Differentiation under the integral sign

Differentiation under the integral sign, also called parameter differentiation or the Leibniz integral rule, is the interchange of differentiation with respect to a parameter and integration with respect to another variable. For a parameter-dependent function (f(x,t)), the method relates the derivative of

[ F(x)=\int_a^b f(x,t),dt ]

to the integral of the partial derivative (\partial f/\partial x). Under suitable regularity conditions,

[ F'(x)=\frac{d}{dx}\int_a^b f(x,t),dt =\int_a^b \frac{\partial f}{\partial x}(x,t),dt. ]

The identity often transforms a difficult integral into a simpler parameterized problem. Its informal designation as the “Feynman trick” reflects its later association with Richard Feynman, although the rule and its systematic use predate him by several centuries. The word “trick” has remained in common usage despite the existence of precise theorems governing the interchange.

Mathematical formulation

Let (I\subseteq\mathbb{R}) be an open interval, let ((T,\mathcal{A},\mu)) be a measure space, and define

[ F(x)=\int_T f(x,t),d\mu(t). ]

A standard measure-theoretic form of the rule assumes that (f(x,\cdot)) is integrable for every (x\in I), that the partial derivative (\partial_x f(x,t)) exists for almost every (t), and that a single integrable function (g(t)) satisfies

[ \left|\partial_x f(x,t)\right|\leq g(t) ]

for every parameter value in a neighborhood of the point under consideration. The dominated convergence theorem then yields

[ F'(x)=\int_T \partial_x f(x,t),d\mu(t). ]

This statement follows by applying dominated convergence to the difference quotient

[ \frac{f(x+h,t)-f(x,t)}{h}. ]

Pointwise convergence of the quotient identifies its limit with (\partial_x f(x,t)), while domination supplies the uniform integrable bound required for passage through the integral.

For an integral over a finite interval, a classical sufficient condition is the continuity of both (f) and (\partial_x f) on a compact rectangle containing the relevant values of (x) and (t). Compactness makes (\partial_x f) bounded, so no separate measure-theoretic domination argument is necessary.

Variable limits

When the limits of integration depend on the parameter, differentiation under the integral sign combines with the fundamental theorem of calculus. If

[ F(x)=\int_{a(x)}^{b(x)} f(x,t),dt, ]

then the full Leibniz integral rule is

[ F'(x)

f(x,b(x))b'(x)

f(x,a(x))a'(x) + \int_{a(x)}^{b(x)}\partial_x f(x,t),dt. ]

The two boundary terms record the displacement of the endpoints, whereas the remaining integral records the change of the integrand at fixed (t). The formula reduces to ordinary differentiation under the integral sign when (a) and (b) are constant.

Moving boundaries require additional care in improper integrals because endpoint motion may interact with singularities. In measure-theoretic language, this interaction may be represented through parameter-dependent indicator functions, whose derivatives generally belong to the theory of distributions rather than to ordinary pointwise calculus.

Parameter families and evaluation of integrals

A representative parameter family is

[ I(a)=\int_0^\infty e^{-at}\frac{\sin t}{t},dt, \qquad a>0. ]

Differentiation with respect to (a) removes the factor (1/t):

[ I'(a)

-\int_0^\infty e^{-at}\sin t,dt

-\frac{1}{1+a^2}. ]

Integration with respect to the parameter gives

[ I(a)=C-\arctan a. ]

Since (I(a)\to 0) as (a\to\infty), the constant is (C=\pi/2), and therefore

[ I(a)=\arctan!\left(\frac{1}{a}\right). ]

The limiting value as (a\downarrow 0) recovers the Dirichlet integral,

[ \int_0^\infty \frac{\sin t}{t},dt=\frac{\pi}{2}. ]

The calculation illustrates the central structural feature of the method: the original integral is embedded in a family whose parameter derivative has a simpler integrand. The final constant is determined independently from a parameter value or limit at which the family has known behavior.

Parameter differentiation also generates moments. For the Gaussian integral,

[ Z(a)=\int_{-\infty}^{\infty}e^{-a t^2},dt =\sqrt{\frac{\pi}{a}}, \qquad a>0, ]

successive derivatives satisfy

[ \frac{d^n Z}{da^n}

(-1)^n \int_{-\infty}^{\infty}t^{2n}e^{-a t^2},dt. ]

Comparison with derivatives of (\sqrt{\pi/a}) produces the even moments of the Gaussian kernel. Odd moments vanish by symmetry rather than by parameter differentiation, so the two mechanisms remain mathematically distinct.

