pith. sign in

arxiv: 2606.23198 · v1 · pith:V4G2LMRNnew · submitted 2026-06-22 · 🧮 math.OC · cs.LG· stat.ML

Incremental Learning in Mirror Flows

Rapha\"el Berthier , Loucas Pillaud-Vivien This is my paper

Pith reviewed 2026-06-26 07:39 UTC · model grok-4.3

classification 🧮 math.OC cs.LGstat.ML
keywords mirror flowsincremental learningconvex optimizationquadratic lossmirror potentialsupport functionsubdifferentialrescaled trajectories
0
0 comments X

The pith

When initialized near the domain boundary, rescaled mirror flow trajectories converge to a limiting flow that performs incremental learning.

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

The paper examines mirror flows driven by a convex quadratic loss and a general convex mirror potential. It proves that initialization near the boundary of the mirror potential's domain causes the rescaled trajectories to converge to a limiting mirror flow using the indicator function as its potential. In this limiting regime, the primal variable minimizes the loss over a time-dependent hypothesis set given by the subdifferential of the support function of the domain at the dual variable. This supplies a general mechanism explaining incremental learning within mirror flows. Readers would care if this unifies how certain optimization trajectories naturally build solutions incrementally.

Core claim

We study mirror flows generated by a convex quadratic loss and a general convex lower semicontinuous mirror potential. We show that, when initialized near the boundary of the domain of the mirror potential, their rescaled trajectories converge to a limiting mirror flow whose potential is the indicator function of the domain. In this limit, the primal variable minimizes the loss over a time-dependent hypothesis set: the subdifferential of the support function of the domain, evaluated at the dual variable. This characterization provides a general mechanism for incremental learning in mirror flows.

What carries the argument

The limiting mirror flow with the indicator function potential, which enforces minimization of the loss over the time-dependent hypothesis set defined by the subdifferential of the domain's support function at the dual variable.

If this is right

  • The mechanism applies to any convex lower semicontinuous mirror potential.
  • The primal variable learns incrementally by minimizing over successively adjusted hypothesis sets driven by the dual variable.
  • Rescaling is essential to observe the convergence to the limiting incremental behavior.
  • Initialization close to the boundary is necessary for the trajectories to enter this regime.

Where Pith is reading between the lines

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

  • If many optimization algorithms can be viewed as mirror flows, this may indicate a broader principle for when they exhibit incremental learning.
  • Future work could test whether the same limit holds for non-quadratic losses.
  • Simulations on polytopes or balls could verify the time-dependent hypothesis sets in practice.

Load-bearing premise

Initialization must be sufficiently close to the boundary of the mirror potential's domain.

What would settle it

A counterexample computation where a mirror flow started near the boundary has rescaled trajectories that do not converge to the indicator-potential flow or do not minimize over the described time-dependent set.

Figures

Figures reproduced from arXiv: 2606.23198 by Loucas Pillaud-Vivien, Rapha\"el Berthier.

