As it turns out, nice spaces like C∞(M), C2(M) almost never contain the solutions to interesting partial differential equations.
Many physical solutions simply aren't differentiable. Which is to say, there is no solution when we interpret derivatives in the traditional way, where we look at derivatives at points and need them to satisfy the equation pointwise.
Instead, we replace these equations with distributional equations
The downside is that many equations no longer have unique solutions
Linear: In the modern landscape, when
we want to solve or formalize linear PDEs, we basically always use
distributions, rather than functions — in other words,
we define differential equations on distributions. Then there are
certain theorems that allow you to show that a distribution-solution to
a differential equation corresponds to an actual function in certain
cases.
Nonlinear: Because distributions don't
multiply nicely together, it's hard to deal with non-linear pdes in a
distribution setting, so people will either use the Colombeau
algebra (which allows you to multiply distributions inside of a
larger algebra at the cost of only allowing ) or viscocity
solutions.
But for this talk, we'll just focus on linear PDEs
Let F be some algebra of integrable functions (compact support, Schwartz, etc.). We'll generally work with Schwartz space for our physics examples since we need solutions that exist as plane waves an thus aren't localized in a compact region
A distribution is a continuous linear functional ϕ : F → ℝ. So the space of distributions in the continuous dual F′.
Schwartz functions: The space of
Schwartz functions is denoted 𝒮(ℝn) and consists of the
smooth functions f : ℝn → ℝ,
f ∈ C∞(ℝn)
such that
supx|xα(∂βf)| < ∞
for all multi-indices α, β.
In other words, you can take any monomial and multiply it by any derivatives of f and the least upper bound is bounded. So as you go off to infinity in some direction in ℝn, your function and all of its derivatives have to decay faster than any $\dfrac{1}{x^{\alpha}}$ (so that there isn't some large x that skyrockets the value)
We call Schwartz functions rapidly decreasing
Examples: Smooth functions with compact
support, and $x^{\alpha }e^{-a \| x
\|^{2}$ where a > 0.
Tempered distributions: The
distributions on Schwartz space are called tempered distributions and
denoted 𝒮(ℝn)′.
These tempered distributions are what we use in our PDEs.
We write the action of distributions on our test functions with a bilinear form ⟨⋅, ⋅⟩ : F′ ⊗ F → ℝ.
Examples: Given any Schwartz function
f, we have an induced
distribution ⟨f, −⟩[g] = ∫f(x)g(x)dx
Delta function δy[f] = f(y). We often see physicists write ∫δ(x − y)f(x)dx = f(y) What this really means is ⟨δy, f⟩ = f(y)
The main philosophy behind distributions is that you push off everything acting on the distribution onto the test functions
Weak derivatives: If we have a ordinary
functions, and we use integration by parts we have ∫(∂f(x))g(x) = −∫f(x)∂g(x) + Boundary
We assume that we're working with functions that go to zero on the boundary of our domain of integration (like Schwartz functions or compactly supported functions), so the boundary term disappears. Then in the language of distributions we have ⟨∂f, g⟩ = −⟨f, ∂g⟩.
For any distribution ϕ, we define the weak derivative ∂ϕ to be a distribution such that ⟨∂ϕ, g⟩ = −⟨ϕ, ∂g⟩ for all test functions g ∈ F.
Multiplication: Given a test function
f and distribution ϕ, we define the distribution fϕ to be such that ⟨fϕ, g⟩ = ⟨ϕ, fg⟩
for all test functions g. Here
we're using the fact that our test functions form an algebra.
Fourier transform: Similarly, we can
define the Fourier transform of a distribution by ⟨ℱ[ϕ]), f⟩ = ⟨ϕ, ℱ[f]⟩
for all f ∈ F.
The Fourier transform on a function is defined to be ℱ[f](x) = ∫ℝnf(x)e−ik ⋅ xdx
And $$ \mathcal{F}^{-1} [f](k) = \dfrac{1}{(2 \pi )^{n}}\int_{\mathbb{R}^{n}} f(k) e^{i k \cdot x} dk $$
where dx is the Lebesgue measure.
This is an automorphism on 𝒮(ℝn).
Uncertainty principle: The HUP and
Paley-Wiener theorems tell us that the Fourier transform of a compactly
supported function doesn't have compact support. This is why we don't
end up using compactly supported functions
Fourier + derivatives: If we view e−ik ⋅ x
as a distribution, we have ℱ[f] = ⟨e−ik ⋅ x, f⟩.
