using Plots, ComplexPhasePortrait, ApproxFun, SingularIntegralEquations, OscillatoryIntegrals,
SpecialFunctions
gr();
Dr Sheehan Olver
Our goal is to solve the integral equation λu(x)+∫∞0K(x−t)u(t)dt=f(x)for0<x<∞.
Outline:
Consider the Fourier transform of sechx=2ex+e−x
xx = -10:0.01:10
plot(xx,sech.(xx))
Now the Fourier transform of sechx is Fsech(s)=∫∞−∞sechte−istdt=πsechπs2
f = Fun(sech, Line())
fourier(f, 2.0)
0.2710149513994194 + 0.0im
π*sech(π*2.0/2)
0.2710149513994184
Note that πsechπz2=2πeπz2+e−πz2
phaseplot(-3..3, -3..3, z -> π*sech(π*z/2))
This is because of the expontial decay.
Theorem (Analyticity of Fourier transforms) Suppose |f(x)eγx|<|M(x)| where M is absolutely integrable for all a<γ<b. Then ˆf(z)=∫∞−∞f(t)e−iztdt
Proof Let z=s+iγ and note that |f(t)eizt|=|f(t)eγt|. Thus for a<γ<b, we can exchange differentiation and integration to get dˆfdz=−izˆf(z)
Remark We don't need f to be analytic at all! Decay in f gives analyticity.
In the case of sechx, we get exponential decay in both directions: that is sechxe|γ|x is absolutely integrable for γ<1.
Another example is e−x2/2, which is absolutely integrable for any γ. Therefore, it's Fourier transform is in fact entire: F[e−⋄2/2](z)=√2πe−z22
A neglected fact of the Fourier transform is that we can think of ˆf(z) living on any line (−∞+iγ,∞+iγ), and in fact we can recover f from the Fourier transform only on this line. This works even if ˆf(s) is not defined on the real-axis, the real-axis is NOT special!
Theorem Suppose f(x)eγx is square integrable. Then f(x)=12π∫∞+iγ−∞+iγˆf(ζ)eixζdζ
Proof Note for g(x)=f(x)eγx ˆg(s)=∫∞−∞f(t)eγt−istdt=ˆf(s+iγ).
Consider now ∫∞0f(t)e−istdt
Because it is identically zero on the negative real axis, we immediately get the following:
Corollary (analyticity of Half-Fourier transform) Suppose f(x) is bounded for x≥0. Then ^fR(z) is analytic in the lower half plane H+={z:ℑz<0}.
More generally, f can even have exponential decay: if f(x)eγx is bounded then ^fR(z) is analytic in {z:ℑz<γ}. As before, the same inversion formula follows:
Corollary (inverting Half-Fourier transform) Suppose f(x)eγx is square integrable for x≥0. Then f(x)=12π∫∞+iM−∞+iM^fR(ζ)eixζdζ
Example Consider f(x)=xe−x for 0≤x<∞. Note that f(x)eγx is square integrable for any γ<1, and we have ^fR(z)=∫∞0te−t−istdt=1(1+is)2
Since x>0, we can use Residue calculus in the upper-half plane, which confirms the result: Resz=−ieixz(1+iz)2=Resz=−ie−x+ixe−x(z+i)+O(z+i)2−(z−i)2=−ixe−x.
Note that the results have corresponding analogues for fL:
Corollary (analyticity of left Half-Fourier transform) Suppose f(x)eγx is bounded for x≤0. Then ^fL(z) is analytic for {z:ℑz>γ}.
Corollary (inverting left Half-Fourier transform) Suppose f(x)eγx is square integrable for x<0. Then f(x)=12π∫∞+iM−∞+iM^fL(ζ)eixζdζ
Now consider the Laplace transform ˇf(z)=∫∞0f(t)e−ztdt
NEVER think of the Laplace transform as a real-valued object: it only makes sense as a complex object. This is seen from the inverse Laplace transform f(x)=12πi∫i∞−M−i∞−Mˇf(ζ)eζxdζ
Consider the following ODE for x≥0: u″(x)+2u′(x)+u(x)=f(x)
We now consider the question of calculating Laplace transforms (or equivalently, half-Fourier transforms) ˇf(s)=∫∞0f(t)e−stdt
Then we have calculated the Cauchy transform: ϕ(z)=C[fe−s⋄](z).
Let's start with f(x)=1 and s=1, that is, what is the Cauchy transform of e−x? Consider the exponential integral: Ei(z)=∫z−∞eζζdζ
Thus C[e−s⋄](z)=ϕ(sz)=−e−szEi(sz)2πi
Let's make sure I didn't make a mistake. Here we first define Ei:
const ei₋₁ = let ζ = Fun(-50 .. -1)
sum(exp(ζ)/ζ)
end
function ei(z)
ζ = Fun(Segment(-1 , z))
ei₋₁ + sum(exp(ζ)/ζ)
end
ei (generic function with 1 method)
φ = (z) -> -exp(-z)*ei(z)/(2π*im)
#9 (generic function with 1 method)
The expression matches the Cauchy transform:
t = Fun(0 .. 50)
s = 2.0
z = 2.0+2.0im
sum(exp(-s*t)/(t-z))/(2π*im)
0.02534853710699088 + 0.017828329563678025im
φ(s*z)
0.02534853710699058 + 0.017828329563677824im
We then recover the Laplace transform by taking the limit: ˇ1(s)=−2πilimz→∞zϕ(sz)=limz→∞zsz=1s
What about rational f? Do the same trick of subtracting off the singularities. For example, consider f(z)=1/(z+1). Then ϕ(sz)−ϕ(−s)z+1
t = Fun(0 .. 50)
s = 2.0
z = 2.0+2.0im
f = 1/(t+1)
sum(exp(-s*t)*f/(t-z))/(2π*im)
0.0174397493357667 + 0.013485355450154557im
(φ(s*z)-φ(-s))/(z+1)
0.017439749335766603 + 0.01348535545015462im
Therefore, ˇf(s)=−2πilimz→∞zϕ(sz)−ϕ(−s)z+1=2πiϕ(−s)=−esEi(−s)
@show sum(exp(-s*t)*f)
@show 2π*im*φ(-s)
@show -exp(s)*ei(-s);
sum(exp(-s * t) * f) = 0.3613286168882254 (2π) * im * φ(-s) = 0.3613286168882283 + 0.0im -(exp(s)) * ei(-s) = 0.3613286168882283