pith. sign in

arxiv: 2605.17590 · v1 · pith:H65MOXJHnew · submitted 2026-05-17 · 💻 cs.LG · math.OC

Form and Function: Machine Unlearning as a Problem of Misaligned States

Kennon Stewart This is my paper

Pith reviewed 2026-05-20 14:32 UTC · model grok-4.3

classification 💻 cs.LG math.OC
keywords machine unlearningonline L-BFGScounterfactual stateoptimizer state alignmentmemory operatordeletion intervention
0
0 comments X

The pith

Machine unlearning for online L-BFGS requires alignment with the counterfactual optimizer state that excludes the deleted data.

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

The paper frames unlearning in online L-BFGS as the problem of reaching the optimizer state that would have existed if the to-be-deleted samples had never entered the stream. It defines separate metrics for parameter mismatch, memory-operator mismatch (via inverse-Hessian actions), combined state error, and update-direction error. A recursive deviation bound is derived under convexity, and interventions are tested against an oracle that possesses the full counterfactual state from the start. The evaluation shows that parameter-only fixes leave residual misalignment that memory corrections can reduce.

Core claim

We formulate machine unlearning for online L-BFGS as a counterfactual state-alignment problem. Given an actual event stream and a deletion-edited counterfactual stream, the target of unlearning is the optimizer state that would have arisen had the deleted samples never been processed. State-aware metrics separately measure parameter error, memory-operator error, combined state error, and update-direction error. Under convexity assumptions, a recursive bound on counterfactual state deviation is derived. Benchmarks of deletion interventions demonstrate that unlearning is not merely a parameter-correction problem: it requires alignment with a realizable counterfactual optimizer state.

What carries the argument

The counterfactual optimizer state, which is the state that would result from processing only the deletion-edited stream and serves as the explicit target for any unlearning intervention.

If this is right

  • Memory-operator error, measured by comparing induced inverse-Hessian actions, captures misalignment invisible to parameter error alone.
  • Under the convexity assumption, counterfactual state deviation admits a recursive bound that limits how much correction is needed.
  • Combined state corrections that address both parameters and memory move closer to the counterfactual oracle than parameter-only fixes.
  • Update-direction error provides an additional diagnostic that parameter or memory corrections can each affect differently.

Where Pith is reading between the lines

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

  • The same state-alignment requirement may appear in other online second-order methods that maintain curvature approximations.
  • A practical system could maintain a lightweight parallel counterfactual optimizer and swap its state upon deletion requests.
  • Relaxing convexity would require either a different bound or empirical verification that the state-alignment gap remains material.

Load-bearing premise

The recursive bound on how far the actual and deletion-edited streams can diverge is derived under convexity assumptions.

What would settle it

A benchmark result in which a parameter-only correction reaches the same or lower combined state error as a memory-inclusive correction, when both are measured against the counterfactual oracle, would falsify the claim that full state alignment is necessary.

read the original abstract

We formulate machine unlearning for online L-BFGS as a counterfactual state-alignment problem. Given an actual event stream and a deletion-edited counterfactual stream, the target of unlearning is the optimizer state that would have arisen had the deleted samples never been processed. We introduce state-aware metrics that separately measure parameter error, memory-operator error, combined state error, and update-direction error. The memory metric compares the inverse-Hessian actions induced by the o-L-BFGS memory, rather than treating curvature pairs as of finite influence. Under convexity assumptions, we derive a recursive bound on counterfactual state deviation. We then evaluate a state-aware benchmark of deletion interventions, including memory-only and parameter-only corrections, against an counterfactual oracle model. These results show that unlearning for online L-BFGS is not merely a parameter-correction problem: it requires alignment with a realizable counterfactual optimizer state.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper formulates machine unlearning for online L-BFGS as a counterfactual state-alignment problem. Given an actual event stream and a deletion-edited counterfactual stream, the target is the optimizer state that would have arisen without the deleted samples. It introduces state-aware metrics measuring parameter error, memory-operator error (via inverse-Hessian actions), combined state error, and update-direction error. Under convexity assumptions a recursive bound on counterfactual state deviation is derived, and deletion interventions (memory-only and parameter-only) are benchmarked against a counterfactual oracle.

Significance. If the central claim holds, the work shows that unlearning for online L-BFGS requires alignment to a realizable counterfactual optimizer state rather than parameter correction alone. The recursive bound under stated convexity assumptions and the oracle benchmark are concrete strengths that make the distinction between state-aware and parameter-only approaches falsifiable and measurable.

major comments (1)
  1. [Abstract] Abstract and derivation of recursive bound: the bound on counterfactual state deviation is explicitly conditioned on convexity assumptions to control deviation between the actual and deletion-edited streams. Online L-BFGS is routinely applied to non-convex problems; the manuscript should either restrict the headline claim to convex regimes or provide evidence (e.g., additional experiments or counter-examples) that the observed gap between parameter-only corrections and full state alignment is not an artifact of convexity.
minor comments (1)
  1. Clarify how the memory metric that compares inverse-Hessian actions induced by the o-L-BFGS memory is computed in practice and whether it reduces to a finite-horizon approximation.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive summary and constructive major comment on the scope of our results. We respond point by point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract and derivation of recursive bound: the bound on counterfactual state deviation is explicitly conditioned on convexity assumptions to control deviation between the actual and deletion-edited streams. Online L-BFGS is routinely applied to non-convex problems; the manuscript should either restrict the headline claim to convex regimes or provide evidence (e.g., additional experiments or counter-examples) that the observed gap between parameter-only corrections and full state alignment is not an artifact of convexity.

