Strict Polyhedral Approximation Theorem

• The strict polyhedral approximation theorem is a schema that enables approximating convex bodies and Banach norms by polyhedral structures within controlled error bounds.
• It establishes explicit quantitative estimates, such as sharp Hausdorff bounds and recession-sensitive metrics, to assess approximation accuracy.
• The theorem extends across disciplines, underpinning advances in convex geometry, renorming theory, and algorithmic multiobjective convex programming.

The literature suggests that the strict polyhedral approximation theorem is best understood as a theorem schema rather than a single statement: it concerns approximation by polyhedra or polyhedral norms with arbitrarily small, and often explicitly quantified, error. In the classical convex-geometric setting, a compact convex set can be approximated arbitrarily well by polyhedra in the Hausdorff distance; in Banach-space renorming, strict approximation means uniform approximation of an equivalent norm by polyhedral norms to arbitrary precision. Recent work has made these themes quantitative, algorithmic, and extensible to unbounded convex sets, spectrahedral shadows, and other geometric categories (Dörfler, 2022, Bible et al., 2015, Löhne et al., 2021, Dörfler et al., 2023).

1. Terminological scope and basic formulations

For compact convex sets, the classical setting is Hausdorff approximation by polyhedra. For unbounded convex sets, the standard Hausdorff framework is often too restrictive, because the Hausdorff metric is typically infinite unless recession cones coincide exactly. In Banach spaces, by contrast, “strict polyhedral approximation theorem” refers to density of polyhedral norms among equivalent norms, with approximation measured by norm equivalence rather than by a set metric (Dörfler, 2022, Bible et al., 2015).

In finite-dimensional convex geometry, one standard model is an outer approximation built from supporting hyperplanes. If $A \subset \mathbb{R}^n$ is convex and compact, its supporting function is

$s(p,A)=\sup_{x\in A}(p,x),$

and the Hausdorff distance can be written as

$h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$

A grid $G$ of step $\Delta\in(0,2)$ on the sphere gives the external polyhedral approximation

$$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$

In renorming theory, a norm $|||\cdot|||$ is $\varepsilon$-equivalent to $\|\cdot\|$ if

$$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$

A polyhedral norm on $s(p,A)=\sup_{x\in A}(p,x),$0 is one whose restriction to every finite-dimensional subspace has a unit ball that is a polytope (Bible et al., 2015).

These two formulations are structurally related by a common principle: polyhedral objects are used as finite-combinatorial surrogates for more general convex or geometric data. This suggests that “strictness” refers not to exact representability, but to approximation with a controlled error model.

2. Sharp Hausdorff theory for compact convex bodies

For strictly convex compacta in finite-dimensional Euclidean spaces, polyhedral approximation admits sharp Hausdorff bounds governed by the modulus of convexity. In finite dimensions, strictly convex and uniformly convex coincide. If $s(p,A)=\sup_{x\in A}(p,x),$1 is compact, convex, and uniformly convex with modulus of convexity $s(p,A)=\sup_{x\in A}(p,x),$2, and $s(p,A)=\sup_{x\in A}(p,x),$3 is a grid with step $s(p,A)=\sup_{x\in A}(p,x),$4, then under the condition

$s(p,A)=\sup_{x\in A}(p,x),$5

the main estimate is

$s(p,A)=\sup_{x\in A}(p,x),$6

This is Theorem 2.2 in the formulation of “Polyhedral approximations of strictly convex compacta” (Balashov et al., 2010).

The same work gives corollaries showing that if

$s(p,A)=\sup_{x\in A}(p,x),$7

then

$s(p,A)=\sup_{x\in A}(p,x),$8

for some constant $s(p,A)=\sup_{x\in A}(p,x),$9. The estimate is described as sharp in order, and the paper also derives related approximation bounds for geometric difference and intersection. For instance, if $h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$0, then

$h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$1

and for uniformly convex $h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$2 and $h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$3,

$h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$4

The same paper also presents an approximate algorithm for convex hulls based on grid sampling and linear programming. At each grid direction $h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$5, one solves

$h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$6

For a uniformly convex set $h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$7 containing a ball of radius $h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$8, one of the resulting estimates is

$h(A,B)=\sup_{p\in B_1(0)} | s(p,A)-s(p,B) |.$9

The significance of these results is that the Hausdorff error is governed explicitly by the modulus of convexity and the grid step, rather than by a purely asymptotic existence statement (Balashov et al., 2010).

