---
title: Predictive Feasibility Value Functions
url: https://www.emergentmind.com/topics/predictive-feasibility-value-functions
type: topic
---

# Predictive Feasibility Value Functions

Predictive feasibility value functions (PFVFs) are mathematical constructs that quantify the extent to which a system state and input plan can satisfy future constraints, as predicted by an underlying model or optimization oracle. PFVFs arise in advanced control, reinforcement learning, and decision-focused predictive analytics, serving as a critical interface for safety, modularity, and policy interpretability. The modern theory unifies polyhedral, convex, and semi-algebraic structures for value function feasibility across constrained stochastic control, model predictive control (MPC), and partially observable Markov decision processes (POMDPs).

## 1. Foundational Definitions of Predictive Feasibility Value Functions

The PFVF is generally a function $h^*$ that, for a given system state $x$ and control plan $r$, returns the minimal possible cumulative violation of operational constraints that would result from following $r$ starting from $x$ over a future time horizon. Formally, in the context of soft-constrained MPC, for $x(k)\in\mathbb{R}^{n_x}$ and reference sequence $r_{\cdot|k}^{\mathrm{H}}\in\mathbb{R}^{N^{\mathrm{H}}n_r}$, the PFVF is defined as

$$
h^*(x(k), r_{\cdot|k}^{\mathrm{H}}) := \min_{x_{\cdot|k}, u_{\cdot|k}, \xi_{\cdot|k}} \sum_{l=0}^{N(k)-1} \|\xi_{l|k}\|_1
$$

subject to system, control, and relaxed (via slack variables $\xi$) constraints. When $h^*=0$, the trajectory is exactly feasible; $h^*>0$ quantifies the magnitude of inevitable constraint violation [2504.12036].

In the infinite-horizon discrete-time (PO)MDP case, the set of feasible value functions $V$ is the set of all $V^\pi\in\mathbb{R}^{|S|}$ such that there exists a memoryless stochastic policy $\pi: O\rightarrow\Delta A$ realizing $V^\pi$, i.e.,

$$
V^{\pi} = (I-\gamma P^\pi)^{-1} r^\pi, \quad \text{where } P^\pi, r^\pi \text{ depend on } \pi.
$$

The predictive feasibility question is, for a candidate value vector $x$: does there exist a $\pi$ such that $x = V^\pi$? The answer depends critically on the underlying system structure [2606.03048].

## 2. Semi-Algebraic and Polyhedral Geometry of Feasibility Sets

The geometry of the predictive feasibility set is governed by system observability, model structure, and constraint couplings:

- **MDP case**: For fully observable Markov decision processes, the feasible set $V$ of value functions forms a polytope in $\mathbb{R}^S$. Explicitly, 
  $$
  x_s \in \mathrm{conv}\{ Q_a^x(s) : a \in A \}, \quad \forall s,
  $$
  where $Q_a^x(s) = r(s,a) + \gamma\sum_{s'}\alpha(s,a;s')x_{s'}$ are the action-value mappings. Thus, feasibility is determined by $2|S|$ linear inequalities—one lower and one upper bound per state [2606.03048].

- **POMDP case**: Under partial observability, the set $V$ is not polyhedral but semi-algebraic, defined as the solution set to a system of polynomial (degree up to $|O|\cdot|A|$) equalities and inequalities arising from the elimination of policy variables in the Bellman equations. Given the augmented Bellman matrix $C(x)$ (encoding both dynamics and normalization constraints), $x\in V$ if and only if specific minors vanish, certain determinants are nonzero, and a finite family of sign conditions is satisfied. Thus, predictive feasibility requires solving a system of polynomial inequalities (i.e., checking membership in a finite union of basic semi-algebraic sets) [2606.03048].

## 3. Algorithmic Approaches for Testing Predictive Feasibility

Two principal classes of algorithms have emerged:

- **Necessary-condition tests**: Enclosure-type tests (e.g., Oettli–Präger bounds) provide necessary conditions via piecewise-linear inequalities; for MDPs these are also sufficient, but for POMDPs, due to nonlinearity induced by observations, they are only necessary [2606.03048].

- **Full semi-algebraic checks**: For complete feasibility verification in general POMDPs, one must enumerate suitable $(\rho, I,B)$ index sets for submatrices of the augmented Bellman matrix, compute associated minors and determinants, and verify sign conditions—an NP-hard problem in general since it reduces to checking the solvability of bilinear systems [2606.03048].

Alternative approaches use function approximation to sidestep explicit feasibility computations—especially in high-dimensional, computationally constrained, or confidential settings. For example, in hierarchical control, neural networks are trained to regress the PFVF over sampled state–plan input pairs, using exact solutions of the underlying optimization as supervision [2504.12036].

