Residual-Controlled Multiplier Learning for Stochastic Constrained Decision-Making

Abstract: Stochastic constrained decision-making requires optimizing performance objectives while enforcing statistical requirements such as safety or fairness. However, standard primal--dual methods struggle to update multipliers robustly under stochastic mini-batch feedback, as the noise of mini-batch gradients and constraint estimates can be directly accumulated into the multiplier memory. To address this issue, we propose Residual-Controlled Multiplier Learning (RCML), which reformulates multiplier updating as projected-pressure feedback. The central idea is to decompose the projected multiplier into an effective pressure signal for primal descent and a pressure-memory residual for finite-gain multiplier tracking. To handle heterogeneous and noisy observations, we further augment this residual-integral backbone with modular stochastic stabilization components. For the convex-affine backbone, we establish finite-gain convergence, derive a stochastic residual bound under mini-batch feedback, and show that the residual feedback law admits a local KKT-residual interpretation near regular KKT points of nonconvex problems. Experiments across optimization, allocation, and fair-ranking tasks show that RCML improves feasibility control and multiplier stability while maintaining competitive objective performance. Code is released at https://anonymous.4open.science/r/RCML-3114/.

• The paper introduces RCML, a novel feedback-based method for stabilizing multiplier updates in stochastic constrained decision-making.
• The methodology leverages projected-pressure residuals, adaptive scaling, and PI-correction to balance noise suppression and feasibility control.
• Empirical results show RCML achieves lower dual variation and improved feasibility across convex, nonconvex, and neural fair ranking tasks.

Residual-Controlled Multiplier Learning for Stochastic Constrained Decision-Making

Motivation

Stochastic constrained optimization is central in machine learning tasks that feature statistical constraints, including safety, fairness, and risk budgeting. The standard primal–dual and augmented Lagrangian methodologies are vulnerable to instability when updated using noisy mini-batch feedback, as stochastic gradient estimates and constraint residuals inject high variance into multiplier updates. When the dual ascent step receives these fluctuating signals without any appropriate adaptation, it commonly yields unstable or slow dynamics, constraint infeasibility, or excessive multiplier drift. Moreover, typical remedies—such as integrating only positive violations or fully replacing multipliers with projected augmented estimates—are either too sluggish to release obsolete dual mass or too aggressive, amplifying variance rather than robustly enforcing feasibility.

The presented work, "Residual-Controlled Multiplier Learning for Stochastic Constrained Decision-Making" (2606.07088), targets this bottleneck by proposing a structured feedback mechanism for multiplier updates. This approach decouples the projected pressure input to the primal from the persistent dual memory, addressing the limitations of direct-violation-driven updates and classical projected ALM replacement.

Methodology

Projected-Pressure Residuals

The key algorithmic construction is the Residual-Controlled Multiplier Learning (RCML) framework, which operates with two distinct multiplier-side signals: the projected effective multiplier $$\lambda_{\boldsymbol\rho}(x,u) = [u + \boldsymbol\rho \odot c(x)]_+$$ and the pressure-memory residual $$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$. The first drives the constraint-corrective direction in the primal variable; the second acts as feedback to update the dual variable.

Unlike classical primal–dual or ALM-type algorithms, RCML employs a finite-gain residual law: instead of full multiplier mass substitution or unfiltered integral error accumulation, the multiplier state is partly updated toward the projected effective pressure with a gain $\beta_k \le 1$, i.e., $u_{k+1} = (1-\beta_k) u_k + \beta_k \lambda_k$. This enables smooth tracking of feasibility pressure (for constraint activation), controlled release of obsolete multiplier memory (when constraints deactivate), and eliminates leakage from inactive (already feasible) constraints.

Algorithmic Stabilization

To cope with heterogeneity and noise, RCML is further equipped with modular stabilization components:

• Constraint Filtering: An exponential moving average is applied to mini-batch constraint residuals, balancing noise suppression and tracking lag.
• Adaptive Coordinate-wise Scaling: The gain $\boldsymbol\rho_k$ is dynamically tuned per constraint coordinate based on online variance estimates to balance feedback efficacy across channels.
• Residual–$\nu$PI Correction: A proportional-integral feedback is introduced in the dual update, leveraging filtered trends in the residuals for improved transient behavior under nonstationary constraint activity.

These enhancements are embedded into a unified algorithmic template, making RCML compatible with a wide suite of stochastic constrained tasks, from convex–affine programs to nonconvex neural objectives.

Theoretical Properties

Optimality and Residual Geometry

The projected-pressure residual $d_{\boldsymbol\rho}(x,u)$ is shown to be an exact complementarity residual: $d_{\boldsymbol\rho}(x,u)=0$ if and only if $(x,u)$ satisfies non-negativity, primal feasibility, and equality (complementarity) slackness for all constraints. When coupled with projected stationarity, this residual forms a natural KKT system for general smooth constrained programs.

