When the toolkit runs out
By now you have a respectable arsenal. You can reach for u-substitution, lean on integration by parts, unfold a rational function with partial fractions, or straighten out a square root with trig substitution. And yet some definite integrals just sit there, arms folded, refusing every one of these. A famous one is the integral from 0 to infinity of (sin x)/x dx — there is no elementary antiderivative for (sin x)/x at all, so the usual move of 'find the antiderivative, plug in the limits' is simply unavailable.
The trick we meet here changes the question instead of forcing the answer. We deliberately spoil the integral by slipping an extra knob into it — a parameter, call it alpha — turning one fixed number into a whole family of integrals I(alpha). The plan sounds almost reckless: make the problem look harder by adding a variable, then differentiate the family with respect to that variable, hoping the derivative is an integral we can actually do.
The Leibniz rule: swapping two operations
At the heart of everything is one clean idea. Suppose I(alpha) = integral from a to b of f(x, alpha) dx, where the limits a and b are fixed and alpha is our parameter. Then, under conditions we will name honestly in a moment, dI/dalpha = integral from a to b of (partial f / partial alpha) dx. In words: the derivative of the integral equals the integral of the derivative. You are allowed to push d/dalpha through the integral sign and let it land on the integrand as a partial derivative in alpha.
Why should swapping these two operations be legal? Picture the integral as an infinite sum of thin slabs f(x, alpha) dx. Nudging alpha changes the height of every slab a little. The total change of the area is just the sum of all those little height-changes — which is exactly the integral of partial f / partial alpha. Each slab responds with its own ordinary derivative in alpha, and adding up the responses is the same whether you add first and then differentiate, or differentiate first and then add. That is the whole intuition.
If the limits also depend on alpha, the full Leibniz rule adds two boundary terms: d/dalpha of integral from a(alpha) to b(alpha) of f dx = f(b, alpha) b'(alpha) - f(a, alpha) a'(alpha) + integral of partial f / partial alpha dx. Those extra pieces are exactly the fundamental theorem of calculus catching the moving endpoints. For the Feynman trick we almost always keep the limits fixed, so they vanish and only the integral-of-the-derivative term survives.
A full worked example, slowly
Let us evaluate I = integral from 0 to 1 of (x^3 - 1) / ln(x) dx. The natural log downstairs is what blocks every elementary method. So we insert a parameter where a power already lives: define I(alpha) = integral from 0 to 1 of (x^alpha - 1) / ln(x) dx, with alpha greater than -1. Notice I(3) is the number we actually want, and I(0) = 0 for free, since x^0 - 1 = 0 makes the whole integrand vanish. That last fact gives us a known anchor to integrate back from.
Now differentiate under the sign. The key cancellation is that partial/partial alpha of x^alpha equals x^alpha ln(x), and that ln(x) annihilates the troublesome ln(x) in the denominator: partial/partial alpha of (x^alpha - 1)/ln(x) = (x^alpha ln x)/ln x = x^alpha. The stubborn integral has collapsed into something a first-year student can do. So I'(alpha) = integral from 0 to 1 of x^alpha dx = 1/(alpha + 1).
I(alpha) = integral_0^1 (x^alpha - 1)/ln(x) dx (want alpha = 3)
I'(alpha) = integral_0^1 d/dalpha[(x^alpha-1)/ln x] dx
= integral_0^1 x^alpha dx (ln x cancels!)
= 1/(alpha + 1)
integrate back, using anchor I(0) = 0:
I(alpha) = integral_0^alpha 1/(t+1) dt = ln(alpha + 1)
=> I(3) = ln(4)The last leg is the integrate-back. We know dI/dalpha = 1/(alpha + 1), and we know the anchor I(0) = 0, so I(alpha) = integral from 0 to alpha of dt/(t+1) = ln(alpha + 1). Setting alpha = 3 gives I = ln(4). A messy integral with a logarithm jammed in the denominator turned out to equal ln 4 — and we never once searched for an antiderivative of the original integrand.
The recipe, and where to hide the knob
Every Feynman-trick problem follows the same five beats. The genuinely creative step is the very first one: deciding where the parameter goes. There is no algorithm for it, only taste sharpened by practice. A good insertion makes the alpha-derivative cancel the obstruction (as ln x cancelled ln x above) or turn an awkward factor into something you can integrate in x.
- Insert a parameter alpha so that I(alpha) reduces to your target at one value, and to something trivially known (often 0 or a standard integral) at another value — that second value is your anchor.
- Differentiate under the sign: I'(alpha) = integral of partial f / partial alpha dx, hoping the result is an integral you can do.
- Evaluate that easier integral in x, leaving a plain function of alpha for I'(alpha).
- Integrate I'(alpha) back up in alpha to recover I(alpha), with one arbitrary constant.
- Pin the constant using the anchor value, then read off the answer at the value of alpha you actually wanted.
The same machine powers many celebrated results. Differentiating the Gaussian integral integral from 0 to infinity of e^{-alpha x^2} dx = (1/2) sqrt(pi/alpha) with respect to alpha instantly produces integral of x^2 e^{-alpha x^2} dx without any new effort — one differentiation gives you a whole ladder of moment integrals. The Frullani integral integral from 0 to infinity of (f(ax) - f(bx))/x dx = (f(0) - f(infinity)) ln(b/a) can likewise be coaxed out by parameter-differentiation. And the (sin x)/x integral that opened this guide yields to inserting e^{-alpha x}, differentiating to kill the sine's awkwardness, and integrating back.
When you are allowed to do it
Be honest: swapping the derivative and the integral is not always legal, and pretending otherwise will eventually bite you. The interchange is justified when f and its partial derivative partial f / partial alpha are continuous in both variables over the region, and — crucially for infinite ranges — when the differentiated integral converges uniformly in alpha. Uniform convergence is the real gatekeeper: it guarantees that no mass sneaks off to infinity faster than your differentiation can track it.
There are real counterexamples where blindly differentiating gives a wrong number, precisely because convergence fails to be uniform. So in serious work you check the hypotheses, or at least confirm your answer numerically. In practice, on the smooth, rapidly-decaying integrands this method is usually aimed at, the conditions hold and the technique is rock-solid — but the discipline of asking 'am I allowed?' is what separates a method from a superstition.
Why it belongs in your hands
Step back and notice what really happened. We refused to attack the integral head-on and instead embedded it in a family, then exploited the fact that a derivative in the parameter is sometimes far gentler than the integral itself. This 'parametrize, differentiate, re-integrate' rhythm echoes across advanced mathematics — it is the same spirit behind generating functions, behind transform methods, and behind much of mathematical physics, where slipping a parameter into a problem to soften it is a reflex, not a stunt.
Treat the trick the way you would a chess opening: study a handful of model insertions until you feel where parameters want to live. Powers x^alpha that meet a logarithm, exponentials e^{-alpha x} that tame oscillation, an alpha inside an arctangent or a square root — these patterns recur. Once your eye is trained, integrals that looked impossible start announcing exactly where they want the knob, and you will reach for this method the way you now reach, without thinking, for integration by parts.