Projection Constant in Banach Spaces

• Projection constant is a norm-theoretic invariant that measures how effectively a subspace is recovered by a bounded linear projection.
• It connects various fields such as Banach space theory, convex geometry, harmonic analysis, and spectral graph theory through equivalent formulations and optimization problems.
• Recent research uses combinatorial-spectral analysis and convex optimization methods to derive precise estimates and structural insights for minimal projections and extremal configurations.

A projection constant is a norm-theoretic invariant that quantifies how efficiently a subspace can be recovered by a bounded linear projection. For a closed subspace $Y$ of a Banach space $X$, the relative projection constant is the infimum of $\|P\|$ over all bounded projections $P:X\to Y$ with $P|_Y=\mathrm{Id}_Y$. Its absolute and maximal variants organize extremal behavior over all ambient spaces or over all spaces of a fixed dimension. Across Banach space theory, convex geometry, harmonic analysis, approximation theory, and more recent work in spectral graph theory, projection constants appear in several equivalent formulations: as norms of minimal projections, as extremal sums attached to tight frames, as combinatorial spectral quantities, and as exact operator norms in structured function spaces (Basso, 2019, Wakhare, 31 Mar 2026).

1. Definitions and basic forms

The standard Banach-space definition is

$$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$

If the infimum is attained, the corresponding operator is a minimal projection. For a finite-dimensional Banach space $E$, the absolute projection constant is obtained by taking the supremum over all isometric embeddings into ambient Banach spaces; the literature in the cited papers uses several notations for this quantity, including $\Pi(E)$, $\lambda(E)$, and $\boldsymbol{\lambda}(E)$ (Kania et al., 12 Apr 2026, Basso, 2019).

A second layer of notation concerns maximal constants in fixed dimension. For real $X$0-dimensional spaces one encounters

$X$1

while related papers write $X$2 for the maximal absolute projection constant among $X$3-dimensional normed spaces over $X$4, or $X$5 for the maximal absolute projection constant over real $X$6-dimensional subspaces (Basso, 2019, Kobos, 30 May 2025, Derȩgowska et al., 2022).

A particularly important reformulation arises from orthogonal projections and tight frames. For integers $X$7,

$X$8

and the real maximal absolute projection constant can be written as

$X$9

In Banach space terminology this equals the quasimaximal absolute projection constant $\|P\|$0; equivalently, it is the supremal normalized absolute inner-product mass of a tight frame with frame operator $\|P\|$1 (Wakhare, 31 Mar 2026).

Two general upper bounds frame much of the subject. The Kadec–Snobar theorem gives $\|P\|$2 for finite-dimensional $\|P\|$3. For hyperplanes $\|P\|$4 in an $\|P\|$5-dimensional normed space, Bohnenblust proved the sharper estimate

$\|P\|$6

which is optimal (Kobos, 2014, Basso, 2019).

2. Extremal finite-dimensional geometry

The maximal projection-constant problem asks for exact formulas for $\|P\|$7, $\|P\|$8, or $\|P\|$9, depending on normalization. One major route is combinatorial-spectral. For every integer $P:X\to Y$0,

$P:X\to Y$1

where $P:X\to Y$2 are the eigenvalues of the associated Seidel matrix. This converts the computation of maximal projection constants into an optimization over switching classes and two-graph spectra (Basso, 2019).

This framework recovers and organizes several low-dimensional values and relative constants. The exact value $P:X\to Y$3 follows from the $P:X\to Y$4-free two-graph analysis, while $P:X\to Y$5, $P:X\to Y$6, and $P:X\to Y$7 arise from equiangular-line configurations in $P:X\to Y$8 and $P:X\to Y$9 and from a detailed six-vertex calculation (Basso, 2019).

A sharper structural description is available when a Gerzon-extremal equiangular tight frame exists. If $P|_Y=\mathrm{Id}_Y$0 is an equiangular set of maximal cardinality $P|_Y=\mathrm{Id}_Y$1 in the real case or $P|_Y=\mathrm{Id}_Y$2 in the complex case, then an $P|_Y=\mathrm{Id}_Y$3-dimensional space $P|_Y=\mathrm{Id}_Y$4 has maximal absolute projection constant if and only if, after a linear isomorphism $P|_Y=\mathrm{Id}_Y$5, its dual unit ball satisfies

$P|_Y=\mathrm{Id}_Y$6

where $P|_Y=\mathrm{Id}_Y$7 is the symmetric zonotope and

$P|_Y=\mathrm{Id}_Y$8

In the same regime, the maximal constant is explicit: $P|_Y=\mathrm{Id}_Y$9 For real spaces, this applies in the known Gerzon-extremal dimensions $$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$0; in the complex setting it applies whenever a SIC-POVM exists (Kobos, 30 May 2025).

