Searcharxiv⌕ Search

arXiv subjects

Egor Shulgin

Publications and source records attributed to Egor Shulgin.

At least 19 recordsLinked to original sources

Understanding Practical Optimization in Machine Learning

Modern machine learning relies on practical optimization methods that often differ from the algorithms studied in classical analyses: they train submodels, clip or normalize gradients, use layer-wise non-Euclidean directions, or approximate costly inner computations. This thesis analyzes the objective, assumptions, update, and computation defined by these procedures, rather than those of simplified proxies. For submodel training, we show that Independent Subnetwork Training can converge to a shifted fixed point because sketching before differentiation biases the update for the original objective. Model-Agnostic Sparsified Training instead incorporates the sketch distribution into its objective, which makes sketched gradients unbiased and admits convergence guarantees. For private optimization, we show that adaptive quantile clipping retains bias at any fixed quantile, while suitable quantile schedules recover near-optimal nonconvex rates under heavy-tailed noise. We then combine smoothed normalization with error compensation and obtain distributed private convergence guarantees under standard assumptions, without assuming bounded gradients. For optimizers based on linear minimization oracles, we develop and validate a layer-wise generalized smoothness model and build on it the Gluon framework for Muon- and Scion-style methods. We also show that the approximate orthogonalization in Muon matters: less precise orthogonalization calls for smaller stepsizes, so its precision should be tuned together with the learning rate. Overall, practical methods are better understood when theory retains the features that determine their optimization behavior.

cs.LG↗

SCAFFOLD under Hessian Similarity: Sharp Convergence Rates

One of the main challenges of federated learning is heterogeneous data split among the clients, leading to biased local updates called client drift. The SCAFFOLD algorithm (Karimireddy et al., 2020) corrects client drift with control variates, but its relation to the similarity of client data is not yet fully understood: the original similarity-based guarantees cover only quadratics with a stronger second-order similarity assumption, and a follow-up weakly convex analysis (Luo et al., 2025) obtains a $\sqrt{Lδ}$ dependence for $L$-smooth losses with averaged Hessian similarity $δ$. We analyze SCAFFOLD with minibatch-based control variates and establish linear dependence on Hessian similarity beyond quadratics, for both convex and nonconvex objectives. In particular, for convex objectives with exact gradients, $R$ SCAFFOLD rounds with $K$ local steps lead to an objective gap of order $(L/K+δ)D_0^2/R$, where $D_0$ is the initial distance to a minimizer. When $K=Θ(L/δ)$, it improves the minibatch SGD rate by a factor $Θ(L/δ)$. For nonconvex objectives, we prove an analogous stationarity guarantee under uniform Hessian similarity for a client-selection variant. In both settings, we extend these guarantees to stochastic gradients and prove matching lower bounds for the analyzed methods with constant local stepsizes. We also show that, under partial participation, even identical clients need not obtain the full communication benefit of additional local steps.

cs.LG↗

A Quadratic Lens on Muon's Polar Update: Benefits, Pitfalls, and the Role of Newton-Schulz

Muon updates matrix-valued parameters along approximately orthogonalized directions and often trains large language models faster than AdamW, yet it remains unclear when and why orthogonalization helps. We study PolarGD, the orthogonalized update at the core of Muon, on matrix quadratics, where optimizers can be analyzed in closed form, and compare update directions under line search to remove the effect of learning-rate tuning. Orthogonalization pays off when the gradient is far from low rank and its largest singular directions see high curvature. PolarGD reaches the minimizer in one step whenever the rows of the error are orthogonal with equal norms, for any Hessian, whereas gradient descent needs all Hessian eigenvalues to be equal; and unlike sign descent, PolarGD is unaffected by rotations of the problem. Over many steps, PolarGD acts as gradient descent preconditioned by the current iterate, so its speed is governed by the conditioning of the iterate rather than of the Hessian. This explains fast regimes as well as a slow one, caused by tiny singular values of the iterate, in which PolarGD is much slower than gradient descent. Finally, the few Newton-Schulz iterations used in Muon treat these tiny singular values as gradient descent does, so fewer iterations can converge faster than exact orthogonalization; in a transformer trained with Muon, about a quarter of the singular values of the attention matrices fall in this regime.

