acceptodds
Under review as a conference paper at ICLR 2027

Curvature as Decoder State: Exact Cache–Update Frontiers for Strongly Convex Unlearning

Abstract

Machine unlearning depends on what task-dependent state survives deployment and what a later deletion request reveals. We separate pre-request decoder state from post-request delivery and study reconstruction of a fixed retraining procedure under an explicit serialized interface. For smooth strongly convex objectives, parameters plus a revealed deletion gradient leave an exact -dimensional compatible-target ball, while exact curvature on orthogonal directions removes precisely dimensions from the high-rate description length. A request-oblivious cache that must hedge future directions instead preserves symmetric degrees of freedom, distinguishing the state-only and request-time coefficients. Ridge regression realizes these statements through an exact Woodbury downdate: retained rows and exact sufficient statistics require no additional transcript under their stated interfaces, whereas parameters alone require bits on a compact family. A fixed float64 audit covering cells, paired requests, and state/interface evaluations confirms that sufficient statistics are dominated by parameters on all nine identifier-only surfaces but become nondominated when requests expose deleted values. Gaussian-design tests recover the predicted request scaling and curvature fluctuations for retained sample size , while the selected curvature caches in our main comparison reduce delivery by – only by increasing serialized state and requiring repeated requests to amortize it. The resulting frontier prices curvature as decoder state, not free compression, and makes no zero-memory, universal finite-bit, or end-to-end neural-converse claim.

Primary area:
learning theory
Keywords:
Continual learning, machine unlearning, behavioral retraction, information theory, rate-distortion theory, memory-efficiency, machine learning theory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_tgEcOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper separates information retained before a deletion request, measured by cache size MM, from information delivered afterward, measured by communication BB. It characterizes ambiguity in exact retraining targets when curvature is missing, studies partial-curvature interfaces, and distinguishes the information required for one realized request from the state needed to anticipate multiple possible requests.

Critical strengths
  1. The paper is generally well written, with a clear problem formulation and an organized presentation of the theoretical results and empirical evaluations.

  2. The problem setup is specified, explicitly distinguishing pre-request storage from post-request communication and clarifying the information available at each stage.

Critical weaknesses
  1. Theorem 3.3 and Appendix Theorem C.3 establish matching leading-order bounds at the endpoints, but leave a gap at storage–communication tradeoffs. Thus, the claimed “exact frontier” is not right.

  2. Corollary 3.2 assumes block-diagonal curvature and exact partial-curvature information. Its dimensional savings therefore have limited direct applicability to practical curvature caches with coupled directions.

  3. The controlled partial-curvature experiment primarily verifies the prescribed achievable construction. Although natural-data witnesses support the ambiguity premise, the experiments provide limited evidence that the worst-case bounds predict practical storage–communication tradeoffs.

Questions for authors

Does the amortization calculation M+TBM+TB assume independent requests against the same initial model or sequential, cumulative deletions?

2 comments
AuthorsOpenReview ↗

Response to Reviewer tgEc

Thank you for the review and the question about amortization. The table summarizes our answers.

Your pointOur answerWhere in the revision
The claimed exact frontier leaves a gap in the interiorThank you. A new construction makes the frontier exact for every state rate above r(r−1)/2r(r-1)/2 and for r≤2r\le2, and the new title states this scope (reply to Reviewer Trgo).Title; abstract; Theorem 3.3; Appendix C.2
Corollary 3.2 assumes block-diagonal curvatureCorrect. Its exact (d−s)(d-s) coefficient applies to exact partial curvature, and the natural-data measurements are reported separately from it.Corollary 3.2; Appendix D.5
The experiments mostly verify the constructionThe natural-data witnesses test the ambiguity premise, and the new entropy-coded comparison tests whether the reductions survive a stronger codec. Most do not: on Tables 12–13 the cache's saving falls to 0.3–1.6%.Appendix D.9
Question: independent or cumulative deletions in M+TBM+TB?Independent requests against the same frozen model and cache. The break-even horizon does not transfer to cumulative deletions, which also change the remaining data and the cache. A new serializer-level run adds a cache refresh every 25 requests: the cache pays for itself after 10 requests, and 100 requests cost 10,208 B against 12,980 B without it.Appendices D.3, D.7 and D.9

Corollary 3.2 requires block-diagonal curvature and exact partial information. Its exact (d−s)(d-s) coefficient cannot be transferred to approximate caches for coupled natural Hessians. Appendix D.5 already treats the nonquadratic measurements separately from a direct test of that corollary. The natural-data reweighting witnesses establish local target ambiguity; their candidate-limited packing does not establish a worst-case coefficient for arbitrary neural models.

