---
title: Loomis–Whitney Type Set Inequality
url: https://www.emergentmind.com/topics/loomis-whitney-type-set-inequality
type: topic
---

# Loomis–Whitney Type Set Inequality

A Loomis–Whitney type set inequality is a projection or section inequality that controls the measure or cardinality of a set by lower-dimensional data. In its classical Euclidean form, if \(S\subset \mathbb R^d\) is a body, then
\[
\mu_d(S)^{d-1}\le \prod_{i=1}^d \mu_{d-1}\!\bigl(\pi_{[d]\setminus\{i\}}(S)\bigr),
\]
and equality holds exactly when \(S\) is a box [1510.00258]. The literature uses “Loomis–Whitney type” for a broad family of related estimates: uniform-cover inequalities, local two-subspace forms, affine-invariant variants, reverse and dual inequalities, non-commutative analogues in Heisenberg groups, finite-field forms, and quantitative stability statements that identify the geometry of near-extremizers [1606.03779].

## 1. Classical formulation and uniform covers

The classical Loomis–Whitney inequality bounds \(d\)-dimensional volume by the product of the \((d-1)\)-dimensional volumes of the coordinate-hyperplane projections. In the discrete setting, the same inequality holds for finite \(S\subset \mathbb Z^d\) after replacing Lebesgue measure by cardinality, and equality again characterizes boxes in \(\mathbb Z^d\) [1510.00258]. This is the basic set-theoretic archetype: global size is constrained by lower-dimensional coordinate shadows.

A standard generalization is the Uniform-Cover inequality. If \(\mathcal G\subset 2^{[d]}\) is a uniform \(m\)-cover, meaning that each coordinate belongs to exactly \(m\) members of \(\mathcal G\), then for every body \(S\subset \mathbb R^d\),
\[
\mu_d(S)^m\le \prod_{g\in\mathcal G}\mu_{|g|}\!\bigl(\pi_g(S)\bigr),
\]
with equality iff \(S\) is a box [1510.00258]. In the coordinate-convex-body setting, Brazitikos, Giannopoulos, and Liakopoulos recovered the Bollobás–Thomason inequality as the special case in which \(I_1,\dots,I_r\subset[n]\) form an \(s\)-uniform cover of \([n]\), so that
\[
|K|_n^s\le \prod_{k=1}^r |\pi_{I_k}(K)|_{|I_k|},
\]
and the classical Loomis–Whitney inequality appears when \(r=n\), \(s=n-1\), and \(I_k=[n]\setminus\{k\}\) [1606.03779].

These inequalities are structurally multiplicative: each coordinate direction contributes through a projection family, and the exponents are determined by exact covering identities. That feature underlies most later generalizations, including local, affine, and non-commutative variants.

## 2. Local, affine-invariant, and specialized Euclidean variants

A major Euclidean refinement is the local Loomis–Whitney inequality of Alonso-Gutiérrez, Artstein-Avidan, González Merino, Jiménez, and Villa. If \(K\subset \mathbb R^n\) is a convex body, \(E\in G_{n,i}\), \(H\in G_{n,j}\), \(2\le i,j\le n-1\), \(i+j\ge n+1\), \(E^\perp\subset H\), and \(k=i+j-n\), then
\[
\vol_k\bigl(P_{E\cap H}K\bigr)\,\vol_n(K)
\le
\binom{n-i}{k}\binom{n-j}{k}\,\vol_i(P_EK)\,\vol_j(P_HK).
\]
The constant is explicit and sharp, and the estimate interpolates between global projection inequalities and Fubini-type bounds [1706.01499]. The proof uses a reverse inequality for sections, Steiner-type symmetrization, and Berwald’s inequality, together with functional Rogers–Shephard inequalities [1706.01499].

A distinct line of development replaces the orthonormal coordinate frame by an arbitrary basis. Alonso-Gutiérrez, Bernués, Brazitikos, and Carbery proved an affine-invariant extension: for any basis \(w_1,\dots,w_n\in\mathbb R^n\), if \(W_j=\mathrm{span}\{w_i:i\neq j\}\), then for every compact \(K\subset \mathbb R^n\),
\[
|K|\le C(w_1,\dots,w_n)\prod_{j=1}^n |P_{W_j^\perp}K|^{1/(n-1)},
\]
where \(C(w_1,\dots,w_n)\) is given explicitly in terms of wedge products of the basis vectors [2002.05794]. The same paper develops restricted local forms, dual section forms, and functional Brascamp–Lieb analogues, obtained by combining affine-invariant Brascamp–Lieb inequalities with Berwald’s inequality [2002.05794].