cs.LG↗

How Does Muon Work in Private Deep Learning? Scalable Differentially Private Training via Orthogonalization

Training deep neural networks requires optimizers that balance computational efficiency with formal data privacy. While orthogonalized update rules, such as the recently introduced Muon optimizer, have shown strong empirical performance and scalability, their behavior under Differential Privacy (DP) is not yet understood. In this paper, we propose DP-Muon, a differentially private variant of the Muon optimizer. By analyzing Muon's orthogonalized updates through the lens of a Linear Minimization Oracle (LMO), we show that the leading effect of per-example clipping is a positive rescaling of the gradient, to which the LMO is blind, so that only a small directional residual controlled by the centered gradient variance enters the optimization error. This yields an $\mathcal O(K^{-1/4})$ convergence rate, up to logarithmic factors, to a clipping-dependent neighborhood at a fixed noise scale and, with noise calibrated to a fixed $(\varepsilon,δ)$ budget, the standard $(\sqrt d/(n\varepsilon))^{1/2}$ privacy--utility bound of DP-SGD, without assuming bounded gradients. Furthermore, we investigate the hyperparameter landscape of DP-Muon and document how the optimal learning rate, momentum, and batch size change under privacy. Empirically, on image classification DP-Muon matches or improves on DP-SGD and DP-Adam, whereas on language modeling the best optimizer depends on the batch size, with DP-Muon best at small batches and DP-Adam best near the privacy-optimal batch size.

cs.LG↗

Optimizers for Diffusion Models: A Controlled Benchmark

Discrete diffusion models now match autoregressive language models on several benchmarks, while the question of how best to train them has received far less attention: the optimizer is inherited from one paper to the next and never compared. New optimizers, meanwhile, are validated almost exclusively on autoregressive pretraining, a different objective on a different loss surface. We present a controlled optimizer benchmark across four diffusion formulations, to our knowledge the first for discrete diffusion: seven optimizers (AdamW, Lion, Muon, SOAP, MARS, MARS-M, Schedule-Free) on masked diffusion (text8), uniform diffusion (QM9, and LM1B through the Gaussian duality) and Gaussian diffusion on images (CelebA-64), each on a task with published reference values. Every optimizer receives the same search protocol, and every winner is retrained at the full budget with three seeds. AdamW is a strong default but not always the right choice: it is beaten by a resolved margin on two of the four tasks, and the winner changes with the formulation, so the optimizer deserves the same care as the rest of the training recipe. Notably, methods validated on autoregressive language model pretraining transfer well: Muon, MARS-M and SOAP each beat the tuned AdamW on at least one diffusion formulation. The benchmark, all runs and every figure are reproducible end to end from the released code at https://github.com/armanbolatov/diffusion-baselines.

cs.LG↗

Deriving Hyperparameter Scaling Laws via Modern Optimization Theory

Hyperparameter transfer has become an important component of modern large-scale training recipes. Existing methods, such as muP, primarily focus on transfer between model sizes, with transfer across batch sizes and training horizons often relying on empirical scaling rules informed by insights from timescale preservation, quadratic proxies, and continuous-time approximations. We study hyperparameter scaling laws for modern first-order optimizers through the lens of recent convergence bounds for methods based on the Linear Minimization Oracle (LMO), a framework that includes normalized SGD, signSGD (approximating Adam), and Muon. Treating bounds in recent literature as a proxy and minimizing them across different tuning regimes yields closed-form power-law schedules for learning rate, momentum, and batch size as functions of the iteration or token budget. Our analysis, holding model size fixed, recovers most insights and observations from the literature under a unified and principled perspective, with clear directions open for future research. Our results draw particular attention to the interaction between momentum and batch-size scaling, suggesting that optimal performance may be achieved with several scaling strategies.