## 4. Predictive Feasibility in Hierarchical and Learning-Based Control

PFVFs are crucial in structuring contract-based interfaces for hierarchical control. In such scenarios:

- **Lower-level controllers** implement soft-constrained MPC, wherein constraint violations are penalized via slack variables.
- **Higher-level planners** utilize the PFVF to assess, for a candidate plan, the minimal unavoidable constraint violation in the lower level.
- **Contract enforcement**: The higher level can either penalize or hard-constrain future reference plans based on the predicted feasibility (i.e., whether the PFVF is zero) [2504.12036].

Function-approximated PFVFs, typically realized as neural networks trained offline, enable efficient and confidential feasibility checking—allowing the lower level to keep its model private while the higher level works only with the PFVF surrogate [2504.12036].

In deep reinforcement learning, the value-shaped running cost (informed by PFVF estimates) serves to lubricate optimization in the presence of sparse or binary rewards and facilitates importance-sampling interpretations of MPC actors, efficiently shaping policies toward feasible and performant domains [1910.03358].

## 5. Predictive Feasibility and Control Barrier/Lyapunov Functions

Recent theoretical developments establish a formal bridge between predictive feasibility value functions, Control Barrier Functions (CBFs), and Control Lyapunov Functions (CLFs):

- **Barrier interpretation**: The value function of a finite-horizon or infinite-horizon MPC, regarded as a CBF, encodes the forward-invariant set of all initial states from which constraints can be met over the horizon—i.e., the predictive feasible set [2502.08400, 2509.22422].
- **CLF–CBF compatibility**: If the terminal MPC cost and set are "well-behaved," the value function and feasible set form a compatible CLF/CBF pair, guaranteeing both stability and constraint satisfaction under appropriate feedback [2509.22422].
- **Approximation robustness**: Even if the value function or feasible set is only approximately known (e.g., from neural regression), theoretical guarantees persist as long as the terminal decrease and invariance conditions hold approximately [2509.22422].

## 6. Learning Predictive Feasibility in Decision-Focused Prediction

The notion of predictive feasibility extends naturally to constrained optimization with uncertain or learned parameters:

- **Decision-focused learning** frameworks balance suboptimality and feasibility through composite loss functions—penalizing both infeasibility (the PFVF being nonzero) and value loss—using softplus surrogates for constraint violations [2510.04951].
- **Trade-off parameterization**: A single tunable parameter $\alpha$ interpolates between strict feasibility (conservative, high regret) and optimality (ambitious, possibly infeasible), enabling practitioners to select the desired point on the feasibility–optimality Pareto frontier [2510.04951].

Empirical results indicate that mid-range $\alpha$ values can simultaneously match or outperform baselines on both regret and feasibility, underscoring the practical utility of explicit predictive feasibility losses in machine-learned decision systems [2510.04951].

## 7. Structural Properties and Geometric Insights

The feasible value function set $V$ in POMDPs exhibits several distinctive geometric properties:

- **Curved, nonconvex boundaries**: Unlike the polyhedral MDP case, partial observability induces genuinely nonlinear, curved feasibility boundaries (semi-algebraic sets), often leading to multiple local maximizers for linear reward objectives.
- **Isolated maximizers and non-uniqueness**: Linear objectives $J_\rho(x)$ can admit multiple, belief-dependent isolated maxima on $\partial V$, resulting in local optima sensitive to the initial belief distribution $\rho$ [2606.03048].
- **Belief sensitivity and policy bifurcation**: Variation of $\rho$ can alter the identity and multiplicity of optimal value functions, causing discontinuous changes in the optimal policy—a qualitative phenomenon absent in fully observable MDPs [2606.03048].
- **Implications**: This complex structure explains the observed challenges in policy optimization and convergence in POMDPs and underscores the necessity of explicit feasibility checks and statistical surrogates in learning-based settings.

---

**References:**

- [2606.03048] "The Value Function Semi-Algebraic Set in Partially Observable Markov Decision Processes"
- [2504.12036] "Contract-based hierarchical control using predictive feasibility value functions"
- [2502.08400] "Predictive Control Barrier Functions: Bridging model predictive control and control barrier functions"
- [1910.03358] "Deep Value Model Predictive Control"
- [2509.22422] "Safe-by-Design: Approximate Nonlinear Model Predictive Control with Real Time Feasibility"
- [2510.04951] "Feasibility-Aware Decision-Focused Learning for Predicting Parameters in the Constraints"

Source: https://www.emergentmind.com/topics/predictive-feasibility-value-functions