Figure 1
Figure 1. Figure 1 [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Positive semidefinite cone. Simulation of the mirror flow for two small values of ε and comparison with the limiting dynamics. The top two plots show the eigenvalue trajectories of X for the mirror flow (solid line) and for the limiting dynamics (dotted line). The top plot corresponds to ε = 10−20 and the middle plot to ε = 10−100. The bottom plot displays the loss. The mirror flow and the limiting dynamic… view at source ↗
Figure 3
Figure 3. Figure 3: Simplex. Simulation of the mirror flow for two small values of ε and comparison with the limiting dynamics. The top two plots show the coordinate trajectories of x for the mirror flow (solid line) and for the limiting dynamics (dotted line). The top plot corresponds to ε = 10−20 and the middle plot to ε = 10−100. The bottom plot displays the excess loss, i.e. the loss minus its minimum over ∆d. The mirror … view at source ↗
Figure 4
Figure 4. Figure 4: (Left) Eigenvalue trajectories of X ε (s) for the matrix mirror flow (orange) and the matrix factorization gradient flow (blue), at ε = 10−100, for four random seeds. (Right) Excess loss ℓ(X ε (s)) − min ℓ for the same flows and seeds. Both flows exhibit incremental rank-increase, but can visit different intermediate saddles in different orders, and pace their activations differently on the rescaled time a… view at source ↗
Figure 5
Figure 5. Figure 5: displays this trajectory. 0.2 0.4 0.6 0.8 1.0 Time s −3 −2 −1 0 1 w ε and w coordinates Limiting dynamics w(s) Mirror flow w ε(s), ε = 10−20 0.2 0.4 0.6 0.8 1.0 Time s −3 −2 −1 0 1 w ε and w coordinates Limiting dynamics w(s) Mirror flow w ε(s), ε = 10−100 [PITH_FULL_IMAGE:figures/full_fig_p041_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: displays its eigenvalues. 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Time s −2 0 W ε and W eigenvalues Limiting dynamics W(s) Mirror flow W ε (s), ε = 10−20 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 Time s −2 0 W ε and W eigenvalues Limiting dynamics W(s) Mirror flow W ε (s), ε = 10−100 [PITH_FULL_IMAGE:figures/full_fig_p042_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Trajectories of the coordinates of the dual variable [PITH_FULL_IMAGE:figures/full_fig_p043_7.png] view at source ↗
read the original abstract

We study mirror flows generated by a convex quadratic loss and a general convex lower semicontinuous mirror potential. We show that, when initialized near the boundary of the domain of the mirror potential, their rescaled trajectories converge to a limiting mirror flow whose potential is the indicator function of the domain. In this limit, the primal variable minimizes the loss over a time-dependent hypothesis set: the subdifferential of the support function of the domain, evaluated at the dual variable. This characterization provides a general mechanism for incremental learning in mirror flows.

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

2 major / 2 minor

Summary. The manuscript studies mirror flows generated by a convex quadratic loss and a general convex lower semicontinuous mirror potential ψ. It shows that, when initialized near the boundary of the domain of ψ, rescaled trajectories converge to a limiting mirror flow whose potential is the indicator function of the domain. In this limit, the primal variable minimizes the loss over a time-dependent hypothesis set given by the subdifferential of the support function of the domain evaluated at the dual variable. This is presented as providing a general mechanism for incremental learning in mirror flows.

Significance. If the stated convergence holds, the result supplies a precise asymptotic characterization linking mirror flows to incremental learning via time-dependent hypothesis sets. The generality of the mirror potential strengthens the scope relative to more restrictive cases. The near-boundary initialization condition, however, is load-bearing for the claimed generality and requires further justification to establish robustness.

major comments (2)
  1. [Abstract] Abstract: the convergence to the indicator-potential limiting flow and the resulting time-dependent hypothesis set ∂σ_D(·) are established only under the near-boundary initialization hypothesis. For the characterization to constitute a 'general mechanism for incremental learning,' either the regime must be shown to arise naturally from the dynamics or the result must be extended to interior initializations; otherwise the scope of the incremental-learning claim is restricted.
  2. [Main theorem (convergence result)] Main convergence statement: the rescaling argument and passage to the indicator limit require explicit control on the error terms and verification that the quadratic loss remains compatible with the indicator potential; without these controls the limiting characterization of the hypothesis set may not hold uniformly.
minor comments (2)
  1. Clarify the precise definition of the support function σ_D and its relation to the domain D throughout the text.
  2. Ensure all technical conditions on ψ (convexity, lower semicontinuity, domain) are stated explicitly before the main convergence result.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful review and constructive suggestions. We address each major comment below. The near-boundary initialization is central to the limiting analysis, and we will revise the abstract and add explicit details to the convergence proof to clarify scope and strengthen the error controls.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the convergence to the indicator-potential limiting flow and the resulting time-dependent hypothesis set ∂σ_D(·) are established only under the near-boundary initialization hypothesis. For the characterization to constitute a 'general mechanism for incremental learning,' either the regime must be shown to arise naturally from the dynamics or the result must be extended to interior initializations; otherwise the scope of the incremental-learning claim is restricted.

