Hölder's Inequality

The statement, precisely

Let (X, Σ, μ) be a measure space and let p, q ∈ [1, ∞] satisfy the conjugate-exponent condition 1/p + 1/q = 1 (with the conventions 1/∞ = 0 and 1/1 + 1/∞ = 1). For any measurable functions f, g : X → ℂ:

∫_X |f(x) g(x)| dμ(x) ≤ (∫_X |f|^p dμ)^(1/p) · (∫_X |g|^q dμ)^(1/q)

In norm notation, ‖fg‖_1 ≤ ‖f‖_p · ‖g‖_q. For finite sums (counting measure on {1, …, n}) this becomes the discrete Hölder inequality:

Σ_{k=1}^n |a_k b_k| ≤ (Σ |a_k|^p)^(1/p) · (Σ |b_k|^q)^(1/q)

The cases at the boundary are special. p = q = 2 is the classical Cauchy-Schwarz inequality |∫fg| ≤ ‖f‖₂ · ‖g‖₂. p = 1, q = ∞ gives the trivial bound ∫|fg| ≤ ‖g‖_∞ · ∫|f|. All the new content of Hölder lives in the intermediate cases — p = 3, q = 3/2; p = 4, q = 4/3; p = 5/3, q = 5/2 — where neither factor is in L² but their product is integrable.

Proof sketch — Young's inequality is the engine

The standard proof rests on Young's inequality for non-negative real numbers a, b:

ab ≤ a^p/p + b^q/q                  whenever 1/p + 1/q = 1, a, b ≥ 0

This is a one-variable inequality that follows from the concavity of log: log(a^p/p + b^q/q) ≥ (1/p) log a^p + (1/q) log b^q = log(ab). Apply Young pointwise to a = |f(x)| / ‖f‖_p and b = |g(x)| / ‖g‖_q:

|f(x) g(x)| / (‖f‖_p · ‖g‖_q) ≤ (1/p) · |f(x)|^p / ‖f‖_p^p + (1/q) · |g(x)|^q / ‖g‖_q^q

Now integrate over X. The right side becomes 1/p + 1/q = 1. Multiplying both sides by ‖f‖_p · ‖g‖_q gives Hölder's inequality. The proof is two lines once Young is established, and Young is one application of concavity.

The equality case in Young is a = b — that is, a^p = b^q. Tracing this back through the scaling, equality in Hölder holds iff |f|^p / ‖f‖_p^p = |g|^q / ‖g‖_q^q almost everywhere, equivalently iff |f|^p and |g|^q are proportional a.e.

Worked examples with numbers

Concrete checks on small sums and integrals make the inequality vivid.

Example 1 (p = q = 2, Cauchy-Schwarz):
  a = (1, 2, 3),  b = (4, 5, 6)
  Σ a·b = 4 + 10 + 18 = 32
  ‖a‖₂ = √14 ≈ 3.7417,  ‖b‖₂ = √77 ≈ 8.7750
  ‖a‖₂ · ‖b‖₂ ≈ 32.8329
  32 ≤ 32.8329     ✓  (close but not equal; a, b not proportional)

Example 2 (p = 3, q = 3/2):
  a = (1, 1, 1),  b = (2, 3, 4)
  Σ a·b = 9
  ‖a‖₃ = 3^(1/3) ≈ 1.4422
  ‖b‖_{3/2} = (2^(3/2) + 3^(3/2) + 4^(3/2))^(2/3) = (2.828 + 5.196 + 8.0)^(2/3) ≈ 6.342
  ‖a‖₃ · ‖b‖_{3/2} ≈ 9.149
  9 ≤ 9.149        ✓

Example 3 (p = ∞, q = 1):
  a = (1, 5, 2),  b = (3, 4, 6)
  Σ |a·b| = 3 + 20 + 12 = 35
  ‖a‖_∞ = 5,  ‖b‖_1 = 13
  5 · 13 = 65
  35 ≤ 65          ✓  (slack: this is the loosest of the Hölder cases)

Example 4 (equality case):
  a = (1, 2, 3),  b = (1, 2, 3)   (b = a, so |a|^p = |b|^q for p = q = 2)
  Σ a·b = 1 + 4 + 9 = 14
  ‖a‖₂ · ‖b‖₂ = √14 · √14 = 14
  14 = 14          ✓  EQUALITY

For a continuous example, take f(x) = x and g(x) = x² on [0, 1] with p = 3, q = 3/2:

∫_0^1 |fg| dx = ∫_0^1 x³ dx = 1/4
‖f‖₃ = (∫_0^1 x³ dx)^(1/3) = (1/4)^(1/3) ≈ 0.6300
‖g‖_{3/2} = (∫_0^1 x³ dx)^(2/3) = (1/4)^(2/3) ≈ 0.3969
‖f‖₃ · ‖g‖_{3/2} ≈ 0.2500          14 ≤ 0.2500 — EQUALITY (because |f|³ = x³ = |g|^{3/2})

The equality case is informative: whenever |f|^p and |g|^q agree (after scaling), Hölder is tight.

