.11 bit

Alina Harbuzova , Or Ordentlich , Yury Polyanskiy This is my paper
classification cs.IT cs.LGmath.ITstat.ML MSC 94A2994A17
keywords vector quantizationrate-distortionwaterfillinguniversal codebookweight-only quantizationHilbert normscovering netregret
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper tries to establish that a universal vector-quantization codebook—one fixed set of codewords, chosen before seeing any data statistics—can serve all possible input covariance matrices nearly as well as a codebook custom-tuned to each covariance. Concretely, for Gaussian weights W and a distortion metric shaped by the covariance Sigma of the activations X, the paper claims the rate overhead of universality over the waterfilling oracle is at most 0.11 bit per dimension, for distortions between 0.005 and 0.995. This matters for hardware: dequantizers must be hard-wired before activation statistics are known, and the result says that constraint costs almost nothing in rate-distortion terms. The paper also phrases the claim geometrically: there exists a net on the unit sphere that is a near-optimal covering simultaneously for all Hilbert norms. The existence proof is non-constructive, and the 0.11 constant is conditional on a numerically verified claim about spectra with at most five distinct eigenvalues.

Core claim

For W ~ N(0,I_n), with distortion d_Sigma(W,What)=(W-What)^T Sigma (W-What), the paper compares two rate-distortion functions: the waterfilling curve R_wf, the benchmark when both encoder and decoder know Sigma, and the random-coding curve R_rc, achieved by an isotropic Gaussian codebook whose overall scale is chosen by the encoder. The main discovery is that sup over spectra (trace n) of R_rc(lambda,D*) - R_wf(lambda,D*) is at most 0.11 bit for every distortion D* in [0.005,0.995], and that a single codebook simultaneously achieves distortion D_rc(spec Sigma, R*) + eta for all Sigma. In other words, the price of being Sigma-oblivious is a bounded, small constant rather than an unbounded los

What carries the argument

The central object is the random-coding rate-distortion function (RDRC): R_rc(lambda,T)= (1/2n) sum log(1+lambda_i T), D_rc(lambda,T)=(1/n) sum lambda_i/(1+lambda_i T). It arises from drawing M iid N(0,I_n) codewords and letting the encoder (who knows Sigma) choose a scalar scale tau to minimize the large-deviations exponent. The proof of Theorem 3 reduces the worst-case gap to spectra with at most five distinct eigenvalues, using Schur-concavity of the gap in the inactive eigenvalues and a KKT analysis of the active ones; the final 0.11 bound is then verified numerically in Claim D.4.

Load-bearing premise

The 0.11-bit constant depends on Claim D.4, a numerical sweep—reported only as a figure, with no code or grid—showing the gap stays below 0.11 for all spectra with at most five distinct eigenvalues; if that sweep is incomplete or wrong, the constant fails, although the existence of a universal codebook survives.

What would settle it

For any spectrum with at most five distinct eigenvalues and trace n, compute R_rc(lambda,D*) - R_wf(lambda,D*) at D* in [0.005,0.995]; if any value exceeds 0.11, Theorem 3 is false. Since the paper reduces the supremum to such spectra, one counterexample (or a reproducible grid showing a gap above 0.11) would settle it.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • There exists a fixed, hardware-friendly codebook whose worst-case distortion is within the random-coding curve D_rc of the optimal covariance-adapted scheme, simultaneously across all Sigma with trace n.
  • If the 0.11-bit bound holds, a single dequantizer design works for all neurons and all activation statistics, removing the need for per-layer calibration.
  • Geometrically, there is a net on the unit sphere that covers the sphere within a constant factor of optimal for every Hilbert norm at once.
  • Lattice-based quantizers cannot achieve this universality: any fixed lattice loses at least about 0.254 bit for some covariance, so the universal object must come from a non-lattice construction.
  • The rate penalty epsilon in Theorem 2 can be made arbitrarily small for large n, so the 0.11-bit price is the dominant asymptotic cost of universality.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: If Claim D.4 is sharpened or reproduced with an openly available exhaustive grid, the worst-case spectra identified by the KKT reduction could guide explicit constructions of universal codebooks.
  • Editorial inference: The equivalence to the additive rate-distortion function for a colored Gaussian source suggests the same single test channel is simultaneously near-optimal for all source covariances under quadratic loss, which may be of independent interest.
  • Editorial inference: An empirical test on real LLM weight matrices—comparing a universal codebook against per-layer calibrated codebooks—would show whether 0.11 bit is also the practical gap, not just the worst-case theoretical one.
  • Editorial inference: The same reduction technique (at most five distinct eigenvalues at the extremum) may apply to other rate-gap problems where two parametric rate-distortion curves are compared.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper studies universal vector quantization for the weighted squared-error metric d_ΣX(W,Ĉ)=(W−Ĉ)^T Σ_X(W−Ĉ), motivated by weight-only LLM quantization. The encoder observes Σ_X, but the decoder/codebook is fixed and Σ_X-oblivious. For W∼N(0,I_n), Theorem 2 asserts the existence of a single random codebook that for every PSD Σ_X with tr(Σ_X)=n achieves distortion D_rc(spec(Σ_X), R*)+η at rate R*+ε, with high probability over the codebook draw. Theorem 3 asserts that the worst-case rate gap sup_Λ{R_rc(Λ,D*)−R_wf(Λ,D*)} is at most 0.11 bit for D*∈[0.005,0.995], conditional on a numerical verification stated as Claim D.4. The proofs use i.i.d. Gaussian codebooks with a data-dependent scale τ, large-deviation bounds for the success probability of one codeword, a covering argument over spectra and orthogonal matrices, and a reduction of the supremum in Theorem 3 to spectra with at most five distinct eigenvalues.

