Knowing a function is integrable to two different powers gives information at every intermediate power. The relevant parameter is the reciprocal of the exponent: it is , not , that lies between and . Hölder's inequality turns that arithmetic into a precise norm estimate. The same calculation also clarifies why the size of the underlying space matters when comparing two classes.
Let be a measure space, not necessarily finite or -finite. Functions may be real or complex, and are identified almost everywhere. We use Hölder's inequality and the standard essential-supremum norm. Write in exponent formulas.
Interpolating a single function
Theorem 4.5.1
(Interpolation of a single function between two Lp norms)
Let , and choose such that
The parameter is uniquely determined by . If , then and
Proof
(Hölder with weights summing to one)
Case . Put and . Both are positive and . Hölder with conjugate exponents and gives
Taking th roots gives (1). The argument also covers the zero function; no division by a norm is needed.
Case . Here . Since almost everywhere,
Taking th roots yields (1), including when either norm is zero.
Proof X-ray
Why the reciprocal exponent is the right parameter
The single move is to split as . This factorization is exact because , and Hölder applies because , which is the defining equation for multiplied by . Both conditions hold only when is the convex combination; a convex combination of itself would satisfy neither. Finiteness of both endpoint norms is used to make the right side finite, so that follows. The case needs no Hölder at all: the bound absorbs the excess power .
For a nonzero with finite endpoint norms, set for with and when . Taking logarithms of (1) gives : convexity in the reciprocal exponent. The theorem concerns a fixed function and a common measure space; it does not assert that every function belongs to both endpoint spaces.
Quick check
(For , and , what is ?)
Solve the reciprocal equation, not a linear equation in the exponents.
Solution · Answer
From , multiply by : , so . Then (1) reads .
Common mistake
(Interpolating the exponent instead of its reciprocal)
Writing gives the wrong weight. For , , it suggests , whereas gives . The wrong version is false: for with , but .
Worked example 4.5.2
(A two-step function between and )
On with Lebesgue measure, let . For , and , . Here , and , so
The constant one in (1) cannot be improved in general: whenever a set has , the choice with gives , equality for every choice of the stated exponents.
Finite total measure and inclusion
Theorem 4.5.3
(Inclusion on a finite measure space)
Assume and . Then
so . A nonzero constant function gives equality.
Proof
(Hölder against the constant function)
For , apply Hölder to with exponents and :
Taking th roots gives (2). For the same formula follows directly from . For both sides equal .
If , all the corresponding equivalence classes are zero and the inclusion is immediate.
Worked example 4.5.4
(The reverse inclusion fails on the unit interval)
On with Lebesgue measure and finite , choose with and set . Then gives , while gives . So . For , any gives an function that is not essentially bounded. The boundary case is excluded because then .
Infinite measure changes the comparison
On infinite-measure spaces the inclusion (2) is unavailable, and the failure is concrete.
Counterexample mode
Without finite measure, neither inclusion holds
Claim: for on every measure space. On with Lebesgue measure and finite , choose and let . Then gives , while gives . So . Near zero the opposite happens: belongs to but not . For , the constant one is in but not . The hypothesis in (2) is what fails.
These are examples, not a claim that every infinite-measure space has precisely the same inclusion behavior. Formula (1), unlike (2), remains valid on an arbitrary measure space because it assumes both endpoint norms are finite.
A warning about operators
Suppose a linear operator is defined on and its outputs satisfy endpoint estimates and . Applying (1) to proves
Common mistake
(Reversing inequality (1) to get an operator bound)
It is tempting to replace the product on the right of (3) by . Inequality (1) points in the opposite direction, so this step is not justified by anything proved here.
Worked example 4.5.5
(The product of endpoint norms can dwarf the middle norm)
On let . Then , so . Also and , so , an unbounded multiple of . This only shows that the displayed proof does not establish an intermediate operator bound; it does not show such a bound fails. Operator interpolation theorems can supply stronger conclusions under their own hypotheses, but they require an additional argument not given here.
Practice
Quick check
(If and , , bound .)
Solution · Answer
Here , so .
Quick check
(Why does on not contradict the interpolation theorem for , ?)
Solution · Answer
It is bounded but not integrable, so the hypothesis fails.
Quick check
(On a probability space, compare and .)
Solution · Answer
With , (2) gives . So and the norm never exceeds the norm; constants give equality.