JOVANA
Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Hölder and Minkowski: The Two Inequalities

Everything that makes Lᵖ a normed space rests on two inequalities. Hölder controls products across conjugate exponents; Minkowski is the triangle inequality in disguise. We prove both, starting from Young's inequality.

Conjugate exponents and Young's inequality

Two exponents p, q in (1, ∞) are conjugate when 1/p + 1/q = 1. So q = p/(p−1); the self-conjugate case is p = q = 2. The whole theory of products in Lᵖ is governed by this pairing, and the seed is a pointwise inequality between two non-negative numbers.

Young's inequality.  For a, b ≥ 0 and conjugate p, q:
        a·b ≤ aᵖ/p + b^q/q.

Proof (via concavity of log).  If a = 0 or b = 0 it is clear.
For a, b > 0, the logarithm is concave, so for weights 1/p + 1/q = 1:
        log( (1/p)·aᵖ + (1/q)·b^q )
              ≥ (1/p)·log(aᵖ) + (1/q)·log(b^q)
              = log(a) + log(b)
              = log(a·b).
Applying the increasing function exp to both sides:
        (1/p)·aᵖ + (1/q)·b^q ≥ a·b. ∎
Young's inequality is just concavity of log, dressed in exponents.

Hölder's inequality

Hölder's inequality says: if f ∈ Lᵖ and g ∈ Lq with conjugate p, q, then fg ∈ L¹ and ‖fg‖₁ ≤ ‖f‖ₚ · ‖g‖q. When p = q = 2 this is exactly the Cauchy–Schwarz inequality. The proof integrates Young's inequality after normalizing both functions to unit norm.

Hölder.  ∫ |fg| dμ ≤ ‖f‖ₚ · ‖g‖q.

If ‖f‖ₚ = 0 or ‖g‖q = 0 then fg = 0 a.e. and both sides are 0.
Otherwise set  F = |f|/‖f‖ₚ ,  G = |g|/‖g‖q ,  so ‖F‖ₚ = ‖G‖q = 1.
Apply Young pointwise with a = F(x), b = G(x):
        F·G ≤ Fᵖ/p + G^q/q.
Integrate over X:
        ∫ F·G dμ ≤ (1/p)∫Fᵖ dμ + (1/q)∫G^q dμ
                  = (1/p)·1 + (1/q)·1 = 1.
Multiply back through by ‖f‖ₚ·‖g‖q:
        ∫ |f g| dμ ≤ ‖f‖ₚ · ‖g‖q. ∎
Normalize to unit norm, integrate Young, multiply back.

Minkowski's inequality is the triangle inequality

Minkowski's inequality states ‖f + g‖ₚ ≤ ‖f‖ₚ + ‖g‖ₚ. This is precisely the triangle inequality for ‖·‖ₚ, the last axiom needed to call Lᵖ a normed vector space. The trick is to split |f+g|ᵖ as |f+g|·|f+g|^(p−1) and apply Hölder to each piece.

Minkowski.  ‖f+g‖ₚ ≤ ‖f‖ₚ + ‖g‖ₚ   (1 < p < ∞).

Write  |f+g|ᵖ = |f+g| · |f+g|^(p−1) ≤ (|f|+|g|)·|f+g|^(p−1).
Integrate and split:
  ‖f+g‖ₚᵖ ≤ ∫ |f|·|f+g|^(p−1) + ∫ |g|·|f+g|^(p−1).
Let q = p/(p−1) be the conjugate. Note |f+g|^(p−1) ∈ Lq because
  ∫ (|f+g|^(p−1))^q = ∫ |f+g|ᵖ = ‖f+g‖ₚᵖ < ∞.
Apply Hölder to each term:
  ∫ |f|·|f+g|^(p−1) ≤ ‖f‖ₚ · ‖ |f+g|^(p−1) ‖q = ‖f‖ₚ · ‖f+g‖ₚ^(p−1).
Same for g. Therefore
  ‖f+g‖ₚᵖ ≤ (‖f‖ₚ + ‖g‖ₚ) · ‖f+g‖ₚ^(p−1).
If ‖f+g‖ₚ > 0, divide by ‖f+g‖ₚ^(p−1):
  ‖f+g‖ₚ ≤ ‖f‖ₚ + ‖g‖ₚ. ∎
Split the power, apply Hölder twice, divide out the common factor.