Feedback Refinement Relation (FRR)

• Feedback Refinement Relation (FRR) is a formal framework that establishes matching constraints between complex systems and symbolic abstractions, ensuring proper controller refinement.
• It is applied in symbolic control synthesis, nonlinear and time-delay control systems, and LLM feedback pipelines to achieve provable system performance under quantization.
• FRR enforces key conditions like input compatibility and successor preservation, allowing systematic construction of transition systems with guaranteed behavioral inclusion.

A Feedback Refinement Relation (FRR) is a formal relational framework used to connect a complex system, often with continuous or high-dimensional dynamics, to a symbolic, quantized, or otherwise simplified abstraction, such that controllers designed for the abstraction provably refine to correct, robust controllers for the original system. FRRs ensure this by enforcing matching constraints on input admissibility and transition preservation, typically within transition systems or control frameworks. The concept originated in symbolic control synthesis, but has since found application in domains including nonlinear control systems—both delay-free and time-delay—and modern LLM feedback-based refinement pipelines. In all contexts, FRR rigorously characterizes the conditions under which structure-preserving abstraction and systematic refinement are possible.

1. Formal Definition and Core Properties

Let $S_1 = (X_1, U_1, F_1)$ and $S_2 = (X_2, U_2, F_2)$ be two simple systems or transition systems, with $U_2 \subseteq U_1$ (often $S_1$ is the "plant" and $S_2$ the abstraction). A relation $\mathcal R \subseteq X_1 \times X_2$ is a feedback refinement relation if for every $(x_1, x_2)\in\mathcal R$:

1. Input compatibility:

$U_{S_2}(x_2) \subseteq U_{S_1}(x_1)$

2. Successor preservation: For all $u\in U_{S_2}(x_2)$ and all $x_1'\in F_1(x_1, u)$,

$S_2 = (X_2, U_2, F_2)$0

This yields $S_2 = (X_2, U_2, F_2)$1. FRRs are reflexive and transitive, and most importantly, they guarantee behavioral inclusion: for any controller $S_2 = (X_2, U_2, F_2)$2 for $S_2 = (X_2, U_2, F_2)$3, the refined controller $S_2 = (X_2, U_2, F_2)$4 for $S_2 = (X_2, U_2, F_2)$5 preserves the traces and specifications that can be established in $S_2 = (X_2, U_2, F_2)$6 (Reissig et al., 2015).

Necessity and sufficiency theorems show $S_2 = (X_2, U_2, F_2)$7 is an FRR if and only if all controllers for $S_2 = (X_2, U_2, F_2)$8 refine correctly to $S_2 = (X_2, U_2, F_2)$9, making FRR the exact relational condition for abstraction-refinement correctness.

2. FRR in Symbolic Abstraction and Control Synthesis

In abstraction-based controller synthesis, FRR serves as the bridge that enables controllers to be synthesized on a finite, quantized (symbolic) abstraction and deployed on the original system with provable correctness. This is critical for practical controller realization when only coarse, quantized, or noisy state observations are available.

Canonical abstraction builds a symbolic model whose states are the nonempty fibers (preimages) of the abstraction map; every FRR factors through this model. Algorithmic construction proceeds by enumerating partition cells and computing transition relations covering the effect of quantization and system evolution (Reissig et al., 2015).

Applications include:

• Nonlinear and time-delayed systems: Symbolic abstractions of nonlinear and time-delay systems via FRR employ logarithmic and zoom quantizers to manage state explosion and focus refinement where needed. The abstract transition system is tractably constructed, and controller synthesis followed by symbolic refinement handles both delay-free and time-delay cases, yielding robustness to quantization error and uncertainty (Ren et al., 2020).
• Robust control: FRRs extend to control in the presence of disturbances, guaranteeing robust performance. Perturbations are absorbed into the transition system, and refinement proceeds analogously (Reissig et al., 2015).

3. Hierarchy and Comparison: FRR, MCR, and ASR

FRR is one element in a hierarchy of relational abstractions:

• Alternating Simulation Relation (ASR): The most permissive; allows the greatest modeling flexibility but does not support direct controller refinement unless restrictive conditions hold.
• Memoryless Concretization Relation (MCR): Intermediate; relaxes the constraint that controller actions be cell-constant, allowing richer, state-dependent policies within partition cells.
• Feedback Refinement Relation (FRR): The most restrictive; requires that abstract and concrete systems share admissible inputs, and that the feedback be "piecewise constant" on quantized cells, preventing state-dependent variation within cells (Calbert et al., 2024).

