L^p Spaces: Hölder, Minkowski, and Riesz-Fischer Completeness

Intuition Beginner

A vector in the plane has a length. We measure it by the Pythagorean rule: square the components, add them, take the square root. A function on an interval is something like an infinite-dimensional vector — its values at each point are the components. We want a length-like measurement for such a function. The first guess is the same idea: square the values, add them up (using an integral instead of a sum), and take the square root. The result is called the $L^2$-norm of the function, and it is one example of a more general family of length measurements.

The family is parameterised by a number $p$ between $1$ and infinity. For each fixed $p$, the $L^p$-norm of a function is the $p$-th root of the integral of the $p$-th power of its absolute value. When $p=1$, the norm is the total absolute area. When $p=2$, the norm is the Pythagorean-style energy norm. When $p$ is very large, the norm focuses on the function's largest values. At $p=\infty$, the norm becomes the essential supremum — the function's largest value, ignoring sets of measure zero.

Why do we want a whole family rather than just one? Different problems ask different questions. Integrating a velocity to get a distance uses $L^1$. Measuring kinetic energy uses $L^2$. Bounding the maximum size of a signal uses $L^{\infty }$. Probability theory uses $L^p$ for various $p$ to compute moments of random variables. Each value of $p$ gives a different finite-versus-infinite verdict for the same function, and the right choice depends on what aspect of the function matters for the problem at hand.

Two key inequalities make $L^p$-norms work like Euclidean lengths. Hölder's inequality says the integral of a product is at most the product of $L^p$-norms with paired exponents. Minkowski's inequality says the norm of a sum is at most the sum of the norms, which is the triangle inequality. Together these two facts let us treat $L^p$-functions as vectors in a normed space, the way physicists treat displacement vectors with the Pythagorean theorem.

The one-sentence takeaway: $L^p$-spaces give functions a length measurement that depends on $p$, and Hölder plus Minkowski make these lengths obey the same rules as Euclidean ones, so we can do geometry on spaces of functions.

Visual Beginner

Imagine three pictures of the same function $f(x)=x$ on the interval $[0,1]$. In the first picture, we measure $f$ in $L^1$: the area under $|f|$ is $1/2$, so the $L^1$-norm of $f$ is $1/2$. In the second, we measure $f$ in $L^2$: we square the function (getting $x^2$), integrate (getting $1/3$), and take a square root (getting $1/\sqrt{3}\approx 0.577$). In the third, we measure $f$ in $L^{\infty }$: the largest value of $|f|$ on $[0,1]$ is $1$.

Three different numbers for the same function, three different ways of summarising its size. The Hölder and Minkowski inequalities tell us how these sizes interact when we add or multiply functions.

Worked example Beginner

We compute the $L^p$-norm of the indicator function of an interval for every $p$ in $[1,\infty ]$.

Step 1. Let $f$ be the function on $[0,1]$ defined by $f(x)=1$ for $x$ in $[0,1]$ and $f(x)=0$ everywhere else. This is the indicator of the unit interval.

Step 2. Compute the $L^p$-norm for finite $p$. The $p$-th power of $|f|$ is the same function (since $1^p=1$ for any $p$). The integral of $f$ over $[0,1]$ is $1$. Taking the $p$-th root gives $1^{1/p}=1$. So the $L^p$-norm of $f$ equals $1$ for every finite $p$ in $[1,\infty )$.

Step 3. Compute the $L^{\infty }$-norm. The essential supremum of $|f|$ is the largest value $f$ takes ignoring sets of measure zero. The largest value is $1$. So the $L^{\infty }$-norm of $f$ also equals $1$.

Step 4. Compare with a different function. Let $g(x)=1/\sqrt{x}$ on the interval $(0,1]$ (and $0$ at $x=0$). The $p$-th power is $1/x^{p/2}$. The integral from $0$ to $1$ of $1/x^{p/2}$ is finite when $p/2<1$, that is when $p<2$. So $g$ is in $L^p((0,1])$ for $p$ less than $2$, and not in $L^2$ or higher. The sharp cutoff is at $p=2$.

What this tells us: indicator functions of bounded intervals have the same $L^p$-norm for every $p$, but functions with mild singularities like $1/\sqrt{x}$ belong to some $L^p$-spaces and not others. The threshold $p=2$ for $1/\sqrt{x}$ on $(0,1]$ is exactly the level where the integral of the square first diverges, marking a phase boundary between integrable and non-integrable.

Check your understanding Beginner

Formal definition Intermediate+

Let $(X,\mathcal{M},\mu )$ be a measure space. For a measurable function $f:X\to \mathbb{C}$ (or $\mathbb{R}$ or $\bar{\mathbb{R}}$) and $p\in [1,\infty )$, define $$\Vert f{\Vert }_p={\left({\int }_X|f{|}^{p}d\mu \right)}^{1/p}.$$ For $p=\infty$, define $$\Vert f{\Vert }_{\infty }=\mathrm{esssup}|f|=\inf \{M\ge 0:\mu (\{x:|f(x)|>M\})=0\}.$$

Definition ($L^p$ space). The space ${\mathcal{L}}^p(X,\mu )$ is the set of measurable functions $f$ with $\Vert f{\Vert }_p<\infty$. Two such functions are identified when they agree almost everywhere; the quotient set $L^p(X,\mu )={\mathcal{L}}^p(X,\mu )/\sim$ (with $\sim$ being a.e.-equality) is a vector space, and the map $f\mapsto \Vert f{\Vert }_p$ descends to a norm on $L^p(X,\mu )$.

Definition (conjugate exponent). For $p\in (1,\infty )$ the conjugate exponent $q$ is defined by $1/p+1/q=1$, equivalently $q=p/(p-1)$. The endpoint pairing is $(p,q)=(1,\infty )$ and $(p,q)=(\infty ,1)$. The pairing is symmetric: if $q$ is conjugate to $p$, then $p$ is conjugate to $q$.

Definition (sequence spaces). For counting measure on $\mathbb{N}$, the $L^p$-space specialises to ${\ell }^p=\{(a_n):{\sum }_n|a_n{|}^p<\infty \}$ with $\Vert (a_n){\Vert }_p=({\sum }_n|a_n{|}^p{)}^{1/p}$ for $p\in [1,\infty )$, and ${\ell }^{\infty }=\{(a_n):{\sup }_n|a_n|<\infty \}$ with $\Vert (a_n){\Vert }_{\infty }={\sup }_n|a_n|$.

