Inhomogeneous PDE

The general idea, when we have an inhomogeneous linear PDE with (in general) inhomogeneous
BC, is to split its solution into two parts, just as we did for inhomogeneous ODEs: u = uh + up .
The first term, uh , is the solution of the homogeneous equation which satisfies the inhomogeneous
boundary conditions (plus the initial conditions, if the time is a variable) of the full problem. We
have already learned how to obtain this solution for all the equations of interest to us.
Then, up must be a solution of the inhomogeneous equation, and satisfy homogeneous BC (plus
homogeneous initial conditions, if time is a variable) because uh has “taken care” of any inhomo-
geneous parts in the BC and IC. So now we need to figure out how to find this solution – we will
use Green’s functions for this, like we did for ODEs. Note that in principle “guessing” is also fine,
but more difficult for these more complicated equations. The textbook shows a solution based on
guessing for the Poisson equation, if the inhomogeneous term is a constant (ie., for a constant density
of charge). Usually one has to deal with problems that are more complicated than that, so I’ll show
you the full solution. I’m sure you’ll run into precisely such types of problems in your E&M course.
I will show you how this works in two cases – the 2D Poisson equation (time is not a variable), and
the 1D heat equation (time is a variable). Once you understand the general idea, the generalizations
to other equations should be straightforward.

0.1 2D Poisson equation

Suppose that we are asked to find the electric potential in a rectangular region of space x ∈ [0, a]; y ∈
[0, b] where there is a known density of charge ρ(x, y) (I will absorb the − 10 factor into it, for simplicity
of notation). In other words, we need to solve for:
∇2 u(x, y) = ρ(x, y) (1)
Let’s say the boundary conditions are of Dirichlet type, u(0, y) = f1 (y); u(a, y) = f2 (y); u(x, 0) =
g1 (x), u(x, b) = g2 (x) where the f’s and g’s are known functions.
As discussed above, we search for u(x, y) = uh (x, y)+up (x, y). Here, uh satisfies the hom. equation:
∇2 uh (x, y) = 0
and the full BC: uh (0, y) = f1 (y); uh (a, y) = f2 (y); uh (x, 0) = g1 (x), uh (x, b) = g2 (x). We know how
to solve for this, by writing it as a sum of simpler problems which have hom. BC on opposite sides
(like we did for the steady-state solution in the 2D rectangle. Note that mathematically that problem
for us and this problem for uh are identical). For each of these simpler problems we can then use the
separation of variables to find the solution, and the total uh is the sum of these solutions.
Next, we need to find the particular solution up = u − uh . The equation satisfied by this is
∇2 up (x, y) = ∇2 u(x, y) − ∇2 uh (x, y) = ρ(x, y)
i.e., it is indeed the full inhomogeneous equation. However, all its BC are homogeneous, because
up (0, y) = u(0, y) − uh (0, y) = 0 and so are up (a, y) = up (x, 0) = up (x, b) = 0. This is a direct parallel
with what we did for ODEs. There we also left uh to take care of any inhomogeneous BCs or ICs,
and we defined the Green’s function such that the BCs or ICs for up are homogeneous.
Note : if the BCs for u are changed – for instance, suppose that on the x = 0 edge we are given
the electric field ∂u |
∂x x=0
= e1 (y) – then we require that ∂u h
|
∂x x=0
= e1 (y), and ∂u p
|
∂x x=0
= 0. So up has
BCs of the same type as the full solution, except they are always homogeneous. In the following, I
will assume Dirichlet-type BCs like those specified above.

1
 Ideally, we’d like to find a Green’s function G(x, y; ξ, χ) such that:
Z a Z b
up (x, y) = dξ dχG(x, y; ξ, χ)ρ(ξ, χ)
0 0

This is just a straightforward generalization of the ODE Green’s function, and it’s easy to show by
substituting into the equation satistied by up that G must satisfy the inhomog. PDE:

