Bernstein Lattice: Theory & Applications

• Bernstein lattice is defined as the set of all polynomials representable via Bernstein basis functions that ensure shape-preserving, convex-hull, and positivity properties.
• Its algebraic structure guarantees nonnegativity, endpoint interpolation, and variation-diminishing effects, enabling accurate approximation and efficient LP relaxations.
• The lattice underpins scalable methods in polynomial optimization and control synthesis, balancing theoretical convergence guarantees (O(1/m)) with computational challenges in high dimensions.

The Bernstein lattice—more precisely, the polynomial representation afforded by the Bernstein basis—serves as a central structure in approximation theory, algebraic geometry, numerical analysis, and, increasingly, the synthesis of polynomial certificates for verification and control of dynamical systems. The term encompasses the ordered set of all polynomials representable as linear combinations of Bernstein basis functions of fixed degree over a standard domain (typically the unit interval or the unit hypercube), where the “lattice” nomenclature reflects the coordinate structure of the basis coefficients and the induced order and enclosure properties. The Bernstein lattice uniquely combines shape-preserving, convex-hull, and positivity properties with strong approximation-theoretic guarantees, enabling both theoretical analysis and efficient computational relaxations in polynomial optimization and control synthesis.

1. Definition and Structure of the Bernstein Lattice

The classical Bernstein basis of degree $n$ on the unit interval $[0,1]$ consists of the $n+1$ polynomials

$$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$

The Bernstein lattice $\mathcal{B}^n$ is the coordinate lattice of all polynomial functions of degree at most $n$,

$p(x) = \sum_{k=0}^n \beta_k\, B_{k,n}(x),$

where $\beta_k \in \mathbb{R}$ are the Bernstein coefficients (also called control points) and $(\beta_0, \ldots, \beta_n) \in \mathbb{R}^{n+1}$. This construction generalizes to the multivariate setting on $[0,1]^D$ via the tensor product or, equivalently, by defining multi-indexed basis functions

$[0,1]$0

with multi-index $[0,1]$1, $[0,1]$2. The Bernstein lattice is the (finite-dimensional) vector lattice $[0,1]$3 of all linear combinations of the $[0,1]$4. Each point in this lattice corresponds bijectively to a unique polynomial of maximal (coordinatewise) degree $[0,1]$5.

2. Algebraic Properties and Convex-hull Structure

Bernstein basis functions satisfy nonnegativity, partition of unity, and endpoint interpolation:

• $[0,1]$6 for all $[0,1]$7;
• $[0,1]$8, $[0,1]$9;
• $n+1$0, $n+1$1, with $n+1$2 the Kronecker delta.

A major structural property of the Bernstein lattice is the convex-hull property: at any $n+1$3, $n+1$4 is a convex combination of its coefficients. Thus,

$n+1$5

The lattice is totally positive (all minors of the associated basis matrix are nonnegative for increasing $n+1$6), implying the variation-diminishing property: the number of sign changes in $n+1$7 does not exceed that in $n+1$8. These properties generalize to the multivariate case on $n+1$9.

3. Approximation Theory and Uniform Convergence

For any continuous function $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$0, define the Bernstein operator

$$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$1

Bernstein's theorem asserts uniform convergence: $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$2 on $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$3 as $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$4. The Bernstein lattice is thus dense in the space of continuous functions—any $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$5 can be approximated arbitrarily well in sup norm using sufficiently high-degree lattice elements. The rate of convergence is $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$6, and further acceleration is possible using local subdivisions (composite lattices), e.g., with error $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$7 for $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$8 subdivisions with degree $$B_{k,n}(x) = \binom{n}{k} x^k (1-x)^{n-k}, \qquad 0 \leq k \leq n.$$9 per subinterval (Hammond et al., 12 Sep 2025).

