Variance-Reduced SLQ

• The paper introduces variance-reduced SLQ algorithms that use SVRG-type estimators within sequential quadratic programming to achieve faster and more robust convergence in constrained problems.
• It applies convex optimization techniques to engineer noise covariance in SLN simulations, significantly reducing statistical variance in quantum dynamic models.
• Empirical evaluations show 10–100× improvements in feasibility and stationarity errors, highlighting scalability for large-scale and long-time simulations.

Variance reduction is a core principle for increasing the statistical efficiency of stochastic optimization and simulation algorithms. In the context of Stochastic Sequential Quadratic Programming (SQP) for constrained optimization (“SVR-SQP”) and Stochastic Liouville–von Neumann (SLN) equation simulations for open quantum systems, variance-reduced approaches have been shown to dramatically improve convergence, robustness, and scalability. This article provides a comprehensive examination of variance-reduced algorithms specifically as applied to SLQ (Sequential Quadratic Programming) and SLN frameworks, detailing problem formulations, algorithmic strategies, theoretical properties, and practical outcomes as established in the literature (Berahas et al., 2022, Schmitz et al., 2018).

1. Problem Setting and Motivation

Variance-reduced SLQ methods address the challenge of solving large-scale, smooth finite-sum, equality-constrained problems of the form

$$\min_{x\in\mathbb{R}^n} f(x) = \frac{1}{N} \sum_{i=1}^N f_i(x), \quad \text{subject to } c(x) = 0,~c:\mathbb{R}^n\to\mathbb{R}^m,$$

where $f_i$ and $c$ are differentiable with Lipschitz gradients. Traditional stochastic SQP methods for such settings often suffer from slow or unstable convergence due to high variance in stochastic gradient estimates, especially as optimization proceeds close to stationary points or constraints become more restrictive. Variance-reduction techniques, such as SVRG-type estimators, have been introduced to control and diminish this variance while maintaining scalability in large $N$ regimes (Berahas et al., 2022).

In the quantum simulation context, the SLN equation models the reduced dynamics of quantum systems coupled to Gaussian environments. Here, the trajectory ensemble’s variance grows rapidly due to nonunitary terms associated with complex noise processes, creating significant computational obstacles. Variance-reduction through convex optimization of noise correlations mitigates this growth, improving sampling efficiency (Schmitz et al., 2018).

2. Variance-Reduced Stochastic SQP: Algorithmic Framework

The SVR-SQP (Stochastic Variance-Reduced Sequential Quadratic Programming) algorithm is based on a nested two-loop structure:

• Outer Loop ($k$): Selects a reference point $x_{k,0}$ and computes the full gradient $g_{k,0} = \nabla f(x_{k,0})$.
• Inner Loop ($s$): Constructs variance-reduced gradient estimates at $x_{k,s}$ based on mini-batches.

At each $(k, s)$, the following subproblem is formulated:

$$\min_{d\in\mathbb{R}^n} \tilde{g}_{k,s}^\top d + \frac{1}{2} d^\top H_{k,s} d,\quad \text{subject to } c_{k,s} + J_{k,s} d = 0,$$

where $H_{k,s}$ is a positive-definite Hessian approximation on the nullspace of the constraint Jacobian $J_{k,s}$ (Berahas et al., 2022).

The variance-reduced estimator is constructed as:

$$\tilde{g}_{k,s} = \bar{g}_{k,s} - \bar{g}_{k,0} + g_{k,0},$$

with $\bar{g}_{k,s}$ and $\bar{g}_{k,0}$ being mini-batch gradients at $x_{k,s}$ and $x_{k,0}$, respectively. This estimator is unbiased and its variance is controlled by the distance $\|x_{k,s} - x_{k,0}\|$ and batch size $b$.

The complete SVR-SQP algorithm incorporates:

• Adaptive or constant step-size selection;
• Nullspace-projected Hessians;
• Merit-parameter updates for penalty-based handling of equality constraints.

3. Convergence Analysis and Theoretical Guarantees

Under the assumptions of Lipschitz continuity, uniform LICQ, and positive-definite Hessian structure on constraint nullspaces, the SVR-SQP algorithm exhibits the following properties:

• Unbiased gradient estimates: $E[\tilde{g}_{k,s}|x_{k,s}] = \nabla f(x_{k,s})$.
• Variance control: $$E[\|\tilde{g}_{k,s} - \nabla f(x_{k,s})\|^2] \leq (L^2/b)\|x_{k,s} - x_{k,0}\|^2$$, diminishing as iterates cluster near the reference point.
• Global expectation convergence: The expected first-order stationarity and feasibility measure $M(x) = \|g(x) + J(x)y\|^2 + \|c(x)\|$ converges to zero at rate $O(1/(KS))$ for both constant and adaptive step-size strategies (Berahas et al., 2022).

