Real-Time Iteration Scheme

• RTI is a numerical algorithm that performs a single local optimization update per sampling instant, ensuring feasibility under strict time constraints.
• It leverages warm-starting, linearization, and a one-shot quadratic programming solve to efficiently approach near-optimal solutions.
• RTI is widely applied in nonlinear model predictive control, quantum simulation, and diffusion policy inference, balancing computation with real-time performance.

A Real-Time Iteration (RTI) scheme is a class of numerical algorithms for solving large-scale, nonlinear, time-dependent optimization and self-consistency problems with tight per-step computational constraints. The defining feature is the execution of exactly one (or a small, fixed number of) local optimization steps—typically a Sequential Quadratic Programming (SQP) update—per sampling instant, with systematic warm-starting from previous iterates or physical time evolution. RTI has become a foundational tool in nonlinear model predictive control (NMPC), robust and distributed MPC, fast trajectory optimization, and, more recently, time-domain many-body quantum simulation and accelerated inference in machine learning. It is particularly effective in applications where large-scale nonlinear programming must be solved with strict real-time deadlines and is thus adopted in both algorithmic theory and embedded/control software stacks.

1. Fundamental Principles and Core Algorithmic Structure

The RTI paradigm replaces full convergence of a nonlinear program (NLP) per time-step with a single local correction evaluated at a shifted or predicted trajectory. In classical optimal control, the underlying discrete-time optimal control problem can be stated (for state $x_k$, input $u_k$):

$$\min_{\{x_k,u_k\}} \sum_{k=0}^{N-1} \ell(x_k,u_k) + V_f(x_N),\quad \text{s.t. } x_{k+1}=f(x_k,u_k),\ h(x_k,u_k)\geq 0,\ x_0=\hat x_0$$

At each sampling instant, RTI executes:

1. Warm-starting: Initialize (shift) the state and control trajectory using the solution from the previous instant.
2. Linearization/Quadratization: Linearize the system dynamics and constraints and (optionally) quadratize the cost around the current trajectory, forming a condensed QP via Gauss–Newton or related Hessian approximations.
3. Quadratic Program Solve: Solve the QP once (full Newton/SQP step, step length $t=1$).
4. Trajectory Update: Update the reference trajectory, apply the new input, and propagate the state forward for the next instant.

This structure contracts toward the locally optimal solution under suitable regularity conditions and enables bounded computation times per sample (Wagner et al., 2024, Wu et al., 2024, Zanon, 2020, Duan et al., 7 Aug 2025).

2. Mathematical Guarantees and Convergence Theory

Convergence analyses of RTI-schemes show that, under standard assumptions—local regularity (LICQ, second-order sufficient conditions), smoothness, boundedness of model and measurement noise—the sequence of iterates converges locally q-linearly to the true solution as the sampling interval $\delta\to0$ and the warm start is adequately close (Wagner et al., 2024, Wu et al., 2024, Stomberg et al., 2024, Zanon et al., 2020, Chen et al., 2018). For distributed or stochastic settings, contractivity of the local map is preserved under mild conditions and guarantees local exponential stability for the closed-loop system.

Recent extensions (AS-RTI, multi-level inexact updates) provide error bounds relating the deviation from exact solution to the number and quality of inner iterations and the accuracy of prediction (Frey et al., 2024). In practical controller deployments, this allows explicit trade-off between computational effort and closed-loop suboptimality. Execution-time certification is feasible using worst-case flop count analysis via Riccati-based IPM factorization (Wu et al., 2024).

3. Algorithmic Variants and Enhancements

Numerous methodological extensions of RTI address computational bottlenecks and robustness:

• Advanced-Step RTI (AS-RTI): Uses predicted future states or outputs to precompute linearization points and QP data, supporting multiple inner updates (A/B/C/D-levels) and further reducing online latency (Frey et al., 2024).
• Partial Sensitivity Updates (CMoN-RTI): Selectively updates Jacobian blocks only where local system nonlinearity is high, controlled by a curvature-like measure, and guarantees bounded suboptimality of the RTI QP (Chen et al., 2018).
• Decentralized RTI: Applies one outer SQP per sample and a fixed number of distributed ADMM updates for large-scale networked systems, preserving local exponential stability with strictly local communications (Stomberg et al., 2024).
• Robust/Stochastic RTI: Augments RTI by alternating nominal and tube-based robustification steps, e.g., via zero-order robust optimization (zoRO) for tube MPC (Frey et al., 2023).
• Sampling Augmentation (SAA-RTI): Warms up the trajectory with sampled feasible candidates to mitigate local minima in constraint adaptation, e.g., for adaptive vehicle control under traction uncertainty (Svensson et al., 2019).
• Real-Time Iteration for Diffusion Inference (RTI-DP): Warm-starts iterative denoising in diffusion policy inference by shifting the last action chunk, yielding order-of-magnitude speedup without distillation (Duan et al., 7 Aug 2025).

