Trigonometric integrals, poles and the Poisson kernel
Derive Fourier moments and a Poisson-kernel consequence using residues.
§8.1Estimated reading time: 14 min
A substitution turns many integrals around a full period into rational contour integrals. The calculation has three distinct parts: translate the integrand, locate every pole relative to the unit circle, and only then use residues. In particular, a pole at the origin can appear because of the substitution even when the original integrand looks harmless.
This chapter develops the Fourier moments of 1/(a+bcost) and a Poisson-kernel consequence with worked examples.
What the unit-circle substitution actually says
Let z=eit for 0≤t≤2π. As t increases, z traverses ∣z∣=1 once counterclockwise, and
dt=izdz,cost=2z+z−1,sint=2iz−z−1.
These equations are identities along the parametrized circle: dz=ieitdt=izdt. Suppose the resulting rational function R(z) has no pole on the unit circle. The residue theorem then gives
This is a finite-contour identity, so there is no large-arc limit to check. It does require us to account for every interior singularity, including z=0, and to verify that no uncancelled pole lies on the contour.
Fourier moments of a reciprocal denominator
Let a,b∈R with a>∣b∣ and b=0, and let n be a nonnegative integer. The denominator a+bcost is at least a−∣b∣>0, so the original integral is continuous and finite. Put
d=a2−b2>0,α=b−a+d=−a+db,β=b−a−d.
The rationalized formula shows ∣α∣=∣b∣/(a+d)<1. Also αβ=1, so ∣β∣>1. These are the two zeros of bz2+2az+b and they are distinct.
Theorem 8.1.1
(Fourier moments of a reciprocal denominator)
Let a,b∈R with a>∣b∣ and b=0, let n≥0 be an integer, and put d=a2−b2 and α=−b/(a+d). Then
The factor z from dt has canceled the z−1 introduced by the cosine. For n≥0 the integrand displayed here has no pole at zero: the quadratic equals b=0 there. Its only pole inside the circle is the simple pole α, with
The last two results follow by taking real and imaginary parts; α is real. The sine integral can also be checked by the change t↦2π−t.
Proof X-ray
Why exactly one pole counts
The condition a>∣b∣ does three jobs: it makes the integrand continuous, it gives ∣α∣<1<∣β∣ so that exactly one root of bz2+2az+b lies inside the circle, and it keeps both roots off the contour.
The key move is the cancellation of the factor z: dt=dz/(iz) against the z−1 from cost=(z+z−1)/2. For n≥0 this leaves no pole at 0, which is why only α is counted.
The proof closes with a single residue, αn/(2bα+2a)=αn/(2d). The factor αn is what makes the moments decay geometrically in n.
For negative n the cancellation fails: zn contributes a pole at 0, and ignoring it gives a wrong value (see the common mistake below).
For negative integer n, conjugation gives Jn=J∣n∣; one should not blindly substitute a negative n into the displayed rational integrand without also accounting for its pole at zero. If b=0, the integral is 2π/a for n=0 and zero for nonzero integer n.
Counterexample mode
The sign condition on $a$ cannot be dropped
Claim: ∫02πa+bcostdt=a2−b22π whenever a2>b2. Take a=−5 and b=4. The denominator −5+4cost≤−1 is negative, so the integral is negative. Substituting t↦t+π turns ∫02π5−4costdt into ∫02π5+4costdt=32π, hence ∫02π−5+4costdt=−32π, but the claimed formula gives +32π. The theorem needs a>∣b∣; for a<−∣b∣ the value is −2π/a2−b2.
Worked example: Fourier moments with alternating signs
Worked example 8.1.2
(Fourier moments with alternating signs)
Find both integrals
∫02π5+4costcos(2t)dt,∫02π5+4costcos(3t)dt.
Here a=5, b=4, d=3 and α=(−5+3)/4=−1/2. Hence the answers are
32π(−21)2=6π,32π(−21)3=−12π.
The denominator is everywhere positive, but a weighted integral need not be positive because cos(3t) changes sign. The sign of αn makes this visible. As a check, the n=0 value is 2π/3; its absolute value bounds both answers since ∣cos(nt)∣≤1.
Why this is also a Poisson-kernel calculation
Definition 8.1.3
(Poisson kernel)
For 0≤r<1, define
Pr(t)=1−2rcost+r21−r2.