4. Representation, Power-basis Conversion, and Degree Elevation

Every polynomial $\mathcal{B}^n$0 of degree $\mathcal{B}^n$1 admits a unique Bernstein representation via a linear change of basis from the monomial (power) basis. The conversion is explicitly given by: $\mathcal{B}^n$2 where

$\mathcal{B}^n$3

Degree-elevation is achieved via

$\mathcal{B}^n$4

and analogous formulas in the multivariate setting. The lattice is stable under degree elevation and subdivision, supporting refinement and grid-based adaptivity (Sassi et al., 2015).

5. Linear Programming over the Bernstein Lattice: Optimization and Verification

The Bernstein lattice underpins LP relaxations for polynomial nonnegativity, optimization, and stochastic safety certification. A key observation: by the convex-hull property, if all Bernstein coefficients obey a desired inequality, the whole polynomial does as well. Thus,

$\mathcal{B}^n$5

This fact allows replacement of infinite-dimensional constraints with finitely many linear inequalities on lattice coordinates. In the context of stochastic barrier functions (SBF) for safety verification, the search for a polynomial $\mathcal{B}^n$6 satisfying nonnegativity and problem-specific inequalities reduces to a large (but finite) LP over the Bernstein lattice coordinates and auxiliary slack variables. The theoretical rate of convergence of the minimum of Bernstein coefficients to the true minimum of $\mathcal{B}^n$7 is $\mathcal{B}^n$8 (or $\mathcal{B}^n$9 in subdivision), and the bounds are tight as $n$0 (Amorese et al., 10 Jun 2025). This enables certified outer approximations in robust and stochastic control with favorable theoretical guarantees (Sassi et al., 2015).

However, practical application is limited by exponential lattice growth with problem dimension: the constraint system is $n$1 for degree $n$2 in $n$3 variables, and scaling to high-dimension or tight relaxation requires either substantial subdivision or very high degree, leading to combinatorial explosion in lattice size.

6. Comparison with Alternative Bases and Hierarchies

Sum-of-Squares (SoS) relaxations based on Gram matrix semidefinite programs (SDPs) take a fundamentally different approach to polynomial positivity and typically deliver compact representations for certain classes of nonnegative polynomials. In contrast, the Bernstein lattice leads to LP relaxations—more scalable in principle for moderate size but requiring many more constraints to match the SoS tightness in high-dimensional instances. Bernstein relaxations converge at an explicit $n$4 rate, whereas SoS relaxations exhibit only logarithmic convergence in multiplier degree $n$5, but empirical studies demonstrate that SoS relaxations can be more computationally tractable for practical problems despite the higher per-constraint cost. The choice between lattice-based (Bernstein) and moment-based (SoS) hierarchies is thus governed by the trade-off between rate of convergence, sparsity, constraint structure, and problem dimension (Amorese et al., 10 Jun 2025).

7. Applications Beyond Optimization: Numerical PDEs, Geometric Modeling, and Constraints

The Bernstein lattice structure extends to spectral collocation and Galerkin methods for PDEs and integral equations, where its endpoint interpolation, total positivity, and explicit derivative formulas yield stable, accurate methods for elliptic, time-fractional, and boundary value problems (Mirkov et al., 2012, Jani et al., 2016, Shirin et al., 2013). In geometric design, the lattice underlies Bézier curve and surface representations, leveraging the coordinate-wise convex hull property for shape control, robust subdivision, and total variation diminution (Khan et al., 2015). Composite and orthonormal versions of the lattice have been described for improved efficiency, spectral convergence, or symmetric functional analysis (Bellucci, 2014, Hammond et al., 12 Sep 2025).

In bounds-preserving approximation theory, imposing constraints directly on the Bernstein coefficients introduces exact enforcement of global polynomial inequalities, enabling certified solution of constrained best-approximation problems through convex quadratic programming over the lattice (Allen et al., 2021). This combination of algebraic structure, bounding guarantees, and computational tractability distinguishes the Bernstein lattice as a uniquely effective tool for certified polynomial approximation and control synthesis.