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

Minkowski Inequality

The Minkowski inequality states that the \(L^p\) norm of a sum is at most the sum of the norms for \(1\le p\le\infty\).

23 keywords6 linked from1 not yet writtenWritten by AI
Triangle Inequal…Lp SpaceMathematical Ana…Norm (mathematic…Complex NumberEuclidean SpaceMeasure TheoryMeasurable Funct…Minkowski…

The Minkowski inequality is the triangle inequality for LpL^p spaces: when 1≤p≤∞1\le p\le\infty, the size of a sum, measured by its pp-norm, does not exceed the sum of the sizes of its terms. Named after Hermann Minkowski, it applies to finite vectors, infinite sequences, and measurable functions. It is a foundational result in mathematical analysis, establishing that pp-norms satisfy one of the defining properties of a norm.

Finite-dimensional form

Let x=(x1,…,xn)x=(x_1,\ldots,x_n) and y=(y1,…,yn)y=(y_1,\ldots,y_n) have real or complex coordinates. For 1≤p<∞1\le p<\infty, the inequality states that

(∑i=1n∣xi+yi∣p)1/p≤(∑i=1n∣xi∣p)1/p+(∑i=1n∣yi∣p)1/p.\left(\sum_{i=1}^{n}|x_i+y_i|^p\right)^{1/p} \le \left(\sum_{i=1}^{n}|x_i|^p\right)^{1/p} + \left(\sum_{i=1}^{n}|y_i|^p\right)^{1/p}.

Writing

∥x∥p=(∑i=1n∣xi∣p)1/p,\|x\|_p=\left(\sum_{i=1}^{n}|x_i|^p\right)^{1/p},

this becomes

∥x+y∥p≤∥x∥p+∥y∥p.\|x+y\|_p\le \|x\|_p+\|y\|_p.

At p=∞p=\infty, the corresponding norm is

∥x∥∞=max⁡1≤i≤n∣xi∣,\|x\|_\infty=\max_{1\le i\le n}|x_i|,

and the same triangle inequality holds.

Important special cases include:

  • p=1p=1: the inequality follows by summing the scalar inequalities ∣xi+yi∣≤∣xi∣+∣yi∣|x_i+y_i|\le |x_i|+|y_i|.
  • p=2p=2: it is the usual triangle inequality for lengths in Euclidean space.
  • p=∞p=\infty: it bounds the largest absolute coordinate of a sum by the sum of the largest absolute coordinates of its terms.

For example, if x=(1,0)x=(1,0), y=(0,1)y=(0,1), and p=2p=2, then

∥x+y∥2=2<2=∥x∥2+∥y∥2.\|x+y\|_2=\sqrt{2}<2=\|x\|_2+\|y\|_2.

Integral and sequence forms

Let (X,Σ,μ)(X,\Sigma,\mu) be a measure space, as studied in measure theory, and let f,g:X→Rf,g:X\to\mathbb{R} or C\mathbb{C} be measurable functions. For 1≤p<∞1\le p<\infty, define

∥f∥p=(∫X∣f∣p dμ)1/p,\|f\|_p= \left(\int_X |f|^p\,d\mu\right)^{1/p},

where the integral is the Lebesgue integral. If ff and gg have finite pp-norms, then

(∫X∣f+g∣p dμ)1/p≤(∫X∣f∣p dμ)1/p+(∫X∣g∣p dμ)1/p.\left(\int_X |f+g|^p\,d\mu\right)^{1/p} \le \left(\int_X |f|^p\,d\mu\right)^{1/p} + \left(\int_X |g|^p\,d\mu\right)^{1/p}.

For p=∞p=\infty, the norm is the essential supremum: the smallest bound on ∣f∣|f| that holds almost everywhere. The inequality becomes

ess sup⁡x∈X∣f(x)+g(x)∣≤ess sup⁡x∈X∣f(x)∣+ess sup⁡x∈X∣g(x)∣.\operatorname*{ess\,sup}_{x\in X}|f(x)+g(x)| \le \operatorname*{ess\,sup}_{x\in X}|f(x)| + \operatorname*{ess\,sup}_{x\in X}|g(x)|.

Functions that agree almost everywhere represent the same element of LpL^p. This identification is necessary because a function can have norm zero without being zero at every point.