cs.LG↗

MAST: Model-Agnostic Sparsified Training

We introduce a novel optimization problem formulation that departs from the conventional way of minimizing machine learning model loss as a black-box function. Unlike traditional formulations, the proposed approach explicitly incorporates an initially pre-trained model and random sketch operators, allowing for sparsification of both the model and gradient during training. We establish the insightful properties of the proposed objective function and highlight its connections to the standard formulation. Furthermore, we present several variants of the Stochastic Gradient Descent (SGD) method adapted to the new problem formulation, including SGD with general sampling, a distributed version, and SGD with variance reduction techniques. We achieve tighter convergence rates and relax assumptions, bridging the gap between theoretical principles and practical applications, covering several important techniques such as Dropout and Sparse training. This work presents promising opportunities to enhance the theoretical understanding of model training through a sparsification-aware optimization approach.

cs.LG↗

First Provable Guarantees for Practical Private FL: Beyond Restrictive Assumptions

Federated Learning (FL) enables collaborative training on decentralized data. Differential privacy (DP) is crucial for FL, but current private methods often rely on unrealistic assumptions (e.g., bounded gradients or heterogeneity), hindering practical application. Existing works that relax these assumptions typically neglect practical FL features, including multiple local updates and partial client participation. We introduce Fed-$α$-NormEC, the first differentially private FL framework providing provable convergence and DP guarantees under standard assumptions while fully supporting these practical features. Fed-$α$-NormE integrates local updates (full and incremental gradient steps), separate server and client stepsizes, and, crucially, partial client participation, which is essential for real-world deployment and vital for privacy amplification. Our theoretical guarantees are corroborated by experiments on private deep learning tasks.

cs.LG↗

Beyond the Ideal: Analyzing the Inexact Muon Update

The Muon optimizer has rapidly emerged as a powerful, geometry-aware alternative to AdamW, demonstrating strong performance in large-scale training of neural networks. However, a critical theory-practice disconnect exists: Muon's efficiency relies on fast, approximate orthogonalization, yet all prior theoretical work analyzes an idealized, computationally intractable version assuming exact SVD-based updates. This work moves beyond the ideal by providing the first analysis of the inexact orthogonalized update at Muon's core. We develop our analysis within the general framework of Linear Minimization Oracle (LMO)-based optimization, introducing a realistic additive error model to capture the inexactness of practical approximation schemes. Our analysis yields explicit bounds that quantify performance degradation as a function of the LMO inexactness/error. We reveal a fundamental coupling between this inexactness and the optimal step size and momentum: lower oracle precision requires a smaller step size but larger momentum parameter. These findings elevate the approximation procedure (e.g., the number of Newton-Schulz steps) from an implementation detail to a critical parameter that must be co-tuned with the learning schedule. NanoGPT experiments directly confirm the predicted coupling, with optimal learning rates clearly shifting as approximation precision changes.

cs.LG↗

Gluon: Making Muon & Scion Great Again! (Bridging Theory and Practice of LMO-based Optimizers for LLMs)

Recent developments in deep learning optimization have brought about radically new algorithms based on the Linear Minimization Oracle (LMO) framework, such as $\sf Muon$ and $\sf Scion$. After over a decade of $\sf Adam$'s dominance, these LMO-based methods are emerging as viable replacements, offering several practical advantages such as improved memory efficiency, better hyperparameter transferability, and most importantly, superior empirical performance on large-scale tasks, including LLM training. However, a significant gap remains between their practical use and our current theoretical understanding: prior analyses (1) overlook the layer-wise LMO application of these optimizers in practice, and (2) rely on an unrealistic smoothness assumption, leading to impractically small stepsizes. To address both, we propose a new LMO-based method called $\sf Gluon$, capturing prior theoretically analyzed methods as special cases, and introduce a new refined generalized smoothness model that captures the layer-wise geometry of neural networks, matches the layer-wise practical implementation of $\sf Muon$ and $\sf Scion$, and leads to convergence guarantees with strong practical predictive power. Unlike prior results, our theoretical stepsizes closely match the fine-tuned values reported by Pethick et al. (2025). Our experiments with NanoGPT and CNN confirm that our assumption holds along the optimization trajectory, ultimately closing the gap between theory and practice.