Significance. If correct, the paper establishes a strong and non-obvious statement: the price of universality in vector quantization is a bounded, small constant (≤0.11 bit/dim) relative to the Σ_X-adapted waterfilling oracle, with no dependence on dimension. This is of genuine interest to information theory and to practical LLM weight quantization. The random-coding analysis is detailed, the nonasymptotic Theorems 4 and 5 are explicit, and the connection to additive rate-distortion (Appendix G) is elegant. The paper is also honest in flagging the numerical Claim D.4 as an assumption of Theorem 3. However, the central contributions are currently not fully established: the proof of Theorem 2 contains a union-bound scaling error, and Theorem 3's 0.11-bit conclusion has an unresolved ε-margin gap in addition to relying on an unreproduced numerical verification.

major comments (3)
  1. [§C.2, after Eq. (32) and Claim C.1] The simultaneous guarantee of Theorem 2 is not established by the written covering argument. Plugging γ=2^{-5nR*} into the bound N N' ≤ (c/γ)^{n^2+n} gives N N' ≈ 2^{5R*(n^3+n^2)}. The union bound then yields success probability at least 1 − exp(−2nε + 5R*(n^3+n^2)), which is vacuous for large n, not the claimed 1 − exp(−2nε(1−c n^{4α−1})). A covering of O(n) at exponentially small resolution necessarily has exp(Θ(n^3)) points, overwhelming the single-exponential probability from Theorem 5. To repair this, the authors need either a net of size 2^{o(n)} (for example by exploiting unitary invariance of the random codebook to avoid covering U at exponential resolution), or a double-exponential probability bound for each fixed Σ_X.
  2. [§D, Claim D.1 and Eq. (34)] The ε-slack bookkeeping in the proof of Theorem 3 is inconsistent. Claim D.1 uses f_n=(D*−ε)/(2n) and concludes sup ≤ restricted_sup + ε. Immediately thereafter the proof says "For ε>0 and f_n = D*/(2n)", which corresponds to ε=0 in Claim D.1; Claim D.4 then asserts only that the restricted supremum is <0.11, with no margin of ε and with ε, f_n not even defined in the claim. Thus the additive ε from Claim D.1 is not absorbed, and the stated 0.11 bound does not follow from the claims as written. The authors must either verify Claim D.4 with the same f_n as Claim D.1 and with a margin strictly below 0.11, or otherwise account for ε explicitly.
  3. [Claim D.4 and Fig. 1] The headline 0.11-bit constant depends entirely on Claim D.4, a numerical verification over spectra with at most five distinct eigenvalues. The manuscript provides no code, no data, no grid resolution, no tolerance, and no description of the optimizer; Figure 1 has no axis labels or numerical values. Since Theorem 3 is explicitly conditional on this claim, the central numerical assertion is not independently verifiable from the manuscript. If this is intended as a computer-assisted proof, the verification artifact (including the search method and tolerances) should be made available; otherwise the theorem should be rephrased as a conditional statement with the numerical evidence supplied.
minor comments (4)
  1. [Abstract and Theorem 1 (informal)] The abstract and the informal Theorem 1 omit the restriction D*∈[δ,1−δ] with δ=0.005 that appears in Theorem 3. As stated, the informal claims suggest the 0.11-bit gap holds for all distortion levels, which is not what is proved.
  2. [Theorem 3 statement] The statement says the supremum is over λ∈R^n_+, but the surrounding section makes clear it must be over spectra with tr(Λ)=n. Without the trace constraint the supremum is not meaningful for a fixed D*.
  3. [Section 3.2, Eq. (24) area] The text says the covering needs "N = exp(−2^{o(n)})" PSD matrices; since N is a cardinality, this should presumably be exp(2^{o(n)}). Please correct the sign/notation.
  4. [Figure 1] The figure caption 'Maximum rate gap found numerically at each R=R_rc(λ,D*)' would benefit from labeled axes, a legend, and a statement of the sweep parameters (grid over λ and D*, number of distinct eigenvalues).