Counterexamples to common slips Intermediate+

• $L^p$-norm is a norm only after identifying a.e.-equivalent functions. On the level of ${\mathcal{L}}^p$ (without quotient), the map $\Vert \cdot {\Vert }_p$ is a seminorm but not a norm: a function vanishing a.e. has $\Vert f{\Vert }_p=0$ without being the zero function pointwise. Passing to the quotient by the a.e.-equivalence kernel produces a genuine normed space.

• $L^p$-spaces are not nested on infinite-measure spaces. On $(\mathbb{R},\mathcal{B},m)$ with Lebesgue measure, the function $f(x)=(1+|x|{)}^{-1}$ is in $L^p$ for $p>1$ but not in $L^1$, while $g(x)={\chi }_{[0,1]}/\sqrt{x}$ is in $L^1$ but not in $L^2$. So neither $L^1\subseteq L^2$ nor $L^2\subseteq L^1$ holds globally.

• Nesting reverses on finite-measure spaces. When $\mu (X)<\infty$, Hölder's inequality applied with $g=1$ and exponents $(p/r,p/(p-r))$ (for $r<p$) gives $\Vert f{\Vert }_r\le \mu (X{)}^{1/r-1/p}\Vert f{\Vert }_p$, so $L^p\subseteq L^r$ for $1\le r\le p\le \infty$.

• The essential supremum is not the pointwise supremum. For $f(x)={\chi }_{\mathbb{Q}}(x)\cdot 10^{100}$ on $[0,1]$, the pointwise supremum is $10^{100}$ but the essential supremum is $0$, since $f$ vanishes outside the measure-zero set $\mathbb{Q}\cap [0,1]$. The $L^{\infty }$-norm respects the a.e.-equivalence.

• Hölder's inequality fails outside the conjugate-exponent constraint. The inequality $\int |fg|\le \Vert f{\Vert }_p\Vert g{\Vert }_q$ requires $1/p+1/q=1$; without this relation the bound can fail by an arbitrarily large factor (consider $f=g$ in $L^p$ with $p$ small versus $\Vert fg{\Vert }_1=\Vert f{\Vert }_2^2$ requiring the pairing $(2,2)$).

Key theorem with proof Intermediate+

Theorem (Riesz-Fischer completeness; Riesz 1907 C. R. Acad. Sci. Paris 144, 615; Fischer 1907 C. R. Acad. Sci. Paris 144, 1022). Let $(X,\mathcal{M},\mu )$ be a measure space and $p\in [1,\infty ]$. Then $L^p(X,\mu )$ is a Banach space: every Cauchy sequence in the $\Vert \cdot {\Vert }_p$-norm converges in $\Vert \cdot {\Vert }_p$ to a function in $L^p$.

Proof. We give the proof for $p\in [1,\infty )$; the case $p=\infty$ is a direct argument via essential-supremum bookkeeping.

Let $\{f_n\}$ be a Cauchy sequence in $L^p$. Extract a subsequence $\{f_{n_k}\}$ with $\Vert f_{n_{k+1}}-f_{n_k}{\Vert }_p\le 2^{-k}$ for each $k\ge 1$. (This is possible: choose $n_1$ such that $\Vert f_n-f_m{\Vert }_p<1/2$ for $n,m\ge n_1$; then choose $n_2>n_1$ such that $\Vert f_n-f_m{\Vert }_p<1/4$ for $n,m\ge n_2$; iterate.)

Step 1 (build an integrable envelope). Define $$G_K(x)=|f_{n_1}(x)|+\sum _{k=1}^K|f_{n_{k+1}}(x)-f_{n_k}(x)|.$$ Each $G_K$ is a non-negative measurable function. By the Minkowski inequality (Theorem 2 below), $$\Vert G_K{\Vert }_p\le \Vert f_{n_1}{\Vert }_p+\sum _{k=1}^K\Vert f_{n_{k+1}}-f_{n_k}{\Vert }_p\le \Vert f_{n_1}{\Vert }_p+\sum _{k=1}^K{2}^{-k}\le \Vert f_{n_1}{\Vert }_p+1.$$ So $\Vert G_K{\Vert }_p$ is bounded uniformly in $K$, say by $M=\Vert f_{n_1}{\Vert }_p+1$.

Step 2 (monotone convergence to a $p$-integrable limit). The sequence $G_K$ is non-decreasing in $K$ (each $G_{K+1}$ adds a non-negative term). Set $G(x)={\lim }_{K\to \infty }{G}_K(x)=|f_{n_1}(x)|+{\sum }_{k=1}^{\infty }|f_{n_{k+1}}(x)-f_{n_k}(x)|$. The sequence $G_K^p$ is non-decreasing in $K$ with pointwise limit $G^p$, so by the monotone convergence theorem 02.07.04, $${\int }_X{G}^{p}d\mu =\underset{K}{\lim }{\int }_X{G}_K^{p}d\mu =\underset{K}{\lim }\Vert G_K{\Vert }_p^p\le M^p.$$ Hence $G\in L^p$, so $G(x)<\infty$ for a.e. $x$.

Step 3 (the series converges a.e.). On the set $\{G<\infty \}$ (which has full measure), the telescoping series ${\sum }_{k=1}^{\infty }(f_{n_{k+1}}-f_{n_k})$ converges absolutely pointwise. Define $f(x)=f_{n_1}(x)+{\sum }_{k=1}^{\infty }(f_{n_{k+1}}(x)-f_{n_k}(x))$ on this set and $f(x)=0$ on the measure-zero exceptional set. The partial sums equal $f_{n_{K+1}}(x)$, so $f_{n_K}(x)\to f(x)$ pointwise a.e.

Step 4 (a.e. limit lies in $L^p$ and the subsequence converges in $L^p$). Since $|f|\le G$ a.e., $f\in L^p$. Also $|f_{n_K}-f{|}^p\to 0$ a.e. and $|f_{n_K}-f{|}^p\le (2G{)}^p\in L^1$. By the dominated convergence theorem 02.07.05, $${\int }_X|f_{n_K}-f{|}^{p}d\mu \to 0,$$ that is $\Vert f_{n_K}-f{\Vert }_p\to 0$.