The real two-dimensional case is exceptional. There the maximizing norm is unique up to isometry, and the dual unit ball is an affine regular hexagon. This is the geometric form of the value $$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$1 and recovers the classical hexagonal extremizer (Kobos, 30 May 2025). By contrast, in the other covered dimensions there are infinitely many non-isometric maximizers because the inclusion between the absolutely convex hull and the scaled zonotope is strict.

Low-dimensional exact values beyond dimension $$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$2 remain difficult. For real maximal absolute projection constants, the only exact value for $$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$3 stated as known in the 2022 paper is $$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$4. The same paper records numerical evidence for

$$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$5

and proves the lower bound

$$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$6

by constructing mutually unbiased equiangular tight frames in $$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$7 (Derȩgowska et al., 2022).

3. Hyperplanes, finite codimension, and minimal projections

For hyperplanes $$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$8, every projection has the form

$$\lambda(Y,X)=\inf\{\|P\|:P:X\to Y\text{ is a bounded linear projection},\ P|_Y=\mathrm{Id}_Y\}.$$9

This makes hyperplane projection constants unusually concrete. Bohnenblust’s bound

$E$0

is sharp, and equality admits a precise convex-geometric characterization. If $E$1 is $E$2-dimensional and $E$3, then $E$4 holds if and only if there are extreme points $E$5 of $E$6 such that $E$7, the points are linearly independent, and the signed facet condition from Kobos’ theorem is satisfied. In particular, every $E$8-dimensional normed space has an $E$9-dimensional subspace $\Pi(E)$0 with

$\Pi(E)$1

and in dimension $\Pi(E)$2 one has the quantitative statement that every three-dimensional space contains a subspace with projection constant less than $\Pi(E)$3 (Kobos, 2014).

The complementary problem is to force all hyperplanes to have projection constant strictly larger than $\Pi(E)$4. For every integer $\Pi(E)$5 there exists an $\Pi(E)$6-dimensional normed space $\Pi(E)$7 such that for every hyperplane $\Pi(E)$8 and every projection $\Pi(E)$9,

$\lambda(E)$0

This gives an explicit uniform lower bound beyond the trivial value $\lambda(E)$1, albeit an extremely small one, and addresses a hyperplane version of a problem of Bosznay and Garay (Kobos, 2015).

In spaces with the Daugavet property, finite-codimensional projection constants satisfy a duality formula that is both exact and structural. If

$\lambda(E)$2

then

$\lambda(E)$3

Minimal projections $\lambda(E)$4 correspond exactly to weak$\lambda(E)$5-continuous minimal projections $\lambda(E)$6 through

$\lambda(E)$7

A complete description follows for hyperplanes: in a Daugavet space,

$\lambda(E)$8

and a minimal projection onto $\lambda(E)$9 exists if and only if $\boldsymbol{\lambda}(E)$0 attains its norm. In the real space $\boldsymbol{\lambda}(E)$1, this duality combines with a transfer principle from duplication-stable subspaces of $\boldsymbol{\lambda}(E)$2 to produce finite-codimensional subspaces $\boldsymbol{\lambda}(E)$3 for which the infimum defining $\boldsymbol{\lambda}(E)$4 is not attained. Indeed, every value $\boldsymbol{\lambda}(E)$5 occurs as $\boldsymbol{\lambda}(E)$6 for some finite-codimensional $\boldsymbol{\lambda}(E)$7 with non-attainment (Kania et al., 12 Apr 2026).

4. Optimization and computation

Projection constants admit several exact optimization formulations. For maximal relative constants in $\boldsymbol{\lambda}(E)$8, Chalmers–Lewicki’s formula, as quoted in the 2022 paper, states that for integers $\boldsymbol{\lambda}(E)$9,

$X$00

The ETF-based upper bound

$X$01

is attained exactly when there exists a real equiangular tight frame of $X$02 unit vectors in $X$03 (Derȩgowska et al., 2022).

