Young's Convolution Inequality

• Young’s convolution inequality is a family of norm estimates that bounds the Lʳ norm of a convolution in terms of Lᵖ and Lᵠ norms.
• The inequality features sharp Euclidean constants with Gaussian extremizers and integrates concepts from Brascamp–Lieb theory and heat-flow monotonicity.
• It also admits reverse forms and extensions to discrete groups, non-unimodular locally compact groups, and weighted function spaces.

Young’s convolution inequality is the family of norm estimates that control the convolution of two functions by their input $L^p$-norms. In its classical Euclidean form, if $f \in L^p(\mathbb{R}^n)$, $g \in L^q(\mathbb{R}^n)$, and the exponents satisfy

$\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$

then

$$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$

The inequality has a sharp Euclidean refinement with nontrivial optimal constants, admits Gaussian extremizers in the interior range $1<p,q,r<\infty$, extends to reverse inequalities for $0Toscani, 2013).

1. Classical formulation and admissible exponents

On $\mathbb{R}^n$, convolution is

$(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$

and Young’s inequality holds for all $1\le p,q,r\le\infty$ with $f \in L^p(\mathbb{R}^n)$0. The endpoint cases include $f \in L^p(\mathbb{R}^n)$1 or $f \in L^p(\mathbb{R}^n)$2, where Minkowski’s inequality yields $f \in L^p(\mathbb{R}^n)$3 or $f \in L^p(\mathbb{R}^n)$4, and $f \in L^p(\mathbb{R}^n)$5 with $f \in L^p(\mathbb{R}^n)$6, where Hölder gives $f \in L^p(\mathbb{R}^n)$7 (Toscani, 2013).

The same exponent relation governs the discrete-group version. For a discrete group $f \in L^p(\mathbb{R}^n)$8 with counting measure,

$f \in L^p(\mathbb{R}^n)$9

and the inequality

$g \in L^q(\mathbb{R}^n)$0

has sharp constant $g \in L^q(\mathbb{R}^n)$1. In torsion-free discrete groups, exact extremizers in the interior range are supported on single points, and the trilinear form formulation

$g \in L^q(\mathbb{R}^n)$2

is equivalent to the norm inequality (Charalambides et al., 2011).

For $g \in L^q(\mathbb{R}^n)$3 with the same balance relation, the inequality reverses: $g \in L^q(\mathbb{R}^n)$4 in the sharp Euclidean theory, and the reverse form also appears in unimodular-group rearrangement results under support restrictions [(Toscani, 2013); (Satomi, 2022)].

2. Sharp Euclidean constants and Gaussian extremizers

The classical constant $g \in L^q(\mathbb{R}^n)$5 is generally not sharp on $g \in L^q(\mathbb{R}^n)$6. In one common normalization, if

$g \in L^q(\mathbb{R}^n)$7

then Beckner’s sharp form on $g \in L^q(\mathbb{R}^n)$8 is

$g \in L^q(\mathbb{R}^n)$9

equivalently $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$0. The constant tensorizes exactly: $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$1 At the endpoints $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$2, $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$3, or $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$4, the sharp constant reduces to $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$5 (Toscani, 2013).

Equality in the sharp Euclidean inequality holds if and only if the inputs are Gaussian, up to the natural symmetries. In the trilinear formulation, extremizing triples have the form

$\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$6

with $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$7 invertible and $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$8. In the bilinear form, this reduces to Gaussian pairs, with covariance compatibility determined by the exponents (Christ, 2011).

The reverse Young inequality for $\frac{1}{p}+\frac{1}{q}=1+\frac{1}{r},$9 has the same sharp Beckner–Brascamp–Lieb factor and the same Gaussian extremizers. This places the direct and reverse theories in a single sharp-constant framework rather than treating the reverse inequality as merely a formal dual statement (Toscani, 2013).

3. Structural proofs and Brascamp–Lieb interpretations

One influential proof strategy uses the heat equation. If $$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$0 solves $$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$1 with initial data $$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$2, then for suitably chosen diffusion coefficients $$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$3, the Lyapunov functional

$$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$4

is monotone in time. Its limit as $$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$5 is governed by Gaussian heat kernels, so the sharp constant and Gaussian extremizers emerge from a monotonicity principle rather than from rearrangement alone. The same mechanism yields the reverse Young inequality, Brascamp–Lieb-type inequalities, the Prékopa–Leindler inequality, and entropy power inequalities [(Toscani, 2013); (Giuseppe, 2012)].

A second structural viewpoint places Young’s inequality inside the Brascamp–Lieb framework. For the linear maps

$$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$6

on $$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$7, with weights $$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$8, $$\|f*g\|_{L^r(\mathbb{R}^n)} \le \|f\|_{L^p(\mathbb{R}^n)}\|g\|_{L^q(\mathbb{R}^n)}.$$9, $1<p,q,r<\infty$0, the Brascamp–Lieb inequality reproduces the trilinear Young form, and the sharp Young constant becomes the corresponding Brascamp–Lieb constant. In this formulation, Gaussian extremizers arise from the Gaussian variational problem for the Brascamp–Lieb datum, and probabilistic proofs based on the Boué–Dupuis–Borell variational formula recover both the direct and reversed forms (Lehec, 2013).

This perspective extends beyond linear Euclidean convolution. A nonlinear Brascamp–Lieb theorem for simple data implies that, in a small neighborhood of the identity of a Lie group, the best local Young constant approaches the Euclidean sharp constant of the same dimension. For a Lie group $1<p,q,r<\infty$1 of dimension $1<p,q,r<\infty$2, the local sharp constant on sufficiently small neighborhoods converges to $1<p,q,r<\infty$3, answering a question of Cowling, Martini, Müller, and Parcet (Bennett et al., 2018).

4. Locally compact groups, modular corrections, and Lie-group geometry

For a locally compact group $1<p,q,r<\infty$4 with left Haar measure $1<p,q,r<\infty$5, convolution is

$1<p,q,r<\infty$6

When $1<p,q,r<\infty$7 is non-unimodular, the modular function $1<p,q,r<\infty$8 must be inserted to obtain the natural Young functional. Satomi defines

$1<p,q,r<\infty$9

where $0Satomi, 2023).

The principal structural result is subgroup monotonicity: $0Satomi, 2023).

For connected Lie groups whose semisimple part has finite center, Satomi derives the global bound

$0

where $0

$0

$\mathbb{R}^n$0

and

$\mathbb{R}^n$1

For simply connected solvable and nilpotent Lie groups, Nielsen’s results give equality with the Euclidean factor raised to the topological dimension, including the Heisenberg groups (Satomi, 2023).

5. Extremizers, near-extremizers, and rigidity phenomena

In Euclidean space, near-extremizers are stable: if

$\mathbb{R}^n$2

with $\mathbb{R}^n$3, then $\mathbb{R}^n$4 and $\mathbb{R}^n$5 are close in norm to Gaussian extremizers, modulo the usual symmetries. Christ’s proof combines the Riesz–Sobolev rearrangement inequality with an approximate inverse Riesz–Sobolev theorem, reducing near-extremality to approximate Gaussian structure of superlevel sets and phases (Christ, 2011).

The discrete theory exhibits a different rigidity. For torsion-free discrete groups, a $\mathbb{R}^n$6-near extremizer triple is close to point masses: each input has nearly all of its $\mathbb{R}^n$7-mass at a single group element, and exact extremizers are singleton-supported with the compatibility relation $\mathbb{R}^n$8. The proof uses uniform convexity in $\mathbb{R}^n$9 via Clarkson inequalities together with Kemperman’s inequality from additive combinatorics (Charalambides et al., 2011).