Step 5 (full sequence converges). The Cauchy hypothesis plus a triangle-inequality argument lifts subsequence convergence to full-sequence convergence: given $\epsilon >0$, choose $N$ with $\Vert f_n-f_m{\Vert }_p<\epsilon /2$ for $n,m\ge N$, then choose $K$ large with $n_K\ge N$ and $\Vert f_{n_K}-f{\Vert }_p<\epsilon /2$. For $n\ge N$: $$\Vert f_n-f{\Vert }_p\le \Vert f_n-f_{n_K}{\Vert }_p+\Vert f_{n_K}-f{\Vert }_p<\epsilon /2+\epsilon /2=\epsilon .$$ Hence $f_n\to f$ in $L^p$.

$\square$

Bridge. The Riesz-Fischer completeness builds toward 02.11.04 the abstract theory of Banach spaces, where the same Cauchy-to-limit construction is the defining feature, and appears again in the Hilbert-space treatment of $L^2$ via Parseval's identity. The central insight is that the Minkowski inequality lets us add absolute values of differences inside the norm, producing a $p$-summable envelope $G$, and this is exactly the structural fact that the monotone convergence theorem turns the envelope into an honest $L^p$-function. The foundational reason completeness reduces to MCT plus DCT is the bracketing: the envelope $G$ from MCT dominates every tail of the sequence, and DCT then identifies the $L^p$-limit with the a.e.-limit. The bridge is between the abstract Cauchy condition and the concrete construction of a candidate limit via pointwise convergence on the full-measure set $\{G<\infty \}$, and putting these together identifies $L^p$ as a Banach space, the pattern generalising to Sobolev spaces, Bochner $L^p$-spaces, and the operator-norm completeness of bounded-linear maps.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: full — Mathlib provides the central $L^p$-space architecture. MeasureTheory.Lp is the quotient ${\mathcal{L}}^p/{\sim }_{a.e.}$ defined for any p : ENNReal and any α →ₘ[μ] E (almost-everywhere measurable maps to a normed group). MeasureTheory.Lp.norm_def gives the norm formula. MeasureTheory.snorm is the unquotiented norm. MeasureTheory.MemLp is the predicate $f\in L^p$. Hölder appears as MeasureTheory.integral_mul_le_Lp_mul_Lq_of_nonneg and MeasureTheory.lintegral_mul_le_Lp_mul_Lq. Minkowski is MeasureTheory.snorm_add_le. Completeness is MeasureTheory.Lp.completeSpace, instantiating CompleteSpace (MeasureTheory.Lp E p μ) for $p\in [1,\infty ]$.

import Mathlib.MeasureTheory.Function.LpSpace
import Mathlib.MeasureTheory.Integral.MeanInequalities

variable {α E : Type*} [MeasurableSpace α] [NormedAddCommGroup E]
variable (μ : MeasureTheory.Measure α) (p : ENNReal) [Fact (1 ≤ p)]

-- L^p as a Banach space
example : CompleteSpace (MeasureTheory.Lp E p μ) := inferInstance

-- Hölder for p, q conjugate
example (f g : α → ℝ≥0∞) (p q : ℝ) (hpq : p.IsConjExponent q) :
    MeasureTheory.lintegral μ (fun x => f x * g x) ≤
    (MeasureTheory.lintegral μ (fun x => f x ^ p)) ^ (1 / p) *
    (MeasureTheory.lintegral μ (fun x => g x ^ q)) ^ (1 / q) :=
  MeasureTheory.ENNReal.lintegral_mul_le_Lp_mul_Lq μ hpq
    (by measurability) (by measurability)

Advanced results Master

The advanced theory of $L^p$-spaces splits across six strands: the Hölder-Minkowski inequalities, Riesz-Fischer completeness, duality $(L^p{)}^{\ast }=L^q$, uniform convexity and reflexivity, density of nice subspaces, and the interpolation and Fourier-analytic extensions.

Theorem 1 (Hölder's inequality; Hölder 1889 Nachr. Göttingen 1889, 38; generalised by Riesz 1910 Math. Ann. 69, 449). Let $p,q\in [1,\infty ]$ with $1/p+1/q=1$. For measurable $f,g:X\to \mathbb{C}$, $${\int }_X|fg|d\mu \le \Vert f{\Vert }_p\Vert g{\Vert }_q.$$ Equality holds iff $|f{|}^p$ and $|g{|}^q$ are linearly dependent (when both $p,q<\infty$); for $(p,q)=(1,\infty )$, equality holds iff $|g|=\Vert g{\Vert }_{\infty }$ a.e. on $\{f\ne 0\}$.

The Hölder inequality generalises the Cauchy-Schwarz inequality (the $p=q=2$ case, due to Bunyakovsky 1859 and Schwarz 1885) and is the load-bearing inequality for $L^p$-duality, interpolation, and the embedding theorems for Sobolev spaces ^{[Hölder 1889]}.

Theorem 2 (Minkowski's inequality; Minkowski 1896 Geometrie der Zahlen). For $p\in [1,\infty ]$ and measurable $f,g$, $$\Vert f+g{\Vert }_p\le \Vert f{\Vert }_p+\Vert g{\Vert }_p.$$ The inequality is the triangle inequality for the $L^p$-norm, identifying $\Vert \cdot {\Vert }_p$ as a genuine norm on the quotient space $L^p(X,\mu )$.

Minkowski's original 1896 statement was for sequences (the ${\ell }^p$-norm triangle inequality for finite-dimensional positive vectors with $p=2k$ even integer); the extension to general $p\in [1,\infty ]$ and integration measures is due to Riesz 1910 ^{[Minkowski 1896]}.

Theorem 3 (Riesz-Fischer completeness; Riesz 1907 C. R. Paris 144, 615; Fischer 1907 C. R. Paris 144, 1022). For any measure space $(X,\mathcal{M},\mu )$ and $p\in [1,\infty ]$, the space $L^p(X,\mu )$ is a Banach space.

Riesz's and Fischer's independent 1907 Comptes Rendus notes proved completeness specifically for $L^2$ in the context of orthogonal function systems (Riesz framed it as completeness of the Hilbert-space spanned by orthonormal sequences; Fischer framed it as convergence-in-mean of square-integrable functions). The extension to general $L^p$ is due to Riesz 1910 ^{[Riesz 1910]}.