2 ∂2 ∂2
∇ G(x, y; ξ, χ) = G(x, y; ξ, χ) + 2 G(x, y; ξ, χ) = δ(x − ξ)δ(y − χ)
∂x2 ∂y
and all the homog. BC. In other words, G(x, y; ξ, χ) is the electric potential at point (x, y) created by
a unit point charge placed at (ξ, χ), if all the BC are homogeneous (you can see from the link between
up and G that up will vanish on all the edges if and only if G vanishes on all the edges; and more
generally that the BCs for G are identical to those for up , and thus homogeneous). Physically, the
Green’s function has precisely the same meaning like for ODEs, except now we have a 2D problem,
not a 1D one. The generalization to 3D problems is hopefully fairly obvious, as well.
With this meaning of G, the equation for up is just the superposition principle, telling us that
the total potential at x, y, i.e. up (x, y), is the “sum” (integral) of the potentials created at this point
by charges ρ(ξ, χ) located at ξ, χ. To find the total potential we have to “sum” over all the possible
locations of these charges, in other words everywhere inside the region of interest.
The question is how to find this G. Here’s the logic. In all problems we’ve solved so far, we always
ended up expressing our solution as a linear combination of eigenfunctions – so that’s probably what
will happen in this case as well. In other words, we expect to find:
X
up (x, y) = cα φα (x, y).
α

where α is a complete set of indexes for the complete set of eigenfunctions φα (x, y). The question
is how to choose which eigenfunctions φα (x, y) to use? To decide this, remember that the equation
we’re trying to solve is:
ρ(x, y) = ∇2 up (x, y) = cα ∇2 φα (x, y).
X

So we should choose the eigenfunctions φα (x, y) such that the expressions ∇2 φα (x, y) are as simple
as possible, but this must also be an eigenproblem, because we know that these are eigenfunctions.
The answer is to choose the eigenfunctions to be the solutions of the eigenproblem:
∇2 φ(x, y) = λφ(x, y) (2)
where the eigenvalues λ are such that the homogeneous BC are satisfied, φ(0, y) = φ(a, y) = φ(x, 0) =
φ(x, b) = 0 (because each φα must satisfy the proper homogeneous BCs if we want their sum up to also
satisfy them). Again, for a problem with a different type of BCs, the BCs for φ change accordingly,
but remain homogeneous.
Solving this eigenproblem is easy, because this is a homog. PDE with homog. BCs. We know
that we can directly use separation of variables φ(x, y) = X(x)Y (y) which leads to:
1 d2 X 1 d2 Y
+ =λ
X dx2 Y dy 2
in other words
d2 X d2 Y
= −AX; = −BY ; λ = −(A + B)
dx2 dy 2

2
with BCs now giving X(0) = X(a) = Y (0) = Y (b) = 0. We’ve solved these many times, as we
 2  2
know we find An = nπ a
, n = 1, 2, ... and X n (x) = sin nπx
a
; and Bm = mπ
b
, m = 1, 2, ... and
mπy
Ym (y) = sin b .
So the eigenvalues and eigenfunctions are characterized by two positive integers α = (n, m), and
2 2 2 2
we have φnm (x, y) = Xn (x)Ym (y) and λnm = −( naπ2 + mb2π ). Note that if we had different BC (von
Neumann, or mixed) this would change the specific forms of Xn , Ym and the specific eigenvalues to
something else. But in all cases, because these are Sturm-Liouville equations, these eigenfunctions
are guaranteed to be orthonormal and complete. Therefore, the particular solution can indeed be
written as: X
up (x, y) = cnm φnm (x, y).
n,m

It certainly satisfies the BC, because each individual φnm does. The final thing we need to do now,
is to find the coefficients cnm so that up satisfies the inhomg. PDE.
Putting this guess into our equation, we find after using Eq. (2) that:

∇2 up (x, y) = cnm ∇2 φnm (x, y) =

X X
cnm λnm φnm (x, y)
n,m n,m

so we must have: X
cnm λnm φnm (x, y) = ρ(x, y)
n,m

Now we use the orthogonality of the eigenfunctions to find (note that here the weight is unity, see
discussion for Sturm-Liouville equations)
Ra
dξ 0b dχρ(ξ, χ)φnm (ξ, χ)
R
0
cnm λnm = Ra Rb
2
0 dξ 0 dχ[φnm (ξ, χ)]

I know you are used to writing these as integrals over x and y. You will see in a bit why I prefer
to change the name of the integration variables to ξ, χ – these are definite integrals so it makes no
difference what names we use for the integration variables. The denominators are some numbers
that are independent of the charge density ρ. For simplicity, let me denote them by
Z a Z b Z a Z b
2 2
hφnm , φnm i = dξ dχ[φnm (ξ, χ)] = dξ[Xn (ξ)] dχ[Ym (χ)|2
0 0 0 0