Using counting measure on the positive integers gives the sequence version:

(∑i=1∞∣ai+bi∣p)1/p≤(∑i=1∞∣ai∣p)1/p+(∑i=1∞∣bi∣p)1/p.\left(\sum_{i=1}^{\infty}|a_i+b_i|^p\right)^{1/p} \le \left(\sum_{i=1}^{\infty}|a_i|^p\right)^{1/p} + \left(\sum_{i=1}^{\infty}|b_i|^p\right)^{1/p}.

Thus the finite-vector and infinite-sequence inequalities are special cases of the same measure-theoretic statement.

A weighted form also follows immediately. For nonnegative weights wiw_i,

(∑iwi∣ai+bi∣p)1/p≤(∑iwi∣ai∣p)1/p+(∑iwi∣bi∣p)1/p,\left(\sum_i w_i|a_i+b_i|^p\right)^{1/p} \le \left(\sum_i w_i|a_i|^p\right)^{1/p} + \left(\sum_i w_i|b_i|^p\right)^{1/p},

provided the norms on the right are finite.

Proof using Hölder’s inequality

For 1<p<∞1<p<\infty, the standard proof uses Hölder’s inequality. Let

q=pp−1,1p+1q=1.q=\frac{p}{p-1}, \qquad \frac1p+\frac1q=1.

First, f+gf+g has finite pp-norm whenever ff and gg do. Indeed,

∣f+g∣p≤(∣f∣+∣g∣)p≤2p−1(∣f∣p+∣g∣p).|f+g|^p \le (|f|+|g|)^p \le 2^{p-1}(|f|^p+|g|^p).

The second inequality follows from the convexity of t↦tpt\mapsto t^p on the nonnegative real numbers.

Set h=f+gh=f+g. The scalar triangle inequality gives

∣h∣p≤(∣f∣+∣g∣)∣h∣p−1.|h|^p\le (|f|+|g|)|h|^{p-1}.

Integrating and applying Hölder’s inequality to each term yields

∥h∥pp≤∫X∣f∣∣h∣p−1 dμ+∫X∣g∣∣h∣p−1 dμ≤(∥f∥p+∥g∥p)(∫X∣h∣(p−1)q dμ)1/q.\begin{aligned} \|h\|_p^p &\le \int_X |f||h|^{p-1}\,d\mu + \int_X |g||h|^{p-1}\,d\mu\\ &\le (\|f\|_p+\|g\|_p) \left(\int_X |h|^{(p-1)q}\,d\mu\right)^{1/q}. \end{aligned}

Since (p−1)q=p(p-1)q=p, this is

∥h∥pp≤(∥f∥p+∥g∥p)∥h∥pp−1.\|h\|_p^p \le (\|f\|_p+\|g\|_p)\|h\|_p^{p-1}.

If ∥h∥p=0\|h\|_p=0, the result is immediate. Otherwise, division by ∥h∥pp−1\|h\|_p^{p-1} proves the inequality.

The cases p=1p=1 and p=∞p=\infty follow directly from the pointwise triangle inequality. For p=2p=2, an alternative proof uses the Cauchy–Schwarz inequality to estimate the cross term in the squared norm.

Equality conditions

The case 1<p<∞1<p<\infty

If neither ff nor gg is zero as an LpL^p element, equality holds precisely when

f=cgalmost everywheref=cg \quad\text{almost everywhere}

for a constant c>0c>0. Equality also holds if either function is zero almost everywhere.

For finite vectors, this means that the two nonzero vectors point in the same direction: one is a positive real multiple of the other. For complex-valued functions, positive proportionality requires both proportional magnitudes and matching complex phases wherever the functions are nonzero.

These conditions express the strict convexity of the LpL^p norm for 1<p<∞1<p<\infty: distinct unit vectors have a midpoint of norm strictly less than one.

The case p=1p=1

Equality holds exactly when

∣f+g∣=∣f∣+∣g∣almost everywhere.|f+g|=|f|+|g| \quad\text{almost everywhere}.

For real-valued functions, this is equivalent to

f(x)g(x)≥0almost everywhere.f(x)g(x)\ge 0 \quad\text{almost everywhere}.

The functions need not be proportional; they need only avoid cancellation at almost every point.

For p=∞p=\infty, equality does not generally imply proportionality, and it need not involve a point at which either essential supremum is attained.

Consequences and applications

Normed spaces and distances

Minkowski’s inequality establishes the triangle axiom needed to make LpL^p, for 1≤p≤∞1\le p\le\infty, a normed vector space. These spaces are also complete and therefore are Banach spaces, although completeness requires a separate proof.