Theorem 4 (Duality; Riesz 1910 Math. Ann. 69, 449). Let $p\in [1,\infty )$ and $q$ the conjugate exponent. On a $\sigma$-finite measure space, every continuous linear functional $T:L^p(X,\mu )\to \mathbb{C}$ has the form $$T(f)={\int }_{X}fgd\mu$$ for a unique $g\in L^q(X,\mu )$, and $\Vert T{\Vert }_{(L^p{)}^{\ast }}=\Vert g{\Vert }_q$. So $(L^p{)}^{\ast }\cong L^q$ as Banach spaces.

The duality $(L^1{)}^{\ast }\cong L^{\infty }$ holds on $\sigma$-finite spaces. The reverse identification $(L^{\infty }{)}^{\ast }\cong L^1$ fails in general: $(L^{\infty }{)}^{\ast }$ is strictly larger than $L^1$, containing finitely additive measures (Banach limits, $\beta \mathbb{N}$ ultrafilter functionals). Riesz's 1910 proof on $[0,1]$ extends to $\sigma$-finite spaces by Radon-Nikodym; the duality fails on non-$\sigma$-finite spaces by counterexamples involving non-$\sigma$-finite ideals ^{[Riesz 1910]}.

Theorem 5 (Uniform convexity and reflexivity; Clarkson 1936 Trans. AMS 40, 396). For $p\in (1,\infty )$, the space $L^p(X,\mu )$ is uniformly convex, hence by the Milman-Pettis theorem (Milman 1938, Pettis 1939) reflexive.

Clarkson's 1936 paper introduced the uniform-convexity inequality $${\Vert \frac{f+g}{2}\Vert }_p^p+{\Vert \frac{f-g}{2}\Vert }_p^p\le \frac{1}{2}(\Vert f{\Vert }_p^p+\Vert g{\Vert }_p^p)(p\ge 2)$$ (with a dual inequality for $p\in (1,2]$), a quantitative refinement of the parallelogram-law-like behaviour of $L^p$-norms. Uniform convexity implies reflexivity (Milman-Pettis), strict convexity, and the existence of best approximations to closed convex sets by unique projections. The endpoints $p=1$ and $p=\infty$ are not uniformly convex and not reflexive: $L^1$ has the strictly larger dual $L^{\infty }$, and $L^{\infty }$ has the strictly larger dual $(L^{\infty }{)}^{\ast }$ containing Banach limits ^{[Clarkson 1936]}.

Theorem 6 (Weak compactness and Banach-Alaoglu). For $p\in (1,\infty )$, every bounded sequence in $L^p$ has a weakly-convergent subsequence (Eberlein-Šmulian theorem applied to the reflexive Banach space $L^p$).

The reflexivity from Theorem 5 plus the Banach-Alaoglu theorem (weak-$\ast$ compactness of the closed unit ball in any dual space) gives the bounded-implies-weakly-precompact conclusion. This is the foundational fact behind weak-solution existence theory in PDE: a bounded sequence of Galerkin approximations in $W^{1,p}$ (Sobolev space, $1<p<\infty$) has a weakly-convergent subsequence, the candidate weak solution.

Theorem 7 (Density of $C_c^{\infty }$ via mollification; Friedrichs 1944 Trans. AMS 55, 132). For $p\in [1,\infty )$, the space $C_c^{\infty }({\mathbb{R}}^n)$ of compactly supported smooth functions is dense in $L^p({\mathbb{R}}^n)$. Density is achieved by convolution with a standard mollifier ${\eta }_{\epsilon }(x)={\epsilon }^{-n}\eta (x/\epsilon )$, where $\eta \in C_c^{\infty }({\mathbb{R}}^n)$ is non-negative with $\int \eta =1$, supported in the unit ball.

For $f\in L^p$, the mollification $f_{\epsilon }=f\ast {\eta }_{\epsilon }$ lies in $C^{\infty }$, and $\Vert f_{\epsilon }-f{\Vert }_p\to 0$ as $\epsilon \to 0$ by uniform continuity of translations in $L^p$. Combining with the truncation argument of Exercise 8 gives $C_c^{\infty }$ density. The density fails for $p=\infty$ (bounded measurable functions are not generally approximable by continuous functions in the essential-supremum norm) ^{[Friedrichs 1944]}.

Theorem 8 (Hausdorff-Young inequality; Young 1912 Proc. London Math. Soc. (2) 11, 357; Hausdorff 1923 Math. Z. 16, 163; sharp constant Beckner 1975 Ann. Math. 102, 159). Let $p\in [1,2]$ and $q$ the conjugate exponent. The Fourier transform $\mathcal{F}:L^p({\mathbb{R}}^n)\to L^q({\mathbb{R}}^n)$ is bounded, with $$\Vert \hat{f}{\Vert }_q\le A_{p,n}\Vert f{\Vert }_p,$$ where $A_{p,n}=(p^{1/p}/q^{1/q}{)}^{n/2}$ is Beckner's sharp constant.

The endpoints $p=1,q=\infty$ and $p=2,q=2$ are due to Young 1912 (the $L^1\to L^{\infty }$ boundedness $\Vert \hat{f}{\Vert }_{\infty }\le \Vert f{\Vert }_1$ is immediate from the definition of $\hat{f}$) and Plancherel 1910 (the $L^2$-isometry). The interpolating range $p\in (1,2)$ is due to Hausdorff 1923 via the Riesz-Thorin convexity theorem. Beckner's 1975 paper identified the sharp constant via Gaussian extremisers; the proof uses the spherical-rearrangement principle of Babenko 1961 and Brascamp-Lieb 1976 ^{[Hausdorff 1923; Young 1912; Beckner 1975]}.

Theorem 9 (Riesz-Thorin interpolation; M. Riesz 1926 Acta Math. 49, 465; Thorin 1948 Medd. Lunds Univ. Mat. Sem. 9). Let $T$ be a linear operator bounded simultaneously $T:L^{p_0}\to L^{q_0}$ with norm $M_0$ and $T:L^{p_1}\to L^{q_1}$ with norm $M_1$, where $p_0,p_1,q_0,q_1\in [1,\infty ]$. For $\theta \in (0,1)$, define $p_{\theta },q_{\theta }$ by $$\frac{1}{p_{\theta }}=\frac{1-\theta }{p_0}+\frac{\theta }{p_1},\frac{1}{q_{\theta }}=\frac{1-\theta }{q_0}+\frac{\theta }{q_1}.$$ Then $T:L^{p_{\theta }}\to L^{q_{\theta }}$ is bounded with norm at most $M_0^{1-\theta }{M}_1^{\theta }$.

