Polynomial-Preserving Property

• Polynomial-preserving property defines criteria under which polynomial maps maintain key structural features—including positivity, monotonicity, and real-rootedness—across matrices, kernels, and operators.
• It plays a crucial role in operator theory, matrix analysis, numerical approximation, and isogeometric analysis by ensuring stability and convergence in computational schemes.
• Rigorous spectral, combinatorial, and symbolic criteria underpin these preservers, offering insights for applications such as nonnegative inverse eigenvalue problems and privacy-preserving machine learning.

The polynomial-preserving property refers to a family of algebraic and analytic criteria under which polynomials, or more generally polynomial maps and linear operators, maintain structural features (such as positivity, nonnegativity, monotonicity, convexity, real-rootedness, total positivity, or membership in geometric cones) when applied to objects like matrices, kernels, stochastic processes, or functions. This property is central to operator theory, matrix analysis, numerical approximation, isogeometric analysis, stochastic process theory, and the structure theory of special functions and orthogonal polynomials.

1. Algebraic Setting: Matrix Polynomial Preserving Property

Let $M^+_n$ denote the cone of $n \times n$ entrywise nonnegative matrices over $\mathbb{R}$. The set of polynomial preservers of nonnegative $n \times n$ matrices is

$$P_n := \{p \in \mathbb{R}[x] : p(A) \ge 0 \ \text{entrywise for all} \ A \in M^+_n\}.$$

Every $p \in P_n$ has real coefficients, though not all are $\geq 0$ (Loewy, 2022). It is immediate that $P_{n+1} \subseteq P_n$. These sets are strictly nested: for all $n \ge 2$, $P_{n+1} \subsetneq P_n$ (Clark–Paparella conjecture, proved in (Loewy, 2022)). Explicit separating polynomials are constructed: $$p_a(x) = 1 + x + x^2 + \ldots + x^{n-1} - a x^n + x^{n+1} + \ldots + x^{2n},$$ where $0 < a \leq a_0$ (with $a_0$ an explicit combinatorial bound). Such $p_a \in P_n$ but $p_a \notin P_{n+1}$, shown via cycle-path combinatorics in matrix powers and explicit matrix counterexamples. These strict inclusions yield dimension-dependent necessary conditions in matrix spectrum problems, notably the Nonnegative Inverse Eigenvalue Problem (NIEP).

Spectral and derivative-based characterizations exist for the $2 \times 2$ case, with $P_3 \subset P_2$ established (Clark et al., 2021). For a polynomial $p$, $p \in P_2$ iff $p'$, the even part, and the odd part all preserve nonnegativity on $[0, \infty)$ subject to specific spectral inequalities.

2. Shape and Real-root Preservation in Polynomial Operators

Operators acting on polynomials, notably those generating classical orthogonal families, may possess real-root preserving properties. For example, the map $x^n \mapsto (-1)^n n! L_n^\alpha(x)$ (Laguerre polynomials) preserves real-rootedness for all $\alpha \ge 0$, proved by constructing the operator exponential $T = \exp(-A)$ with $A = xD^2 + (\alpha + 1)D$ and continuity/root-location lemmas (Venkataramana, 2015). Similar results are established for Hermite polynomial maps (Cardon et al., 2019). In these families, the coefficient polynomials $Q_k(x)$ arising in generalized differential representations are real-rooted and exhibit strict interlacing, a phenomenon revealing deep ties to zero distributions and the Riemann Hypothesis landscape.

For Chebyshev and Legendre systems, polynomial preservers are highly rigid, with only trivial structures preserving real-rootedness interior to the interval.