4. Domain-Specific Implementations

Nonlinear Model Predictive Control (NMPC)

RTI is the workhorse for embedded NMPC in automotive, robotics, and power systems, enabling real-time feasible updates for horizons $N=10\ldots100$, state/input dimensions up to hundreds, and sub-millisecond feedback (Carlos et al., 2020, Frey et al., 2023). Multiple shooting, banded sparse QP solvers, and partial condensing are used for implementation. Gauss–Newton Hessians enable avoidance of online second-order derivative calculations (Zanon, 2020). Implicit discretizations (e.g., implicit Euler) are deployed for stiff dynamics (Wagner et al., 2024). Execution-time certification is realized using factorized Riccati recursion (Wu et al., 2024).

Distributed and Robust MPC

Decentralized RTI enables scalable and stabilizing distributed control with neighbor-only communication (Stomberg et al., 2024). Tube-based RTI with efficient BLASFEO-backed code generation allows robust NMPC on embedded processors (Frey et al., 2023).

Quantum Simulation and Many-Body Physics

In dynamical mean-field theory (DMFT), a real-time RTI scheme can self-consistently solve for retarded Green’s functions on quantum hardware, mapping the impurity problem to a finite bath chain and iteratively updating the hybridization via time-domain fitting—thus circumventing the need for analytic continuation and unstable inversions endemic to classical methods (Rangi et al., 27 Jan 2026).

Machine Learning: Diffusion Policy Inference

RTI-DP accelerates inference in diffusion policies by leveraging time-history as an initial guess, pruning the iterative procedure from $K_\text{full}\sim200-400$ steps (800–1000 ms latency) to $K\sim3$–10 steps (25–150 ms latency) with negligible performance loss, as proven in contractivity analysis for the reverse diffusion chain (Duan et al., 7 Aug 2025).

5. Numerical Performance and Practical Considerations

RTI demonstrates consistent real-time feasibility:

• Automotive NMPC: Achieves step times of 5–15 ms per iteration for $N=15$ and $\sim$12 states/inputs, with 7–10$u_k$0 speedup over full NLP solvers (Wagner et al., 2024).
• Robust MPC: Drastically reduces the computation for uncertainty-aware problems, e.g., 10$u_k$1 speedup for differential-drive robot control (Frey et al., 2023).
• Multi-agent CAV control: Enables solution of large intersection problems ($u_k$2 up to 600, dozens of vehicles) in under 4 ms, with sub-2.3% loss in optimality (Zhao et al., 2024).
• Quantum/DMFT: 5-site RTI schemes recover full three-peak spectral structures and Mott gaps in DMFT, compatible with $u_k$310–12 qubit devices and time discretization $u_k$4 (Rangi et al., 27 Jan 2026).
• Diffusion Policy Inference: In RL/robotics, RTI-DP yields speedups of 5–30$u_k$5 in cost per chunk, retaining or exceeding the task completion performance of full diffusion policies (Duan et al., 7 Aug 2025).