The Riesz-Thorin theorem is the complex interpolation theorem: M. Riesz 1926 stated it in the bilinear form and proved the case where the dual exponents lie in the convexity-allowed region; Thorin 1948 extended to the full range via Hadamard's three-lines theorem. The proof is a single application of the three-lines lemma to the analytic family $z\mapsto \int (T{f}_z)g_{z}d\mu$ where $f_z,g_z$ are analytic interpolations between simple-function dyadic decompositions of $f,g$. The Hausdorff-Young inequality (Theorem 8) is the canonical application: interpolate between Plancherel's $L^2$-isometry of the Fourier transform and Young's $L^1\to L^{\infty }$ bound to obtain the $L^p\to L^q$ boundedness for the entire range $p\in [1,2]$ ^{[Riesz 1910]}.

Theorem 10 (Marcinkiewicz interpolation; Marcinkiewicz 1939 C. R. Acad. Sci. Paris 208, 1272; Zygmund 1956 J. Math. Pures Appl. 35). Let $T$ be a sublinear operator of weak types $(p_0,q_0)$ and $(p_1,q_1)$ with $1\le p_i\le q_i\le \infty$ and $q_0\ne q_1$. Then for every $\theta \in (0,1)$, $T$ is of strong type $(p_{\theta },q_{\theta })$ with the same convexity exponent relations.

The Marcinkiewicz theorem is the real interpolation theorem: it allows the endpoints to be weak-$L^q$ bounds (sublinear-typed control by distribution-function inequalities $\mu (\{|Tf|>\lambda \})\le (C\Vert f{\Vert }_{p_i}/\lambda {)}^{q_i}$) rather than strong $L^q$ bounds, and the conclusion is a strong $L^q$ bound in the interior. The theorem applies to operators that fail to be of strong type at one endpoint — e.g., the Hardy-Littlewood maximal operator is of weak type $(1,1)$ but not strong type $(1,1)$ — and gives boundedness on the entire interior of the convexity diagram. Marcinkiewicz's 1939 paper announced the result; the detailed proof and applications are due to Zygmund 1956 J. Math. Pures Appl. 35 ^{[Riesz 1910]}.

Theorem 11 (Sobolev embedding; Sobolev 1938 Mat. Sb. 4, 471; Gagliardo 1958 Ricerche Mat. 7, 102; Nirenberg 1959 Ann. Sc. Norm. Sup. Pisa 13). Let $\Omega \subseteq {\mathbb{R}}^n$ be a bounded open set with Lipschitz boundary. For $1\le p<n$ and $p^{\ast }=np/(n-p)$ (the Sobolev conjugate), $$W^{1,p}(\Omega )\hookrightarrow L^{p^{\ast }}(\Omega )$$ is a continuous embedding with $\Vert f{\Vert }_{L^{p^{\ast }}}\le C(n,p,\Omega )\cdot \Vert \nabla f{\Vert }_{L^p}$ for $f\in W_0^{1,p}(\Omega )$.

The Sobolev embedding theorem promotes integrability of derivatives to higher integrability of the function itself, via the Sobolev exponent $p^{\ast }=np/(n-p)$. The proof for $p=1$ is due to Gagliardo 1958 and Nirenberg 1959 via the iterated-slicing inequality $\Vert f{\Vert }_{n/(n-1)}\le C{\prod }_{i=1}^n\Vert {\partial }_{i}f{\Vert }_1^{1/n}$ for compactly supported $f$. The case $p>1$ follows by interpolation. The borderline $p=n$ embeds into the Bounded Mean Oscillation (BMO) space (John-Nirenberg 1961) rather than into $L^{\infty }$, and $p>n$ embeds into Hölder-continuous functions $C^{0,\alpha }$ with $\alpha =1-n/p$ (Morrey 1938) ^{[Schwartz 1950]}.

Synthesis. The $L^p$-space architecture is the foundational reason that modern analysis can treat function spaces as honest Banach (and at $p=2$, Hilbert) spaces with the full apparatus of linear functional analysis available. The central insight is the Hölder-Minkowski-Riesz-Fischer triple: Hölder provides the pairing inequality identifying the dual; Minkowski provides the triangle inequality identifying the norm; Riesz-Fischer provides completeness identifying the Banach-space structure. This is exactly the structure that generalises across the entire family $L^p$, with the duality $(L^p{)}^{\ast }\cong L^q$ providing the load-bearing connection between $p$ and its conjugate.

The pattern recurs through three escalations. First, $L^2$ acquires Hilbert-space structure via the inner product $⟨f,g⟩=\int f\bar{g}d\mu$, supporting orthonormal expansions, Bessel inequality, Parseval identity, and the spectral theorem for self-adjoint operators. Second, $L^p$ for $1<p<\infty$ acquires uniform convexity (Clarkson 1936) and reflexivity (Milman-Pettis), enabling weak compactness arguments (Banach-Alaoglu + Eberlein-Šmulian) that drive weak-solution existence theory in PDE. Third, the interpolation theorems of Marcinkiewicz 1939 and Riesz-Thorin 1956 promote inequalities on the endpoints $(L^1,L^{\infty })$ and $(L^2,L^2)$ to inequalities on the entire range $L^p$ for $p$ in the convex hull, the foundational fact behind Fourier-analytic boundedness theorems (Hausdorff-Young 1923, Hardy-Littlewood-Sobolev 1928, Calderón-Zygmund 1952). Putting these together identifies $L^p$-theory as the bridge between abstract Banach-space theory, harmonic analysis, distribution theory (Schwartz 1950), Sobolev embeddings, and the modern theory of weak solutions to PDE. The pattern generalises to Lorentz spaces $L^{p,q}$ (Lorentz 1950 Pac. J. Math. 1, 411), Orlicz spaces (Orlicz 1932), Besov and Triebel-Lizorkin spaces, and the modern framework of function spaces in harmonic analysis.

Full proof set Master

Proposition 1 (Young's inequality for non-negative reals). For $a,b\ge 0$ and $p,q\in (1,\infty )$ with $1/p+1/q=1$, $$ab\le \frac{a^p}{p}+\frac{b^q}{q},$$ with equality iff $a^p=b^q$.

