Mirror Descent Beyond Euclidean Stability: An Exponential Separation in Initialization Sensitivity

• URL: http://arxiv.org/abs/2606.11431v1
• Date: Tue, 09 Jun 2026 20:33:01 GMT
• Title: Mirror Descent Beyond Euclidean Stability: An Exponential Separation in Initialization Sensitivity
• Abstract summary: Mirror Descent (MD) extends Gradient Descent (GD) beyond Euclidean geometry.<n>How sensitive are MD dynamics to their inputs?<n>For canonical KL-regularized MD on the simplex, we show that even linear objectives can amplify an initial $varepsilon$ perturbation exponentially fast.
• Score: 23.519129043938758
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Mirror Descent (MD) extends Gradient Descent (GD) beyond Euclidean geometry and has recently reappeared as a lens for KL-regularized policy optimization in reinforcement learning and LLM post-training. This raises a basic robustness question, crucial to reproducibility and reliability: how sensitive are MD dynamics to their inputs? We focus on initialization, often itself a pretrained or previously aligned model. Quadratic-regularized MD, including GD and Mahalanobis geometries, is well-known to be stable for convex smooth objectives. We show a sharp contrast: once the regularizer is non-quadratic, MD can be exponentially more sensitive to initialization than GD, even with a well-conditioned regularizer in Euclidean norm. We give a three-dimensional construction with a convex, smooth objective and a strongly convex, smooth, well-conditioned regularizer where an initial $\varepsilon$ perturbation is quickly amplified to $\min\{\text{polylog}^{-1}(1/\varepsilon), \varepsilon e^{Ω(ηT)}\}$ after $T$ iterations of MD with step size $η$. For canonical KL-regularized MD on the simplex, we show that even linear objectives can amplify an initial $\varepsilon$ perturbation exponentially fast in high-dimensional or near-boundary regimes. Finally, we show that adding a Bregman regularization term toward an anchor point can stabilize the dynamics while largely preserving the optimization guarantees, and that the choice of anchor is crucial: anchoring at the initialization only partially mitigates the instability, whereas anchoring at a fixed point yields a more stable mechanism.