There are also specialized symmetry-driven refinements. For an unconditional, permutationally invariant convex body \(K\subset \mathbb R^n\), Nayar and Tkocz proved that if \(K_j\) denotes the projection onto \(\mathrm{span}\{e_1,\dots,e_j\}\), then the sequence \(a_j=|K_j|\) is log-concave:
\[
a_j^2\ge a_{j-1}a_{j+1},\qquad 2\le j\le n-1.
\]
This is not a direct corollary or strengthening of the classical inequality; rather, it is a flag-type refinement available in a highly symmetric class [1103.6232].

## 3. Reverse and dual inequalities

The classical Loomis–Whitney inequality is an upper bound on volume in terms of projections. Reverse and dual theories ask for lower bounds, typically under convexity hypotheses and often after allowing an adapted frame.

Campi, Gritzmann, and Gronchi introduced the reverse Loomis–Whitney constant
\[
\Lambda(K;F)=\frac{\vol_n(K)^{n-1}}{\prod_{i=1}^n \vol_{n-1}(K|u_i^\perp)},
\]
where \(F=\{u_1,\dots,u_n\}\) is an orthonormal basis, and studied the universal number \(\lambda(n)\) obtained by optimizing over both \(F\) and \(K\). They proved \(\lambda(2)=1/2\), showed the general lower bound
\[
\lambda(n)\ge \binom{2n}{n}(2n)^{-n},
\]
and recorded the asymptotic behavior \(\lambda(n)=\Theta(n^{-n/2})\) as \(n\to\infty\) [1607.07891]. A key structural fact is that for a polytope \(P\), any best frame has at least \(n-1\) vectors parallel to facets of \(P\) [1607.07891].

The dual direction replaces projections by sections. In coordinate form, Meyer’s dual Loomis–Whitney inequality states that for every convex body \(K\subset\mathbb R^n\),
\[
|K|_n^{\,n-1}\ge \frac{n!}{n^n}\prod_{i=1}^n |K\cap e_i^\perp|_{n-1},
\]
with equality exactly for cross-polytopes up to diagonal linear images [1606.03779]. Hao and Jog gave a broader information-theoretic framework: for subspaces \(E_j\subset\mathbb R^n\) with weights \(c_j>0\) satisfying the Brascamp–Lieb conditions, if \(S_{\max}(j)\) denotes the maximal slice of \(K\) parallel to \(E_j^\perp\), then
\[
|K|^{C-1}\ge e^{n+M_g}\prod_{j=1}^m S_{\max}(j)^{c_j},
\qquad C=\sum_{j=1}^m c_j,
\]
and in the special case \(\sum c_jP_{E_j}=I_n\) one has \(M_g=0\) [1901.06789].

A precise generalized dual inequality was obtained in 2025 for compact convex bodies \(K\subset\mathbb R^n\) containing the origin in their interior. If \(E_1,\dots,E_k\) are non-trivial linear subspaces of dimensions \(d_i\) and \(\sum_{i=1}^k c_iP_{E_i}=I_n\), then
\[
|K|\ge \frac{\prod_{i=1}^k d_i!}{n!}\prod_{i=1}^k |K\cap E_i|^{c_i}.
\]
Equality holds if and only if \(\mathbb R^n\) admits a direct orthogonal decomposition into certain independent subspaces \(F_1\oplus\cdots\oplus F_\ell\), each \(F_j\) being an intersection of a choice \(E_i\) or \(E_i^\perp\), and
\[
K=\conv\{K\cap F_j: j=1,\dots,\ell\}.
\]
In the classical coordinate case this recovers the direct-product characterization of the equality case [2507.13462].

Intrinsic-volume analogues extend the reverse and dual program beyond \(V_n\). Campi, Gardner, and Gronchi proved sharp results for \(V_1\) and \(V_{n-1}\), including
\[
V_{n-1}(K)^2\ge \sum_{i=1}^n V_{n-1}(K|e_i)^2
\quad\text{and}\quad
V_{n-1}(K)^2\ge \sum_{i=1}^n V_{n-1}(K\cap e_i)^2,
\]
with equality in the projection inequality characterizing orthogonal cross-polytopes [1311.7076].