Proof. If $a=0$ or $b=0$ both sides are non-negative and the inequality is direct. Assume $a,b>0$. The exponential function is convex; equivalently, the logarithm is concave. Write $ab=\exp (\log a+\log b)$ and use Jensen's inequality on the two-point convex combination with weights $1/p$ and $1/q$ (which sum to $1$): $$\log (ab)=\frac{1}{p}\log (a^p)+\frac{1}{q}\log (b^q)\le \log \left(\frac{a^p}{p}+\frac{b^q}{q}\right).$$ Exponentiating gives $ab\le a^p/p+b^q/q$. Equality in Jensen requires the two points to coincide, i.e., $a^p=b^q$.

$\square$

Proposition 2 (Hölder via Young). Hölder's inequality follows from Young's inequality by scaling and integration.

Proof. As in Exercise 7. Scale to $\Vert f{\Vert }_p=\Vert g{\Vert }_q=1$, apply Young pointwise with $a=|F(x)|,b=|G(x)|$, integrate to get $\int |FG|\le 1/p+1/q=1$, and undo the scaling.

$\square$

Proposition 3 (Minkowski via Hölder). Minkowski's inequality for $p\in (1,\infty )$ follows from Hölder.

Proof. Write $|f+g{|}^p\le |f+g{|}^{p-1}(|f|+|g|)$ and apply Hölder to each of the two resulting integrals with conjugate exponents $(p,q)$ where $q=p/(p-1)$ (so $(p-1)q=p$): $$\int |f+g{|}^{p-1}|f|d\mu \le {\left(\int |f+g{|}^{p}d\mu \right)}^{1/q}\cdot \Vert f{\Vert }_p,$$ and similarly for $|g|$. Adding and dividing both sides by $\Vert f+g{\Vert }_p^{p/q}=\Vert f+g{\Vert }_p^{p-1}$: $$\Vert f+g{\Vert }_p^p\cdot \Vert f+g{\Vert }_p^{-(p-1)}\le \Vert f{\Vert }_p+\Vert g{\Vert }_p,$$ i.e., $\Vert f+g{\Vert }_p\le \Vert f{\Vert }_p+\Vert g{\Vert }_p$. The endpoints $p=1$ and $p=\infty$ are direct.

$\square$

Proposition 4 (uniform integrability and DCT control of $L^p$-convergence). Let $f_n\to f$ a.e. on a measure space with $|f_n|\le g\in L^p$ a.e. Then $f_n\to f$ in $L^p$ for any $p\in [1,\infty )$.

Proof. Apply the dominated convergence theorem 02.07.05 to the non-negative sequence $|f_n-f{|}^p$, which converges to $0$ a.e. and is dominated by $(2g{)}^p=2^p{g}^p\in L^1$ (using $|f|\le g$ a.e., from passing to the limit in $|f_n|\le g$). DCT gives $\int |f_n-f{|}^{p}d\mu \to 0$, i.e., $\Vert f_n-f{\Vert }_p\to 0$.

$\square$

Proposition 5 (the dual pairing in $\sigma$-finite duality). For $p\in (1,\infty )$ and $g\in L^q$, the functional $T_g(f)={\int }_{X}fgd\mu$ is bounded on $L^p$ with operator norm $\Vert T_g{\Vert }_{(L^p{)}^{\ast }}=\Vert g{\Vert }_q$.

Proof. Hölder gives $|T_g(f)|\le \int |fg|\le \Vert f{\Vert }_p\Vert g{\Vert }_q$, so $\Vert T_g\Vert \le \Vert g{\Vert }_q$. For the reverse, choose $f_0=|g{|}^{q-1}\mathrm{sgn}(\bar{g})/\Vert g{\Vert }_q^{q/p}$ when $g\ne 0$; then $\Vert f_0{\Vert }_p=1$ (computation using $(q-1)p=q$) and $T_g(f_0)=\int |g{|}^q/\Vert g{\Vert }_q^{q/p}d\mu =\Vert g{\Vert }_q^q/\Vert g{\Vert }_q^{q/p}=\Vert g{\Vert }_q^{q-q/p}=\Vert g{\Vert }_q$ (using $q-q/p=q(1-1/p)=q/q=1$). The inequality $\Vert T_g\Vert \ge \Vert g{\Vert }_q$ follows. Combining, $\Vert T_g\Vert =\Vert g{\Vert }_q$.

$\square$

Proposition 6 (Riesz representation surjectivity on $\sigma$-finite spaces). Every bounded linear functional $T\in (L^p{)}^{\ast }$ for $p\in [1,\infty )$ on a $\sigma$-finite measure space has the form $T=T_g$ for some $g\in L^q$.

Proof sketch. For $\sigma$-finite $\mu$, decompose $X={⨆}_k{X}_k$ with $\mu (X_k)<\infty$. On each $X_k$, the bounded functional $T$ restricts to $T_k:L^p(X_k)\to \mathbb{C}$. The signed measure ${\nu }_k(E)=T({\chi }_{E\cap X_k})$ is absolutely continuous with respect to $\mu {|}_{X_k}$ (by the bound $|T({\chi }_E)|\le \Vert T\Vert \cdot \mu (E{)}^{1/p}$). The Radon-Nikodym theorem 02.07.07 gives a density $g_k\in L^1(X_k)$. Patching the $g_k$ together gives $g$ on $X$. The bound $\Vert T\Vert \le \Vert g{\Vert }_q$ from Proposition 5 plus Hölder gives $g\in L^q$. The functional $T-T_g$ vanishes on the dense subspace of simple functions, hence on all of $L^p$ by continuity, so $T=T_g$.

The full proof handles the endpoint $p=1$ (giving $g\in L^{\infty }$) and the $\sigma$-finiteness assumption (without which $(L^1{)}^{\ast }\ne L^{\infty }$ can fail).

$\square$

Proposition 7 (Clarkson's inequality for $p\ge 2$). For $p\in [2,\infty )$ and $f,g\in L^p$, $${\Vert \frac{f+g}{2}\Vert }_p^p+{\Vert \frac{f-g}{2}\Vert }_p^p\le \frac{1}{2}(\Vert f{\Vert }_p^p+\Vert g{\Vert }_p^p).$$

Proof. The pointwise inequality $|a+b{|}^p+|a-b{|}^p\le 2^{p-1}(|a{|}^p+|b{|}^p)$ for $a,b\in \mathbb{C}$ and $p\ge 2$ follows from convexity of $t\mapsto t^{p/2}$ for $p\ge 2$: write $|a+b{|}^p=(|a+b{|}^2{)}^{p/2}$ and use $|a+b{|}^2+|a-b{|}^2=2(|a{|}^2+|b{|}^2)$ (the parallelogram law in $\mathbb{C}$), then apply $(x+y{)}^{p/2}\le 2^{p/2-1}(x^{p/2}+y^{p/2})$ to $x=|a+b{|}^2$ and $y=|a-b{|}^2$. Integrating gives Clarkson with the constant $2^{p-1}$ on the right (after dividing by $2^p$).