Variants and generalizations

• Discrete Hölder. Σ|aₖbₖ| ≤ (Σ|aₖ|^p)^(1/p) · (Σ|bₖ|^q)^(1/q). Used to compare ℓ^p sequence norms.
• Generalized n-function Hölder. Σ 1/pᵢ = 1 ⇒ ∫|f₁ … fₙ| ≤ Π‖fᵢ‖_{pᵢ}. Routine in nonlinear PDE estimates.
• Interpolation form (Hölder with three exponents). If 1/r = θ/p + (1−θ)/q with θ ∈ [0, 1], then ‖f‖_r ≤ ‖f‖_p^θ · ‖f‖_q^(1−θ). A direct consequence of two-function Hölder applied to |f|^θ and |f|^(1−θ).
• Reverse Hölder. For 0 < p < 1 (so q < 0), the inequality flips: ∫|fg| ≥ ‖f‖_p · ‖g‖_q. Used in Gehring's lemma and Muckenhoupt A_p weight theory.
• Young's convolution inequality. ‖f * g‖_r ≤ ‖f‖_p · ‖g‖_q when 1/p + 1/q = 1 + 1/r. Reduces to Hölder when r = ∞.
• Hausdorff-Young inequality. For 1 ≤ p ≤ 2 with conjugate q, the Fourier transform satisfies ‖F̂‖_q ≤ ‖f‖_p. A Hölder-type estimate adapted to Fourier duality.
• Hölder on Lorentz spaces. ‖fg‖_{L^{1,1}} ≤ ‖f‖_{L^{p, r}} · ‖g‖_{L^{q, r′}}, refining the L^p form using Lorentz exponents.

Hölder proves Minkowski (the triangle inequality)

This is the most important corollary. To show ‖f + g‖_p ≤ ‖f‖_p + ‖g‖_p for 1 ≤ p < ∞, write:

∫ |f + g|^p ≤ ∫ |f + g|^(p−1) (|f| + |g|)
            = ∫ |f| · |f + g|^(p−1) + ∫ |g| · |f + g|^(p−1)

Apply Hölder to each term with conjugate exponent q = p/(p−1):

∫ |f| · |f + g|^(p−1) ≤ ‖f‖_p · ‖(f + g)^(p−1)‖_q = ‖f‖_p · (∫|f + g|^p)^((p−1)/p)

Combining, dividing by (∫|f + g|^p)^((p−1)/p), and recognising the resulting exponent gives ‖f + g‖_p ≤ ‖f‖_p + ‖g‖_p. Without Hölder, the L^p norm wouldn't satisfy the triangle inequality. Hölder is logically prior to Minkowski.

Hölder and L^p–L^q duality

For 1 ≤ p < ∞ with q its conjugate, every g ∈ L^q defines a bounded linear functional T_g : L^p → ℂ by T_g(f) = ∫ fg dμ. Hölder bounds its norm:

‖T_g‖ = sup {|T_g(f)| : ‖f‖_p = 1} ≤ ‖g‖_q

The reverse inequality (‖T_g‖ ≥ ‖g‖_q) follows by choosing f to be a normalized rearrangement of |g|^(q−1) sgn(g). So ‖T_g‖ = ‖g‖_q exactly, and the map g ↦ T_g is an isometric isomorphism L^q ≅ (L^p)*. This is the Riesz representation theorem for L^p, and Hölder is half its proof.

The case p = q = 2 is the self-dual one: (L²)* = L² via the inner product. The case p = 1 is special: (L¹)* = L^∞ but (L^∞)* ⊋ L¹ — L^∞ is not reflexive, because the dual of L^∞ contains "finitely additive measures" beyond L¹.

Common pitfalls

• Forgetting the conjugate condition. Hölder requires 1/p + 1/q = 1. Plugging arbitrary exponents (e.g. p = 2, q = 3) gives a wrong, non-rescalable inequality.
• Using Hölder when one of the norms is infinite. If ‖f‖_p = ∞ the inequality is trivially true; it gives no information. The useful regime is when both norms are finite.
• Forgetting the absolute values. The statement controls ∫|fg|, not ∫fg. For complex-valued or sign-changing f, g you take absolute values throughout.
• Equality requires |f|^p ∝ |g|^q, not f ∝ g. For p = q = 2 these coincide, but for other (p, q) they differ — easy to misremember.
• Confusing Hölder with Minkowski. Hölder bounds ∫|fg| by a product of norms; Minkowski bounds the norm of a sum by a sum of norms. Different statements with different roles.
• Believing Hölder fails for p = ∞. The case p = ∞, q = 1 is included via the conventions ‖f‖_∞ = ess sup |f| and gives the elementary bound ∫|fg| ≤ ‖f‖_∞ · ‖g‖_1.

Where Hölder shows up

• L^p space theory. Hölder proves Minkowski's inequality (triangle for ‖·‖_p), identifies (L^p)* = L^q, bounds the inclusion L^p ⊂ L^r on a finite-measure space (for r < p), and underlies every L^p estimate in functional analysis.
• PDE energy estimates. Bounding a nonlinear term like ‖u² ∂_x u‖_{L^1} by Hölder gives ‖u‖_{L^4}² · ‖∂_x u‖_{L^2}, which couples to Sobolev embedding to close the estimate. Standard machinery for Navier-Stokes, KdV, NLS.
• Probability theory. E|XY| ≤ (E|X|^p)^(1/p) · (E|Y|^q)^(1/q) is the probabilistic Hölder. Special case: Jensen's-like bounds, moment inequalities, Lyapunov's inequality (‖X‖_p increases in p on a probability space).
• Statistics and machine learning. Bias-variance decompositions in high dimensions use Hölder to bound cross-terms. Generalization-error bounds for L^p-regularized models use it to relate empirical to population norms.
• Harmonic analysis. The Hausdorff-Young inequality (‖f̂‖_q ≤ ‖f‖_p, 1 ≤ p ≤ 2) is Hölder-flavoured; Strichartz estimates for dispersive PDE chain Hölder with space-time L^p^q^r norms.
• Number theory. Discrete Hölder bounds character sums, exponential sums (Vinogradov's mean-value theorem), and partial sums in analytic number theory.
• Optimization. Cauchy-Schwarz (p = q = 2) is the foundational duality in linear regression, principal components, and gradient methods. General Hölder extends to ℓ^p regularization and L^p-norm minimization.
• Information theory. Rényi entropies of conjugate orders interact via Hölder-type identities; Bregman divergences include Hölder as a building block.