    Authors: We acknowledge that the recursive bound is derived under convexity assumptions to control stream deviation, as already stated in the manuscript. The core contribution is the general formulation of unlearning as counterfactual state alignment together with the state-aware metrics; these are not restricted to convex settings. The empirical gap between parameter-only and full-state interventions is demonstrated in the evaluated (convex) regimes. To address the comment directly, we will revise the abstract and introduction to more explicitly qualify the theoretical bound and headline claims as holding under the stated convexity assumptions, while noting that extensions to non-convex regimes remain open. This clarification incorporates the referee's point without requiring additional experiments. revision: yes

Circularity Check

0 steps flagged

No significant circularity: target state and bound defined externally

full rationale

The paper defines the target counterfactual optimizer state directly from the deletion-edited stream as an external reference, then derives a recursive deviation bound under explicit convexity assumptions to bound errors between streams. State-aware metrics and the oracle comparison are constructed from these definitions and evaluated empirically against interventions. No equation or claim reduces the central result to a fitted input, self-citation, or definitional equivalence; the derivation remains self-contained against the stated assumptions and external oracle.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim depends on convexity to bound state deviation and on the existence of a well-defined counterfactual stream that produces a realizable optimizer state.

axioms (1)
  • domain assumption Convexity assumptions
    Invoked to derive the recursive bound on counterfactual state deviation.
invented entities (1)
  • counterfactual optimizer state no independent evidence
    purpose: The target state the unlearned model should reach, defined as the state arising from a deletion-edited event stream.
    Introduced as the alignment objective; no independent falsifiable prediction outside the formulation is given.

pith-pipeline@v0.9.0 · 5676 in / 1317 out tokens · 43492 ms · 2026-05-20T14:32:26.300083+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages · 1 internal anchor

  1. [1]

    Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable, May 2024

    Martin Bertran, Shuai Tang, Michael Kearns, Jamie Morgenstern, Aaron Roth, and Zhi- wei Steven Wu. Reconstruction Attacks on Machine Unlearning: Simple Models are Vulnerable, May 2024. arXiv:2405.20272

    work page arXiv 2024

  2. [2]

    A., Jia, H., Travers, A., Zhang, B., Lie, D., and Papernot, N

    Lucas Bourtoule, Varun Chandrasekaran, Christopher A. Choquette-Choo, Hengrui Jia, Adelin Travers, Baiwu Zhang, David Lie, and Nicolas Papernot. Machine Unlearning, December 2020. arXiv:1912.03817

    work page arXiv 2020

  3. [3]

    Boyd and Lieven Vandenberghe.Convex optimization

    Stephen P. Boyd and Lieven Vandenberghe.Convex optimization. Cambridge University Press, Cambridge New York Melbourne New Delhi Singapore, version 29 edition, 2023

    work page 2023

  4. [4]

    Byrd, Peihuang Lu, Jorge Nocedal, and Ciyou Zhu

    Richard H. Byrd, Peihuang Lu, Jorge Nocedal, and Ciyou Zhu. A Limited Memory Algorithm for Bound Constrained Optimization.SIAM Journal on Scientific Computing, 16(5):1190–1208, September 1995

    work page 1995

  5. [5]

    Cambridge University Press, Cambridge New York, 2006

    Nicolò Cesa-Bianchi and Gábor Lugosi.Prediction, learning, and games. Cambridge University Press, Cambridge New York, 2006

    work page 2006

  6. [6]

    When Machine Unlearning Jeopardizes Privacy, September 2021

    Min Chen, Zhikun Zhang, Tianhao Wang, Michael Backes, Mathias Humbert, and Yang Zhang. When Machine Unlearning Jeopardizes Privacy, September 2021. arXiv:2005.02205

    work page arXiv 2021

  7. [7]

    Feder Cooper, Christopher A

    A. Feder Cooper, Christopher A. Choquette-Choo, Miranda Bogen, Kevin Klyman, Matthew Jagielski, Katja Filippova, Ken Liu, Alexandra Chouldechova, Jamie Hayes, Yangsibo Huang, Eleni Triantafillou, Peter Kairouz, Nicole Elyse Mitchell, Niloofar Mireshghallah, Abigail Z. Jacobs, James Grimmelmann, Vitaly Shmatikov, Christopher De Sa, Ilia Shumailov, Andreas ...

    work page arXiv 2025

  8. [8]

