March 6, 2019

DISTRIBUTIONS

RODICA D. COSTIN

Contents
1. Setting the stage 2
1.1. Test functions 2
1.2. Distributions 3
1.3. Examples of distributions 3
1.4. Dirac’s delta function 3
2. Operations with distributions 4
2.1. Linear combinations 4
2.2. Multiplication by C ∞ functions 4
2.3. Any distribution is differentiable 4
2.4. Convolution of distributions with test functions 5
2.5. The Fourier transform of distributions 6
3. Calculus with distributions 7
3.1. Limits of distributions 7
3.2. Fundamental sequences for δ 7
3.3. A sequence of step functions 7
3.4. Other sequences 7
4. Green’s function 8
4.1. Non-homogeneous conditions can be reduced to homogeneous
conditions, but for a non-homogeneous equation 8
4.2. Superposition 9
4.3. First examples 10
4.4. Autonomous equations 12
5. Green’s function of self-adjoint Sturm-Liouville problems 13
5.1. Liouville transformation 13
5.2. Existence and uniqueness of the Greens’s function for
non-homogeneous problems are not guaranteed 16
6. Conditions for existence and uniqueness of the solution to
real-valued, non-homogeneous Sturm-Liouville problems 17
6.1. Existence of eigenvalues for real-valued Sturm-Liouville
problems 17
7. Green’s function for self-adjoint Sturm-Liouville problems 18
7.1. When the Green’s function of Sturm-Liouville problems exists
and is unique 18
7.2. When the Green’s function for self-adjoint Sturm-Liouville
problems exists and it is not unique 22
7.3. The resolvent 22
1
2 RODICA D. COSTIN

8. Continuous spectrum 24
8.1. Example: Bessel operator 26
d
8.2. A cautionary tale: −i dx on L2 ([0, ∞) Formally self-adjoint is
not the same as self-adjoint! 27

1. Setting the stage

For a long time mathematicians, physicists, engineers worked, particularly
in problems in differential equations, by differentiating functions which are
not differentiable, integrating against ”functions” which are zero everywhere
except at one point, and obtaining a nonzero integral, integrating functions
whose integral is not defined, and, in the end, obtaining valid results!
The ideas and techniques have been gathered and set on a rigorous math-
ematical ground by Sobolev (1930’s), and, separately, by Laurent Schwartz
(1940’s). The main idea in working with generalized functions is that these
are only needed and defined in integrals, that is, they are linear functionals.
Distributions are continuous linear functionals on spaces of test functions.

1.1. Test functions. These are defined on a domain of interest for a partic-
ular problem: on R, or R2 , Rn or a subset (a, b) ⊂ R, Ω ⊂ Rn ) and are very,
very, nice functions, having all the properties one may wish for in tackling
a problem at hand. Namely:
- they are infinitely many times differentiable (we say for short that they
are class C ∞ ), and
- they vanish, together with all their derivatives, at the boundary of the
domain (so that when we integrate by parts the boundary terms vanish).
Here are the most popular spaces of test functions.
S(R), the rapidly decaying functions (the Schwartz space) contains func-
tions which decay, towards ±∞, more rapidly than any power, and so do all
2
their derivatives. For example, φ(x) = e−x ∈ S(R).
It can be shown that the Fourier transform F : S → S is one-to-one and
onto. This requires a proof, but here is the intuitive reason: f has decay
↔ Ff is smooth, and f is smooth ↔ Ff has decay, and functions in S are
both smooth and have decay.
C0∞ (Ω), the functions with compact support on Ω ⊂ Rn are functions
which are non-zero only on a bounded set, and they are identically zero
before the boundary of Ω. For example,
(
− 12
φ(x) = e 1−x if |x| < 1 ∈ C0∞ (R)
0 otherwise
and in fact φ(x) ∈ C0∞ (−a, a) for any a > 1.
Note: the spaces of test functions are linear subspaces of L2 (and they
are dense in L2 ).
 DISTRIBUTIONS 3

A sequence of test functions φn is said to converge to a test function φ

(in the same space) if it converges in the strongest sense you may imagine
(k)
(not only point-wise, but also all the derivatives φn (x) → φ(k) (x), and,
moreover, they converge uniformly on any compact subset... this ensures
that one may pass to the limit freely all the operations needed: summation,
differentiation, integration, without worry).