$\square$

Proposition 8 (uniform convexity from Clarkson). Clarkson's inequality implies $L^p$ is uniformly convex for $p\in (1,\infty )$.

Proof. Uniform convexity means: for every $\epsilon >0$ there is $\delta >0$ such that $\Vert f{\Vert }_p,\Vert g{\Vert }_p\le 1$ and $\Vert f-g{\Vert }_p\ge \epsilon$ imply $\Vert (f+g)/2{\Vert }_p\le 1-\delta$. Clarkson's inequality (Proposition 7 for $p\ge 2$) gives: $${\Vert \frac{f+g}{2}\Vert }_p^p\le \frac{1}{2}(\Vert f{\Vert }_p^p+\Vert g{\Vert }_p^p)-{\Vert \frac{f-g}{2}\Vert }_p^p\le 1-(\epsilon /2{)}^p.$$ So $\Vert (f+g)/2{\Vert }_p\le (1-(\epsilon /2{)}^p{)}^{1/p}$, and we can take $\delta =1-(1-(\epsilon /2{)}^p{)}^{1/p}>0$. For $p\in (1,2]$ the dual Clarkson inequality (replacing the exponents $p\mapsto q$, $p$-power $\mapsto q$-power) closes the case via a parallel argument.

$\square$

Connections Master

• Lebesgue integral and the monotone convergence theorem 02.07.04. The direct prerequisite for the Riesz-Fischer proof: MCT applied to the partial sums $G_K\uparrow G$ of the absolutely-summed envelope converts a Cauchy sequence into a $p$-integrable bound, the load-bearing step of the completeness argument. Without MCT the envelope construction loses control of the $L^p$-norm at the limit.

• Fatou's lemma and the dominated convergence theorem 02.07.05. The just-shipped peer: DCT applied to $|f_n-f{|}^p$ with dominator $(2g{)}^p$ gives the $L^p$-convergence step (Step 4 of the Riesz-Fischer proof). The same DCT-with-$L^p$-domination argument identifies $L^p$-convergence with a.e.-convergence plus $L^p$-domination, the standard tool for parameter integrals in $L^p$.

• Lebesgue outer measure and Carathéodory construction 02.07.02. Supplies the measure-theoretic foundation on which $L^p$-spaces are built. The completeness of Lebesgue measure (every subset of a null set is measurable) underlies the a.e.-equivalence quotient defining $L^p$ as a true normed space rather than a seminormed one.

• Banach spaces and completeness 02.11.04. The abstract framework. $L^p$ for $p\in [1,\infty ]$ is the canonical example of a Banach space arising from analysis: the completeness theorem of Riesz-Fischer realises an explicit infinite-dimensional Banach space via integration theory. For $p=2$ the inner product $⟨f,g⟩=\int f\bar{g}d\mu$ promotes the Banach structure to a Hilbert-space structure, the prototype for spectral theory and quantum mechanics.

• Fourier transform and Plancherel identity 02.10.04. $L^2$ is the natural domain for the Fourier transform as a unitary operator (Plancherel 1910); the Hausdorff-Young inequality (Theorem 8) extends Fourier-boundedness to the entire range $L^p\to L^q$ for $p\in [1,2]$. The interpolation theorems of Marcinkiewicz 1939 and Riesz-Thorin 1956 use the $L^p$ scale as a smooth one-parameter family of Banach spaces between the $L^1$-endpoint and the $L^2$- (or $L^{\infty }$-) endpoint, the foundational fact behind modern Fourier analysis.

Historical & philosophical context Master

Hölder's 1889 Nachrichten von der Königlichen Gesellschaft der Wissenschaften zu Göttingen paper ^{[Hölder 1889]} introduced the inequality $\sum a_k{b}_k\le (\sum a_k^p{)}^{1/p}(\sum b_k^q{)}^{1/q}$ for finite-dimensional positive vectors and conjugate exponents $1/p+1/q=1$. Hölder framed the inequality as a generalisation of the Cauchy-Schwarz inequality (Bunyakovsky 1859 C. R. Acad. Sci. St. Petersburg; Schwarz 1885) from the special case $p=q=2$ to the entire range $p>1$. The extension to integration measures and infinite-dimensional sequence spaces is due to F. Riesz 1910.

Minkowski's 1896 monograph Geometrie der Zahlen ^{[Minkowski 1896]} introduced the triangle inequality for ${\ell }^p$-norms in the context of his geometry-of-numbers programme: the lattice-point counting problems Minkowski was studying required convexity-based estimates of which the Minkowski inequality is the prototype. The extension to integration measures is again due to Riesz 1910.

Frigyes Riesz's and Ernst Fischer's independent 1907 Comptes Rendus de l'Académie des Sciences à Paris notes ^{[Riesz 1907; Fischer 1907]} proved completeness of $L^2$ in the context of orthonormal expansions. Riesz's framing (15 March 1907) was via systems of orthogonal functions: the closed span of an orthonormal sequence in $L^2([a,b])$ is itself complete. Fischer's framing (7 May 1907) was via convergence-in-mean: a sequence of $L^2$-functions converging in $L^2$-norm has an $L^2$-limit. The two formulations are equivalent. The result resolved David Hilbert's question (asked in 1906 lectures) about whether the completeness of ${\ell }^2$ (Hilbert sequence space) had an analogue for function spaces.

Riesz's 1910 Mathematische Annalen paper ^{[Riesz 1910]} introduced $L^p$-spaces for general $p\in [1,\infty )$ (Riesz used the notation $L_p$), proved Hölder and Minkowski in this generality, established Riesz-Fischer completeness, and identified the duality $(L^p{)}^{\ast }=L^q$ on $[0,1]$. Riesz's framework became the canonical formulation; the modern textbook treatment (Halmos 1950, Folland 1984, Rudin 1966, Brezis 2010) is a direct descendant of Riesz 1910 with the addition of the $\sigma$-finite generalisation via Radon-Nikodym (Radon 1913, Nikodym 1930).