Finite-Gain Convergence under Convexity

RCML's backbone is analyzed in the convex–affine setting. It is shown that for stepsizes $\alpha_k$ and gain ratios $$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$0, the sequence $$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$1 converges to the unique primal solution and the KKT multiplier set. The primal–dual Lyapunov construction confirms that finite-gain tracking suffices—full ALM-style replacement is unnecessary for convergence.

Stochastic Residual Bounds

Under stochastic mini-batch feedback, the algorithm achieves a weighted residual bound:

$$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$2

where $$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$3 is batch size, $$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$4 is constraint noise variance, and $$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$5 is horizon length. The dominant term for fixed $$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$6 is the pressure-noise floor $$d_{\boldsymbol\rho}(x,u) = \lambda_{\boldsymbol\rho}(x,u) - u$$7; exact convergence to KKT points is precluded in the presence of persistent stochasticity, but the iterates remain in a controllable residual neighborhood. The theoretical analysis further quantifies how filtering, scaling, and dynamic correction modules act as structured perturbations to the core feedback channels.

Local Nonconvex Characterization

While global convergence in the nonconvex case is not shown, the projected-pressure residual is locally equivalent to traditional stationarity and complementarity conditions near regular KKT points (e.g., under LICQ, strict complementarity, and second-order sufficiency), providing a diagnostic map for empirical constraint satisfaction.

Empirical Evaluation

RCML and its variants are rigorously benchmarked across four scenarios:

• Signal-Level Diagnostics: On synthetic convex and mildly nonconvex problems, RCML consistently controls multiplier memory with substantially lower dual variation (DualTV) than both violation-based primal–dual baselines and projected-ALM. The residual interface achieves high reliability without the oscillatory or sticky behaviors typical of alternatives.
• Tolerance-Aware Stochastic Resource Allocation: On reserve allocation tasks admitting constraint violation within engineering tolerances, RCML variants achieve lower tolerance-excess and improved cost–reliability trade-offs relative to both aggressive feasible baselines and slack-accepting primal or penalty alternatives.
• Nonconvex Solution Discovery: In stochastic pricing-inventory allocation, RCML yields reliable solution discovery (high feasible rate across random initializations and seeds) where violation-driven algorithms, constraint extrapolation, and single-loop nonconvex methods often fail.
• Large-Scale Fair Ranking: When deployed for neural fair ranking under soft exposure constraints, RCML robustly satisfies admissible fairness levels with utility (NDCG@10) competitive with, or exceeding, ALM, standard SGDA, penalty, and barrier baselines, and lower tolerance violation.

Across all experiments, RCML's finite-gain mechanism stabilizes multiplier dynamics and improves feasibility control. In high-noise and heterogeneous-scale regimes, the importance of constraint filtering and adaptive scaling is also empirically validated.

Implications and Outlook

The RCML framework introduces a feedback-structured, interpretable, and modular mechanism for updating multipliers in large-scale stochastic constrained problems. By splitting the dual update into projected pressure and finite-gain memory, RCML avoids both the variance amplification of ALM-style substitution and the slow memory clearance of sign- or positive-violation integration.

Theoretical Implications:

• RCML offers a principled interface for analysis, allowing the study of stochastic convergence, stability, and trade-offs between tracking lag and noise robustness in a unified framework.
• The results clarify the role of finite gains and structured feedback in maintaining primal–dual stability, providing guidance on algorithm design and hyperparameter selection in practical settings.

Practical Implications:

• RCML is broadly applicable in scenarios where constraint estimates are obtained from sampled or mini-batch data, and statistical feasibility is non-negotiable (e.g., fairness, safety, resource planning).
• The stabilization modules (filtering, adaptive scaling, PI-correction) can be selectively activated based on data or task geometry, supporting automatic or adaptive controller deployment.

Limitations and Future Directions:

• The method introduces extra hyperparameters related to gain, filtering, and scaling; practical efficacy may depend on tuning, although hierarchical guidance is provided.
• The approach relies on passive stabilization near non-smooth projection boundaries (active-set kinks); sharper noise robustness could be achieved by incorporating margin-aware or adaptive projection schemes.
• Extension to global nonconvex convergence and full automation of hyperparameter selection are promising directions, as is integration with variance-reduced or robust sampling pipelines.

Conclusion

Residual-Controlled Multiplier Learning (RCML) advances the stochastic constrained optimization toolkit by coupling projected pressure-based primal updates with a feedback-structured, finite-gain-controlled memory mechanism. RCML's theoretical guarantees extend the paradigm of projected augmented Lagrangian to stochastic, high-variance regimes, and its empirical advantages are demonstrated across convex, nonconvex, and neural settings. This work sets a foundation for feedback-oriented constrained learning under realistic data-driven settings and calls for deeper studies into robust, adaptive dual control in large-scale, noisy environments.

---

Reference:

"Residual-Controlled Multiplier Learning for Stochastic Constrained Decision-Making" (2606.07088)