Circularity Check

0 steps flagged

No circular reduction: the universal-codebook guarantee and the 0.11-bit gap are derived from closed-form rate-distortion formulas; the numerical Claim D.4 is an evidence gap, not a circular step.

full rationale

I walked the derivation chain. Theorem 2's codebook is drawn iid from N(0, tau^2 I), and the target curve D_rc is the closed-form parametric curve in (RDRC), not a parameter fitted to the codebook's empirical distortion. Achievability of D_rc is proved through large-deviation estimates (Lemma B.2), so the theorem is a genuine random-coding bound. The benchmark R_wf is the independent waterfilling formula (WF), and Theorem 3 compares two closed-form rate functions. The reduction of the worst-case gap to spectra with at most five distinct eigenvalues is an analytic compactness/KKT argument (Claims D.1-D.3). The final 0.11 constant rests on Claim D.4, stated as 'numerical' with only Figure 1 and no code, grid, or tolerance, and the paper explicitly makes Theorem 3 conditional on that claim; this is a reproducibility/correctness concern, not a circular reduction. The epsilon bookkeeping in Claim D.1 (epsilon slack, then setting f_n = D*/(2n) without an explicit 0.11-epsilon margin) is likewise a potential proof gap, not an identity. Self-citations to [OP25], [OP26], and [PW24] are contextual, standard textbook material, or side remarks; none is used to define the target curve, fit a constant, or forbid alternatives. No step exhibits a prediction equivalent to its input by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities. Its central assumptions are the Gaussian-weight model and trace normalization; the most fragile input is the numerical Claim D.4, which supplies the 0.11 constant without an analytic proof or reproducible code.

free parameters (1)
  • δ = 0.005 (D* range lower bound) = 0.005
    Chosen arbitrarily in Theorem 3 'to verify numerical claim in a stable way'; the 0.11-bit bound is only claimed for D* ∈ [δ, 1−δ].
axioms (5)
  • domain assumption W ~ N(0, I_n)
    Modeling assumption stated in Section 2 and abstract; the result is specialized to Gaussian weights.
  • domain assumption Σ_X ∈ S^n_+ with tr(Σ_X) = n
    Normalization convention in Theorems 2 and 3; scale is fixed but the rank/profile varies.
  • standard math Waterfilling rate-distortion lower bound (Prop. 2.1) is the oracle benchmark
    The oracle lower bound is standard rate-distortion theory for Gaussian sources under weighted squared error.
  • standard math Probabilistic method: random iid Gaussian codebook of size 2^{nR} with concentration and covering
    Main technical engine; standard in random coding and high-dimensional geometry.
  • ad hoc to paper Claim D.4 numerical verification is correct and exhaustive
    The 0.11 bound depends on an unproved numerical sweep over spectra with at most 5 distinct eigenvalues; only Fig. 1 is provided.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Price of metric universality in vector quantization is at most 0.11 bit." pith.science (2026). https://pith.science/paper/5FXYJD7P

@misc{pith2026260205790,
  author       = {Pith},
  title        = {Pith review of: Price of metric universality in vector quantization is at most 0.11 bit},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5FXYJD7P}},
  note         = {Machine review of arXiv:2602.05790}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Fast computation of a matrix product $W^\top X$ is a workhorse of modern LLMs. To make their deployment more efficient, a popular approach is that of using a low-precision approximation $\widehat W$ in place of true $W$ (``weight-only quantization''). Information theory demonstrates that an optimal algorithm for reducing precision of $W$ depends on the (second order) statistics of $X$ and requires a careful alignment of vector quantization codebook with PCA directions of $X$ (a process known as ``waterfilling allocation''). Dependence of the codebook on statistics of $X$, however, is highly impractical. This paper proves that there exist a universal codebook that is simultaneously near-optimal for all possible statistics of $X$, in the sense of being at least as good as an $X$-adapted waterfilling codebook with rate reduced by 0.11 bit per dimension in the case when $W$ is Gaussian. Such universal codebook would be an ideal candidate for the low-precision storage format, a topic of active modern research, but alas the existence proof is non-constructive. Equivalently, our result shows existence of a net in $\mathbb{R}^n$ that is a nearly-optimal covering of a sphere simultaneously with respect to all Hilbert norms.