3. Multiobjective convex programming and Benson-type algorithms

A quantitatively strict version of polyhedral approximation for arbitrary convex bodies is obtained by reducing the problem to multiobjective convex programming. In “On the approximation error for approximating convex bodies using multiobjective optimization,” the convex body is represented as

$G$0

where $G$1 has full row rank and the $G$2 are convex. The associated MOCP is

$G$3

with

$G$4

Polyhedral approximations of $G$5 are then obtained from outer or inner polyhedral approximations of the upper image $G$6 by intersecting with

$G$7

and projecting to the first $G$8 coordinates (Löhne et al., 2021).

The primal Benson type algorithm produces a shrinking sequence of outer polyhedral approximations to $G$9; the dual Benson type algorithm produces a shrinking sequence of inner polyhedral approximations via the lower image of the geometric dual problem. Termination is controlled by a tolerance $\Delta\in(0,2)$0. The paper establishes tight bounds for the upper image and for the original convex body: $\Delta\in(0,2)$1 and, for the induced approximation of the convex body,

$\Delta\in(0,2)$2

These are identified in the paper as Proposition 6 and 11 for the MOCP upper image, and Theorems 8 and 14 for the convex body approximation. The bounds hold for both outer and inner approximations, are attained for both primal and dual Benson algorithms, and scale linearly with the stopping tolerance $\Delta\in(0,2)$3. The multiplicative factor $\Delta\in(0,2)$4 depends only on the dimension $\Delta\in(0,2)$5, and the projection step increases the worst-case error bound because

$\Delta\in(0,2)$6

This formulation is a central modern realization of a strict polyhedral approximation theorem: it converts approximation of a convex body into approximation of an upper image in vector optimization, and then relates the final Hausdorff error directly to the algorithmic stopping rule (Löhne et al., 2021).

4. Unbounded convex sets, recession cones, and homogenization

For unbounded convex sets, strict Hausdorff approximation generally fails unless recession behavior is matched exactly. One extension introduces $\Delta\in(0,2)$7-approximation for a nonempty closed, convex, line-free set $\Delta\in(0,2)$8. A line-free polyhedron $\Delta\in(0,2)$9 is an $$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$0-approximation of $$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$1 if it satisfies three conditions: outer approximation $$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$2; vertex proximity

$$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$3

and recession cone proximity

$$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$4

For compact $$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$5, this reduces to classical $$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$6-Hausdorff approximation. If $$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$7 is a sequence of $$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$8-approximations with $$\widehat{A}=\{x\in\mathbb{R}^n \mid (p,x)\le s(p,A),\ \forall p\in G\}.$$9, then $|||\cdot|||$0 in the sense of Painlevé–Kuratowski. The paper also gives the local quantitative estimate

$|||\cdot|||$1

for some $|||\cdot|||$2 a convex combination of vertices of $|||\cdot|||$3 (Dörfler, 2022).

A related but more symmetric framework is homogeneous $|||\cdot|||$4-approximation via homogenization. For a convex set $|||\cdot|||$5,

$|||\cdot|||$6

A polyhedron $|||\cdot|||$7 is a homogeneous $|||\cdot|||$8-approximation of $|||\cdot|||$9 if

$\varepsilon$0

where for closed convex cones,

$\varepsilon$1

The key convergence statement is that

$\varepsilon$2

The same framework is compatible with polarity: $\varepsilon$3 and if $\varepsilon$4 is a homogeneous $\varepsilon$5-approximation of $\varepsilon$6 and both contain the origin, then $\varepsilon$7 is a homogeneous $\varepsilon$8-approximation of $\varepsilon$9 (Dörfler et al., 2023).

These extensions are algorithmic for semidefinite-representable sets. One line of work gives finite algorithms for $\|\cdot\|$0-approximations of spectrahedra (Dörfler, 2022); another gives finite algorithms for homogeneous $\|\cdot\|$1-approximations of spectrahedral shadows via homogenization (Dörfler et al., 2023). Closely related recession-cone algorithms for spectrahedra and spectrahedral shadows compute inner and outer $\|\cdot\|$2-approximations in truncated Hausdorff distance and terminate finitely under stated assumptions (Dörfler et al., 2022). The conceptual shift is that strictness is preserved, but the error model is moved from ordinary Hausdorff distance to recession-sensitive or conic metrics.

5. The Banach-space renorming theorem

In renorming theory, the strict polyhedral approximation theorem has a different formal content. A Banach norm is polyhedral if, on every finite-dimensional subspace $\|\cdot\|$3, the corresponding unit ball is a polytope. The separable version quoted in “Smooth and polyhedral approximation in Banach spaces” is:

Let $\|\cdot\|$4 be a separable Banach space with a polyhedral norm. Then any equivalent norm on $\|\cdot\|$5 can be approximated by polyhedral norms.

This is given there as Theorem 1.4 and described as the strict polyhedral approximation theorem in the separable case (Bible et al., 2015).