Table: RTI Computational Performance in Representative Domains

Domain	State/Input Dim	Horizon ($u_k$6)	RTI Step Time	Scaling
Automotive NMPC (Wagner et al., 2024)	$u_k$712	15	15 ms	$u_k$8
Robust MPC, zoRO (Frey et al., 2023)	5/2	20	1.5 ms (robust)	$u_k$9
CAV Coordination (Zhao et al., 2024)	$$\min_{\{x_k,u_k\}} \sum_{k=0}^{N-1} \ell(x_k,u_k) + V_f(x_N),\quad \text{s.t. } x_{k+1}=f(x_k,u_k),\ h(x_k,u_k)\geq 0,\ x_0=\hat x_0$$015–30	50–600	5 ms – 3.6 s	$$\min_{\{x_k,u_k\}} \sum_{k=0}^{N-1} \ell(x_k,u_k) + V_f(x_N),\quad \text{s.t. } x_{k+1}=f(x_k,u_k),\ h(x_k,u_k)\geq 0,\ x_0=\hat x_0$$1 (sparse QP, OSQP)
DMFT real-time (Rangi et al., 27 Jan 2026)	Bath sites = 5	$$\min_{\{x_k,u_k\}} \sum_{k=0}^{N-1} \ell(x_k,u_k) + V_f(x_N),\quad \text{s.t. } x_{k+1}=f(x_k,u_k),\ h(x_k,u_k)\geq 0,\ x_0=\hat x_0$$2	Quantum gate depth	$$\min_{\{x_k,u_k\}} \sum_{k=0}^{N-1} \ell(x_k,u_k) + V_f(x_N),\quad \text{s.t. } x_{k+1}=f(x_k,u_k),\ h(x_k,u_k)\geq 0,\ x_0=\hat x_0$$3
Diffusion Policy (Duan et al., 7 Aug 2025)	Action Chunks	$$\min_{\{x_k,u_k\}} \sum_{k=0}^{N-1} \ell(x_k,u_k) + V_f(x_N),\quad \text{s.t. } x_{k+1}=f(x_k,u_k),\ h(x_k,u_k)\geq 0,\ x_0=\hat x_0$$4	25–150 ms (RTI-DP)	$$\min_{\{x_k,u_k\}} \sum_{k=0}^{N-1} \ell(x_k,u_k) + V_f(x_N),\quad \text{s.t. } x_{k+1}=f(x_k,u_k),\ h(x_k,u_k)\geq 0,\ x_0=\hat x_0$$5 steps per chunk

6. Advantages, Limitations, and Trade-offs

Advantages:

• Enables MPC-quality control and self-consistency calculations on embedded, distributed, or quantum hardware.
• Warm-starting and partial update variants exploit temporal coherence for maximum efficiency.
• Theoretical local stability and contractivity available under standard conditions.
• Empirical studies substantiate order-of-magnitude reductions in wall-clock latency with <2% loss of optimality across domains (Zhao et al., 2024, Duan et al., 7 Aug 2025).

Limitations:

• Convergence is only guaranteed when the initialization remains sufficiently close to the local optimum; failure to detect and recover from poor initial guesses may cause performance degradation.
• Features at fine spectral or geometric scales (e.g., narrow metallic bands in DMFT at low $$\min_{\{x_k,u_k\}} \sum_{k=0}^{N-1} \ell(x_k,u_k) + V_f(x_N),\quad \text{s.t. } x_{k+1}=f(x_k,u_k),\ h(x_k,u_k)\geq 0,\ x_0=\hat x_0$$6 (Rangi et al., 27 Jan 2026), sharp decision boundaries in NMPC near constraint activation) may require more iterations or finer discretizations, which then stress hardware resources.
• For highly nonconvex/process-limited or discrete decision problems, integration with augmented sampling, identification, or learning layers is necessary to achieve global performance (Svensson et al., 2019).

7. Cross-Domain Impact and Future Directions

RTI algorithms have become standard in open-source MPC and optimization frameworks (e.g., acados, HPIPM+BLASFEO (Frey et al., 2023, Carlos et al., 2020)), quantum simulation workflows (Rangi et al., 27 Jan 2026), and control-oriented policy inference in robotics (Duan et al., 7 Aug 2025). Current and future trends center on:

• Rigorous worst-case analysis for certifiable real-time control (Wu et al., 2024).
• Seamless integration of RTI loops with machine learning agents, online identification, and robustification layers (Duan et al., 7 Aug 2025, Zanon et al., 2020).
• Adaptive meta-heuristics for trajectory sampling and constraint selection (Svensson et al., 2019).
• Scalable, parallel and distributed implementations for networked dynamical systems (Stomberg et al., 2024).
• Application to emerging hardware-limited domains such as near-term quantum devices and resource-constrained robotics (Rangi et al., 27 Jan 2026, Duan et al., 7 Aug 2025).

RTI scheme thus provides a versatile and computationally efficient backbone across modern computational control, optimization, simulation, and machine learning applications, systematically trading off iteration depth against real-time feasibility while retaining stability, performance, and extensibility.