Hölder’s inequality is repeated Cauchy–Schwarz

Cauchy–Schwarz and Hölder’s inequality are the basic tools for controlling the interaction, or correlation, of two functions. For finite sequences and , their correlation is measured by the inner product

The triangle inequality reduces the problem to estimating Thus the central question is this: how can one control the total interaction using only separate information about the sizes of and Cauchy–Schwarz answers this when both sequences are measured using squares:

Hölder’s inequality is the general version. It says that one may measure and using different powers, provided the powers fit together correctly. If and then

Equivalently, in norm notation,

The deeper point of this discussion is that Hölder is not an independent miracle that sits above Cauchy–Schwarz. Morally, it is Cauchy–Schwarz repeated at finer and finer interpolation scales. Cauchy–Schwarz controls a midpoint. Repeating midpoint estimates reaches dyadic fractions such as , while finite chains of such steps already produce every rational proportion. Continuity then fills in every real interpolation parameter.

Examples:

For Hölder is precisely Cauchy–Schwarz. The next case already shows the main idea.. Let

We want to move from the mixed product toward the two pure endpoint expressions and The first split is

Applying Cauchy–Schwarz gives

This does not finish the proof, because the first factor is still mixed. But it is less mixed than before: the power of has moved from to while the remaining power of has been reduced. We apply Cauchy–Schwarz once more:

Combining the two inequalities gives and hence

Thus Hölder for is simply Cauchy–Schwarz applied twice. The proof works because each application moves the mixed expression one step nearer to a pure endpoint.

The case is the same story, with one additional intermediate stage. Define

and let Cauchy–Schwarz supplies the chain

The final inequality reaches the pure endpoint. Working backwards through the chain gives

and therefore taking eighth roots gives Hölder for

The pattern is now visible. Every Cauchy–Schwarz step doubles the current power of while the unassigned power of is moved into the fixed endpoint quantity These are the easiest cases because the relevant interpolation parameters are dyadic.

The pair has a different appearance. The proof no longer moves along one straight ladder toward an endpoint. Instead, two mixed sums arise and control each other. Repeated Cauchy–Schwarz forms a short cycle. Let

We begin from the product and split it as so Cauchy–Schwarz gives

The first step has created a new mixed quantity It is not an endpoint norm, but it admits a complementary decomposition: and hence applying Cauchy–Schwarz again gives

The two estimates form a loop: The first inequality carries us from to while the second carries us back from to Eliminating the intermediate quantity gives

Dividing by and taking cube roots gives

The two Cauchy–Schwarz estimates form the loop which closes to give Hölder for The point of this example is not merely that it proves one more case. It shows that “repeated Cauchy–Schwarz” need not mean a rigid one-directional iteration. It may produce a small system of inequalities between several mixed sums; solving that system gives the final exponent.

Every rational Hölder exponent

The preceding examples are all instances of one finite construction. Fix an integer , let and define

The endpoints are the two pure sums are The intermediate sums are obtained by transferring one power at a time from to The crucial observation is that each intermediate monomial is exactly the geometric mean of its two neighbors:

Therefore Cauchy–Schwarz gives, for every

Thus the list is log-convex. This is the finite-chain version of midpoint convexity. Writing this says

Equivalently, the discrete slopes increase:

Defining the successive slopes , This ordered list of slopes is the whole mechanism. To estimate , split the total change from to at :

Every slope in the first sum is at most every slope in the second sum. Hence the average slope on the left is at most the average slope on the right:

Rearranging gives . In geometric language, the points lie below the line segment joining the endpoint points and Thus, for every

Exponentiating gives In other words,

This formula already contains Hölder. Choose Then while The chain estimate becomes

This is Hölder with Every rational exponent arises in this way. Indeed, write with integers Taking gives

There is also a completely multiplicative elimination procedure, which shows exactly how the local Cauchy–Schwarz inequalities combine. Define

We want to multiply powers of the so that every interior cancels except . The required powers are not guessed. Set and require that the exponent of in the product vanish for every . Since , this cancellation condition is Thus the weights must change at a constant rate on either side of , so they are forced to be the piecewise-linear sequence

At , the slope has a jump of size ; this produces precisely the desired power . Expanding the product now gives the exact identity

Since every and every the product is at most , proving again that .

The cycle above is exactly the case and Indeed, taking and gives

The two local midpoint inequalities are precisely

Eliminating yields

Likewise, and give

We want to start from and remove every quantity except itself and the endpoints . First remove using , obtaining . The only remaining unwanted term is now . To remove it, use its local relation ; this introduces , but only as a temporary obstruction. Remove that new term immediately with , which follows from . Hence , so . Substituting this into leaves only : . Thus , and therefore . Equivalently,

which is Hölder for

Thus finite repeated Cauchy–Schwarz arguments prove Hölder for every rational conjugate pair.

Cauchy–Schwarz as midpoint interpolation

The general mechanism is clearest when one removes the specific exponents. Let and define

The endpoints are and The parameter measures how much of the exponent has been transferred from to Now take two parameters At their midpoint,

Cauchy–Schwarz therefore gives the midpoint inequality

Whenever is positive, this is equivalent to

Thus is convex: the logarithm of the mixed sum lies below the straight line joining its endpoint values. This is the exact analytic content of repeated Cauchy–Schwarz.

Applying the midpoint estimate first to and then to and and so on, gives

for every dyadic rational Since is continuous, dyadic rationals approximate every and hence

This is the basic interpolation inequality. Its proof is nothing more than Cauchy–Schwarz at all dyadic midpoint scales, followed by a limiting argument.

Now let satisfy Set

Then and the mixed monomial becomes

The interpolation inequality is therefore exactly Hölder’s inequality:

Thus Hölder is the continuous completion of repeated Cauchy–Schwarz: Cauchy–Schwarz gives midpoint control, finite chains give rational positions, and continuity gives every real position.

Convexity proof

The standard proof uses convexity. For and conjugate exponents weighted arithmetic-geometric mean gives

This is Young’s inequality, and it follows from convexity of the exponential function. Normalize the sequences by setting

Then Summing Young’s inequality gives

This proves Hölder immediately. It is often the quickest route, but the interpolation proof explains the deeper geometry: Hölder is the log-convex continuation of Cauchy–Schwarz from one midpoint to every point between two endpoints.

The same argument works for integrals. One replaces finite sums by integrals, applies the finite or simple-function case first, and then passes to general measurable functions by approximation. Thus Hölder is the fundamental principle that a product can be controlled by assigning complementary portions of its exponent to the two factors.