cs.LG↗

Smoothed Normalization for Efficient Distributed Private Optimization

Federated learning enables training machine learning models while preserving the privacy of participants. Surprisingly, there is no differentially private distributed method for smooth, non-convex optimization problems. The reason is that standard privacy techniques require bounding the participants' contributions, usually enforced via $\textit{clipping}$ of the updates. Existing literature typically ignores the effect of clipping by assuming the boundedness of gradient norms or analyzes distributed algorithms with clipping but ignores DP constraints. In this work, we study an alternative approach via $\textit{smoothed normalization}$ of the updates motivated by its favorable performance in the single-node setting. By integrating smoothed normalization with an error-feedback mechanism, we design a new distributed algorithm $α$-$\sf NormEC$. We prove that our method achieves a superior convergence rate over prior works. By extending $α$-$\sf NormEC$ to the DP setting, we obtain the first differentially private distributed optimization algorithm with provable convergence guarantees. Finally, our empirical results from neural network training indicate robust convergence of $α$-$\sf NormEC$ across different parameter settings.

cs.LG↗

On the Convergence of DP-SGD with Adaptive Clipping

Stochastic Gradient Descent (SGD) with gradient clipping is a powerful technique for enabling differentially private optimization. Although prior works extensively investigated clipping with a constant threshold, private training remains highly sensitive to threshold selection, which can be expensive or even infeasible to tune. This sensitivity motivates the development of adaptive approaches, such as quantile clipping, which have demonstrated empirical success but lack a solid theoretical understanding. This paper provides the first comprehensive convergence analysis of SGD with quantile clipping (QC-SGD). We demonstrate that QC-SGD suffers from a bias problem similar to constant-threshold clipped SGD but show how this can be mitigated through a carefully designed quantile and step size schedule. Our analysis reveals crucial relationships between quantile selection, step size, and convergence behavior, providing practical guidelines for parameter selection. We extend these results to differentially private optimization, establishing the first theoretical guarantees for DP-QC-SGD. Our findings provide theoretical foundations for widely used adaptive clipping heuristic and highlight open avenues for future research.

cs.LG↗

Towards a Better Theoretical Understanding of Independent Subnetwork Training

Modern advancements in large-scale machine learning would be impossible without the paradigm of data-parallel distributed computing. Since distributed computing with large-scale models imparts excessive pressure on communication channels, significant recent research has been directed toward co-designing communication compression strategies and training algorithms with the goal of reducing communication costs. While pure data parallelism allows better data scaling, it suffers from poor model scaling properties. Indeed, compute nodes are severely limited by memory constraints, preventing further increases in model size. For this reason, the latest achievements in training giant neural network models also rely on some form of model parallelism. In this work, we take a closer theoretical look at Independent Subnetwork Training (IST), which is a recently proposed and highly effective technique for solving the aforementioned problems. We identify fundamental differences between IST and alternative approaches, such as distributed methods with compressed communication, and provide a precise analysis of its optimization performance on a quadratic model.

cs.LG↗

SPAM: Stochastic Proximal Point Method with Momentum Variance Reduction for Non-convex Cross-Device Federated Learning

Cross-device training is a crucial subfield of federated learning, where the number of clients can reach into the billions. Standard approaches and local methods are prone to issues such as client drift and insensitivity to data similarities. We propose a novel algorithm (SPAM) for cross-device federated learning with non-convex losses, which solves both issues. We provide sharp analysis under second-order (Hessian) similarity, a condition satisfied by a variety of machine learning problems in practice. Additionally, we extend our results to the partial participation setting, where a cohort of selected clients communicate with the server at each communication round. Our method is the first in its kind, that does not require the smoothness of the objective and provably benefits from clients having similar data.

math.OC↗

Towards accelerated rates for distributed optimization over time-varying networks