1.2. Distributions. Given a space of test functions, distributions are con-

tinuous linear functionals, that is, if u is a distribution
we denote u(φ) := (φ, u) for any test function φ
and
(cφ + dψ, u) = c(φ, u) + d(ψ, u) for test functions φ, ψ and constants c, d
whenever φn → φ as test functions, then (φn , u) → (φ, u)

1.3. Examples of distributions. In what follows we assume our test func-

tions are C0∞ (R). For other spaces of test functions minimal extra care may
be needed.

1.3.1. Function type distributions. Any function which is not too wild (i.e. it
is integrable on compact sets) can be regarded as a distribution: say, f (x) =
1, or f (x) = sin x, or f (x) = eix , or f (x) = H(x), defines a distribution by
the formula
Z ∞
(1) (φ, f ) = f (x)φ(x) dx
−∞

The integral converges since any test function φ(x) vanishes outside some
interval [−M, M ].1 The functional defined by (1) is clearly linear; a bit of
mathematical argumentation would also show continuity (which we do not
pursue here).

Note the underlying idea: (1) is, essentially, the inner product of L2 . We
want to use it for more general functions f , like the examples above, which
are not in L2 ; we do that, but the price we pay is that we can only do pairing
with very rapidly decaying functions φ.

1.4. Dirac’s delta function. δ is defined as

(2) (φ, δ) = φ(0) for any test functions φ
and shifted delta, δ(x − a) is defined as
(3) (φ, δ(x − a)) = φ(a)

1Of course, if the test functions are S, then we should require f not to increase faster
than some power.
4 RODICA D. COSTIN

Note that (2),(3) are commonly written, in the spirit of (1), using an
integral sign:
Z ∞
(φ, δ) = φ(x)δ(x) dx = φ(0)
−∞
and
Z ∞ Z ∞
(φ, δ(x − a)) = φ(x)δ(x − a) dx = φ(t + a)δ(t) dt = φ(a)
−∞ −∞

These are not bonafide integrals, but they work like integrals.

2. Operations with distributions

2.1. Linear combinations. Any linear combination of distributions is a
distribution (why?).

2.2. Multiplication by C ∞ functions. We can multiply distributions u

with C ∞ functions f (x) and we obtain another distribution f u, defined as
(φ, f u) = (f φ, u) for any test function φ
(the definition is, of course, inspired by what happens for function type
distributions). Note that the definition makes sense since if φ ∈ C0∞ (R),
and f ∈ C ∞ (R), hence φf ∈ C0∞ (R).
Note: When we work with distributions over the test functions S (these
care called tempered distributions) we also need to require that the function
f increases no faster than some power at ±∞, to ensure that for whatever
φ ∈ S then also f φ ∈ S.

2.3. Any distribution is differentiable. If our distribution is a nice, dif-

ferentiable function f (x), what should its derivative be? As a distribution,
let us see how it acts on test functions:
Z ∞
0 0
(φ, f ) = [since f is a function] = φ(x)f 0 (x)dx
−∞

and integrating by parts

Z ∞
∞
= φ(x)f (x) −∞ − φ0 (x)f (x)dx = [since φ(±∞) = 0] = −(φ0 , f )
−∞

This justifies
Definition 1. The derivative u0 of a distribution u is defined by the rule
(φ, u0 ) = −(φ0 , u) for any test function φ
Theorem 2. If u is a function-type distribution, and u is differentiable,
then its derivative in the sense of distributions coincides with its derivative
as a function.
 DISTRIBUTIONS 5

