Piecewise Multi-Linear Representation

• Piecewise Multi-Linear Representation is a framework for modeling multivariate functions by partitioning the domain into hyperrectangular cells where each function is affine.
• It employs a tensor-product construction of canonical bases, ensuring efficient computation, exact interpolation, and straightforward differentiation via block-Kronecker mechanics.
• PMLR is applied in control systems, optimization, and neural network design to achieve zero-interpolation error and real-time performance in complex environments.

A piecewise multi-linear representation (PMLR) is a structural framework for modeling, interpolating, and computing with multivariate functions that are linear in each input variable when restricted to subdomains, with sharp transitions ("kinks") at prescribed breakpoints. PMLR generalizes the concept of canonical piecewise-linear models to higher dimensions via a tensor-product construction, enabling exact, computationally efficient representation and manipulation of functions critical in control, optimization, approximation, and neural network design. It has particular importance in nonlinear control allocation, as it provides zero-interpolation-error representations of complex aerodynamics or utility functions, and underpins efficient convex hull formulations or neural architectures for both theoretical and practical applications (Rajput et al., 2022).

1. Canonical Construction of Piecewise Multi-Linear Representations

Any continuous piecewise-linear function of one variable can be written uniquely as

$$f(z) = \gamma\,\hat{z},\qquad \hat{z} = [1,\ z,\ |z-\mu(2)|,\ \dotsc,\ |z-\mu(L-1)|]$$

where $\gamma\in\mathbb{R}^L$ and breakpoints $\{\mu(1),\mu(2),\dotsc,\mu(L)\}$ define the subdomains. The first two terms produce a global affine baseline; absolute value terms encode the transitions at breakpoints (Rajput et al., 2022).

For higher dimensions, if $g(z_1,\dots,z_k)$ is multi-linear in each input (on $\prod_j (L_j-1)$ cell grid), PMLR constructs its representation by tensorizing the 1D bases:

$$g(z_1,\ldots,z_k) = \Gamma\,( \hat{z}_1 \otimes \hat{z}_2 \otimes \cdots \otimes \hat{z}_k )$$

with $\Gamma$ a coefficient tensor and $\hat{z}_j$ the canonical basis for the $j$-th variable. This ensures that $g$ is affine (multi-linear) in each variable within its cell, and matches exactly on the specified rectilinear grid (Rajput et al., 2022).

2. Block-Kronecker and Kronecker Product Mechanics

The multi-dimensional structure of PMLR necessitates generalizations of the classic Kronecker product. The block-Kronecker product, denoted $\gamma\in\mathbb{R}^L$0, is defined for a block-partitioned $\gamma\in\mathbb{R}^L$1 and a block $\gamma\in\mathbb{R}^L$2 as

$\gamma\in\mathbb{R}^L$3

preserving the nested structure critical to efficient matrix assembly in high-dimensional cases (Rajput et al., 2022). Using associativity and bilinearity, the block-Kronecker allows recursive lifting of coefficient tensors across dimensions and underpins recursive evaluation, differentiation, and efficient model fitting.

Equivalently, the full model can be written using the standard Kronecker product:

$\gamma\in\mathbb{R}^L$4

noting that both formulations are mathematically equivalent.

3. Differentiation and Analytical Jacobians

Because PMLR models are affine with respect to the canonical basis vectors, computation of derivatives is efficient and explicit. The partial derivative with respect to $\gamma\in\mathbb{R}^L$5 is:

$\gamma\in\mathbb{R}^L$6

where

$\gamma\in\mathbb{R}^L$7

This structure allows for closed-form, efficient evaluation of gradients, enabling real-time feedback and optimization in embedded and control systems (Rajput et al., 2022).

4. Exactness and Grid Coverage Properties

PMLR achieves exact multilinear interpolation on rectilinear grids: if the data at the grid points comes from linear interpolation, then PMLR recovers the function values with zero error. The domain is partitioned into $\gamma\in\mathbb{R}^L$8 hyperrectangular "cells" where the function is uniquely affine. There is no approximation error and no Runge phenomenon, which means that PMLR avoids the oscillatory interpolation artifacts possible with high-degree polynomials (Rajput et al., 2022).

5. Learning and Fitting PMLR from Data

Given data at all grid points, the coefficient tensor $\gamma\in\mathbb{R}^L$9 can be recovered without large matrix inversion. If the 1D basis matrices $\{\mu(1),\mu(2),\dotsc,\mu(L)\}$0 are formed by evaluating $\{\mu(1),\mu(2),\dotsc,\mu(L)\}$1 at the breakpoints, then an efficient recursive algorithm performs small $\{\mu(1),\mu(2),\dotsc,\mu(L)\}$2 matrix inversions, stacking and reshaping submatrices at each step. The process is efficient and numerically stable compared to naively solving a large system (Rajput et al., 2022).

6. Applications in Control, Optimization, and Learning

PMLR is especially suited for on-board, real-time nonlinear control allocation, where nonlinear dynamic inversion demands a high-fidelity, evaluable, and differentiable surrogate of complex models—such as aircraft aerodynamics over a flight envelope. In this context, PMLR provides perfect allocation performance and avoids approximation errors incurred by polynomial-based models (Rajput et al., 2022).

Beyond control, the PMLR framework generalizes to mixed-integer linear programming relaxations for multilinear terms (Sundar et al., 2020), interpretable neural network and piecewise-affine architectures (Braniff et al., 11 May 2025), and the exact neural representation of continuous piecewise (multi-)linear functions (McCarty, 2023, Chen et al., 2022, Zanotti, 17 Mar 2025). In all cases, compact, low-complexity, and exact representations facilitate fast inference, certified optimization, and interpretability crucial for safety-critical or real-time applications.

7. Variants, Extensions, and Related Frameworks

Several closely related representations arise across computational mathematics, optimization, and machine learning:

• Piecewise polyhedral relaxations: The PPR framework creates convex hull relaxations over refined partitions, capturing nonconvexity with MILP formulations (Sundar et al., 2020).
• Neural architectures: Exact (train-free) networks such as YANNs implement piecewise affine or multi-linear maps with explicit, explainable architectures, inheriting all mathematical guarantees from their underlying PMLR (Braniff et al., 11 May 2025).
• Hybrid and interpretable models: PMLR components form modular subnetworks within interpretable neural systems, providing explicit feature contribution functions and capturing nonlinearities in a transparent manner (Guo et al., 2020).

The general paradigm of PMLR thus serves as a unifying tool, providing a principled approach for modeling, fitting, and deploying piecewise multi-linear functions in both theoretical analysis and engineering practice.