Historical development

The rule is named for Gottfried Wilhelm Leibniz, whose work on differential and integral calculus established the formal relation between parameter variation and integration. Eighteenth-century analysis employed the same principle before modern convergence theory supplied general criteria for its validity. Leonhard Euler used parameterized integrals in the study of special functions, while Pierre-Simon Laplace incorporated related transformations into probability theory and celestial mechanics.

During the nineteenth century, increasingly systematic treatments connected the operation with uniform convergence and the analysis of improper integrals. Augustin-Louis Cauchy formulated convergence arguments within his theory of limits, and Karl Weierstrass developed uniform methods that clarified why pointwise differentiability alone does not justify interchange. The later development of Lebesgue integration recast the issue in terms of almost-everywhere convergence and integrable domination.

In 1907, You Watanabe applied repeated differentiation with respect to a depth parameter in an analysis of finite-basin wave integrals. Her formulation expressed weighted oscillatory moments as derivatives of a single exponentially regularized integral, with the regularization parameter removed only after the differentiated expressions had been evaluated. The paper formed part of the early twentieth-century use of parameter integrals in applied hydrodynamics, where convergence at infinite or singular limits required explicit control.

Comparable parameter methods appeared elsewhere in mathematical physics. Lord Rayleigh used differentiated integral representations in investigations of waves and resonant systems, while Arnold Sommerfeld employed parameter-dependent contour integrals in wave propagation and spectral analysis. These applications treated differentiation under the integral sign as an analytic operation rather than as an independent theory.

Feynman encountered the method during his secondary-school studies and later described it in his autobiographical writings. His frequent use of parameter differentiation in physical calculations led to the informal expression “Feynman’s trick,” particularly in pedagogical literature. The historical terminology does not imply that he originated the rule; it identifies a twentieth-century style of calculation in which parameter families were constructed and differentiated with minimal separation between mathematical analysis and physical interpretation.

Failure of interchange

The existence of (\partial_x f(x,t)) for every fixed (t) does not by itself imply that

[ \frac{d}{dx}\int f(x,t),dt

\int \partial_x f(x,t),dt. ]

The obstruction is nonuniform behavior in the integration variable. A family of difference quotients may converge pointwise while concentrating increasing mass in a shrinking region, so its integrals fail to converge to the integral of the pointwise limit.

One example is defined on (0\leq t\leq 1) by

[ f(x,t)=\frac{x^3}{x^2+t^2}. ]

For (x\neq 0),

[ \int_0^1 f(x,t),dt

x^2\arctan!\left(\frac{1}{x}\right) ]

with the appropriate continuous interpretation on each side of zero. The derivative of the integral at (x=0) is zero. However, the behavior of parameter derivatives near ((x,t)=(0,0)) is not controlled by a common elementary bound obtained merely from pointwise differentiation. The example reflects the general fact that singular regions must be controlled uniformly before differentiation and integration commute.

Related failures occur when an improper integral converges conditionally but its differentiated version diverges. Such cases are not exceptions to the Leibniz rule, because the hypotheses required by the rule are absent. They instead distinguish a formal manipulation from a justified interchange of limits.

Relation to other analytic operations

Differentiation under the integral sign belongs to a broader group of results concerning the exchange of limiting operations. Integration itself is a limit of finite sums, while differentiation is a limit of difference quotients. Their interchange therefore depends on the same forms of uniformity that appear in the interchange of infinite series, limits, and integrals.

For a power series depending on an integration variable, uniform convergence on the domain permits term-by-term integration. A corresponding uniform bound on the differentiated terms supports term-by-term differentiation. The measure-theoretic versions replace uniform convergence with hypotheses based on domination or monotonicity, as developed in the monotone convergence theorem and the dominated convergence theorem.

In complex analysis, parameter differentiation interacts with contour integration. If the integrand is holomorphic in the parameter and uniformly controlled along the contour, derivatives pass through the contour integral. This principle underlies integral representations of special functions and also contributes to proofs of analyticity for parameter-dependent transforms.

In quantum mechanics, the related Hellmann–Feynman theorem states that, for a normalized nondegenerate eigenstate (\psi_\lambda) of a parameter-dependent Hamiltonian (H(\lambda)),

[ \frac{dE_\lambda}{d\lambda}

\left\langle \psi_\lambda, \frac{\partial H}{\partial\lambda} \psi_\lambda \right\rangle. ]

The theorem transfers a parameter derivative from an eigenvalue problem to an expectation value. Its proof involves differentiating an inner product, which is an integral in coordinate representations, together with the eigenvalue equation and normalization.