Stochastic Sequential Quadratic Programming

• Stochastic Sequential Quadratic Programming (SSQP) is a family of methods for solving constrained stochastic optimization problems using momentum debiasing and adaptive stepsizes.
• SSQP constructs local quadratic models with linearized constraints to efficiently handle noisy gradients and Hessians while ensuring global convergence.
• The approach guarantees asymptotic normality and supports online covariance estimation, making it valuable in machine learning, statistics, and finance.

Stochastic Sequential Quadratic Programming (SSQP) encompasses a family of algorithms for solving constrained stochastic optimization problems. These methods extend classical SQP to settings where the objective or constraints involve expectations over random variables, and where only stochastic (noisy) first- and possibly second-order information is accessible. SSQP schemes have become central in machine learning, statistics, operational research, and online inference, enabling direct handling of constraints without the need for computationally intractable projections or excessively conservative penalty methods. They provide rigorous convergence properties, support statistical inference, and maintain robustness to noise, bias, and high-dimensionality under diverse regularity conditions.

1. Problem Formulation and Core Algorithmic Structure

SSQP methods are designed for problems of the form: $$\begin{aligned} &\min_{\theta\in\mathbb{R}^d} && f(\theta) := \mathbb{E}[\ell(\theta; \xi)] \ &\text{subject to} && h(\theta) = 0 \in \mathbb{R}^m,\quad g(\theta) \leq 0 \in \mathbb{R}^r, \end{aligned}$$ where $\ell(\theta;\xi)$ is a sample loss, $h$ and $g$ are smooth (nonlinear) constraint functions, and $\xi$ is a random variable. The Lagrangian is given by $$L(\theta,\lambda,\mu) = f(\theta) + \lambda^\top h(\theta) + \mu^\top g(\theta)$$ with $\mu\geq0$.

At each iteration, the SSQP algorithm constructs a local quadratic approximation of the objective and a linearization of the constraints at the current parameter $\theta_k$, then solves a stochastic convex quadratic program: $$\begin{aligned} &\min_{d \in \mathbb{R}^d}\quad && \bar{g}_k^\top d + \frac{1}{2} d^\top W_k d \ &\text{subject to}\quad && h(\theta_k) + \nabla h(\theta_k)^\top d = 0,\quad g(\theta_k) + \nabla g(\theta_k)^\top d \leq 0, \end{aligned}$$ where $\bar{g}_k$ is an averaged estimate of $\ell(\theta;\xi)$0, $\ell(\theta;\xi)$1 an estimated or regularized Hessian, and all constraint information is local to the step.

The algorithm updates the primal and dual variables according to: $\ell(\theta;\xi)$2 where $\ell(\theta;\xi)$3 is a carefully chosen stepsize, and $\ell(\theta;\xi)$4 are the dual multipliers from the SQP subproblem (Gao et al., 27 Nov 2025).

2. Debiasing and Stochastic Approximation Techniques

A central difficulty in stochastic constrained optimization is that naive stochastic estimates induce bias into the search direction, especially in the presence of inequality constraints. SSQP overcomes this via momentum-type averaging: $\ell(\theta;\xi)$5 for appropriate decay schedules $\ell(\theta;\xi)$6, allowing the algorithm to debias the step direction and suppress the variance induced by stochastic oracles. The subproblem Hessian is then constructed as $\ell(\theta;\xi)$7, where $\ell(\theta;\xi)$8 regularizes the Hessian to ensure uniform positive definiteness (Gao et al., 27 Nov 2025).

Adaptive line search or model-based stepsize selection can further mitigate stochastic errors, either by backtracking (using batch or single-sample inexact Armijo conditions) or by analytic rules based on estimated local Lipschitz constants and penalized model reduction metrics (Gao et al., 27 Nov 2025, Na et al., 2021, Berahas et al., 2023).

3. Theoretical Guarantees: Convergence and Asymptotic Inference

Under standard regularity conditions, including Lipschitz gradients, constraint qualification (e.g., LICQ), and second-order sufficient conditions (SOSC), SSQP methods exhibit the following properties:

• Global Almost-Sure Convergence: For suitably diminishing stepsizes ($\ell(\theta;\xi)$9), and consistent decays of $h$0, the algorithm almost surely drives the composite KKT residual to zero:

$h$1

with any limit point satisfying first-order stationary conditions (Gao et al., 27 Nov 2025).

• Local Asymptotic Normality: Near a strict local solution $h$2, the joint sequence

$h$3

satisfies

$h$4

where $h$5 is given by the inverse KKT matrix “sandwich” formula, and for $h$6, minimax-efficient $h$7-consistency is attained, matching the Hájek–Le Cam lower bound for joint inference (Gao et al., 27 Nov 2025).