The amortization study uses independent deletion requests against the same frozen reference model and cache, as specified in Appendix D.3 and the neural restart protocol in D.7. Under a common request interface, M+TBM+TB accounts for one-time state placement and repeated additional messages; comparisons across interfaces must also include QQ. These results do not model cumulative deletions.

Sequential deletion changes the remaining data, retraining target, and possibly the cache after each request. Its accounting must include state growth, request bytes, additional messages, cache-refresh delivery, and refresh computation. The fixed-reference break-even horizon therefore applies to repeated requests against a frozen reference, not automatically to a cumulative-deletion service.

The revision adds a serializer-level run that charges refresh delivery and refresh computation (Appendix D.9). It uses d=100d=100, r=10r=10, state rate RM=300R_M=300, and ε=10−3\varepsilon=10^{-3}. The serialized cache state is 392 B, and each request costs 86.4 B with the cache and 129.8 B without it. The cache is refreshed every 25 requests, and each refresh takes 0.41 ± 0.03 ms over five traces.

RequestsRefreshes after the initial cacheWith cacheWithout cacheSaving
1001,256 B1,298 B42 B
2502,552 B3,245 B693 B
5015,104 B6,490 B1,386 B
100310,208 B12,980 B2,772 B

The cache pays for itself after ⌈392/(129.8−86.4)⌉=10\lceil392/(129.8-86.4)\rceil=10 requests. In this run, each request's message has the same size between refreshes. The run therefore prices the refresh schedule but does not model how cumulative deletions change the data between refreshes.

Please let us know whether these answers address your question. We will reply to any further point during the discussion period.

thanks for your reply. it mainly solve my concern, and i will raise my score

Reviewer_62epOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies machine unlearning by using a pre-request cached decoder state and post-request update message to save memory and exploiting curvature structure.

Critical strengths

The separation of pre and post deletion request management itself appears to be interesting and unlearning is an important problem.

Critical weaknesses
  1. This paper is very poorly written and difficult to understand the content. Many notations in the main paper were not carefully defined/explained: examples RY+(D+)R_{Y^+}(D_{+}) are introduced, but the underlying topology wasn't introduced until Appendix B. Similarly B∗(ϵ)B^*(\epsilon), JQJ_Q are not defined. Anything appear in the main text should either be defined or clearly referenced if the definition exists later in the paper.

  2. Related to the previous note, mm is both strong convexity and sample size.

  3. Some terms are used without explanation. What do the authors mean by "data-native witness" or "value-exposing requests".

  4. The paper seems unfinished. On page 10 it says ""Experiments were run using [ACTUAL CPU/GPU HARDWARE] with [ACTUAL RAM/VRAM], using Python [VERSION]" without specifying the details.

  5. On the math front, I do not see much novelty in the analysis. Also unclear how the deleted gradient can be received by the decoder (without communicating a d−d-dimensional object.

  6. Numerical experiment shows small improvements.

Questions for authors

Please consider carefully rewrite the paper before resubmitting.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 62ep

Thank you for the review. We agree that the submission was hard to read and that it left the hardware sentence unfilled. The table lists each point and what the revision does.

Your pointOur answerWhere in the revision
Notation used before it is defined, such as the covering topology, B⋆(ε)B^\star(\varepsilon), and JQJ_QEach is defined where it is first used. The covering topology moves from Appendix B into Section 2.Sections 2 and 3
One symbol for both strong convexity and sample sizeSeparate symbols: μ\mu is now the strong-convexity constant, while mm remains the retained sample size and nn the number of records.Throughout
"Data-native witness" and "value-exposing request"A value-exposing request supplies the deleted records' values, not only their identifiers. A data-native witness uses existing rows and positive reweightings that keep the deployed optimum but change the compatible curvature (Proposition A.1).Section 2
The unfilled hardware sentence on page 10Thank you for catching it. The revision removes this unfilled template sentence.Reproducibility statement
How the decoder receives the deleted gradient without a dd-dimensional messageThe gradient is part of the request, whose bits QQ are counted separately. BB is the additional message, so total request communication is Q+BQ+B.Section 2; Theorem 3.1; Appendices C.1 and D.1
Novelty of the analysisThe analysis gives the exact dd-dimensional ambiguity ball for retraining targets, the exact removal of ss dimensions by partial curvature, and the prp_r hedging dimension for request-oblivious caches, with matching endpoint bounds. The revision adds an exact frontier above the shared-state threshold. Reviewer Trgo checked the original arguments and found no gap.Section 3; Appendix C.2
Small numerical improvementsCorrect, and entropy coding shrinks them to 0.3–1.6% on Tables 12–13 (reply to Reviewer Trgo). The paper prices curvature as decoder state rather than claiming a compression gain.Abstract; Appendix D.9