    Authors: We agree that the stated convergence and the time-dependent hypothesis set characterization hold specifically under the near-boundary initialization. The manuscript already states this condition in the abstract and theorem statements; the generality claimed refers to the arbitrary convex lsc mirror potential ψ rather than to arbitrary initializations. The near-boundary regime arises naturally in applications where trajectories approach the boundary of dom(ψ), which is relevant for modeling incremental learning under domain constraints. We will revise the abstract to foreground this initialization hypothesis more explicitly. Extending the result to interior initializations would require a separate analysis and is outside the current scope, as the limiting indicator flow does not generally emerge from interior starts. revision: partial

  2. Referee: [Main theorem (convergence result)] Main convergence statement: the rescaling argument and passage to the indicator limit require explicit control on the error terms and verification that the quadratic loss remains compatible with the indicator potential; without these controls the limiting characterization of the hypothesis set may not hold uniformly.

    Authors: The proof of the main convergence result does contain rescaling estimates that control the error terms between the original flow and the limiting indicator flow, together with a verification that the quadratic loss passes to the limit compatibly (via weak convergence arguments and subdifferential inclusions). However, these controls can be made more prominent. We will revise the manuscript to include an explicit lemma or remark isolating the error bounds and confirming uniform compatibility of the loss with the indicator potential, ensuring the hypothesis-set characterization holds in the stated topology. revision: yes

Circularity Check

0 steps flagged

No circularity; conditional convergence result derived from dynamics under explicit assumption

full rationale

The paper states and proves a convergence theorem: rescaled trajectories of the mirror flow converge to the indicator-potential flow precisely when initialized near the boundary of dom(ψ). This is presented as a derived property of the ODE dynamics rather than presupposed. No equations reduce a prediction to a fitted input by construction, no self-citation chains justify uniqueness or ansatzes, and the initialization condition is stated openly as the operative hypothesis rather than smuggled in. The result is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Ledger extracted from abstract only; full paper may introduce additional technical assumptions.

axioms (2)
  • domain assumption Loss is convex quadratic and mirror potential is convex lower semicontinuous.
    Explicitly stated as the setting for the mirror flows under study.
  • domain assumption Initialization is near the boundary of the domain of the mirror potential.
    Required condition for the rescaled trajectories to converge to the claimed limit.

pith-pipeline@v0.9.1-grok · 5610 in / 1209 out tokens · 25359 ms · 2026-06-26T07:39:41.660739+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Effective dynamics of the Sinkhorn algorithm in the regime of low entropy regularization

    math.OC 2026-07 unverdicted novelty 7.0

    Derives the cold Sinkhorn limiting dynamics as tau approaches zero, proving finite-time convergence to unregularized OT and improved O(tau^{-1}) iteration complexity for dual suboptimality.

Reference graph

Works this paper leans on

33 extracted references · 5 canonical work pages · cited by 1 Pith paper

  1. [1]

    Sgd learning on neural networks: leap complexity and saddle-to-saddle dynamics

    Emmanuel Abbe, Enric Boix-Adsera, and Theodor Misiakiewicz. Sgd learning on neural networks: leap complexity and saddle-to-saddle dynamics. In Conference on Learning Theory , pages 2552--2623. PMLR, 2023

    2023

  2. [2]

    Implicit regularization in deep matrix factorization

    Sanjeev Arora, Nadav Cohen, Wei Hu, and Yuping Luo. Implicit regularization in deep matrix factorization. In Advances in Neural Information Processing Systems , volume 32, 2019

    2019

  3. [3]

    A closer look at memorization in deep networks

    Devansh Arpit, Stanis aw Jastrz e bski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon Lacoste-Julien. A closer look at memorization in deep networks. In International Conference on Machine Learning , pages 233--242. PMLR, 2017

    2017

  4. [4]

    Convergence de fonctionnelles convexes

    Hedy Attouch. Convergence de fonctionnelles convexes. In Journ \'e es d’Analyse Non Lin \'e aire: Proceedings, Besan c on, France, June 1977 , pages 1--40. Springer, 1977

    1977

  5. [5]

