Controlling products and sums
Integrable functions can be too large to multiply: on , is integrable, but its square is not. The scale gives a precise way to distribute integrability between two factors. Throughout, is a measure space, functions are measurable and equal almost everywhere functions are identified.
Definition 4.1.1
(The norm and conjugate exponents)
For , set
The conjugate exponent to is defined by . The endpoint convention is when , with the essential supremum of .
Young and Hölder
Lemma 4.1.2
(Young's inequality)
If and are conjugate, then , with equality exactly when .
Proof
(Weighted arithmetic–geometric mean)
The weighted arithmetic–geometric mean inequality with weights and gives , with equality exactly when , including the zero case.
Young's inequality is pointwise; Hölder's inequality integrates it after a normalization.
Theorem 4.1.3
(Hölder's inequality)
If and , where , then and
At the endpoint , on any measure space.
Proof
(Normalize, apply Young pointwise, integrate)
If either norm is zero, the corresponding function vanishes almost everywhere and the assertion is immediate. Otherwise apply Young pointwise to and . Since , integrating gives . Multiply by the norms. At the endpoint, almost everywhere by the essential bound; integrate, with no finite-measure assumption.
For nonzero in the non-endpoint case, equality in Hölder occurs precisely when the normalized powers and agree almost everywhere: the nonnegative Young gap must have integral zero. For the estimate , equality additionally depends on the phase of ; equality in Hölder for absolute values alone does not impose that phase condition.
Minkowski and the triangle inequality
Theorem 4.1.4
(Minkowski's inequality)
For and , .
Proof
(Split and apply Hölder to each piece)
The case follows from the pointwise triangle inequality. Let . First , by convexity of , shows that . Put . Then , and since , Hölder applied to each term yields
If there is nothing to prove; otherwise divide by .
Proof X-ray
Why the finiteness step comes first
The proof has one idea and one precaution. The idea is to write and put the factor in : the identity makes its norm a power of itself, so the inequality becomes one about a single number. The precaution is the crude bound . Without it, might be , and dividing by would prove nothing. Hölder is used twice, once for each summand, and the conclusion closes by dividing out the common power. Together with homogeneity and the fact that precisely when almost everywhere, this makes a normed vector space of equivalence classes.
Density of simple functions with finite-measure support
Throughout this subsection, is an arbitrary measure space, , and is or .
Definition 4.1.5
(Simple functions with finite-measure support)
A simple function is a measurable that takes only finitely many values. Equivalently, with and , and every value of is a finite number. Let be the set of simple with .
If , then , so .
Theorem 4.1.6
(Density of )
is dense in for .
Proof
(Truncate to a set of finite measure, then quantize)
Fix . The set is -null because , and setting there does not change the class. So we may assume is finite at every point.
Step 1 (truncation). Put . Since on ,
If , then for every . If , then for all large , because is finite. Hence pointwise. Moreover , so dominated convergence gives . The argument uses only , so it applies equally to real and complex .
Step 2 (quantization). Let . For and real , put , so that . For complex , put , so that . Take in the real case and in the complex case, and define .
Since is Borel, is measurable. On the real and imaginary parts of lie in , so only finitely many grid values occur. Thus , with and . Therefore
By Minkowski's inequality, .
Common mistake
(Approximating uniformly without controlling the support)
A uniform bound gives only . This is useless when has infinite measure: the constant on with Lebesgue measure is uniformly small but is not in . For this reason the proof first truncates to and then scales by . No -finiteness of is used: finiteness of comes from integrability of alone, so is automatically -finite even when is not.
Worked example 4.1.7
(Simple approximants of a decaying exponential)
On with Lebesgue measure, let . For an integer , define for and for . Its values lie in , it is measurable, and , so .
On we have , so . On we have . Minkowski's inequality then gives .
The theorem genuinely needs . On with Lebesgue measure let . If , then has finite measure, so has infinite, and hence positive, measure. On that set , so for every . The density theorem holds only for .
Quick check
(Let be measurable with , and let with . Show that . Then, on with Lebesgue measure, explain why does not belong to .)
Bound by a constant on .
Solution · Answer
. The function is simple and bounded, but its nonzero set has infinite Lebesgue measure. In fact , so it is not even in for . Boundedness alone is not enough on an infinite-measure set.
Why the space is complete
Theorem 4.1.8
(Riesz–Fischer theorem)
For , every Cauchy sequence in converges in the norm to a member of .
Proof
(A fast subsequence, a dominating function and Fatou)
Let be Cauchy. Choose a subsequence so that . Select measurable representatives and define
Minkowski gives . As , monotone convergence yields . Thus almost everywhere and the telescoping series converges absolutely there. Let be its measurable sum (assign any value on the null set); then , so .
For , the finite-tail triangle inequality gives . For fixed , apply Fatou's lemma to as to obtain
Finally the Cauchy property carries convergence from the subsequence to the full sequence: choose a large and use for all sufficiently large .
Proof X-ray
Where the limit comes from
A Cauchy sequence in need not converge at any point, so the proof cannot start from pointwise limits. The key move is the subsequence with : Minkowski then bounds the partial sums of absolute differences uniformly in , and monotone convergence passes the bound to . Finiteness of is what makes finite almost everywhere, which is the only source of a pointwise limit. Fatou's lemma, not dominated convergence, transfers the norm estimate to the limit, because it needs no integrable majorant for in advance. Minkowski is used twice: for and for the finite tails. The proof closes with the general fact that a Cauchy sequence with a convergent subsequence converges.
Common mistake
(Expecting the whole sequence to converge pointwise)
The proof extracts a subsequence before taking pointwise limits, and this is necessary. Index the dyadic intervals as for , , on . Then , so in ; but each lies in one interval of every generation , so for infinitely many and for infinitely many . The sequence converges pointwise nowhere, although a subsequence converges to almost everywhere.
Examples and limits of the statements
Worked example 4.1.9
(A product of two integrable powers)
On , exactly when , so exactly when . Thus belong to but is not integrable. Hölder is not violated: with , neither factor is in , since .
Worked example 4.1.10
(Equality in Hölder)
On , take and with and . The normalized powers agree because and are proportional. Hence Hölder is an equality: , , and .
Whether one space contains another depends on the measure.
Counterexample mode
Lp inclusion depends on the measure space
Claim: , or , on every measure space. On , but not , since and . On , but not , since and . Thus no universal inclusion holds. On a finite measure space Hölder does give for , but the reverse inclusion can fail.
Practice and solutions
Quick check
(Determine for which the function lies in .)
Compute the exponent of in .
Solution · Answer
The integral of over is finite exactly when , that is, for .
Quick check
(On a space with finite measure, prove .)
Apply Hölder to and .
Solution · Answer
Hölder with applied to and gives .
Quick check
(Let on . Does converge to zero in ? Does it converge pointwise almost everywhere?)
Compute , then fix .
Solution · Answer
for every , so does not converge to zero in norm. For each fixed , once , so it does converge to zero pointwise on . Pointwise convergence does not imply norm convergence.
Quick check
(If in , show .)
Use the triangle inequality in both directions.
Solution · Answer
Minkowski gives and , so .