(you will encounter similar notation in advanced courses of quantum mechanics). For our simple sine
solutions obtained for the Dirichlet boundary conditions, these numbers are hφnm , φnm i = ab
4
.
Now we substitute these values for cnm in up :
Ra Rb
X 1 0 dξ 0 dχρ(ξ, χ)φnm (ξ, χ)
up (x, y) = φnm (x, y)
n,m λnm hφnm , φnm i
and now you see why I needed to save x, y and couldn’t use them as as the dummy variables inside
the integrals. But remember that we’re looking for:
Z a Z b
up (x, y) = dξ dχG(x, y; ξ, χ)ρ(ξ, χ)
0 0

so by comparing the two, the Green’s function is:

X φnm (x, y)φnm (ξ, χ)
G(x, y; ξ, χ) =
n,m λnm hφnm , φnm i

3
 And there it is: we have the Green’s functions in terms of eigenvalues and eigenfunctions, which
are already known. We can now calculate up (x, y) for any desired ρ(x, y). For other types of BCs,
the particular expressions for λnm , φnm will change, but the general expression for G stays the same.
If you have to solve a 3D Poisson equation, then there will be 3 sets of eigenvalues φn,m,k (x, y, z) =
Xn (x)Yn (y)Zn (z) where λnmk = −An − Bm − Ck is the sum of eigenvalues for each simple ODE; and
going through the same steps, you’ll find:
X φnmk (x, y, z)φnmk (ξ, χ, ζ)
G(x, y, z; ξ, χ, ζ) =
n,m,k λnmk hφnmk , φnmk i

to be the potential at (x, y, z) if a point charge is located at (ξ, χ, ζ), and the proper hom. BC are
satisfied on all outside surfaces. Again, we can calculate up (x, y, z) for any density of charge ρ(x, y, z)
by just integrating over all space where charges are placed:
Z a Z b Z c
up (x, y, z) = dξ dχ dζG(x, y, z; ξ, χ, ζ)ρ(ξ, χ, ζ)
0 0 0

Of course, let’s not forget that for the full solution, we then need to add uh to up to get the full u.
In other coordinates things change accordingly, for instance in 2D, if we choose to work with ρ, θ
instead of x, y, then we need tolook for the eigenvalues ∇2 φ(ρ, θ) = λφ(ρ, θ) with the proper hom.
2
BC, where now ∇2 φ = ρ1 ∂ρ ∂
ρ ∂φ
∂ρ
+ ρ12 ∂∂θφ2 (this is what the Laplacean looks in polar coordinates, I may
give this to you as a homework to check). So this will change the expressions of the eigenfunctions,
but the final expression for G is of the same type as above.

0.2 1D heat equation

Let us now consider an example where time is also a variable. Consider the inhomogeneous equation:
∂ 2 u(x, t) 1 ∂u(x, t)
− = ρ(x, t)
∂x2 κ ∂t
where x ∈ [0, a] and we have the IC u(x, t = 0) = f (x) and the BCs u(0, t) = TL , u(a, t) = TR .
Again, we split the problem in 2 parts, u(x, t) = uh (x, t) + up (x, t). The homog. solution satisfies
the homog. equation, and the IC and BC of the full problem:
∂ 2 uh (x, t) 1 ∂uh (x, t)
2
=
∂x κ ∂t
where uh (x, t = 0) = f (x) and uh (0, t) = TL , uh (a, t) = TR . We know how to solve for uh in 2 steps,
first the steady-state solution and then the transient solution and then uh = us + ut , so I won’t say
anymore about it.
The second part up (x, t) = u − uh satisfies the full inhomogeneous equation:
∂ 2 up (x, t) 1 ∂up (x, t)
− = ρ(x, t)
∂x2 κ ∂t
but homog. IC and BC: up (x, t = 0) = 0 and up (0, t) = up (a, t) = 0.
Again, we’re aiming to find a Green’s function such that:
Z a Z ∞
up (x, t) = dξ dτ G(x, t; ξ, τ )ρ(ξ, τ )
0 0