The strict chain is:

$U_2 \subseteq U_1$0

FRR thus supports the simplest symbolic interface, but at the cost of prohibiting within-cell state dependence in controllers.

Relation	Input Matching	Controller Interface	Symbolic Complexity
FRR	$U_2 \subseteq U_1$1	Piecewise constant on cells	Minimal
MCR	$U_2 \subseteq U_1$2 abstract	State-dependent within cells	Higher
ASR	Most permissive	Stateful	Highest

MCR allows for abstraction completion (transition "enrichment") to meet its conditions, increasing non-determinism. FRR cannot exploit inside-cell measurements, unlike MCR (Calbert et al., 2024).

4. FRR in Feedback-Driven LLM Refinement

FRR principles have generalized beyond control theory into structured LLM output refinement. In this context, an FRR models the mapping from a raw model output and feedback to an improved output that better satisfies external criteria—e.g., factuality, faithfulness, or other dimensions (Wadhwa et al., 2024, Yun et al., 27 Mar 2025).

• DCR pipeline (Wadhwa et al., 2024): Decomposes refinement into DETECT (error localization via discriminative model), CRITIQUE (fine-grained NL feedback), and REFINE (output correction via seq2seq LLM). This tripartite factoring improves factual alignment and avoids over-correction.
• Multi-dimensional FRR (Yun et al., 27 Mar 2025): Generalizes the relation to $U_2 \subseteq U_1$3, where $U_2 \subseteq U_1$4 refines $U_2 \subseteq U_1$5 according to multi-objective feedback, respecting trade-offs (faithfulness, completeness, conciseness). Reflective reasoning and robust validation steps are built into the refinement operator.

These feedback-refinement pipelines exploit the mathematical discipline of FRR to systematically align LLM-generated content to external evaluators' feedback, and have demonstrated empirical superiority to direct or one-step baselines.

5. Algorithmic Construction and Pseudocode

Algorithmic instantiation of FRR appears in both control and LLM contexts. A typical abstraction process for continuous/discrete dynamical systems includes:

• Quantization (static/logarithmic, dynamic/zoom, spline for histories)
• Construction of transition relations respecting over-approximation of dynamics and quantization error
• Establishment of the inclusion relation $U_2 \subseteq U_1$6 mapping original to abstract states (Ren et al., 2020)

Example pseudocode in control (Ren et al., 2020):

$U_2 \subseteq U_1$7

In LLM feedback-driven refinement (Wadhwa et al., 2024):

$U_2 \subseteq U_1$8

6. Empirical Metrics and Case Studies

In the context of control, FRR-based abstraction and controller refinement are validated by trace inclusion and robust specification satisfaction, demonstrated in high-dimensional vehicle and aircraft maneuvering tasks (Reissig et al., 2015). Symbolic controllers synthesized via FRR operate with quantized state information and can be realized with static quantizers.

In LLM pipelines, metrics include alignment scores, factuality Likert ratings, and discriminative win rates, all showing improvement via FRR-based multi-step feedback-refinement. Ablations in ReFeed document superior robustness to feedback noise, order variation, and trade-offs between objectives compared to naive or sequential baselines (Yun et al., 27 Mar 2025, Wadhwa et al., 2024).

7. Limitations and Extensions

FRR abstraction is weaker than bisimulation, not requiring strong stability assumptions. However, symbolic controllers built via FRR are limited to cell-wise constant feedback and may be conservative in highly nonlinear or complex state spaces. MCR-based approaches can exploit within-cell state dependence for more expressive controllers, though at the cost of increased model complexity (Calbert et al., 2024). In LLM domains, the practicality of reflective feedback-refinement is contingent on high-quality automated feedback and accurate error localization.

Potential advancements include compositional abstractions for networks of systems, automated quantizer parameter tuning, probabilistic FRR generalizations for stochastic systems, and integration with sophisticated feedback evaluators in LLM-driven pipelines (Ren et al., 2020, Yun et al., 27 Mar 2025).