For specific function spaces, convex optimization becomes computationally effective. In univariate polynomial spaces $X$04, every projection onto $X$05 can be represented by finitely many bounded linear functionals, hence by signed measures. The 2018 paper develops an LP upper bound by discretizing those measures with Dirac masses and an SDP lower bound by passing to trigonometric moments and Toeplitz positive semidefinite constraints. For $X$06, the space of algebraic polynomials of degree at most $X$07, the method reproduces

$X$08

and gives the estimates $X$09, $X$10, and $X$11. It also provides tight intervals up to degree $X$12, contests the belief that minimal projections are unique for algebraic polynomial spaces, and gives computational evidence against $X$13-convexity preservation for $X$14 (Foucart et al., 2018).

A different LP framework arises in Lipschitz-free spaces. For a finite metric space $X$15,

$X$16

so the absolute Lipschitz extendability constant is exactly a supremum of relative projection constants of Lipschitz-free spaces. The same paper gives LP formulations for $X$17, proves

$X$18

and establishes

$X$19

This turns a nonlinear extension invariant into a projection-constant computation (Basso, 2021).

5. Explicit formulas in structured spaces

Many classical spaces admit closed formulas or asymptotically sharp expressions for their projection constants. For trigonometric polynomials on a compact abelian group $X$20, if $X$21 is finite, then

$X$22

On $X$23, this yields product formulas in terms of one-dimensional Lebesgue constants; for analytic polynomial spaces on $X$24, it gives asymptotics of order $X$25 for degree-truncated spaces (Defant et al., 2022).

For homogeneous polynomials on Hilbert space, Ryll–Wojtaszczyk’s formula gives

$X$26

and more general $X$27-invariant index sets admit disk-integral formulas involving explicit coefficient sequences. The same general framework relates degree-$X$28 polynomial projection constants to powers of $X$29, up to polarization and annihilating-projection factors (Defant et al., 2022).

For the trace class $X$30, the absolute projection constant is exactly

$X$31

where $X$32 is Haar probability measure on the unitary group. Moreover,

$X$33

This formula is obtained by embedding $X$34 into $X$35, identifying the unique minimal projection via Rudin averaging, and then evaluating the $X$36-norm of the kernel $X$37 (Defant et al., 2023).

These examples show that projection constants are not confined to abstract existence theory. In many structured settings they become exact integrals, explicit Gamma-function ratios, or asymptotic laws.

6. New connections, variants, and terminological distinctions

A recent development links projection constants directly to universal graph eigenvalue bounds. If $X$38 is a graph of order $X$39 and $X$40 is its $X$41-th largest adjacency eigenvalue, then for every $X$42,

$X$43

The proof passes through Ky Fan’s minimum principle, a weighted inequality for the sum of the smallest eigenvalues of symmetric matrices with off-diagonal entries in $X$44, and Weyl’s inequality applied to the complement graph. In dimensions where $X$45 is known, this gives explicit coefficients: $X$46 and

$X$47

from the currently best upper bound on $X$48. The $X$49 bound is sharp by Linz’s closed blowups of the icosahedral graph (Wakhare, 31 Mar 2026).

The term also appears in decomposition theory in a related but not identical sense. For a Schauder basis or a finite-dimensional decomposition $X$50, one studies the supremum of norms of canonical coordinate or interval projections. The strong bimonotonicity projection constant is

$X$51

and the 2020 embedding theorem shows that if $X$52 has separable dual, then for every $X$53 it embeds isometrically into a space with a shrinking $X$54-monotone basis. If $X$55 originally has an FDD with $X$56, the new basis has strong bimonotonicity projection constant at most $X$57; analogous preservation holds for unconditional constants (Barroso, 2020). This usage retains the core theme of controlling canonical projection norms, but the object is now a decomposition rather than a single complemented subspace.

A further terminological distinction is necessary outside Banach space theory. In the Archimedean projection property for hypersurfaces $X$58, the “projection constant” is a geometric measure-scaling factor $X$59 defined by

$X$60

For the smooth compact Archimedean spherical arrays constructed in that setting,

$X$61

This is a measure-theoretic constant attached to an orthogonal projection of a hypersurface, not the operator norm of a bounded linear projection between Banach spaces (Coll et al., 2015).

Taken together, these developments show that projection constants now sit at the intersection of several mature theories. In Banach space geometry they govern complementability and extremal norm structure; in computational work they admit LP, SDP, and combinatorial spectral formulations; in modern applications they control adjacency eigenvalues, Lipschitz extension constants, and canonical decomposition estimates. The common thread is the same extremal question: how large must a projection be when one asks it to preserve a prescribed structure exactly.