Young's 1912 Proceedings of the London Mathematical Society paper ^{[Young 1912]} introduced the Young inequality $ab\le a^p/p+b^q/q$ in the context of Fourier-coefficient estimates, motivating Hausdorff's 1923 Mathematische Zeitschrift extension ^{[Hausdorff 1923]} of Plancherel's identity to the entire $L^p\to L^q$ range for $p\in [1,2]$. The sharp constant in the Hausdorff-Young inequality was open for over fifty years until Beckner's 1975 Annals of Mathematics paper ^{[Beckner 1975]} identified it via Gaussian extremisers and Babenko's earlier 1961 spherical-rearrangement principle.

Clarkson's 1936 Transactions of the American Mathematical Society paper ^{[Clarkson 1936]} introduced the uniform-convexity inequality bearing his name and used it to prove $L^p$ uniformly convex for $1<p<\infty$. The reflexivity consequence (uniform convex $⟹$ reflexive) was proved by Milman 1938 C. R. Acad. Sci. URSS and independently by Pettis 1939 Bull. AMS 45. Clarkson's work was motivated by James's 1950 reflexivity-characterisation programme and forms the analytical backbone of weak-compactness arguments in modern Banach-space theory.

Schwartz's 1950 monograph Théorie des Distributions ^{[Schwartz 1950]} developed distribution theory on tempered functions, in which the $L^p$-spaces appear as concrete instances of locally convex spaces of integrable functions. Schwartz's framework promotes $L^p$ from a Banach-space-theoretic object to a building block of modern PDE theory: weak derivatives, Sobolev spaces, and tempered distributions all rest on the $L^p$-machinery.

The structural arc is a one-hundred-year story: Hölder 1889 and Minkowski 1896 (the two inequalities), Riesz 1907 and Fischer 1907 (completeness of $L^2$), Riesz 1910 (general $L^p$ and duality), Young 1912 and Hausdorff 1923 (Fourier extension), Clarkson 1936 (uniform convexity), Schwartz 1950 (distribution-theoretic framework), Beckner 1975 (sharp Fourier constants). Each step extends the previous to a wider domain or a sharper inequality, and the result is the canonical infinite-dimensional analogue of the Euclidean norm, the bedrock of twentieth-century analysis.

Bibliography Master

@article{Holder1889,
  author  = {H\"older, Otto},
  title   = {Ueber einen {M}ittelwertsatz},
  journal = {Nachrichten von der K\"oniglichen Gesellschaft der Wissenschaften zu G\"ottingen},
  year    = {1889},
  pages   = {38--47}
}

@book{Minkowski1896,
  author    = {Minkowski, Hermann},
  title     = {Geometrie der Zahlen},
  publisher = {Teubner},
  address   = {Leipzig},
  year      = {1896}
}

@article{Riesz1907,
  author  = {Riesz, Frigyes},
  title   = {Sur les syst\`emes orthogonaux de fonctions},
  journal = {Comptes Rendus de l'Acad\'emie des Sciences \`a Paris},
  volume  = {144},
  year    = {1907},
  pages   = {615--619}
}

@article{Fischer1907,
  author  = {Fischer, Ernst},
  title   = {Sur la convergence en moyenne},
  journal = {Comptes Rendus de l'Acad\'emie des Sciences \`a Paris},
  volume  = {144},
  year    = {1907},
  pages   = {1022--1024}
}

@article{Riesz1910,
  author  = {Riesz, Frigyes},
  title   = {Untersuchungen \"uber {S}ysteme integrierbarer {F}unktionen},
  journal = {Mathematische Annalen},
  volume  = {69},
  year    = {1910},
  pages   = {449--497}
}

@article{Young1912,
  author  = {Young, William Henry},
  title   = {On the multiplication of successions of {F}ourier constants},
  journal = {Proceedings of the London Mathematical Society},
  series  = {2},
  volume  = {11},
  year    = {1912},
  pages   = {357--366}
}

@article{Hausdorff1923,
  author  = {Hausdorff, Felix},
  title   = {Eine {A}usdehnung des {P}arsevalschen {S}atzes \"uber {F}ourierreihen},
  journal = {Mathematische Zeitschrift},
  volume  = {16},
  year    = {1923},
  pages   = {163--169}
}

@article{Clarkson1936,
  author  = {Clarkson, James A.},
  title   = {Uniformly convex spaces},
  journal = {Transactions of the American Mathematical Society},
  volume  = {40},
  year    = {1936},
  pages   = {396--414}
}

@article{Friedrichs1944,
  author  = {Friedrichs, Kurt O.},
  title   = {The identity of weak and strong extensions of differential operators},
  journal = {Transactions of the American Mathematical Society},
  volume  = {55},
  year    = {1944},
  pages   = {132--151}
}

@book{Schwartz1950,
  author    = {Schwartz, Laurent},
  title     = {Th\'eorie des {D}istributions},
  publisher = {Hermann},
  address   = {Paris},
  year      = {1950}
}

@article{Beckner1975,
  author  = {Beckner, William},
  title   = {Inequalities in {F}ourier analysis},
  journal = {Annals of Mathematics},
  volume  = {102},
  year    = {1975},
  pages   = {159--182}
}

@book{Folland1999,
  author    = {Folland, Gerald B.},
  title     = {Real Analysis: Modern Techniques and Their Applications},
  edition   = {2},
  publisher = {Wiley},
  year      = {1999}
}

@book{Rudin1987,
  author    = {Rudin, Walter},
  title     = {Real and Complex Analysis},
  edition   = {3},
  publisher = {McGraw-Hill},
  year      = {1987}
}

@book{Brezis2010,
  author    = {Brezis, Ha\"im},
  title     = {Functional Analysis, {S}obolev Spaces and Partial Differential Equations},
  publisher = {Springer},
  year      = {2010}
}

@book{Bogachev2007,
  author    = {Bogachev, Vladimir I.},
  title     = {Measure Theory, Volume 1},
  publisher = {Springer},
  year      = {2007}
}

@book{RoydenFitzpatrick2010,
  author    = {Royden, H. L. and Fitzpatrick, P. M.},
  title     = {Real Analysis},
  edition   = {4},
  publisher = {Pearson},
  year      = {2010}
}

@book{Tao2016,
  author    = {Tao, Terence},
  title     = {Analysis II},
  edition   = {3},
  publisher = {Hindustan Book Agency},
  year      = {2016}
}