Why: denoting for the moment by Du the derivative of u as a function,

and by u0 its derivative as a distribution, we have
Z
(φ, u0 ) = [by definition] −(φ0 , u) = [since u is a function] − φ0 (x)u(x) dx
R
Z
[by parts] = φ(x)Du(x) dx = function-type distribution Du applied to φ 2
R
Exercises.
1. Show that |x| is differentiable as a distribution and

d 1 if x > 0
|x| =
dx −1 if x < 0
2. Show that the derivative of

1 if x > 0
H(x) =
0 if x < 0
is δ.
3. Show that (φ, δ 0 ) = −φ0 (0).
4. More generally, show that if f (x) has continuous derivatives on (−∞, x0 ]
and on [x0 , +∞) and (possibly) has a jump discontinuity at x = x0 , then its
derivative (in distribution sense) is fc0 (x)+[f (x0 +)−f (x0 −)]δ(x−x0 ) where
fc0 (x) denotes the derivative of f (x) as a function on (−∞, x0 ) ∪ (x0 , ∞).

2.4. Convolution of distributions with test functions. If g is a test

function, then for each fixed x, the function g(x − .) defined as g(x − .)(t) =
g(x − t) (as a function of t) is also a test function (it is a flip and a shift of
g). Note that, if φ is another test function, then
Z Z
ψ(t)g(x − t)dt = ψ(x − t)g(t)dt
R R
It is natural to define ψ ∗ u, for any distribution u, by
(ψ ∗ u)(x) = (ψ(x − .), u(.))
It is not hard to see (with some work) that ψ ∗ u is a function-type distri-
bution, even ψ ∗ u ∈ C ∞ ! (The intuitive reason is that the dependence on x
is only through φ, which is C ∞ .) Convolution is regularizing (smoothing).

Very important example: ψ ∗ δ = ψ, or,

Z Z
(4) ψ(x − t)δ(t)dt = ψ(x) = ψ(t)δ(x − t)dt
R R
Why: (ψ ∗ δ)(x) = (ψ(x − .), δ(.)) = ψ(x). 2
Equality (4) is the continuous analogue of the finite dimensional fact
x = nk=1 xk ek where x = (x1 , . . . , xn ) and e1 , . . . , en is the standard basis.
P
And (4) can be intuitively interpreted as: a function ψ(x) can be decom-
posed as a superposition of impulses located at t, of amplitude ψ(t).
6 RODICA D. COSTIN

2.5. The Fourier transform of distributions. Let us see how the Fourier
transform Ff of a super-nice function (from the point of view of the Fourier
transform), say, f ∈ S acts on test functions:
Z ∞ Z ∞ Z ∞
1
(φ, Ff ) = (Ff )(ξ)φ(ξ)dξ = dξ √ dx e−ixξ f (x)φ(ξ)
−∞ −∞ 2π −∞
Z ∞ Z ∞ Z ∞
1
= dx f (x) √ dξ e−ixξ φ(ξ) = dx f (x) (Fφ)(x) = (Fφ, f )
−∞ 2π −∞ −∞
It is natural then to define:
Definition 3. For any distribution u on S (i.e. tempered distribution)
define its Fourier transform Fu by the rule
(φ, Fu) = (Fφ, u) for any test function φ
Examples.
1. The Fourier transform of δ is 1 (well, up to a normalization constant...
you need to check the normalization you are using). For the one here
1
Fδ = √
2π
(check!) Also
1
Fδ(x − a) = √ e−iaξ
2π
Exercise.
Show that δ(−x) = δ(x). Find the Fourier transforms of 1, and of eibx , sin x, cos x.
 DISTRIBUTIONS 7

3. Calculus with distributions

3.1. Limits of distributions.
Definition 4. A sequence of distributions {un }n converges to the distribu-
tion u (all defined on the same space of test functions) if (φ, un ) → (φ, u)
for all test functions φ.
The point is that distributions can be obtained as limits of functions.
3.2. Fundamental sequences for δ. Intuitively, one may think of δ(x) as
a ”function” which
R ∞ is zero at all points except for x = 0, where ”it is infinite”,
and such that −∞ δ(x)dx = 1. Of course, there are no such functions, but
δ(x) is a limit (in distribution sense) of functions which are zero outside
intervals around 0), of lengths shrinking to zero, while the total integral of
these functions remains one. Of course, in this case the maximum of these
functions must go to infinity.
There are many sequences of functions which converge, in the distribution
sense, to Dirac’s δ function. They are called fundamental sequences for δ.
3.3. A sequence of step functions. Consider a narrower and narrower
step, centered at 0, of total area 1:
 n
 2 if x ∈ − n1 , n1
 