The deletion gradient is an explicit revealed input in Theorem 3.1, lines 173–177. The additional message BB excludes declared request contents; total request communication is Q+BQ+B. Appendix C.1 distinguishes these interfaces, and Appendix D.1 gives request sizes of 1,992/2,512/3,168 bits for identifiers versus 9,760/31,840/59,368 bits when record values are exposed. Thus B=0B=0 under a value-exposing interface does not mean zero communication.

“Value-exposing” means that the request supplies deleted record values, not just identifiers. The “data-native” construction in Proposition A.1 uses existing rows and positive reweightings that preserve the deployed optimum while changing compatible curvature. The revision introduces these concrete meanings before the terminology, defines B⋆(ε)B^\star(\varepsilon) and JQJ_Q where first used, and separates the symbols for strong convexity and retained sample count.

The revision removes the unfilled template sentence on page 10. The reported 6.8–19.6% delivery reductions are fixed-width savings, and they require larger retained states and repeated requests for amortization. The revision states the state cost and request interface beside each reduction, rather than describing additional-message savings as an end-to-end compression guarantee.

Please let us know whether the revision addresses your concerns. We will reply to any further point during the discussion period.

Thank you for the point-by-point response by the authors. The added the explanation does clarify some of my concerns. I still found the paper overall to be written in a confusing way. I'm increasing my score from 1 to 2.

Reviewer_TrgoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies machine unlearning from information-theoretic viewpoint. It separates the decoder state kept before a deletion request (M bits) from the update message sent after the request (B bits), and gives converse bounds for both by data processing and conditional covering number.

The main results concern curvature. For smooth strongly convex objective, a decoder that holds only the deployed parameters and the deleted gradient faces an exact d-dimensional ball of possible retraining targets, so the request-time description length has coefficient d (Theorem 3.1). Caching exact curvature on an s-dimensional subspace removes exactly s dimensions (Corollary 3.2). A request-oblivious cache that hedges r directions must keep p_r = rd − r(r−1)/2 degrees of freedom (Theorem 3.3). Ridge regression realizes these statements exactly: sufficient statistics give B = 0 under value-exposing request, while parameters alone need Θ(d log 1/ε) bits.

The experiments confirm the predicted coefficients, the √k request scaling and the m^{−3/2} curvature fluctuation, and measure delivery reduction of 6.8% to 19.6% on ridge, logistic, frozen deep features and two neural networks together with the state growth and crossover horizon. The numbers I recomputed are internally consistent.

Critical strengths

The lower and upper bounds in Theorems 3.1, 3.3 and C.2 match in the leading coefficient, and the family that attains the ball is built explicitly by Householder map (Lemma J.1). Each theorem states its interface assumption inside the statement, so it is clear what is public and what is charged to M. The proofs in Appendices H to J are complete and use only standard tools, which the authors say honestly. I checked the main arguments and did not find a gap. The novelty is modest, but the precision is high.

Critical weaknesses

W1. No entropy coding in the residual codec. The update packet in Sections D.3, D.5 and D.7 uses uniform quantization, then picks one integer bit width b that covers the largest index, and packs d indices with fixed length. So B is decided by the dynamic range of the single worst coordinate, not by the information in the residual. This matters because the main claim is how much a curvature cache reduces B. If the cache makes the residual distribution narrower but the maximum does not change, B stays the same. The K-FAC result in Table 16 shows this clearly: the 10.0% reduction for both MLP and ResNet-18 is exactly a width change from 10 bits to 9 bits (Table 15). Also, the fixed 18 byte header is two thirds of the 27 byte packet when d = 8, so the headline range 6.8% to 19.6% mixes tasks with very different header share. The authors already implemented a transform codec with reverse water-filling and entropy coding in Section D.4. I do not understand why it is not used in the other experiments. With that codec, the reported reductions may change a lot.