Figures

Figures reproduced from arXiv: 2602.05790 by Alina Harbuzova, Or Ordentlich, Yury Polyanskiy.

Figure 1
Figure 1. Figure 1: Maximum rate gap found numerically at each [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. High-Rate Quantized Matrix Multiplication II

    cs.LG 2026-05 unverdicted novelty 6.0

    Waterfilling rate allocation makes quantized matrix multiplication for LLMs near information-theoretically optimal, with WaterSIC being basis-free and within 0.25 bits per entry of the limit.

  2. High-Rate Quantized Matrix Multiplication II

    cs.LG 2026-05 unverdicted novelty 5.0

    With known covariance, waterfilling improves GPTQ and WaterSIC reaches within 0.25 bit/entry of the rate-distortion limit while being basis-independent.

Reference graph

Works this paper leans on

27 extracted references · 12 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Croci, Bo Li, Pashmina Cameron, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman

    Saleh Ashkboos, Amirkeivan Mohtashami, Maximilian L. Croci, Bo Li, Pashmina Cameron, Martin Jaggi, Dan Alistarh, Torsten Hoefler, and James Hensman. Quarot: Outlier-free 4-bit inference in rotated llms, 2024. URL: https://arxiv.org/abs/2404.00456, https://arxiv.org/abs/2404.00456 arXiv:2404.00456

    Pith/arXiv arXiv 2024

  2. [2]

    On lov \'a sz’lattice reduction and the nearest lattice point problem

    L \'a szl \'o Babai. On lov \'a sz’lattice reduction and the nearest lattice point problem. Combinatorica , 6(1):1--13, 1986

    1986

  3. [3]

    The lattice geometry of neural network quantization--a short equivalence proof of gptq and babai's algorithm

    Johann Birnick. The lattice geometry of neural network quantization--a short equivalence proof of gptq and babai's algorithm. arXiv preprint arXiv:2508.01077 , 2025

    arXiv 2025

  4. [4]

    Half-quadratic quantization of large machine learning models, November 2023

    Hicham Badri and Appu Shaji. Half-quadratic quantization of large machine learning models, November 2023. URL: https://mobiusml.github.io/hqq_blog/

    2023

  5. [5]

    Quip: 2-bit quantization of large language models with guarantees, 2024

    Jerry Chee, Yaohui Cai, Volodymyr Kuleshov, and Christopher De Sa. Quip: 2-bit quantization of large language models with guarantees, 2024. URL: https://arxiv.org/abs/2307.13304, https://arxiv.org/abs/2307.13304 arXiv:2307.13304

    Pith/arXiv arXiv 2024

  6. [6]

    Wush: Near-optimal adaptive transforms for llm quantization

    Jiale Chen, Vage Egiazarian, Torsten Hoefler, and Dan Alistarh. Wush: Near-optimal adaptive transforms for llm quantization. arXiv preprint arXiv:2512.00956 , 2025

    Pith/arXiv arXiv 2025

  7. [7]

    Fast quantizing and decoding and algorithms for lattice quantizers and codes

    John Conway and Neil Sloane. Fast quantizing and decoding and algorithms for lattice quantizers and codes. IEEE Transactions on Information Theory , 28(2):227--232, 1982

    1982

  8. [8]

    The geometry of llm quantization: Gptq as babai's nearest plane algorithm

    Jiale Chen, Yalda Shabanzadeh, Elvir Crn c evi \'c , Torsten Hoefler, and Dan Alistarh. The geometry of llm quantization: Gptq as babai's nearest plane algorithm. arXiv preprint arXiv:2507.18553 , 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale. Advances in Neural Information Processing Systems , 35:30318--30332, 2022

    2022

  10. [10]

    Gptq: Accurate post-training quantization for generative pre-trained transformers, 2023

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers, 2023. URL: https://arxiv.org/abs/2210.17323, https://arxiv.org/abs/2210.17323 arXiv:2210.17323

    Pith/arXiv arXiv 2023

  11. [11]

    Vector quantization and signal compression , volume 159

    Allen Gersho and Robert M Gray. Vector quantization and signal compression , volume 159. Springer Science & Business Media, 2012

    2012

  12. [12]

    Spinquant: Llm quantization with learned rotations, 2025

    Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. Spinquant: Llm quantization with learned rotations, 2025. URL: https://arxiv.org/abs/2405.16406, https://arxiv.org/abs/2405.16406 arXiv:2405.16406