Then ℱ[∂μf] = ⟨e−ik ⋅ x, ∂μf⟩ = −⟨∂μe−ik ⋅ x, f⟩ = ikμ⟨e−ik ⋅ x, f⟩.
So
ℱ[∂μf] = ikμℱ[f]
What you can't do
Can't multiply distributions: The main
example is xδx−1,
which can't be defined associatively.
No evaluation at a point: You also
can't evaluate distributions at individual points; only on test
functions. This is related to the Heisenberg Uncertainty Principle; we
can't really talk about individual points any more than we can talk
about infinite momenta or infinite energy.
PDE: We want to look at the equation
(−∂t2 + ∂x2 + ∂y2 + ∂z2 + m2)ϕ = 0
where ϕ is a distribution. So
this is really ⟨ϕ, Df⟩ = 0 for
all test functions f, where
D = −∂t2 + ∂x2 + ∂y2 + ∂z2 + m2.
We apply the Fourier transform. Note that because the Fourier transform is pushed onto the test functions, it's always the case that ⟨ϕ, f⟩ = ⟨ℱ[ϕ], ℱ−1[f]⟩ = ⟨ℱ−1[ϕ], ℱ[f]⟩
We'll look at ⟨ℱ[ϕ], Df⟩ = ⟨ϕ, ℱ[Df]⟩ = 0. Now ℱ[Df] = (i2(−k02 + k12 + k22 + k32) + m2)ℱ[f]
So ℱ[Df] = (k02 − k̄ ⋅ k̄ + m2)ℱ[f].
So now ⟨ϕ, ℱ[Df]⟩ = ⟨ϕ, (k02 − k̄ ⋅ k̄ + m2)ℱ[f]⟩ = 0
Now whenever we have a distributional equation of the form ⟨ϕ, λ(k)f(k)⟩ = 0 for all f, we get a delta function as our solution.
Support: Let U be an open set such that λ(k) ≠ 0 for all k ∈ U. Then since the above
equation has to hold for all f, we can take f = λ−1(k),
and we find that ϕ must be
zero in this region. So non-trivial solutions must only have support in
the surface where λ(k) = 0. More precisely,
we can take f = λ−1(k)h(k)
for any test function h(k), which shows that
⟨ϕ, h⟩ = 0 for all
h with support in U. This shows that ϕ goes to zero outside of test
functions with support in λ(k) = 0.
It turns out that this forces ϕ = δ(λ(k))ψ(k) where ψ(k) ∈ 𝒮(ℝn) is a test function
Composition: And where δ(λ(k)) is the
distribution defined by $$
\delta (\lambda(k))[f] \equiv \int_{k \mid \lambda(k)=0}
\dfrac{1}{|\nabla \lambda(k)|} f(k) \psi(k)
$$
where ψ is a test function.
In the discrete case (when n = 1 in ℝn) we have $\langle \delta (\lambda(k)) \psi, \lambda f \rangle = \sum_{i} \lambda(k_{i}) \psi(k_{i}) f(k_{i}) \dfrac{1}{|\lambda'(k_{i})|} = 0 $.
If λ has higher order roots (meaning λ′(ki) = 0), we have to be more careful, and we get higher and higher order derivatives showing up. But for our case we can stop at order one.
In our case λ(k) = (k02 − k̄ ⋅ k̄ + m2), so
ϕ = δ(k02 − k̄ ⋅ k̄ + m2)ψ(k)
Note that ϕ is living in momentum space; it acts on test functions in momentum space, like ℱ[Df] , where f is in position space.
Hyperboloid: The condition k02 − k̄ ⋅ k̄ + m2 = 0
defines two sheet hyperboloid, where the sheets
correspond to the sign of k0. Actually, now that I
look at this, you want to define things so that m2 and k̄ ⋅ k̄ have the same
sign; otherwise you get imaginary mass. That's fine; we just
replace m → im.
Delta function: Using our identity for
delta function composition, we have (for any h(k) in momentum space)
$$
[\phi](h) = \int_{\Sigma} \dfrac{1}{\sqrt{4 \sum_{i} k_{i}^{2}}} h(k)
\psi(k) dk
$$
where $\Sigma = \{k \mid k_{0}^{2} - \sum_{i=1}^{3} k_{i}^{2} = m^{2} \} $. In particular we can define $E_{k} \equiv |k_{0}| = \sqrt{ \bar{k} \cdot \bar{k} + m^{2}}$. Then we want to do a change of variables.