The analysis employs a Lyapunov function involving the $\ell_1$ penalty merit function and a quadratic term in iterate distance. Telescoping arguments over the inner and outer loops yield the non-vanishing step-size convergence rate.

4. Implementation and Empirical Performance

SVR-SQP was implemented and evaluated on constrained logistic regression with random linear constraints and norm equality constraints, across several LIBSVM datasets (with $N$ ranging from hundreds to $50,000$). Key implementation parameters included mini-batch sizes $b \in \{16, 128\}$ and inner loop length $S \approx N/(2b)$.

Performance benchmarks include Sto-SQP (without variance reduction) and Sto-Subgrad-VR (SVRG on the merit function). After 30 epochs, SVR-SQP with adaptive step-size (β=1, no tuning) achieved feasibility and stationarity errors that were 10–100× smaller than Sto-SQP on small batches, and exceeded or matched Sto-SQP on large batches. Sto-Subgrad-VR was consistently outperformed. The algorithm requires at most one full gradient per outer pass, effectively controlling the overall stochastic gradient computation load (Berahas et al., 2022).

5. Variance Reduction in Stochastic Liouville–von Neumann Simulations

In quantum system simulations, the SLN equation governs the dynamics of a system coupled to a Gaussian reservoir, utilizing stochastic samples of complex noise processes. The naive SLN direct sampling experiences exponential variance growth in strong-coupling or long-time regimes, traceable to nonunitary evolution driven by imaginary noise components.

A convex optimization method for noise covariance engineering minimizes the spectral power in these nonunitary (imaginary) noise components, subject to matching the required noise correlations for exact SLN dynamics and maintaining positive semidefiniteness. This is formulated, for each frequency $\omega$, as a constrained semidefinite program:

$$\min_{C(\omega)\succeq 0} ~ C_{\xi''\xi''}(\omega) + C_{\nu''\nu''}(\omega) ~~\text{subject to noise correlation constraints }.$$

Block-diagonal structure and closed-form solutions can be exploited in practice. The methodological steps include precomputing reservoir kernels, solving the semidefinite program per frequency, FFT-based noise synthesis, propagation of the SLN equation, and ensemble averaging (Schmitz et al., 2018).

6. Numerical Results and Scaling

Empirical evaluation in the SLN context, using the spin-boson model with Ohmic spectral density, demonstrates significant improvements in variance scaling and computational cost. At moderate to strong dissipation and finite temperature, the optimized covariance approach reduced the number of samples required to achieve a given statistical precision by several orders of magnitude. For example, at $\beta\Delta = 5$ and $t=40$, variance growth rates and sample variances were decreased by $\sim 2\times$ to $>5\times$ compared to unoptimized schemes, and the number of required samples dropped accordingly. The exponential scaling persisted in the extreme strong-coupling limit, but with much reduced prefactors (Schmitz et al., 2018).

	var-growth rate (no VR)	var-growth rate (VR)	N₄₀ (no VR)	N₄₀ (VR)
βΔ = 5	0.20	0.10	$3\times10^7$	$5.2\times10^6$
βΔ = 2	0.086	0.047	$1.8\times10^5$	$1.8\times10^5$
βΔ = 1	0.055	0.032	$5.5\times10^4$	$5.5\times10^4$
βΔ = 0.5	0.025	0.013	$1.7\times10^4$	$1.7\times10^4$

This suggests that for physically relevant parameter regimes, variance-reduced SLN techniques enable long-time and strong-coupling simulations that would otherwise be computationally prohibitive (Schmitz et al., 2018).

7. Broader Significance and Research Directions

Variance-reduced SLQ formulations, exemplified by SVR-SQP for constrained optimization and the convex covariance approach for SLN, establish a robust foundation for scalable stochastic computation in both classical machine learning and quantum dynamics. In both domains, the primary advantage is the decoupling of sample complexity from problem size or integration time via improved variance control, enabling the achievement of high statistical accuracy with practical computational resources.

Further research may focus on:

• Extending variance-reduction frameworks to more general constraint structures or nonconvex objectives in SQP.
• Adapting covariance optimization techniques to nonlinear or non-Gaussian quantum environments.
• Investigating hybrid stochastic-deterministic architectures leveraging variance reduction within parallel and distributed settings.

A plausible implication is that future stochastic algorithms for optimization and simulation will increasingly leverage variance-reduction as a standard component, especially in large-scale, high-precision settings.

---

References:

• [Accelerating Stochastic Sequential Quadratic Programming for Equality Constrained Optimization using Predictive Variance Reduction, (Berahas et al., 2022)]
• [A Variance Reduction Technique for the Stochastic Liouville-von Neuman Equation, (Schmitz et al., 2018)]