3. Total Positivity, Pólya Frequency, and Rigidity of Polynomial Preservers

Entrywise polynomial actions on totally positive (TP) or totally nonnegative (TN) kernels and Pólya frequency (PF) functions are subject to extreme rigidity. For higher-order or infinite-dimensional kernels, the only nontrivial polynomial preservers are positive homotheties $f(x) = c x$ (and in some TN settings, step functions) (Belton et al., 2021). Classification theorems show:

• For TN$_p$ kernels, entrywise powers $x^a$ preserve TN$_p$ iff $a \in \mathbb{N}_0 \cup (p-2, \infty)$.
• For TP$_p$ kernels, only $a=1$ survives for $p \ge 4$.
• For PF functions and sequences, polynomial preservation implies $f(x) = c x$ (possibly $f \equiv 0$ in discrete cases).

This rigidity results from deep combinatorial, analytic, and representation theory arguments, including Descartes' rule of signs and spectral separation phenomena, universally splitting admissible exponents into discrete and continuous sets.

4. Polynomial-Preserving Maps in Numerical Approximation and Isogeometric Analysis

Numerical schemes, such as the Bernstein operator, may or may not preserve geometric features upon integer coefficient modification. For the modified Bernstein operators $\widetilde{B}_n$ and $\widehat{B}_n$ (Draganov, 2020), monotonicity and convexity are not strictly preserved for every $n$, but uniform asymptotic preservation holds: for monotone (resp. convex) functions $f$, small perturbations restore monotonicity (resp. convexity) for large $n$. Quantitative bounds on perturbations are given, and exact sufficient conditions are detailed in terms of increment and second-difference estimates.

In isogeometric analysis, polynomial preserving recovery (PPR) operators for splines (notably PHT-splines) are constructed to exactly reproduce polynomials up to degree 4, yielding superconvergent gradient recovery (order $O(h^4)$), and driving efficient adaptive error estimation (Cai et al., 15 Aug 2025).

5. Symbol-Criterion For Preservation in Algebraic Cones

For geometric and combinatorial classes like volume polynomials (encoding intersection numbers/mixed volumes), linear operator preservation is governed by the symbol-criterion (Grund et al., 23 Mar 2025): a linear operator $T$ on polynomials of bounded multidegree preserves the volume property iff its symbol $S_T(w, u)$ is itself a volume polynomial in combined variables. This extends analogous theorems for stability-preserving (Borcea–Brändén) and Lorentzian-preserving (Brändén–Huh) operators. The symbol-criterion underpins preservation by normalization, diagonalization, antiderivative, truncation, and polarization—and is constructively applied to matroid generating polynomials and polymatroid theory.

6. Polynomial Preserving Property in Stochastic Processes

The classical property for Markov or Lévy processes is polynomial-preservation of conditional expectation: for an adapted process $(X(t))$ in finite dimensions, $\mathbb{E}[p(X(t)) | \mathcal{F}_s]$ is a polynomial in $X(s)$ of degree at most that of $p$. This generalizes to infinite-dimensional Banach space processes (Benth et al., 2018), where multilinear forms replace polynomials. If the Banach space is commutative, independent increment processes are polynomial processes (binomial expansion for conditional expectations survives); in non-commutative contexts, only a multilinearity-preserving property holds, with no general binomial-type structure.

7. Applications to Privacy-Preserving Machine Learning

Polynomial-preserving architectures are essential for compatibility with homomorphic encryption (HE). Poly-GRACE (Pandey, 19 Sep 2025) constructs all network layers, similarity computations, and loss functions strictly from additions and multiplications, ensuring that pre-training and evaluation, under leveled HE schemes, remain entirely polynomial maps. Degree and multiplicative depth bounds ensure practicality, with no need for bootstrapping or transcendental approximations.

8. Further Directions and Open Problems

Outstanding questions include complete classifications of real-root and shape-preserving operator families (especially outside classical orthogonal polynomial systems), the behavior under infinite-dimensional and non-commutative extensions, and the full characterization of preservation in algebraic cones such as those arising in matroid and polymatroid theory (Grund et al., 23 Mar 2025, Cardon et al., 2019). The discovery and analysis of explicit counterexamples, symbol-criteria, and combinatorial interpretations continue to sharpen the understanding of polynomial-preserving phenomena across algebra, analysis, and applied computation.