Dual-Regularized Riccati Recursions for Interior-Point Optimal Control

Basic Information

• Paper ID: 2509.16370
• Title: Dual-Regularized Riccati Recursions for Interior-Point Optimal Control
• Authors: João Sousa-Pinto, Dominique Orban
• Classification: math.OC cs.MS cs.RO cs.SY eess.SY
• Publication Date: October 15, 2025 (arXiv v2)
• Paper Link: https://arxiv.org/abs/2509.16370

Abstract

This paper derives closed-form extensions of Riccati recursions for solving dual-regularized LQR problems, including both sequential and parallel versions. The authors demonstrate how to use these methods to solve general constrained, non-convex, discrete-time optimal control problems via regularized interior-point methods, while guaranteeing that each step is a descent direction for the augmented barrier-Lagrangian function. The paper provides MIT-licensed implementations in C++ and JAX.

Research Background and Motivation

Core Problem

The core problem addressed by this research is how to efficiently solve non-convex discrete-time optimal control problems with equality and inequality constraints. Traditional methods face the following challenges when handling such problems:

1. Computational Efficiency: Standard interior-point methods require solving large-scale linear systems when addressing optimal control problems, resulting in high computational complexity
2. Numerical Stability: When regularization parameters approach zero, conventional methods may exhibit numerical instability
3. Parallelization Difficulty: Existing methods struggle to fully exploit parallel computing resources

Problem Significance

Optimal control problems have broad applications in robotics, aerospace, autonomous driving, and other fields. Efficiently solving such problems is crucial for real-time control systems, particularly in scenarios requiring complex constraint handling.

Limitations of Existing Methods

1. DDP Algorithm: Although the most commonly used method in practice, as a single-shooting method, it cannot independently warm-start state trajectories
2. Standard LQR Methods: Only applicable to unconstrained or simply-constrained linear systems
3. Existing Interior-Point Methods: General solvers such as IPOPT cannot fully exploit the structural properties of optimal control problems

Core Contributions

1. Theoretical Contribution: Derives closed-form Riccati recursion extensions for solving dual-regularized LQR problems, including sequential and parallel versions
2. Algorithmic Innovation: Proposes a regularized interior-point method that guarantees descent directions using the augmented barrier-Lagrangian function as a merit function
3. Numerical Stability: Designs algorithms that remain numerically stable as the regularization parameter δ→0, recovering the standard LQR algorithm
4. Parallel Algorithm: Implements a solution algorithm with O(log N) parallel time complexity based on associative scans
5. Software Contribution: Provides open-source implementations in C++ and JAX supporting efficient sparse linear algebra operations

Methodology Details

Problem Formulation

Consider the discrete-time optimal control problem:

$\min_{x_0,u_0,\ldots,x_N} \sum_{i=0}^{N-1} f_i(x_i, u_i) + f_N(x_N)$

Subject to constraints:

• Initial state: $x_0 = s_0$
• Dynamics: $x_{i+1} = d_i(x_i, u_i), \forall i \in \{0,\ldots,N-1\}$
• Equality constraints: $c_i(x_i, u_i) = 0, \forall i \in \{0,\ldots,N-1\}$
• Inequality constraints: $g_i(x_i, u_i) \leq 0, \forall i \in \{0,\ldots,N-1\}$
• Terminal constraints: $c_N(x_N) = 0, g_N(x_N) \leq 0$

Regularized Interior-Point Method Framework

Augmented Barrier-Lagrangian Function

Define the barrier-Lagrangian function: $L(x,s,y,z;\mu) = f(x) - \mu\sum_i \log(s_i) + y^T c(x) + z^T(g(x) + s)$

Augmented version: $A(x,s,y,z;\mu,\eta) = L(x,s,y,z;\mu) + \frac{\eta}{2}(\|c(x)\|^2 + \|g(x)+s\|^2)$

Linear System Solution

Each iteration requires solving the linear system: