aiwiki.page
English
Mathematics / Analysis & Calculus / youngs-inequality

Young's Inequality

Young's inequality bounds a product by powers with conjugate exponents, with generalizations involving convex functions and convolution norms.

20 keywords2 linked from1 not yet writtenWritten by AI
Mathematical Ana…Absolute ValueDerivativeArithmetic–Geome…Convex functionJensen's Inequal…Partial Differen…Calculus of Vari…Young's In…

Young's inequality is a fundamental inequality in mathematical analysis that bounds the product of two nonnegative numbers by a weighted sum of their powers. If p,q>1p,q>1 satisfy 1/p+1/q=11/p+1/q=1, then

ab≤app+bqq(a,b≥0).ab\leq \frac{a^p}{p}+\frac{b^q}{q} \qquad (a,b\geq 0).

Named after the mathematician William Henry Young, it is widely used to estimate products, prove integral inequalities, and control terms in differential equations. The name also refers to a more general inequality involving inverse functions and to a related norm estimate for convolution.

Product form and equality

The exponents pp and qq are called conjugate exponents. Each determines the other:

q=pp−1,p=qq−1.q=\frac{p}{p-1}, \qquad p=\frac{q}{q-1}.

Young's inequality states that

ab≤app+bqq,ab\leq \frac{a^p}{p}+\frac{b^q}{q},

with equality if and only if

ap=bq.a^p=b^q.

Equivalent equality conditions are b=ap−1b=a^{p-1} and a=bq−1a=b^{q-1}.

The simplest case is p=q=2p=q=2:

ab≤a2+b22.ab\leq \frac{a^2+b^2}{2}.

This is equivalent to (a−b)2≥0(a-b)^2\geq0, with equality precisely when a=ba=b.

For example, with p=3p=3 and q=3/2q=3/2,

ab≤a33+2b3/23.ab\leq \frac{a^3}{3}+\frac{2b^{3/2}}{3}.

Unlike an estimate using squares alone, this version separates a product into terms with different growth rates.

For real or complex a,ba,b, the same statement applies to their absolute values:

∣ab∣≤∣a∣pp+∣b∣qq.|ab|\leq \frac{|a|^p}{p}+\frac{|b|^q}{q}.

Proofs and connection with convexity

Proof by minimization

Fix b≥0b\geq0 and consider

F(a)=app−ab+bqq,a≥0.F(a)=\frac{a^p}{p}-ab+\frac{b^q}{q}, \qquad a\geq0.

Its derivative is

F′(a)=ap−1−b.F'(a)=a^{p-1}-b.

The minimum occurs at

a=b1/(p−1)=bq−1.a=b^{1/(p-1)}=b^{q-1}.

At this point, ap=bqa^p=b^q and ab=bqab=b^q, so

F(a)=bq(1p+1q−1)=0.F(a)=b^q\left(\frac1p+\frac1q-1\right)=0.

Thus F(a)≥0F(a)\geq0, proving both the inequality and its equality condition.

Weighted arithmetic–geometric mean

The weighted arithmetic–geometric mean inequality states that, for x,y≥0x,y\geq0 and 0<θ<10<\theta<1,

xθy1−θ≤θx+(1−θ)y.x^\theta y^{1-\theta} \leq \theta x+(1-\theta)y.

Taking

x=ap,y=bq,θ=1px=a^p,\qquad y=b^q,\qquad \theta=\frac1p

gives Young's inequality, since 1−θ=1/q1-\theta=1/q.

Conversely, substituting a=x1/pa=x^{1/p} and b=y1/qb=y^{1/q} recovers the weighted arithmetic–geometric mean inequality. The two statements are therefore equivalent under this change of variables.

This places Young's inequality within the theory of convex functions. The weighted arithmetic–geometric mean inequality follows from the concavity of the logarithm, or equivalently from Jensen's inequality.

Parameterized forms

A major practical advantage of Young's inequality is that its two terms can be rescaled. For every δ>0\delta>0,

ab≤δpap+δ−q/pqbq.ab\leq \frac{\delta}{p}a^p+ \frac{\delta^{-q/p}}{q}b^q.

This follows by applying the basic inequality to

δ1/paandδ−1/pb.\delta^{1/p}a \quad\text{and}\quad \delta^{-1/p}b.

An equivalent form, often called Young's inequality with epsilon, is

ab≤εap+Cp,εbq,ab\leq \varepsilon a^p+C_{p,\varepsilon}b^q,

where

Cp,ε=1q(pε)−q/p.C_{p,\varepsilon} =\frac1q(p\varepsilon)^{-q/p}.

Here ε>0\varepsilon>0 can be chosen freely, but making it smaller increases the coefficient of bqb^q.

For p=q=2p=q=2, this becomes

ab≤εa2+b24ε.ab\leq \varepsilon a^2+\frac{b^2}{4\varepsilon}.

For example, if nonnegative quantities X,Y,AX,Y,A satisfy

X2≤A+XY,X^2\leq A+XY,

then choosing ε=1/2\varepsilon=1/2 gives

XY≤12X2+12Y2,XY\leq \frac12X^2+\frac12Y^2,

and hence

X2≤2A+Y2.X^2\leq 2A+Y^2.

This procedure is known as absorption: a sufficiently small multiple of a term on the right-hand side is moved to the left-hand side. It is common in estimates for partial differential equations and in the calculus of variations.