    Online Learning and Unlearning, May

    Yaxi Hu, Bernhard Schölkopf, and Amartya Sanyal. Online Learning and Unlearning, May

    work page

  9. [9]

    An Information Theoretic Evaluation Metric For Strong Unlearning, 2024

    Dongjae Jeon, Wonje Jeung, Taeheon Kim, Albert No, and Jonghyun Choi. An Information Theoretic Evaluation Metric For Strong Unlearning, 2024

    work page 2024

  10. [10]

    A discrete Grönwall inequality with application to numerical schemes for subdiffusion problems, November 2018

    Hong-lin Liao, William McLean, and Jiwei Zhang. A discrete Grönwall inequality with application to numerical schemes for subdiffusion problems, November 2018. arXiv:1803.09879

    work page arXiv 2018

  11. [11]

    Threats, Attacks, and Defenses in Machine Unlearning: A Survey.IEEE Open Journal of the Computer Society, 6:413–425, 2025

    Ziyao Liu, Huanyi Ye, Chen Chen, Yongsen Zheng, and Kwok-Yan Lam. Threats, Attacks, and Defenses in Machine Unlearning: A Survey.IEEE Open Journal of the Computer Society, 6:413–425, 2025

    work page 2025

  12. [12]

    Hessian-Free Online Certified Unlearn- ing, February 2025

    Xinbao Qiao, Meng Zhang, Ming Tang, and Ermin Wei. Hessian-Free Online Certified Unlearn- ing, February 2025. arXiv:2404.01712

    work page arXiv 2025

  13. [13]

    Remember What You Want to Forget: Algorithms for Machine Unlearning, July 2021

    Ayush Sekhari, Jayadev Acharya, Gautam Kamath, and Ananda Theertha Suresh. Remember What You Want to Forget: Algorithms for Machine Unlearning, July 2021. arXiv:2103.03279. 17

    work page arXiv 2021

  14. [14]

    Machine Unlearning for Streaming Forgetting, July 2025

    Shaofei Shen, Chenhao Zhang, Yawen Zhao, Alina Bialkowski, Weitong Chen, and Miao Xu. Machine Unlearning for Streaming Forgetting, July 2025. arXiv:2507.15280

    work page arXiv 2025

  15. [15]

    Shape of Memory: a Geometric Analysis of Machine Unlearning in Second-Order Optimizers

    Kennon Stewart. Shape of Memory: a Geometric Analysis of Machine Unlearning in Second- Order Optimizers, April 2026. arXiv:2604.23046

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    Suriyakumar and Ashia C

    Vinith M. Suriyakumar and Ashia C. Wilson. Algorithms that Approximate Data Removal: New Results and Limitations, September 2022. arXiv:2209.12269

    work page arXiv 2022

  17. [17]

    Heng Xu, Tianqing Zhu, Lefeng Zhang, Wanlei Zhou, and Philip S. Yu. Machine Unlearning: A Survey.ACM Computing Surveys, 56(1):1–36, January 2024. 18 Quantity Default value Dimension d= 25 Stream length, finite-memory decay T= 700 Deletion time, finite-memory decay tdel = 300 Post-deletion horizon, finite-memory decay H= 250 Stream length, state-aware benc...

    work page 2024

  18. [18]

    Generate an online event stream and train the actual o-LBFGS optimizer on the prefixe1:tdel

    work page

  19. [19]

    Select a deletion setUfrom the prefix using the specified deletion mode

    work page

  20. [20]

    Construct the oracle counterfactual stateθ−U tdel by replaying the prefix while skipping all events inU

    work page

  21. [21]

    Apply each unlearning intervention to the actual stateθtdel, producing an intervened state ˜θ(r) tdel for methodr

    work page

  22. [22]

    Propagate every intervened state and the oracle state on the same future event stream

    work page

  23. [23]

    C Experimental Data 22 Intervention Description Oracle Replay This is the gold standard

    Record initial, final, and cumulative trajectory discrepancies relative to the oracle. C Experimental Data 22 Intervention Description Oracle Replay This is the gold standard. The model is retrained from scratch without the offending data, serving as a baseline. No-Op Deletion The deletion is registered and excluded from future loss evaluations, but the p...

    work page

  24. [24]

    P1 is the post-deletion phase where the deleted data remains in the range of curvature pairs

    work page

  25. [25]

    This is considered to be some period of reasonable indirect influence

    P2 is the post-deletion phase where the deleted data has passed from direct to indirect memory, but still remains within2τof the time of deletion. This is considered to be some period of reasonable indirect influence

    work page

  26. [26]

    25 Figure 7.Phase-specific exponential decay rates for the quadratic stream

    P3 is the post-deletion phase that goes from2τto T and encompasses the decay of deleted information within the indirect memory. 25 Figure 7.Phase-specific exponential decay rates for the quadratic stream. Although the quadratic loss surface is more controlled than the logistic setting, the direct post-deletion phase is not uniformly contractive. For inter...

    work page