W2. “Exact” in the title holds only at the two endpoints. The lower bound (6) is M + rB ≥ p_r log(·), but the achievable code in (8) and (9) gives M + rB = α p_r log + (1 − α) r d log. Since rd > p_r, there is a gap of factor 2d/(d+1) for every interior α, as the authors note in Appendix C.2. The two endpoints where the bounds meet are the trivial cases: store the whole inverse curvature, or send the whole answer. I suggest to change the title and abstract to say that the frontier is exact at the endpoints and within factor two inside, or to close the gap.

Questions for authors

Q1. lease apply the transform codec from Section D.4 (reverse water-filling and entropy coding) to the experiments in Tables 3, 12, 13 and 16. How do the reductions change? Even a parameter-free signed Exp-Golomb code on the quantized residual indices, as used for se(v) in H.264, would be informative. It needs no tuning and its cost grows only logarithmically for the few large coordinates that currently decide the fixed bit width.

Q2. Why was the fixed-width packing chosen as the main codec, while the theory assumes an ε-cover index? Is there a reason the D.4 codec cannot be used for the neural residuals?

2 comments
AuthorsOpenReview ↗

Response to Reviewer Trgo - Part 1

Thank you for checking the proofs and for the precise codec questions. The table summarizes our answers, and the text after it gives the results and the frontier statement.

Your pointOur answerWhere in the revision
First question: entropy coding and the transform codecRun on Tables 12, 13 and 16 and on California. On the four Table 12–13 settings, the cache's saving falls from 6.8–19.6% under fixed-width coding to 0.5–1.6% with signed Exp-Golomb and 0.3–0.6% with the retuned D.4 transform. On the nonconvex models, a public transform keeps a 5–6% saving.Appendix D.9 (new)
The 18-byte header dominates small packetsEvery arm of the new comparison carries the same 18-byte header, so its body cost is the total minus 18 bytes.Appendix D.9 (new)
Second question: why fixed-width packing is the main codecIt gives an exact byte count and an error bound for every request, which is the criterion behind the theory's ε\varepsilon-cover. The D.4 codec targets an average Gaussian distortion, which does not bound each request's error, so it cannot replace the fixed-width codec without a direct check. Exp-Golomb coding of the same indices keeps the reconstruction exact, so it is a valid replacement, and its results are now reported.Appendix D.4
"Exact" holds only at the two endpointsThank you for pointing to the gap. A new construction closes it above a state rate of r(r−1)/2r(r-1)/2 and for r≤2r\le2, and the new title states this scope.Title; abstract; Theorem 3.3; Appendix C.2

The fixed-width serializer is explicit but not optimal. Appendix D.3 uses an 18-byte header and ⌈db/8⌉\lceil db/8\rceil body bytes, with width bb determined by the largest quantized coordinate. The 27-byte California baseline is therefore two-thirds header. Tables 15–16 also show that the neural 10% reduction is a width change from ten bits to nine. These are valid packet measurements under the declared codec, but they do not carry over to entropy coding, as the comparison below shows.

The Gaussian transform experiment in Appendix D.4 does not recode the natural-data or neural residuals. The comparison below applies signed Exp-Golomb coding to the same quantized indices, preserving reconstruction exactly, and separately tests a public orthogonal transform with matched distortion. A learned basis, scale, or codebook would have to be included in decoder state.

Setting and codecUnitParameters-only messageCurvature-cache message
California, signed Exp-Golomb indicesbytes22.422.9
MLP, signed Exp-Golomb indicesMiB0.2250.228
MLP, public transform and signed codingMiB0.2140.201
ResNet-18, public transform and signed codingMiB10.8010.21

Under signed Exp-Golomb coding of the same indices, the curvature-cache message is slightly larger than the parameters-only message on California (22.9 vs 22.4 bytes) and on the MLP (0.228 vs 0.225 MiB). With the public transform, it is smaller: 0.201 vs 0.214 MiB on the MLP and 10.21 vs 10.80 MiB on ResNet-18, about 6% and 5% less. The fixed-width gains therefore do not carry over to entropy-coded indices, while a smaller gain remains after the transform.

The comparison retains the 18-byte header and separates it from body cost. The transform arms match distortion on average, so unlike the fixed-width and Exp-Golomb arms, they do not bound the error of each request. Requests are paired within a base model; independent base fits, not residual coordinates, provide independent replication.