Vector calculus: Whenever we integrate
over a surface defined by g(k0, …, k3) = 0,
we can change the variables to get
$$ d \sigma = \dfrac{\| \nabla g \|}{|\partial_{k_{0}} g|} d^{3} k $$
In our case $d \sigma = \dfrac{2\sqrt{k_{0}^{2} + \bar{k} \cdot \bar{k}}}{2 |k_{0}|} d^{3}k$
Thus our integral can be expressed as
$$ \int_{\mathbb{R}^{3}} \dfrac{1}{2 \sqrt{k_{0}^{2} + \bar{k} \cdot \bar{k}}} }\dfrac{2\sqrt{k_{0}^{2} + \bar{k} \cdot \bar{k}}}{2 |k_{0}|} ( h(E_{k} , k) \psi(E_{k}, k) + h(-E_{k}, k) \psi(-E_{k}, k) ) $$
So $$ \phi(h) = \int_{\mathbb{R}^{3}} \dfrac{1}{2 E_{k}} (h(E_{k}, k) \psi(E_{k}, k) + h(-E_{k}, k) \psi(-E_{k}, k)) $$
Transform back: Now if we want to view
things in position space, we apply the identity ⟨ϕ, h⟩ = ⟨ℱ[ϕ], ℱ−1[h]⟩.
Now if $$
f(x) = \mathcal{F}^{-1} [h] = \dfrac{1}{(2 \pi )^{4}} \int h(k) e^{i k
\cdot x} dk
$$ then we have h(k) = ∫f(x)e−i(k ⋅ x)dx
where here I mean k ⋅ x = k0x0 + k1x1 + k2x2 + k3x3.
And since our h has a fixed value of k0 = Ek, we have h(Ek, k) = ∫f(x, t)e−i(Ekt + k̄ ⋅ x̄)dxdt
So we get $$ \mathcal{F}[\phi](f) = \int_{\mathbb{R}^{4}} dxdt \int_{\mathbb{R}^{3}} \dfrac{1}{2 E_{k}} (e^{-i(E_{k}t + \bar{x}\cdot \bar{k})}f(x,t) \psi(E_{k}, k) + e^{-i(-E_{k}t + \bar{x}\cdot \bar{k})}f(x,t) \psi(-E_{k}, k)) $$
Since this holds for all f, we get that ℱ[ϕ] as a distribution has an associated function:
$$ \int_{\mathbb{R}^{3}} \dfrac{1}{2 E_{k}} (e^{-i(E_{k}t + \bar{x}\cdot \bar{k})} \psi(E_{k}, k) + e^{-i(-E_{k}t + \bar{x}\cdot \bar{k})} \psi(-E_{k}, k)) $$
The solution to the Klein Gordon equation is an actual function: $$ \int_{\mathbb{R}^{3}} \dfrac{1}{2 E_{k}} (e^{-i(E_{k}t + \bar{x}\cdot \bar{k})} \psi(E_{k}, k) + e^{-i(-E_{k}t + \bar{x}\cdot \bar{k})} \psi(-E_{k}, k)) $$
We usually denote ψ(−Ek, k) = ψ†(Ek, k) and call psi a(k), so our solution is
$$ \int_{\mathbb{R}^{3}} \dfrac{1}{2 E_{k}} (e^{-i(E_{k}t + \bar{x}\cdot \bar{k})} a(k) + e^{i(E_{k}t - \bar{x}\cdot \bar{k})} a^{\dagger}(k)) $$
Setup: ℝ4 ≅ ℝ3, 1 as both
smooth manifolds and topological vector spaces. They only differ in
terms of the geometry (metric) laid on top.
You can add in a friction term ε in order to make the classical equations actually have a solution. You then pick the weak solution which the classical solutions fε converge to as ε → 0.
Jean Leray showed in 1934 that weak solutions exist. The problem is that we don't know how these solutions behave over time.
∂tu + (u ⋅ ∇)u = νΔu − ∇p
where νΔu is the viscosity (u ⋅ ∇)u is the advection/convection — this is what givese turbulence.
Solutions u are known to be bounded in L2 ∫|u(x, t)|2dx ≤ M but the topology on L2, despite being Hilbert, isn't good for PDEs(!)
When it comes to PDEs, you actually want a topology that isn't too coarse; if it's too weak, you allow all kinds of singular solutions
A coarse/weak topology will have very large open sets, relatively speaking, and so it won't be able to regulate the behaviour of functions on a microscopic level — this is why you end up with non-singular things
Consider f′ = g. If g has a jump discontinuity, then this has no traditional solution. We instead replace it with the weak equation ∫uϕ′ = −∫gϕ for all ϕ ∈ Cc∞.
That is, we ask for f to be the weak derivative of g.