• Online Plug-in Covariance Estimation: SSQP supports consistent online estimation of the limiting covariance, enabling valid inference for constrained parameters in streaming and online settings, without requiring projections onto constraint sets (Gao et al., 27 Nov 2025).

4. Computational Characteristics and Practical Implementation

Each iteration of SSQP solves a (stochastic) convex quadratic program—whose size grows with the number of constraints—using only local gradient and Hessian (or Hessian-vector product) information. Notable computational features include:

• Avoidance of Global Projections: At no point is a global projection onto the feasible set $h$8 required; all constraint enforcement is local and implicit via the SQP subproblem.
• Debiasing Momentum: The use of Polyak-type momentum in gradient and Hessian estimation is essential for correcting the bias induced by stochastic truncation, particularly under inequality constraints (Gao et al., 27 Nov 2025).
• Adaptive Stepsizes: Stepsizes are selected from a bracket $h$9, with $g$0 sufficient to maintain inference guarantees.
• Covariance Estimation: On-the-fly plug-in estimators (sample covariance of stochastic gradients over the iterates) enable direct monitoring of asymptotic uncertainty, supporting sequential hypothesis testing and confidence interval construction.

A high-level implementation sketch is:

$g$3

5. Empirical Performance and Applications

SSQP methods have been validated on a diverse range of benchmark and real-world problems:

• CUTEst Nonlinear Test Problems: On noisy nonlinear optimization test sets, SSQP achieves global convergence even under substantial gradient and Hessian noise, outperforming alternatives that require large batches or rely on active-set heuristics (Gao et al., 27 Nov 2025).
• Constrained Generalized Linear Models (GLMs): On synthetic and real datasets (e.g., Chicago pollution), SSQP produces parameter confidence intervals with empirical coverage close to 95%, demonstrating both calibration and practical inferential utility.
• Portfolio Optimization: In portfolio allocation models involving multiple constraint types (GMV, MV, log-utilities), SSQP matches the out-of-sample Sharpe and Sortino ratios of "offline" $g$1-estimation solutions (Gao et al., 27 Nov 2025).
• In all cases, the KKT residual declines to zero globally, coverage for key parameters is maintained at the nominal level, and the momentum scheme significantly shrinks the stochastic variance of estimates over time.

6. Limitations and Future Directions

SSQP inherits several open issues and limitations:

• Constraint Qualification Dependence: All global and local guarantees require standard constraint qualification, typically LICQ near limit points. Relaxation or extension to broader classes of constraint degeneracy remain open.
• Per-Iteration Complexity: Solving a large convex QP at each step can be computationally intensive in high dimensions or with many constraints, which necessitates further development of efficient, possibly inexact, large-scale QP solvers.
• Extension to High-Dimensional and Statistical Regimes: Non-asymptotic rates, extensions to high-dimensional settings (including sparse or low-rank constraints), and integration with non-Euclidean or structured stochastic settings represent active areas of research (Gao et al., 27 Nov 2025).
• Online Inference Beyond Minimax Settings: While current methods attain minimax-optimal covariance, further work on finite-sample adaptivity, non-parametric uncertainty quantification under weak constraint structure, and higher-order inferential goals is underway.

7. Distinctiveness and Relation to Prior Approaches

The key aspects distinguishing the SSQP method (Gao et al., 27 Nov 2025):

• Fully Online and Projection-Free: Unlike projected stochastic gradient or primal-dual averaging schemes, SSQP requires no projection onto possibly intractable constraint sets.
• Primal-Dual Optimality and Inference: It is the first fully online method to guarantee both almost-sure convergence and asymptotic normality with sharp (Hájek–Le Cam) covariance for the joint $g$2 variables, accessible via a plug-in estimator updated on the fly.
• Variance Debiasing via Momentum: The step direction is actively debiased using a gradient moving-average (“Polyak momentum”), which is necessary for accurate inference in the presence of inequality-induced truncation.
• No Penalty or Augmented Lagrangian Tuning: SSQP is formulated entirely in primal-dual coordinates without requiring explicit penalty parameter tuning or the use of augmented Lagrangians or intermediate merit functions for globalization.

The SSQP methodology therefore represents the current state-of-the-art in fully online, theoretically principled, and inferentially calibrated methods for constrained stochastic nonlinear optimization. Its design, theoretical framework, and empirical performance set a rigorous foundation for further developments in large-scale, online, and high-dimensional stochastic optimization with constraints (Gao et al., 27 Nov 2025).