Related papers

• Adaptive Bregman Proximal Stochastic Gradient with a Stabilized Barzilai--Borwein Step Size [43.46286813814826]
  Bregman finite proximal ergo gradient (BPSG) methods bring variance-reduced composite size to objectives whose geometry is poorly captured by Euclidean smoothness.<n>Their performance remains sensitive to stepwise: raw curvature estimates can fluctuate, whereas line searches add repeated searches.<n>We introduce AdaBPSG, a line-search method that combines sparse values and a non-negative matrix factorization rate.
  arXiv  Detail & Related papers  (2026-08-12T12:46:05Z)
• Phasor Attention: Mean Root Square Normalization for Phase Manifold Preservation [3.0459410574367314]
  We introduce Mean Root Square Normalization (MRSNorm)<n>MRSNorm mathematically inverts the traditional scaling paradigm.<n>We analytically demonstrate that this constraint yields a built-in, trigonometric gradient clipper governed by the Pythagorean identity.
  arXiv  Detail & Related papers  (2026-07-20T11:10:51Z)
• Expanding SPHERE-JEPA: A Family of Statistical Regularizers for the Hypersphere [11.087393998848254]
  In Self-Supervised Learning (SSL), preventing representation collapse by explicitly enforcing a uniform distribution on the unit hypersphere has proven to be effective.<n>Current frameworks typically rely on sliced statistical regularizers such as SIGReg (used in LeJEPA) and SUSReg (used in SPHEREJEPA)<n>This injects projection variance into the training gradients, destabilizing optimization, and hindering convergence.<n>In this work, we first show that analytically integrating out these random projections yields a deterministic Maximum Mean Discrepancy (MMD), bypassing the variance of sliced methods.
  arXiv  Detail & Related papers  (2026-06-16T07:10:05Z)
• MoSSP: A Momentum-Based Single-Loop Stochastic Penalty Method for Nonconvex Constrained DC-Regularized Optimization [6.024178662558234]
  We show a class of non constrained problems with difference-of-an- oracle (DC regularization) regularization.<n>We propose a Momentum Mo Mo Penalty method for such problems with provable complexity guarantees.
  arXiv  Detail & Related papers  (2026-05-28T09:06:26Z)
• Regularized Online RLHF with Generalized Bilinear Preferences [68.44113000390544]
  We consider the problem of contextual online RLHF with general preferences.<n>We adopt the Generalized Bilinear Preference Model to capture preferences via low-rank, skew-symmetric matrices.<n>We prove that the dual gap of the greedy policy is bounded by the square of the estimation error.
  arXiv  Detail & Related papers  (2026-02-26T15:27:53Z)
• Stability and Generalization of Push-Sum Based Decentralized Optimization over Directed Graphs [55.77845440440496]
  Push-based decentralized communication enables optimization over communication networks, where information exchange may be asymmetric.<n>We develop a unified uniform-stability framework for the Gradient Push (SGP) algorithm.<n>A key technical ingredient is an imbalance-aware generalization bound through two quantities.
  arXiv  Detail & Related papers  (2026-02-24T05:32:03Z)
• Majorization-Minimization Networks for Inverse Problems: An Application to EEG Imaging [4.063392865490957]
  Inverse problems are often ill-posed and require optimization schemes with strong stability and convergence guarantees.<n>We propose a learned Majorization-Minimization (MM) framework for inverse problems within a bilevel optimization setting.<n>We learn a structured curvature majorant that governs each MM step while preserving classical MM descent guarantees.
  arXiv  Detail & Related papers  (2026-01-23T10:33:45Z)
• Precise Bayesian Neural Networks [0.0]
  We develop a lightweight, implementation-ready variational unit that fits modern normalized architectures and improves calibration without sacrificing accuracy.<n>In short, by aligning the variational posterior with the network's intrinsic geometry, BNNs can be simultaneously principled, practical, and precise.
  arXiv  Detail & Related papers  (2025-06-24T15:42:00Z)
• Revisiting Glorot Initialization for Long-Range Linear Recurrences [35.51491162256673]
  Glorot is designed to ensure stable signal propagation.<n>Small positive deviations in the spectral radius are amplified through time and cause the hidden state to explode.<n>We propose a simple, dimension-aware rescaling of Glorot that shifts the spectral radius slightly below one.
  arXiv  Detail & Related papers  (2025-05-26T11:04:59Z)
• Euclidean Distance Matrix Completion via Asymmetric Projected Gradient Descent [25.846262685970164]
  This paper proposes and analyzes a gradient-type algorithm based on Burer-Monteiro factorization.<n>It reconstructs the point set configuration from partial Euclidean distance measurements.
  arXiv  Detail & Related papers  (2025-04-28T07:13:23Z)
• Learning with Norm Constrained, Over-parameterized, Two-layer Neural Networks [54.177130905659155]
  Recent studies show that a reproducing kernel Hilbert space (RKHS) is not a suitable space to model functions by neural networks. In this paper, we study a suitable function space for over- parameterized two-layer neural networks with bounded norms.
  arXiv  Detail & Related papers  (2024-04-29T15:04:07Z)
• Improved Convergence Rate of Stochastic Gradient Langevin Dynamics with Variance Reduction and its Application to Optimization [50.83356836818667]
  gradient Langevin Dynamics is one of the most fundamental algorithms to solve non-eps optimization problems. In this paper, we show two variants of this kind, namely the Variance Reduced Langevin Dynamics and the Recursive Gradient Langevin Dynamics.
  arXiv  Detail & Related papers  (2022-03-30T11:39:00Z)
• Benign Overfitting of Constant-Stepsize SGD for Linear Regression [122.70478935214128]
  inductive biases are central in preventing overfitting empirically. This work considers this issue in arguably the most basic setting: constant-stepsize SGD for linear regression. We reflect on a number of notable differences between the algorithmic regularization afforded by (unregularized) SGD in comparison to ordinary least squares.
  arXiv  Detail & Related papers  (2021-03-23T17:15:53Z)
• Direction Matters: On the Implicit Bias of Stochastic Gradient Descent with Moderate Learning Rate [105.62979485062756]
  This paper attempts to characterize the particular regularization effect of SGD in the moderate learning rate regime. We show that SGD converges along the large eigenvalue directions of the data matrix, while GD goes after the small eigenvalue directions.
  arXiv  Detail & Related papers  (2020-11-04T21:07:52Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.