Heisenberg groups provide a third pattern. Christ showed that the sharp Young constant on $(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$0 equals the Euclidean sharp constant in the topological dimension $(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$1,

$(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$2

yet no nonzero extremizing triple exists. The obstruction is a symplectic functional equation forced by equality, which has no solution. Nonetheless, near-extremizers exist and, after Heisenberg symmetries, are close to compatible diffuse Gaussian triples. The contrast with $(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$3 shows that equality of sharp constants does not imply existence of extremizers, and the relevant dimension for the constant is the topological dimension $(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$4, not the homogeneous dimension $(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$5 (Christ, 2017).

6. Extensions, variants, and current directions

Several modern developments reinterpret Young’s inequality as a template rather than a single estimate. On unimodular locally compact groups, Satomi proved a rearrangement inequality for convex functionals of convolution under the support condition

$(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$6

from which one obtains

$(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$7

and, when $(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$8, corresponding bounds for the Hausdorff–Young constant $(f*g)(x)=\int_{\mathbb{R}^n} f(x-y)g(y)\,dy,$9 when $1\le p,q,r\le\infty$0 (Satomi, 2022).

Weighted and function-space variants are also well developed. Sharp regions for convolution and multiplication in weighted Lebesgue, Fourier Lebesgue, modulation, and Wiener amalgam spaces were established for polynomial weights $1\le p,q,r\le\infty$1 and $1\le p,q,r\le\infty$2, with necessity and sufficiency of the weight conditions expressed through the Young functional

$1\le p,q,r\le\infty$3

and the balance inequalities involving $1\le p,q,r\le\infty$4 or $1\le p,q,r\le\infty$5 (Toft et al., 2013). A separate interpolation-based theory gives

$1\le p,q,r\le\infty$6

for exact interpolation functors $1\le p,q,r\le\infty$7, leading to Young inequalities in Orlicz, Lorentz–Zygmund, Lorentz–Karamata, and grand Lebesgue scales, and to bilinear multiplier bounds on $1\le p,q,r\le\infty$8 (Fernández-Martínez et al., 2017).

The Young paradigm has also been transplanted to modified convolutions. For the linear canonical transform, Huo introduced a canonical convolution $1\le p,q,r\le\infty$9 satisfying

$f \in L^p(\mathbb{R}^n)$00

together with an exact transform-side multiplication law up to a unimodular phase (Huo, 2018). In Frobenius von Neumann $f \in L^p(\mathbb{R}^n)$01-algebras, a quantum Young inequality

$f \in L^p(\mathbb{R}^n)$02

holds, with associated entropic convolution inequalities and extremizer characterizations in the subfactor case (Huang et al., 2022).

Multilinear and combinatorial extensions remain active. Fractional hypergraph formulations conjecture sharp constants of the form

$f \in L^p(\mathbb{R}^n)$03

linking generalized Young inequalities to entropy power and Brunn–Minkowski inequalities (Bobkov et al., 2010). A multilinear embedding framework connects new Young-type forms to Hardy’s inequality and multilinear Hardy–Littlewood–Sobolev inequalities, with sharp constants in several regimes and realizations on hyperbolic space (Beckner, 2013). On the discrete hypercube $f \in L^p(\mathbb{R}^n)$04, a recent sharp diagonal result replaces the classical exponent $f \in L^p(\mathbb{R}^n)$05 by

$f \in L^p(\mathbb{R}^n)$06

showing that strong support restrictions can improve the admissible exponents while keeping constant $f \in L^p(\mathbb{R}^n)$07 (Beltran et al., 8 Jul 2025).

Young’s convolution inequality therefore occupies a central position in modern analysis: it is simultaneously a sharp Euclidean theorem with Gaussian extremizers, a group-theoretic invariant sensitive to modular structure and subgroup geometry, a stability problem with sharply different behavior in Euclidean, discrete, and nilpotent settings, and a prototype for extensions in interpolation theory, time–frequency analysis, quantum harmonic analysis, multilinear inequalities, and additive combinatorics (Satomi, 2023).