    Variational convergence for functions and operators

    Hedy Attouch. Variational convergence for functions and operators . Applicable Mathematics Series. Pitman, 1984

    1984

  6. [6]

    On learning G aussian multi-index models with gradient flow

    Alberto Bietti, Joan Bruna, and Loucas Pillaud-Vivien. On learning G aussian multi-index models with gradient flow. arXiv preprint arXiv:2310.19793 , 2023

    work page arXiv 2023

  7. [7]

    Convex analysis and monotone operator theory in Hilbert spaces

    Heinz Bauschke and Patrick Combettes. Convex analysis and monotone operator theory in Hilbert spaces . Springer, 2nd edition, 2017

    2017

  8. [8]

    Incremental learning in diagonal linear networks

    Rapha \"e l Berthier. Incremental learning in diagonal linear networks. Journal of Machine Learning Research , 24(171):1--26, 2023

    2023

  9. [9]

    Diagonal linear networks and the lasso regularization path

    Rapha \"e l Berthier. Diagonal linear networks and the lasso regularization path. arXiv preprint arXiv:2509.18766 , 2025

    work page arXiv 2025

  10. [10]

    Positive Definite Matrices

    Rajendra Bhatia. Positive Definite Matrices . Princeton University Press, 2009

    2009

  11. [11]

    An adaptive inverse scale space method for compressed sensing

    Martin Burger, Michael M \"o ller, Martin Benning, and Stanley Osher. An adaptive inverse scale space method for compressed sensing. Mathematics of Computation , 82(281):269--299, 2013

    2013

  12. [12]

    Nonlinear inverse scale space methods for image restoration

    Martin Burger, Stanley Osher, Jinjun Xu, and Guy Gilboa. Nonlinear inverse scale space methods for image restoration. In International Workshop on Variational, Geometric, and Level Set Methods in Computer Vision , pages 25--36. Springer, 2005

    2005

  13. [13]

    Gradient flow dynamics of shallow relu networks for square loss and orthogonal inputs

    Etienne Boursier, Loucas Pillaud-Vivien, and Nicolas Flammarion. Gradient flow dynamics of shallow relu networks for square loss and orthogonal inputs. Advances in Neural Information Processing Systems , 35:20105--20118, 2022

    2022

  14. [14]

    Opérateurs maximaux monotones et semi-groupes de contractions dans les espaces de Hilbert

    Haim Br \'e zis. Opérateurs maximaux monotones et semi-groupes de contractions dans les espaces de Hilbert . Elsevier, 1973

    1973

  15. [15]

    Error estimation for bregman iterations and inverse scale space methods in image restoration

    Martin Burger, Elena Resmerita, and Lin He. Error estimation for bregman iterations and inverse scale space methods in image restoration. Computing , 81(2):109--135, 2007

    2007

  16. [16]

    Local minima and plateaus in hierarchical structures of multilayer perceptrons

    Kenji Fukumizu and Shun-ichi Amari. Local minima and plateaus in hierarchical structures of multilayer perceptrons. Neural Networks , 13(3):317--327, 2000

    2000

  17. [17]

    Implicit regularization of discrete gradient dynamics in linear neural networks

    Gauthier Gidel, Francis Bach, and Simon Lacoste-Julien. Implicit regularization of discrete gradient dynamics in linear neural networks. In Advances in Neural Information Processing Systems , volume 32, 2019

    2019

  18. [18]

    The implicit bias of depth: How incremental rank learning helps generalize

    Daniel Gissin, Shai Shalev-Shwartz, and Amit Daniely. The implicit bias of depth: How incremental rank learning helps generalize. In International Conference on Learning Representations , 2020

    2020

  19. [19]

    Implicit regularization in matrix factorization

    Suriya Gunasekar, Blake Woodworth, Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Implicit regularization in matrix factorization. In Advances in Neural Information Processing Systems , volume 30, 2017

    2017

  20. [20]

    Saddle-to-saddle dynamics in deep linear networks: Small initialization training, symmetry, and sparsity