## 4. Heisenberg-group and finite-field analogues

In the first Heisenberg group \(\mathbb H\cong(\mathbb R^3,\cdot)\), with group law
\[
(x,y,t)\cdot(x',y',t')=(x+x',\,y+y',\,t+t'+\tfrac12(xy'-yx')),
\]
the relevant maps are the vertical projections
\[
\pi_x(x,y,t)=(x,0,t-\tfrac12xy),\qquad
\pi_y(x,y,t)=(0,y,t+\tfrac12xy).
\]
For measurable \(K\subset\mathbb H\),
\[
|K|\lesssim |\pi_x(K)|^{2/3}|\pi_y(K)|^{2/3}.
\]
This is a genuine Loomis–Whitney-type bound with two projections and exponent \(2/3\), rather than the Euclidean three-projection exponent \(1/2\) in \(\mathbb R^3\) [2003.05862]. The proof is not purely geometric; it is deduced from a planar \(\delta\)-incidence theorem asserting that if \(P\) and \(\mathcal L\) are finite \(\delta\)-separated sets of points and lines in \(\mathbb R^2\), then
\[
|\mathcal I_\delta(P,\mathcal L)|\lesssim |P|^{2/3}|\mathcal L|^{2/3}\delta^{-1/3},
\]
obtained by polynomial partitioning and rich-point counting [2003.05862].

Fässler and Pinamonti extended the Heisenberg theory to \(\mathbb H^n\). If \(\pi_j\), \(j=1,\dots,2n\), are the vertical Heisenberg projections onto the hyperplanes \(\{x_j=0\}\), then for every set \(K\subset\mathbb H^n\),
\[
|K|\lesssim \prod_{j=1}^{2n} |\pi_j(K)|^{\frac{n+1}{n(2n+1)}}.
\]
They also proved the multilinear strong-type estimate
\[
\int_{\mathbb H^n}\prod_{j=1}^{2n} f_j(\pi_j(p))\,dp
\lesssim
\prod_{j=1}^{2n}\|f_j\|_{\frac{n(2n+1)}{n+1}},
\]
deduced inductively from the \(n=1\) case, whose base step is an \(L^{3/2}\to L^3\) improving property of the planar Radon transform [2104.06684]. The comparison with Euclidean space is explicit: in topological dimension \(2n+1\), only \(2n\) vertical projections are used, and the exponents differ from the Euclidean \(1/(d-1)\) because the fibers are left-translates of horizontal one-dimensional subgroups rather than orthogonal coordinate lines [2104.06684].

Finite Heisenberg groups exhibit an additional arithmetic factor. For \(K\subset \mathbb H^n(\mathbb F_q)\),
\[
|K|\le C\,q^{1/(2n+1)}\prod_{j=1}^{2n} |\pi_j(K)|^{\frac{n+1}{n(2n+1)}},
\]
with \(C\) absolute, and the extra factor \(q^{1/(2n+1)}\) is unavoidable [2510.05022]. In the planar case \(n=1\), when the set is not too small, the paper gives the sharper estimate
\[
|K|\lesssim \frac{|\pi_1(K)|\,|\pi_2(K)|}{q}
+q^{1/2}\sqrt{|\pi_1(K)|\,|\pi_2(K)|},
\]
derived from a point-line incidence bound over \(\mathbb F_q\) [2510.05022].

## 5. Equality, near-equality, and stability

Exact equality cases are rigid and often characterize product-type or cross-polytope geometry. In the classical Euclidean inequality, equality holds exactly for boxes [1510.00258]. In the sharp local two-subspace inequality, equality occurs precisely when \(K\) splits relative to
\[
\mathbb R^n=E^\perp\oplus(E\cap H)\oplus H^\perp
\]
into a “two-point join” configuration; Hanner polytopes aligned with this splitting form a concrete equality family [1706.01499]. For reverse inequalities, triangles are exactly the planar minimizers of the reverse Loomis–Whitney constant, while in the dual generalized setting equality is governed by orthogonal decompositions and convex-hull splittings by lower-dimensional sections [1607.07891], [2507.13462].

A deeper issue is quantitative rigidity. Ellis, Friedgut, Kindler, and Yehudayoff proved stability for the classical Loomis–Whitney and Uniform-Cover inequalities by an information-theoretic method. If a body \(S\subset\mathbb R^d\) satisfies
\[
\mu_d(S)^{d-1}\ge (1-\varepsilon)\prod_{i=1}^d
\mu_{d-1}\!\bigl(\pi_{[d]\setminus\{i\}}(S)\bigr),
\]
then there exists a box \(B\subset\mathbb R^d\) such that
\[
\mu_d(S\triangle B)\le c(d)\,\varepsilon\,\mu_d(S),
\qquad c(d)\le 36\,d^3.
\]
For a general family \(\mathcal G\subset 2^{[d]}\) with \(\sigma(\mathcal G)>0\), they also obtained Uniform-Cover stability with
\[
|S\triangle B|\le b(d)\,\rho(\mathcal G)\,\varepsilon\,|S|,
\qquad b(d)=4d^2+64d,
\]
in both discrete and continuous settings [1510.00258]. The same framework yields a stability theorem for the edge-isoperimetric inequality in \(\mathbb Z^d\): nearly optimal edge boundary forces closeness to an axis-parallel cube [1510.00258].