The same paper extends the approximation phenomenon to nonseparable spaces of the form $\|\cdot\|$6. Its Theorem 1.7 states:

Let $\|\cdot\|$7 be an arbitrary set, and let $\|\cdot\|$8 be any equivalent norm on $\|\cdot\|$9. Then $$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$0 can be approximated by both $$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$1 smooth norms and polyhedral norms.

The mechanism passes through boundary theory. If $$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$2 has a strong Markushevich basis and

$$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$3

is a boundary of $$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$4, where

$$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$5

then the norm can be approximated by norms having a boundary of finite-support functionals, hence by polyhedral and smooth norms.

The paper also gives a necessary condition for polyhedrality in weakly compactly generated spaces. If $$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$6 is WCG and $$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$7 is polyhedral, then

$$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$8

where each $$\forall x\in X:\quad \|x\|\le |||x||| \le (1+\varepsilon)\|x\|.$$9 is relatively discrete in the weak* topology. This extends an earlier separable countability result. In this branch of the subject, strict approximation means that for every $s(p,A)=\sup_{x\in A}(p,x),$00 there is an $s(p,A)=\sup_{x\in A}(p,x),$01-equivalent polyhedral norm; the relevant “distance” is multiplicative norm distortion, not Hausdorff distance between convex bodies (Bible et al., 2015).

6. Limits, rigidity, and adjacent variants

Several later developments clarify what strict polyhedral approximation can and cannot mean. One limitation is that approximation does not imply exact finite representability. For the semidefinite cone $s(p,A)=\sup_{x\in A}(p,x),$02, sparse polyhedral inner approximations generated by expanded SD bases satisfy

$s(p,A)=\sup_{x\in A}(p,x),$03

with

$s(p,A)=\sup_{x\in A}(p,x),$04

The same work states that no finite polyhedral cone equals $s(p,A)=\sup_{x\in A}(p,x),$05 except for $s(p,A)=\sup_{x\in A}(p,x),$06. This distinguishes strict inner approximation from exact conic representation (Wang et al., 2019).

Another limitation is metric incompatibility. “One polytope fits all” shows a rigidity phenomenon for smooth strictly convex bodies: if a single sequence of inscribed polytopes is asymptotically best for volume and mean width difference simultaneously, then the body must be a Euclidean ball. More generally, the Euclidean ball is the unique $s(p,A)=\sup_{x\in A}(p,x),$07 convex body for which one sequence of polytopes can approximate all intrinsic volumes simultaneously at the optimal asymptotic rate. The same rigidity persists in a probabilistic form for sampling densities and, by polarity, for circumscribed polytopes with a restricted number of facets (Hoehner, 20 Feb 2026). This directly counters a common overgeneralization: a strict approximation theorem does not provide one universally optimal polyhedral scheme across distinct geometric error functionals.

In adjacent geometric areas, the polyhedral approximation paradigm persists but the convergence mode changes. Any length surface homeomorphic to a $s(p,A)=\sup_{x\in A}(p,x),$08-manifold with boundary is the Gromov–Hausdorff limit of polyhedral surfaces with controlled geometry; there exists an approximately isometric sequence of topological embeddings $s(p,A)=\sup_{x\in A}(p,x),$09, and for each compact $s(p,A)=\sup_{x\in A}(p,x),$10,

$s(p,A)=\sup_{x\in A}(p,x),$11

for an absolute constant $s(p,A)=\sup_{x\in A}(p,x),$12 (Ntalampekos et al., 2021). The same polyhedral approximation theorem was extended to arbitrary metric surfaces with locally finite Hausdorff $s(p,A)=\sup_{x\in A}(p,x),$13-measure, with approximately isometric retractions added to compensate for the absence of a length structure (Ntalampekos et al., 2022). For Riemannian manifolds, a sequence of $s(p,A)=\sup_{x\in A}(p,x),$14-dimensional $s(p,A)=\sup_{x\in A}(p,x),$15-polyhedral spaces with curvature $s(p,A)=\sup_{x\in A}(p,x),$16 that Lipschitz-converges to a manifold $s(p,A)=\sup_{x\in A}(p,x),$17 forces the curvature-tensor condition $s(p,A)=\sup_{x\in A}(p,x),$18, and $s(p,A)=\sup_{x\in A}(p,x),$19 is sufficient for local polyhedral approximation; global results are proved under a stably trivial tangent bundle hypothesis and conjectured more generally (Petrunin, 2022).

Taken together, these developments show that the strict polyhedral approximation theorem is a unifying motif across convex geometry, optimization, and renorming theory, but only after the ambient error notion is specified: Hausdorff distance for compact convex bodies, recession-sensitive or homogenized distances for unbounded convex sets, multiplicative equivalence for Banach norms, and Gromov–Hausdorff or Lipschitz control for metric and Riemannian spaces.