The revision also recodes the settings of Tables 12 and 13 at their parameter-RMSE target of at most 10−410^{-4}, with the 18-byte header included. The D.4 quantizers are retuned to the same target. Four rows of Table 3 (MNIST logistic, ResNet-feature logistic, and the CIFAR-10 ResNet and CIFAR-100 ViT ridge heads) are these same settings.

SettingCodecParameters-only (KiB)Curvature cache (KiB)Cache saving
MNIST 3 vs 8 (Table 12)Fixed-width1.481.3012.2%
Signed Exp-Golomb0.940.9251.6%
D.4 + signed Exp-Golomb0.8350.8300.6%
CIFAR-10 cat vs dog, ResNet-18 features (Table 12)Fixed-width1.120.9019.6%
Signed Exp-Golomb0.7160.7081.1%
D.4 + signed Exp-Golomb0.6420.6400.3%
CIFAR-10 / ResNet-18 head (Table 13)Fixed-width7.87.010.3%
Signed Exp-Golomb5.024.980.8%
D.4 + signed Exp-Golomb4.534.510.4%
CIFAR-100 / ViT-B/16 head (Table 13)Fixed-width116.2108.36.8%
Signed Exp-Golomb73.873.40.5%
D.4 + signed Exp-Golomb67.166.80.4%
AuthorsOpenReview ↗

Response to Reviewer Trgo - Part 2

The fixed-width rows reproduce Tables 12 and 13. With raw Exp-Golomb indices, the ResNet-18 head gives 5.44 against 5.41 KiB, a 0.6% saving. Under entropy coding, the cache's saving falls to 0.3–1.6%, consistent with entropy coding removing much of the same redundancy. The revision therefore states the abstract's 6.8–19.6% as fixed-width savings.

The precise frontier statement retains both converses. For 1≤r≤d1\le r\le d, requests q=1,…,rq=1,\ldots,r, targets w+aHeqw+aHe_q, a>0a>0, symmetric HH with h1I⪯H⪯h2Ih_1I\preceq H\preceq h_2I and 0<h1<h20<h_1<h_2, and pr=rd−r(r−1)2,δ=h2−h12,p_r=rd-\frac{r(r-1)}2,\qquad \delta=\frac{h_2-h_1}2, fixed-length state and message budgets satisfying uniform error ε<δa/(22r)\varepsilon<\delta a/(2\sqrt{2r}) obey

\qquad M+B\ge d\log_2\frac{\delta a}{\varepsilon}.$$ The first follows by covering the $p_r$-dimensional operator slice with the state and all potential messages; the second covers the $d$-dimensional target ball for one fixed request. With $d,r,a,h_1,h_2$ fixed as $\varepsilon\to0$, the submitted coarse-state/refined-column construction achieves leading coefficients $M/\log_2(1/\varepsilon)\le\alpha p_r$ and $B/\log_2(1/\varepsilon)\le(1-\alpha)d$. At state coefficient $\alpha p_r$, the message coefficient is consequently bounded below by $$\max\{(1-\alpha)p_r/r,\ d-\alpha p_r,\ 0\}$$ and above by $(1-\alpha)d$. Endpoints match, and $r=1$ matches throughout at leading order. For $\alpha<1$, the ratio to the joint-bound-only coefficient is $rd/p_r=2d/(2d-r+1)<2$; $2d/(d+1)$ is the special case $r=d$. The submitted construction therefore leaves the interior open, and the revision closes most of it. Write $R_M=M/\log_2(1/\varepsilon)$ and $R_B=B/\log_2(1/\varepsilon)$ for the state and message rates. The paper keeps $m$ for the retained sample size. The first $r$ columns of $H$ contain a shared symmetric block $H_{[r],[r]}$ and $r(d-r)$ unshared entries, $d-r$ in each column. Each of the $r(r-1)/2$ off-diagonal entries of $H_{[r],[r]}$ enters two requests, and each of its $r$ diagonal entries enters one. Thus $p_r=r(r+1)/2+r(d-r)$. Under the same fixed-length coding and uniform-error criterion, successive refinement of the bounded entries gives a construction that caches shared entries first: - For $r(r-1)/2\le R_M\le r(r+1)/2$, cache every off-diagonal entry of $H_{[r],[r]}$ at full precision and each diagonal entry at fractional precision $\kappa_1=(R_M-r(r-1)/2)/r$. On request $q$, refine $H_{qq}$ and send the unshared part of column $q$. Then $R_B=d-r+1-\kappa_1$. - For $r(r+1)/2\le R_M\le p_r$, cache all of $H_{[r],[r]}$, and the unshared entries at fractional precision $\kappa_2=(R_M-r(r+1)/2)/(r(d-r))$ when $r<d$. Then $R_B=(1-\kappa_2)(d-r)$. In both regions $R_M+rR_B=p_r$, which meets the first converse at leading order. The frontier is therefore $$R_B^\star=\frac{p_r-R_M}{r},\qquad \frac{r(r-1)}2\le R_M\le p_r .$$ Below the threshold, caching the off-diagonal entries at fractional precision gives $R_B=d-2R_M/r$. This matches the converse for $r\le2$, and for $r\ge3$ it leaves a message-rate gap of at most $r/2-1$. At $d=100$, $r=10$ and $R_M=45$, for example, $p_r=955$ and the frontier value is $R_B=91$, against about 95.3 for the submitted construction. The revision states this result as Proposition C.6, with its proof in Appendix C.2, and summarizes it after Theorem 3.3. The new title is "Curvature as Decoder State for Strongly Convex Unlearning: Exactness Above the Shared-State Threshold". Please let us know whether these results answer your questions. We will reply to any further point during the discussion period.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We thank the reviewers. Reviewer Trgo checked the main arguments, found no gap, and recomputed the reported numbers as internally consistent. Reviewer tgEc found the problem formulation clear and the information available at each stage well specified. Two concerns were shared: the word "exact" in the title, and the codec behind the reported reductions. The revision addresses both, and it rewrites the notation and terminology that Reviewer 62ep found unclear.

