[Paper Review] Noisy Linear Convergence of Stochastic Gradient Descent for CV@R Statistical Learning under Polyak-Łojasiewicz Conditions

Dionysios S. Kalogerias|arXiv (Cornell University)|Dec 14, 2020

Statistical Methods and Inference23 references4 citations

TL;DR

This paper establishes noisy linear convergence of stochastic gradient descent (SGD) for conditional value-at-risk (CVaR) statistical learning under the Polyak-Łojasiewicz (PŁ) condition, proving that CVaR learning is computationally as efficient as risk-neutral learning for a broad class of loss functions—including smooth and strongly convex ones. The key result is that SGD achieves fixed-accuracy linear convergence even when the loss is not convex, disproving the common belief that CVaR optimization is inherently difficult.

ABSTRACT

Conditional Value-at-Risk ($\mathrm{CV@R}$) is one of the most popular measures of risk, which has been recently considered as a performance criterion in supervised statistical learning, as it is related to desirable operational features in modern applications, such as safety, fairness, distributional robustness, and prediction error stability. However, due to its variational definition, $\mathrm{CV@R}$ is commonly believed to result in difficult optimization problems, even for smooth and strongly convex loss functions. We disprove this statement by establishing noisy (i.e., fixed-accuracy) linear convergence of stochastic gradient descent for sequential $\mathrm{CV@R}$ learning, for a large class of not necessarily strongly-convex (or even convex) loss functions satisfying a set-restricted Polyak-Lojasiewicz inequality. This class contains all smooth and strongly convex losses, confirming that classical problems, such as linear least squares regression, can be solved efficiently under the $\mathrm{CV@R}$ criterion, just as their risk-neutral versions. Our results are illustrated numerically on such a risk-aware ridge regression task, also verifying their validity in practice.

Motivation & Objective

• To challenge the widely held belief that CVaR-based statistical learning leads to computationally difficult optimization problems.
• To establish convergence guarantees for stochastic gradient descent (SGD) in sequential CVaR learning under minimal assumptions.
• To demonstrate that CVaR learning can achieve the same convergence rates as risk-neutral learning, even for non-convex or non-strongly convex losses.
• To validate the theoretical findings through numerical experiments on risk-aware ridge regression.

Proposed method

• The authors analyze sequential SGD for CVaR learning using a variational reformulation of CVaR, reducing the problem to a standard stochastic optimization with an augmented loss function.
• They introduce a set-restricted Polyak-Łojasiewicz (PŁ) inequality, which generalizes the classical PŁ condition to allow non-convex and non-strongly convex loss functions.
• The method leverages gradient estimates derived from the variational form of CVaR, enabling stable and efficient SGD updates in a streaming data setting.
• Convergence is proven under fixed-accuracy (noisy) conditions, showing linear convergence to a neighborhood of the optimal solution.
• The analysis accounts for the impact of the CVaR confidence level α and the smoothing parameter σ in the gradient approximation.
• Numerical validation is conducted on a risk-aware ridge regression task, comparing CVaR-SGD with standard LMS (risk-neutral) SGD.

Experimental results

Research questions

• RQ1 Can stochastic gradient descent achieve linear convergence for CVaR-based statistical learning under weak assumptions?
• RQ2 Does the presence of non-convexity or lack of strong convexity in the loss function prevent fast convergence in CVaR learning?
• RQ3 Is the computational complexity of CVaR learning comparable to that of risk-neutral learning, particularly for classical problems like ridge regression?
• RQ4 Can the PŁ condition be extended to a set-restricted form to cover non-convex losses in CVaR optimization?
• RQ5 How does the choice of the CVaR confidence level α affect the convergence rate and solution stability?

Key findings

• SGD achieves noisy linear convergence for CVaR learning under the set-restricted Polyak-Łojasiewicz condition, even when the loss function is not convex or strongly convex.
• The convergence rate is linear up to a fixed accuracy, with the error bound scaling as O(1/α²) in the worst case, depending on problem parameters.
• All smooth and strongly convex losses satisfy the set-restricted PŁ condition, confirming that classical problems like linear least squares and ridge regression can be solved efficiently under the CVaR criterion.
• Numerical results on risk-aware ridge regression show that CVaR-SGD converges at a rate comparable to LMS (risk-neutral) SGD, with significantly reduced prediction error volatility.
• The risk-aware solution trades off mean performance for improved stability in worst-case error, with the trade-off customizable via the α parameter.
• The theoretical error bound for the fixed-accuracy convergence is of the order O(√(max{β,γ}²/min{β,γ})/α²), confirming robustness to parameter choices.