    Pith/arXiv arXiv 2025

  13. [13]

    Wornell, and Ram Zamir

    Emin Martinian, Gregory W. Wornell, and Ram Zamir. Source coding with distortion side information. IEEE Transactions on Information Theory , 54(10):4638--4665, 2008. https://doi.org/10.1109/TIT.2008.928983 doi:10.1109/TIT.2008.928983

    arXiv 2008

  14. [14]

    Pretraining large language models with NVFP4

    NVIDIA et al. Pretraining large language models with NVFP4 . arXiv preprint arXiv:2509.25149 , 2025

    arXiv 2025

  15. [15]

    Optimal quantization for matrix multiplication

    Or Ordentlich and Yury Polyanskiy. Optimal quantization for matrix multiplication. IEEE Transactions on Information Theory , 2025

    2025

  16. [16]

    High-rate quantized matrix multiplication: Theory and practice

    Or Ordentlich and Yury Polyanskiy. High-rate quantized matrix multiplication: Theory and practice. arXiv preprint arXiv:2601.17187 , 2026

    Pith/arXiv arXiv 2026

  17. [17]

    OCP microscaling formats ( MX ) specification

    Open Compute Project . OCP microscaling formats ( MX ) specification. Technical report, Open Compute Project, 2023. URL: https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf

    2023

  18. [18]

    New bounds on the density of lattice coverings

    Or Ordentlich, Oded Regev, and Barak Weiss. New bounds on the density of lattice coverings. Journal of the American Mathematical Society , 35(1):295--308, 2022

    2022

  19. [19]

    Incremental refinement using a gaussian test channel

    Jan Ostergaard and Ram Zamir. Incremental refinement using a gaussian test channel. In 2011 IEEE International Symposium on Information Theory Proceedings , pages 2233--2237, 2011. https://doi.org/10.1109/ISIT.2011.6033957 doi:10.1109/ISIT.2011.6033957

    arXiv 2011

  20. [20]

    Information theory: From coding to learning

    Yury Polyanskiy and Yihong Wu. Information theory: From coding to learning . Cambridge university press, 2024

    2024

  21. [21]

    Nestquant: Nested lattice quantization for matrix products and llms, 2025

    Semyon Savkin, Eitan Porat, Or Ordentlich, and Yury Polyanskiy. Nestquant: Nested lattice quantization for matrix products and llms, 2025. URL: https://arxiv.org/abs/2502.09720, https://arxiv.org/abs/2502.09720 arXiv:2502.09720

    Pith/arXiv arXiv 2025

  22. [22]

    Quip\#: Even better llm quantization with hadamard incoherence and lattice codebooks, 2024

    Albert Tseng, Jerry Chee, Qingyao Sun, Volodymyr Kuleshov, and Christopher De Sa. Quip\#: Even better llm quantization with hadamard incoherence and lattice codebooks, 2024. URL: https://arxiv.org/abs/2402.04396, https://arxiv.org/abs/2402.04396 arXiv:2402.04396

    Pith/arXiv arXiv 2024

  23. [23]

    Qtip: Quantization with trellises and incoherence processing, 2025

    Albert Tseng, Qingyao Sun, David Hou, and Christopher De Sa. Qtip: Quantization with trellises and incoherence processing, 2025. URL: https://arxiv.org/abs/2406.11235, https://arxiv.org/abs/2406.11235 arXiv:2406.11235

    Pith/arXiv arXiv 2025

  24. [24]

    Model-preserving adaptive rounding, 2025

    Albert Tseng, Zhaofeng Sun, and Christopher De Sa. Model-preserving adaptive rounding, 2025. URL: https://arxiv.org/abs/2505.22988, https://arxiv.org/abs/2505.22988 arXiv:2505.22988

    Pith/arXiv arXiv 2025

  25. [25]

    Smoothquant: Accurate and efficient post-training quantization for large language models, 2024

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: Accurate and efficient post-training quantization for large language models, 2024. URL: https://arxiv.org/abs/2211.10438, https://arxiv.org/abs/2211.10438 arXiv:2211.10438

    Pith/arXiv arXiv 2024

  26. [26]

    The rate loss in the wyner-ziv problem

    Ram Zamir. The rate loss in the wyner-ziv problem. IEEE Transactions on Information Theory , 42(6):2073--2084, 2002

    2073

  27. [27]

    Multiterminal source coding with high resolution

    Ram Zamir and Toby Berger. Multiterminal source coding with high resolution. IEEE Transactions on Information Theory , 45(1):106--117, 2002

    2002

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.