  • The frontier is now exact above a shared-state threshold. A new construction caches the entries shared between requests before request-specific ones. It meets the converse at leading order for every state rate of at least r(r−1)/2r(r-1)/2, so the frontier is exact there, and everywhere when r≤2r\le2. For r≥3r\ge3, a message-rate gap of at most r/2−1r/2-1 remains below the threshold. The revision retitles the paper "Curvature as Decoder State for Strongly Convex Unlearning: Exactness Above the Shared-State Threshold".
  • Entropy coding removes most of the fixed-width gains. On the four Table 12–13 settings, the cache's saving falls from 6.8–19.6% under fixed-width coding to 0.5–1.6% under signed Exp-Golomb coding, at the same parameter-RMSE target. With the retuned D.4 transform it is 0.3–0.6%. On California and the MLP, the cache message is slightly larger than parameters-only under Exp-Golomb coding. With a public orthogonal transform, the nonconvex MLP and ResNet-18 keep a 5–6% saving. The paper's claim is a price for curvature as decoder state, not a compression gain. The revision states the abstract's reductions as fixed-width savings, with the codec and the state cost beside each one.
  • Presentation. Every symbol is defined where it is first used. "Value-exposing" and "data-native" are explained before the terms appear, the strong-convexity constant and the sample size get separate symbols, and the unfilled template sentence on page 10 is removed.
  • Refresh accounting (tgEc). In a serializer-level run with a cache refresh every 25 requests, the cache pays for itself after 10 requests. Over 100 requests it costs 10,208 B, against 12,980 B without it.

Revisions at a glance

ChangeRequested byWhere in the revision
Shared-entry construction, exact above a state rate of r(r−1)/2r(r-1)/2, with a matching titleTrgo, tgEcTitle; abstract; Theorem 3.3; Appendix C.2
Entropy-coded and transform-coded update messagesTrgoAppendix D.9 (new)
Notation defined at first use, and the two terms explained62epSections 2 and 3
Abstract's delivery reductions stated as fixed-width savingsTrgo, 62epAbstract
Unfilled template sentence removed62epReproducibility statement
Assumptions behind the amortization horizon, and a run with periodic cache refreshtgEcAppendix D.3 (assumptions); Appendix D.9 (refresh run)

Scope of the exact results

The exact results concern compatible-target geometry and the leading-order storage–message frontier above the shared-state threshold. Below that threshold, with three or more hedged directions, the tradeoff is bounded rather than exactly characterized.

A curvature cache trades retained state against additional communication. Its benefit depends on the request interface, serializer, distortion target, and number of requests. The reported delivery reductions are interface- and codec-specific, not a claim of optimal compression or free state.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.