Prerequisites: differentiation and the mean value theorem, Riemann integration and uniform convergence.
Setting and notation
Throughout, and are real numbers, and is a compact rectangle. All integrals are proper Riemann integrals over .
Definition 17.1.1
(Partial derivative in the parameter)
For a function , write
At the limit is taken over only, and at over only. These are one-sided partial derivatives.
Define
If is continuous on , then for each fixed the map is continuous on . It is therefore Riemann integrable, so is well defined.
The theorem
Theorem 17.1.2
(Differentiation under the integral sign)
Suppose is continuous on , exists at every point of (one-sided on the edges and ), and is continuous on . Then is differentiable at every , and
The same formula holds for the right derivative at and the left derivative at . Moreover, is continuous on .
Proof
Proof
(Uniform continuity and the mean value theorem)
Step 1: the integrand on the right is integrable. For fixed , the map is continuous on , so exists.
Step 2: the uniform estimate. The set is compact and is continuous on , so is uniformly continuous on . Hence, given , there is such that
The crucial point is that depends only on , not on . Pointwise continuity in for each fixed would not give this.
Step 3: the difference quotient. Fix and let with and . Fix . The function is continuous on the closed interval between and and differentiable on its interior. By the Mean Value Theorem there is strictly between and with
Since , Step 2 gives
The point depends on , but we use it only to obtain this pointwise bound. Next, note that
is continuous in , hence integrable, and satisfies for every . By linearity and monotonicity of the Riemann integral,
Step 4: conclusion. Since was arbitrary, the difference quotient tends to as .
- For , both signs of are allowed, so .
- At only is admissible, giving the right derivative.
- At only is admissible, giving the left derivative.
The MVT step at the edges uses only points strictly inside , so it is valid there as well.
Step 5: continuity of . If , then by Step 2,
So is (uniformly) continuous.
Proof X-ray
Where uniformity enters and what closes the argument
Each hypothesis does one job. Continuity of makes well defined, and together with continuity of it makes every integrable. Continuity of on the compact set gives uniform continuity, which is the only place where the size of matters: one must serve every at once. The mean value theorem is applied separately for each , so the intermediate point varies with in an uncontrolled way. This is harmless because is never integrated; only the bound that it delivers is.
The non-obvious move is to compare the difference quotient with at the same , not with a limit taken after integrating. The proof closes with the elementary fact that on implies . Equivalently, the difference quotients converge to uniformly in , and uniform limits may be passed through Riemann integrals. No dominated convergence theorem is involved.
Without uniformity the closing estimate has no to use: the warning section on pointwise limits gives an example in which the quotient at equals , so no single exists.
Worked examples
Worked example 17.1.3
(Derivative of the integral of arctan(tx))
Let on for some . Then
Both and are continuous on the rectangle. The theorem therefore gives, for ,
If : substitute , so . This gives
If : the integrand is , so .
This agrees with the limit , as Step 5 predicts.
Independent check. For , integrate by parts with and :
Differentiating,
This matches the result from the theorem.
Worked example 17.1.4
(A one-sided derivative at the edge of the parameter interval)
Let on , so . Here , and both functions are continuous on . For , integration by parts gives
At the edge the theorem gives the right derivative . Check: for with , so the right derivative at is . The same series shows that the closed formula above tends to as , as continuity of requires.
Worked example 17.1.5
(A hypothesis that fails at one corner of the rectangle)
Let on . Since , both and are continuous on . The theorem gives
which agrees with differentiating directly. If the parameter interval were instead, the point would lie in the rectangle, where is not even defined, and the hypotheses fail. The failure is real: diverges, so is not defined at .
Warning: pointwise limits are not enough
Counterexample mode
A derivative that exists at every point but cannot be moved inside
Claim: if is continuous for each and exists at every point of the rectangle, then . Define on
The partial derivative at is zero for every .
- For : as .
- For : for all , so the quotient is .
So for every , and .
The integral itself has derivative at . For ,
and . Hence .
What went wrong. The difference quotients converge to for each , but not uniformly. For , the point is in the domain and the quotient equals
which is unbounded as . Neither nor is continuous at . Indeed, along the admissible diagonal , . Moreover, away from ,
which is unbounded as , whereas . In particular the continuity hypothesis used in Step 2 fails. The repair is the theorem's hypothesis: continuity of on the whole compact rectangle.
Sufficient, not necessary
The hypotheses of the theorem are sufficient, not necessary. Let
and set on . Then is differentiable everywhere but is discontinuous at , so is not continuous on the rectangle. Nevertheless,
for every . The interchange holds even though the theorem does not apply. Failure of a hypothesis means the theorem is silent; it does not mean the conclusion is false.
Situations requiring separate care
Variable limits. For , the theorem alone does not apply. Additional assumptions are needed:
- and are differentiable;
- their values stay in .
Then one combines the theorem with the Fundamental Theorem of Calculus and the chain rule applied to . This yields
This requires checking that is differentiable as a function of three variables, which is a separate argument.
Common mistake
(Dropping the boundary terms when the limits move)
Writing ignores the two boundary terms. Take and . Then , so the incorrect formula gives , whereas . The correct formula supplies the missing term .
Improper integrals. On or with an unbounded integrand, there is no compact rectangle. Uniform continuity in Step 2 can then fail, and the bound becomes useless.
For example, equals for every , so its derivative is . Yet the formally differentiated integrand is not improperly integrable on . Such cases need extra hypotheses, for instance uniform convergence of , proved by separate arguments.
Quick checks
Quick check
(Why must the number in Step 2 be independent of ?)
Look at which estimate is integrated over .
Solution · Answer
The bound is integrated over all for one fixed . This needs a single , so that every with gives the bound at every at once. If depended on , the infimum of over could be zero and no admissible would remain. Compactness of is what turns continuity of into such a uniform .
Quick check
(Let for . Use the theorem to find , and check it against the closed form of .)
Compute and integrate by parts.
Solution · Answer
Here is continuous on , so . Integration by parts gives , so . Since , differentiating directly gives the same expression.
Quick check
(Let for . Compute from the variable-limit formula, and verify it by evaluating first.)
Here , and .
Solution · Answer
The formula gives . Directly, , whose derivative is . The two agree; the boundary term accounts for two thirds of the answer.