The formula

d(f,g)=∥f−g∥pd(f,g)=\|f-g\|_p

defines a metric on LpL^p. In particular, Minkowski’s inequality gives

d(f,h)≤d(f,g)+d(g,h).d(f,h)\le d(f,g)+d(g,h).

It also implies the reverse triangle inequality,

∣∥f∥p−∥g∥p∣≤∥f−g∥p,\bigl|\|f\|_p-\|g\|_p\bigr| \le \|f-g\|_p,

so convergence in norm implies convergence of the norms themselves.

Sums and approximation errors

Repeated application gives, for any finite collection f1,…,fmf_1,\ldots,f_m,

∥∑k=1mfk∥p≤∑k=1m∥fk∥p.\left\|\sum_{k=1}^{m}f_k\right\|_p \le \sum_{k=1}^{m}\|f_k\|_p.

This provides a basic estimate in approximation theory: if an error is decomposed into several terms, its total norm is bounded by the sum of their norms. Unlike squared-norm identities involving orthogonality, this bound requires no cancellation or independence assumptions.

Random variables

On a probability space, the integral form becomes

(E∣X+Y∣p)1/p≤(E∣X∣p)1/p+(E∣Y∣p)1/p.\left(\mathbb{E}|X+Y|^p\right)^{1/p} \le \left(\mathbb{E}|X|^p\right)^{1/p} + \left(\mathbb{E}|Y|^p\right)^{1/p}.

Here XX and YY are random variables with finite pp-th absolute moments, and E\mathbb{E} denotes expectation. No independence assumption is needed.

For p=2p=2, applying the inequality to centered variables gives

SD⁡(X+Y)≤SD⁡(X)+SD⁡(Y),\operatorname{SD}(X+Y) \le \operatorname{SD}(X)+\operatorname{SD}(Y),

where SD⁡\operatorname{SD} is the standard deviation.

Minkowski’s integral inequality

A related form moves an integral outside an LpL^p norm. For suitable jointly measurable F:X×Y→RF:X\times Y\to\mathbb{R} or C\mathbb{C}, on sigma-finite measure spaces, and 1≤p<∞1\le p<\infty,

[∫X(∫Y∣F(x,y)∣ dν(y))pdμ(x)]1/p≤∫Y[∫X∣F(x,y)∣p dμ(x)]1/pdν(y).\left[ \int_X \left(\int_Y |F(x,y)|\,d\nu(y)\right)^p d\mu(x) \right]^{1/p} \le \int_Y \left[ \int_X |F(x,y)|^p\,d\mu(x) \right]^{1/p} d\nu(y).

In particular, if the right-hand side is finite, then

∥∫YF( ⋅ ,y) dν(y)∥Lp(X)≤∫Y∥F( ⋅ ,y)∥Lp(X) dν(y).\left\|\int_Y F(\,\cdot\,,y)\,d\nu(y)\right\|_{L^p(X)} \le \int_Y \|F(\,\cdot\,,y)\|_{L^p(X)}\,d\nu(y).

This is the integral counterpart of the triangle inequality for a finite sum. It is useful in estimating convolutions; for example, on Rn\mathbb{R}^n,

∥f∗g∥p≤∥f∥1∥g∥p,1≤p≤∞.\|f*g\|_p\le \|f\|_1\|g\|_p, \qquad 1\le p\le\infty.

Why the range p\ge 1 matters

For 0<p<10<p<1, the expression

∥x∥p=(∑i∣xi∣p)1/p\|x\|_p=\left(\sum_i |x_i|^p\right)^{1/p}

does not generally satisfy the triangle inequality. Taking x=(1,0)x=(1,0) and y=(0,1)y=(0,1) gives

∥x+y∥p=21/p>2=∥x∥p+∥y∥p.\|x+y\|_p=2^{1/p}>2=\|x\|_p+\|y\|_p.

It therefore defines a quasi-norm, not a norm.

Nevertheless, the scalar inequality

∣a+b∣p≤∣a∣p+∣b∣p|a+b|^p\le |a|^p+|b|^p

implies

∥f+g∥pp≤∥f∥pp+∥g∥pp.\|f+g\|_p^p\le \|f\|_p^p+\|g\|_p^p.

Consequently,

d(f,g)=∫X∣f−g∣p dμd(f,g)=\int_X |f-g|^p\,d\mu

does define a metric on LpL^p equivalence classes for 0<p<10<p<1, even though the usual pp-norm triangle inequality fails.