    Arthur Jacot, Fran c ois Ged, Berfin S im s ek, Cl \'e ment Hongler, and Franck Gabriel. Saddle-to-saddle dynamics in deep linear networks: Small initialization training, symmetry, and sparsity. arXiv preprint arXiv:2106.15933 , 2021

    work page arXiv 2021

  21. [21]

    Sgd on neural networks learns functions of increasing complexity

    Dimitris Kalimeris, Gal Kaplun, Preetum Nakkiran, Benjamin Edelman, Tristan Yang, Boaz Barak, and Haofeng Zhang. Sgd on neural networks learns functions of increasing complexity. Advances in Neural Information Processing Systems , 32, 2019

    2019

  22. [22]

    Towards resolving the implicit bias of gradient descent for matrix factorization: Greedy low-rank learning

    Zhiyuan Li, Yuping Luo, and Kaifeng Lyu. Towards resolving the implicit bias of gradient descent for matrix factorization: Greedy low-rank learning. In International Conference on Learning Representations , 2021

    2021

  23. [23]

    Implicit regularization for group sparsity

    Jiangyuan Li, Thanh Nguyen, Chinmay Hegde, and Raymond Wong. Implicit regularization for group sparsity. arXiv preprint arXiv:2301.12540 , 2023

    work page arXiv 2023

  24. [24]

    Implicit bias of gradient descent on reparametrized models: On equivalence to mirror descent

    Zhiyuan Li, Tianhao Wang, Jason Lee, and Sanjeev Arora. Implicit bias of gradient descent on reparametrized models: On equivalence to mirror descent. Advances in Neural Information Processing Systems , 2022

    2022

  25. [25]

    Problem Complexity and Method Efficiency in Optimization

    Arkadij Semenovi c Nemirovskij and David Borisovich Yudin. Problem Complexity and Method Efficiency in Optimization . Wiley-Interscience, 1983

    1983

  26. [26]

    Sparse recovery via differential inclusions

    Stanley Osher, Feng Ruan, Jiechao Xiong, Yuan Yao, and Wotao Yin. Sparse recovery via differential inclusions. Applied and Computational Harmonic Analysis , 41(2):436--469, 2016

    2016

  27. [27]

    Saddle-to-saddle dynamics in diagonal linear networks

    Scott Pesme and Nicolas Flammarion. Saddle-to-saddle dynamics in diagonal linear networks. Advances in Neural Information Processing Systems , 36:7475--7505, 2023

    2023

  28. [28]

    Implicit regularization in deep learning may not be explainable by norms

    Noam Razin and Nadav Cohen. Implicit regularization in deep learning may not be explainable by norms. In Advances in Neural Information Processing Systems , volume 33, pages 21174--21187, 2020

    2020

  29. [29]

    Convex analysis

    Ralph Tyrrell Rockafellar. Convex analysis . Princeton Mathematical Series, 1970

    1970

  30. [30]

    Exact solutions to the nonlinear dynamics of learning in deep linear neural networks

    Andrew Saxe, James McClelland, and Surya Ganguli. Exact solutions to the nonlinear dynamics of learning in deep linear neural networks. In International Conference on Learning Representations , 2014

    2014

  31. [31]

    Learning in-context n -grams with transformers: Sub- n -grams are near-stationary points

    Aditya Varre, Gizem Y\" u ce, and Nicolas Flammarion. Learning in-context n -grams with transformers: Sub- n -grams are near-stationary points. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors, Proceedings of the 42nd International Conference on Machine Learning , volume...

    2025

  32. [32]

    Kernel and rich regimes in overparametrized models

    Blake Woodworth, Suriya Gunasekar, Jason Lee, Edward Moroshko, Pedro Savarese, Itay Golan, Daniel Soudry, and Nathan Srebro. Kernel and rich regimes in overparametrized models. In Conference on Learning Theory , pages 3635--3673. PMLR, 2020

    2020

  33. [33]

    Saddle-to-saddle dynamics explains a simplicity bias across neural network architectures

    Yedi Zhang, Andrew Saxe, and Peter Latham. Saddle-to-saddle dynamics explains a simplicity bias across neural network architectures. arXiv preprint arXiv:2512.20607 , 2025

    work page arXiv 2025