We study the problem of decentralized optimization over time-varying networks with strongly convex smooth cost functions. In our approach, nodes run a multi-step gossip procedure after making each gradient update, thus ensuring approximate consensus at each iteration, while the outer loop is based on accelerated Nesterov scheme. The algorithm achieves precision $\varepsilon > 0$ in $O(\sqrt{κ_g}χ\log^2(1/\varepsilon))$ communication steps and $O(\sqrt{κ_g}\log(1/\varepsilon))$ gradient computations at each node, where $κ_g$ is the global function number and $χ$ characterizes connectivity of the communication network. In the case of a static network, $χ= 1/γ$ where $γ$ denotes the normalized spectral gap of communication matrix $\mathbf{W}$. The complexity bound includes $κ_g$, which can be significantly better than the worst-case condition number among the nodes.

math.OC↗

Certified Robustness in Federated Learning

Federated learning has recently gained significant attention and popularity due to its effectiveness in training machine learning models on distributed data privately. However, as in the single-node supervised learning setup, models trained in federated learning suffer from vulnerability to imperceptible input transformations known as adversarial attacks, questioning their deployment in security-related applications. In this work, we study the interplay between federated training, personalization, and certified robustness. In particular, we deploy randomized smoothing, a widely-used and scalable certification method, to certify deep networks trained on a federated setup against input perturbations and transformations. We find that the simple federated averaging technique is effective in building not only more accurate, but also more certifiably-robust models, compared to training solely on local data. We further analyze personalization, a popular technique in federated training that increases the model's bias towards local data, on robustness. We show several advantages of personalization over both~(that is, only training on local data and federated training) in building more robust models with faster training. Finally, we explore the robustness of mixtures of global and local~(i.e. personalized) models, and find that the robustness of local models degrades as they diverge from the global model

cs.LG↗

Shifted Compression Framework: Generalizations and Improvements

Communication is one of the key bottlenecks in the distributed training of large-scale machine learning models, and lossy compression of exchanged information, such as stochastic gradients or models, is one of the most effective instruments to alleviate this issue. Among the most studied compression techniques is the class of unbiased compression operators with variance bounded by a multiple of the square norm of the vector we wish to compress. By design, this variance may remain high, and only diminishes if the input vector approaches zero. However, unless the model being trained is overparameterized, there is no a-priori reason for the vectors we wish to compress to approach zero during the iterations of classical methods such as distributed compressed {\sf SGD}, which has adverse effects on the convergence speed. Due to this issue, several more elaborate and seemingly very different algorithms have been proposed recently, with the goal of circumventing this issue. These methods are based on the idea of compressing the {\em difference} between the vector we would normally wish to compress and some auxiliary vector which changes throughout the iterative process. In this work we take a step back, and develop a unified framework for studying such methods, conceptually, and theoretically. Our framework incorporates methods compressing both gradients and models, using unbiased and biased compressors, and sheds light on the construction of the auxiliary vectors. Furthermore, our general framework can lead to the improvement of several existing algorithms, and can produce new algorithms. Finally, we performed several numerical experiments which illustrate and support our theoretical findings.

cs.LG↗

Adaptive Catalyst for Smooth Convex Optimization

In this paper, we present a generic framework that allows accelerating almost arbitrary non-accelerated deterministic and randomized algorithms for smooth convex optimization problems. The main approach of our envelope is the same as in Catalyst (Lin et al., 2015): an accelerated proximal outer gradient method, which is used as an envelope for a non-accelerated inner method for the $\ell_2$ regularized auxiliary problem. Our algorithm has two key differences: 1) easily verifiable stopping criteria for inner algorithm; 2) the regularization parameter can be tunned along the way. As a result, the main contribution of our work is a new framework that applies to adaptive inner algorithms: Steepest Descent, Adaptive Coordinate Descent, Alternating Minimization. Moreover, in the non-adaptive case, our approach allows obtaining Catalyst without a logarithmic factor, which appears in the standard Catalyst (Lin et al., 2015, 2018).

math.OC↗