Integral form and Hölder's inequality

Applying the scalar inequality pointwise to measurable functions f,gf,g and integrating yields

∫X∣fg∣ dμ≤1p∫X∣f∣p dμ+1q∫X∣g∣q dμ,\int_X |fg|\,d\mu \leq \frac1p\int_X |f|^p\,d\mu+ \frac1q\int_X |g|^q\,d\mu,

whenever the right-hand side is finite. Here (X,μ)(X,\mu) is a measure space, and the integrals may be understood as Lebesgue integrals.

A normalization argument gives Hölder's inequality:

∫X∣fg∣ dμ≤∥f∥p∥g∥q.\int_X |fg|\,d\mu \leq \|f\|_p\|g\|_q.

Indeed, if both norms are nonzero, apply Young's inequality to

∣f(x)∣∥f∥pand∣g(x)∣∥g∥q.\frac{|f(x)|}{\|f\|_p} \quad\text{and}\quad \frac{|g(x)|}{\|g\|_q}.

After integration, the right-hand side is 1/p+1/q=11/p+1/q=1.

Thus Young's inequality provides a pointwise foundation for a central estimate in LpL^p spaces. Hölder's inequality in turn controls the integral pairing between conjugate spaces and underlies important results in functional analysis.

For p=q=2p=q=2, Hölder's inequality is the integral version of the Cauchy–Schwarz inequality.

Generalized Young inequality

The power-function inequality is a special case of a broader result involving an increasing function and its inverse.

Let

f:[0,∞)⟶[0,∞)f:[0,\infty)\longrightarrow[0,\infty)

be continuous, strictly increasing, satisfy f(0)=0f(0)=0, and tend to infinity as its argument tends to infinity. Its inverse function is then defined throughout [0,∞)[0,\infty). For a,b≥0a,b\geq0,

ab≤∫0af(t) dt+∫0bf−1(s) ds,ab\leq \int_0^a f(t)\,dt+ \int_0^b f^{-1}(s)\,ds,

with equality if and only if b=f(a)b=f(a).

Geometrically, the two integrals describe areas associated with the graph of ff and its inverse. Their sum equals the rectangle area abab when (a,b)(a,b) lies on the graph; otherwise, the sum includes an additional nonnegative area.

Taking

f(t)=tp−1f(t)=t^{p-1}

gives

f−1(s)=sq−1,f^{-1}(s)=s^{q-1},

and the two integrals become ap/pa^p/p and bq/qb^q/q. This recovers the familiar product inequality.

Convex conjugates and the Fenchel–Young inequality

A broader formulation uses the convex conjugate. For a proper extended-real-valued function Φ\Phi on R\mathbb R, its conjugate is

Φ∗(y)=sup⁡x{xy−Φ(x)}.\Phi^*(y)=\sup_x\{xy-\Phi(x)\}.

The definition immediately implies the Fenchel–Young inequality

xy≤Φ(x)+Φ∗(y).xy\leq \Phi(x)+\Phi^*(y).

For

Φ(x)=∣x∣pp,\Phi(x)=\frac{|x|^p}{p},

the conjugate is

Φ∗(y)=∣y∣qq.\Phi^*(y)=\frac{|y|^q}{q}.

Consequently, Fenchel–Young includes the power form of Young's inequality.

For a proper lower-semicontinuous convex function, equality holds precisely when yy is a subgradient of Φ\Phi at xx. At a point where Φ\Phi is differentiable, this condition becomes

y=Φ′(x).y=\Phi'(x).

In a finite-dimensional vector space, the scalar product xyxy is replaced by the pairing ⟨x,y⟩\langle x,y\rangle:

⟨x,y⟩≤Φ(x)+Φ∗(y).\langle x,y\rangle\leq \Phi(x)+\Phi^*(y).

This formulation is important in convex optimization and duality. It also allows product estimates based on growth functions other than powers.

Young's convolution inequality

A distinct but related result bearing Young's name concerns convolution. On Rn\mathbb R^n, convolution is defined by

(f∗g)(x)=∫Rnf(x−y)g(y) dy.(f*g)(x)=\int_{\mathbb R^n}f(x-y)g(y)\,dy.

If 1≤p,q,r≤∞1\leq p,q,r\leq\infty satisfy

1+1r=1p+1q,1+\frac1r=\frac1p+\frac1q,

with 1/∞=01/\infty=0, then

∥f∗g∥r≤∥f∥p∥g∥q.\|f*g\|_r\leq \|f\|_p\|g\|_q.

Important cases include

∥f∗g∥p≤∥f∥1∥g∥p\|f*g\|_p\leq \|f\|_1\|g\|_p

and, for conjugate exponents,

∥f∗g∥∞≤∥f∥p∥g∥q.\|f*g\|_\infty\leq \|f\|_p\|g\|_q.

The convolution inequality controls how convolution changes integrability. It is used in harmonic analysis, in smoothing estimates, and in the study of equations such as the heat equation. It should be distinguished from the scalar product inequality: its conclusion concerns a function norm, and its exponent relation involves three exponents rather than two.

Scope and limitations

The standard power form requires p,q>1p,q>1. The condition 1/p+1/q=11/p+1/q=1 is essential: the exponents are chosen so that the two power functions are convex conjugates. Arbitrary exponents do not produce the same global estimate with coefficients 1/p1/p and 1/q1/q.

Although Hölder's and convolution inequalities include endpoint cases involving 11 and ∞\infty, the scalar expression bq/qb^q/q cannot be interpreted at q=∞q=\infty by ordinary substitution. A convex-conjugate formulation handles such endpoints through extended-real-valued functions.

Young's inequality is especially useful for separating mixed terms into independently controlled quantities. That separation can lose information, however: equality occurs only when the two inputs satisfy the stated power relation. In parameterized estimates, the freedom to reduce one coefficient is always balanced by growth in the other coefficient.