4
Note that the time interval runs to infinity, as that’s the interval of time we are concerned with.
Putting this guess into the equation satisfied by up , we find that the equation satisfied by the
Green’s function must be (as you hopefully expect):
∂ 2 G(x, t; ξ, τ ) 1 ∂G(x, t; ξ, τ )
− = δ(x − ξ)δ(t − τ )
∂x2 κ ∂t
P
The approach to finding G is just as before: again, we expect that up (x, t) = n an (t)Xn (x),
where Xn (x) are a complete set of eigenfunctions. Note that because t is a variable, the coefficients
must depend on it. We know that each Xn (x) must satisfy the homog. BC, so that their sum up also
does. Still, we have the freedom to choose for which eigenproblem are these Xn (x) the eigenfunctions
– of course, we will do this so as to simplify the spatial derivatives as much as possible.
For this particular problem, we know that spatial part involves the 2nd derivative with respect to
2
x, so we will choose the eigenproblem ddxX2 = λX. Together with the hom. BC X(0) = 0, X(a) = 0 we
 2
nπ
have λn = − a
, Xn (x) = sin nπx
a
, n = 1, 2, .... So with this choice, we can now put the expression
∞
X
up (x, t) = an (t)Xn (x)
n=1

into the inhomog. PDE, to find:

∞
" 2 #
nπ 1 dan (t)
X 
− an (t) − Xn (x) = f (x, t).
n=1 a κ dt
Because the Xn (x) are orthogonal (and for this simple equation the weight is 1) we can find the
coefficients of the expansion as usual:
2 Ra
nπ 1 dan (t) dxXn (x)f (x, t)

0
− an (t) − = = gn (t)
a κ dt hXn , Xn i
where again I denoted hXn , Xn i = 0a dx[Xn (x)]2 – these are some numbers independent of f (x, t).
R

For our BC and their corresponding sine functions, hXn , Xn i = a2 , the usual value for Fourier series.
We can now solve for an (t) – they satisfy simple inhomog. first order ODEs. The hom. solution
2 2
is e−κkn t so we use variation of parameters to find a particular solution an (t) = v(t)e−κkn t . After a
bit of work, we find the general solution:
Z t " Z t Ra #
dξXn (ξ)f (ξ, τ )
 
2t
−κkn 2τ
κkn 2t
−κkn 2τ
κkn 0
an (t) = e C1 − κ dτ e gn (τ ) = e C1 − κ dτ e
0 0 hXn , Xn i
Remember that up must also satisfy a homogeneous IC (the uh part took care of any inhomogeneity),
i.e. up (x, t = 0) = 0 → an (0) = 0 → C1 = 0, and so we have our final solution:
∞ Z t Ra Z t Z a
X 2t
−κkn 2τ 0
κkn dξXn (ξ)f (ξ, τ )
up (x, t) = −κ e dτ e Xn (x) = dτ dξG(x, t; ξ, τ )f (ξ, τ )
n=1 0 hXn , Xn i 0 0

where ∞
2 Xn (x)Xn (ξ)
e−κkn (t−τ )
X
G(x, t; ξ, τ ) = −κ if t > τ
n=1 hXn , Xn i
and G = 0 if t < τ . Different BC will change the specific Xn (x), kn2 values, but the general expression
remains the same.

5
 Two final notes:
(1) although in the general expression of up that we started from, we allowed the integral over
the time τ to run to infinity, you can see that after solving the problem, we find that in practice the
integral over τ only runs up to the time t of interest. This is again a consequence of causality: the
solution at time t cannot depend on what happens after time t. This is directly analogous to what
we found for Green’s functions for time-dependent ODEs.
(2) if this was an inhomog. wave equation we would proceed similarly, however we would now
have a second order inhomog. ODE for the an (t) coefficients (because the wave eq. has ∂ 2 u/∂t2 , not
∂u/∂t, like the heat equation). Not a problem, because we know how to solve 2nd order inhomog.
ODES with constant coefficients! Of course, we now find 2 constants of integration but we also have
2 hom IC, so we can find them and then extract the Green’s function. This will have a somewhat
more complicated expression for the time-depending part. Since I do not expect that you will need
to solve such an equation during your undergrad studies, I will stop here. But you are now fully
equiped to actually solve such problems, if need be. They have all been reduced to simple ODEs and
eigenproblems – each individual part of the calculation is quite simple, and you simply need to put
together all these pieces.