The denominator is ∣1−reit∣2≥(1−r)2>0. Taking a=1+r2 and b=−2r in the previous calculation gives d=1−r2 and α=r when r>0; the same formulas at r=0 follow directly, with r0=1. Thus
In particular, Pr≥0 and its integral is 2π. These facts have a geometric consequence. For a continuous 2π-periodic real-valued function g, define its Poisson average
Let g be continuous, real-valued and 2π-periodic. Then for each fixed θ, u(r,θ)→g(θ) as r↑1.
Proof
(Near and far parts of the kernel)
Let ε>0. Continuity gives η∈(0,π) such that ∣g(θ−t)−g(θ)∣<ε when ∣t∣<η. Since the kernel has total mass 2π, the contribution of this near interval to ∣u−g(θ)∣ is at most ε. On η≤∣t∣≤π,
1−2rcost+r2=(1−r)2+2r(1−cost).
For r≥1/2 this is at least 1−cosη>0, so Pr(t)≤(1−r2)/(1−cosη) there. The function g is bounded, and this bound tends to zero uniformly on the remaining interval. The far contribution therefore tends to zero. Letting ε tend to zero proves the claim.
This is an approximation to the boundary value; no claim that P1 is an ordinary continuous kernel is being made.
Worked example 8.1.5
(Poisson average of a trigonometric polynomial)
For example, with g(θ)=2+3cos(2θ)−sinθ, the addition formulas and the moment identities give
u(r,θ)=2+3r2cos(2θ)−rsinθ.
The radial factors show how the kernel damps higher-frequency terms before the boundary is reached.
Solved practice: a squared denominator
Worked example 8.1.6
(A squared denominator)
Problem. Evaluate ∫02π(5+4cost)−2dt, and justify any differentiation under the integral.
Solution. For a>4, the section on Fourier moments of a reciprocal denominator, with b=4 and n=0, gives
F(a)=∫02πa+4costdt=a2−162π.
For a in the compact interval [9/2,11/2], the denominator is at least 1/2. The integrand and its a-derivative −(a+4cost)−2 are continuous there, and the derivative is bounded in absolute value by 4. Thus the standard differentiation-under-the-integral theorem on a compact interval applies; equivalently, the mean-value theorem supplies a uniformly bounded difference quotient and uniform continuity gives its limit. Consequently
−∫02π(a+4cost)2dt=F′(a)=−(a2−16)3/22πa.
At a=5 the required value is 10π/27. The neighborhood of a=5 matters: differentiating a formula at a boundary where the denominator can vanish would require a separate argument.
Final check. Verify the original denominator and every contour pole first, specify the orientation and the integer range, and distinguish an ordinary integral from a principal value whenever a boundary singularity is present. The formula derived here assumes a>∣b∣; it does not automatically apply when this hypothesis fails.
Common mistakes
Common mistake
(Dropping the pole at the origin for negative n)
For n=−1 the integrand z−1/(bz2+2az+b) has a pole at 0 with residue 1/b as well as the pole at α. Counting only α gives 2πβ/d with ∣β∣>1, which exceeds J0=2π/d in modulus, impossible because ∣cost∣≤1 and the denominator is positive. Including the residue at 0 gives 4π(b1+2dβ)=2πα/d=J1, as conjugation requires.
Common mistake
(Letting r↑1 inside the integral)
For each fixed t=0, Pr(t)→0 as r↑1, yet 2π1∫Prdt=1 for every r. Passing the limit inside would give 0 instead of the boundary value g(θ). The mass of the kernel concentrates at t=0, which is why the proof splits near and far parts.
Quick checks
Quick check
(Evaluate ∫02π5+4costcostdt.)
Use the theorem with a=5, b=4, n=1.
Solution · Answer
Here d=3 and α=−21, so the integral is 32π⋅(−21)=−3π. As a check, 5+4costcost=41(1−5+4cost5) gives 41(2π−5⋅32π)=−3π.
Quick check
(Evaluate ∫02π2+costdt.)
Identify a, b and d.
Solution · Answer
With a=2>1=∣b∣ we get d=4−1=3, so the integral is 32π.
Quick check
(For P1/2(t)=5/4−cost3/4, what is 2π1∫02πP1/2(t)cos(3t)dt?)
Use the moment identities for the Poisson kernel.
Solution · Answer
The moment identity gives rn with r=21 and n=3, so the value is 81.