A complementary near-extremal theorem was proved by Zhang in a finite combinatorial setting. If \(E\subset B^n\) and every coordinate-omission projection satisfies \(|\pi_{\hat k}(E)|\le N\), then there exists a refinement \(E'\subset E\) with
\[
|E'|\ge |E|/(C_nN)
\]
such that fibers over any \(m\) coordinates satisfy uniform lower bounds of order \(N^{\,n-m-1}/C_n\), and certain two-coordinate images have size at most \(C_nN^{n-3}\) [1309.2372]. The structural interpretation given in the paper is that near-extremizers of Loomis–Whitney must, after discarding a small exceptional set, resemble approximate Cartesian products [1309.2372].

## 6. Functional, information-theoretic, and analytic consequences

Several recent papers sharpen set inequalities by importing entropy or submodularity. In 2026, a refined finite-set projection inequality was derived from a special case of the strong Madiman–Tetali inequality. If \(S\subset \mathbb R^d\) is a finite set of \(n\) points, \(n_{-i}\) counts the distinct coordinate-deletion projections for \(i\ge 2\), and
\[
n^*_{-1|x_1}=\max_{a\in\mathbb R}
\bigl|\{(x_2,\dots,x_d):(a,x_2,\dots,x_d)\in S\}\bigr|,
\]
then
\[
n^{d-1}\le \Bigl(\prod_{i=2}^d n_{-i}\Bigr)\,n^*_{-1|x_1}.
\]
Since \(n^*_{-1|x_1}\le n_{-1}\), this strictly strengthens the classical bound whenever slice information is nontrivial [2601.15723].

The information-theoretic program also yields dual consequences. Hao and Jog used generalized subadditivity of entropy and a new \(L_1\)-Fisher information to derive lower bounds on both volume and surface area from sections. For a polyconvex set \(K\subset\mathbb R^n\), if \(\alpha_{ij}\) are sampled slice-areas by coordinate hyperplanes, then
\[
\Vol_{n-1}(\partial K)\ge
\frac1{\sqrt n}\sum_{i=1}^n\sum_{j=0}^{M_i}
|\alpha_{i,j+1}-\alpha_{ij}|,
\]
which is a dual section-based surface-area inequality rather than a projection inequality [1901.06789].

In analysis, Loomis–Whitney-type bounds function as multilinear convolution estimates. Kinoshita and Schippa proved a nonlinear Loomis–Whitney estimate on transverse \(C^{1,3}\) hypersurfaces \(S_1,S_2,S_3\subset\mathbb R^3\):
\[
\|f*g\|_{L^2(S_3)}\le C\,A^{1/2}\|f\|_{L^2(S_1)}\|g\|_{L^2(S_2)},
\]
where \(A^{-1}\) is the lower bound for the transversality determinant of the normals, and established a thickened-surface and discrete-lattice version suited to Fourier analysis on \(\mathbb R\times (\tfrac1N\mathbb Z)^2\) [2003.02694]. That estimate is the key new ingredient in their low-regularity local well-posedness result for the fully periodic two-dimensional Zakharov–Kuznetsov equation [2003.02694].

The Heisenberg inequalities have comparable analytic consequences. In \(\mathbb H\), the set inequality implies the sharpened Sobolev–BV estimate
\[
\|f\|_{4/3}\lesssim \sqrt{\|Xf\|\|Yf\|},
\qquad f\in BV(\mathbb H),
\]
which sharpens the classical geometric Sobolev inequality and yields Pansu’s isoperimetric inequality [2003.05862]. In \(\mathbb H^n\), the higher-dimensional analogue is
\[
\|u\|_{\frac{2n+2}{2n+1}}
\lesssim
\prod_{j=1}^{2n}\|X_j u\|^{1/(2n)},
\qquad u\in BV(\mathbb H^n),
\]
again deduced from the corresponding Heisenberg Loomis–Whitney inequality [2104.06684]. Beyond Lebesgue measure, Hosle extended both Loomis–Whitney and Ball’s inequality to measures with \(p\)-concave, \(r\)-homogeneous density, replacing Euclidean projection volume by a measure-projection operator \(P_{\mu,K}(u)\) and introducing the symmetrized density \(\breve g\) into the constants [1908.01257].

Across these directions, the central theme remains unchanged: global size is encoded by a carefully chosen family of lower-dimensional shadows, but the relevant shadows may be coordinate projections, nonlinear group-theoretic projections, local overlaps of subspaces, maximal sections, or slice-sensitive functional marginals. The modern theory is therefore less a single inequality than a framework connecting convex geometry, incidence geometry, entropy, submodularity, harmonic analysis, and sub-Riemannian geometry.

Source: https://www.emergentmind.com/topics/loomis-whitney-type-set-inequality