∆n (x) =
0 otherwise


Then limn→∞ ∆n = δ. Indeed, for any test function φ,

Z 1  
n n 1 1
(φ, ∆n ) = φ(x) dx = φ(cn ) for some cn ∈ − ,
2 −1 n n
n

(by the mean value theorem) which converges to φ(0) = (φ, δ).
3.4. Other sequences. We could smooth out the corners of our step func-
tions to make them even C ∞ .
In fact, one can find many fundamental sequences for δ by rescaling, as
follows. Pick a function f which is C ∞ on R Rwhich is zero outside some
∞
interval [−M, M ] and has total integral one: −∞ f (x)dx = 1. For each
 > 0 use rescaling to shrink [−M, M ] to [−M, M ] while dialting f so that
the total area remains one: define
1 x
f (x) = f
 
R∞
Then f (x) = 0 outside [−M, M ] and −∞ f (x)dx = 1 (check!). We have
lim→0 f = δ since
Z ∞ Z ∞ Z M
(φ, f ) = f (x)φ(x)dx = f (y)φ(y)dy = f (y)φ(y)dy
−∞ −∞ −M
Z M
= {f (y) [φ(y) − φ(0)] + f (y)φ(0)} dy
−M
8 RODICA D. COSTIN

M M
φ(y) − φ(0)
Z Z
= φ(0) f (y) dy + yf (y) dy
−M −M y
and using the mean value theorem for the function φ(y)
Z M
= φ(0) +  yf (y)φ0 (c) for some c ∈ (0, M )
−M

hence the last term in the sum goes to 0 as  → 0 and the limit is φ(0).

4. Green’s function
4.1. Non-homogeneous conditions can be reduced to homogeneous
conditions, but for a non-homogeneous equation. If the initial and
boundary conditions are not zero, the problem is first turned into one with
homogeneous conditions (but with a non-homogeneous term) by subtract-
ing from the unknown function a function witch has the prescribed initial
and/or boundary conditions. This is easiest to see when illustrated on a few
examples.

4.1.1. First order equations. Consider the simplest example

dy
(5) dx + a(x)y = g(x), y(x0 ) = y0 , for x ≥ x0

We find a function h(x) so that h(x0 ) = y0 , then substitute y = h + u.

For example, we can take h(x) = y0 . Problem (5) becomes
du
dx + a(x)u = g(x) − a(x)y0 , u(x0 ) = 0, for x ≥ x0

which is the type (11) studied before.

4.1.2. Second order equations. An equation with non-homogeneous bound-

ary conditions, say:

P (x)y + Q(x)y 0 + R(x)y = 0, y(0) = A, y(1) = B

can be transformed into a second-order non-homogeneous equation, but with

homogeneous boundary conditions: just find a function g(x) so that g(0) = A
and g(1) = B, then substitute y = g + u. (For example, one could take
g(x) = (B − A)x + A). The problem becomes

P (x)y 00 + Q(x)y 0 + R(x)y = f (x), y(0) = A, y(1) = B

where f (x) = −(P (x)g 00 + Q(x)g 0 + R(x)g).

So from now on, when dealing with non-homogeneous equations, we can

restrict our considerations to homogeneous conditions.
 DISTRIBUTIONS 9

4.2. Superposition. Suppose we have a linear, non-homogeneous differen-

tial equation, Lu = f where f (x) is a given function (the forcing term, or
d
the non-homogeneity), and L = L(x, dx ) is a linear differential operator, for
example,
d du
(6) L= dx + α for equation dx + αu = f (x)
d du
(7) L= dx + a(x) for equation dx + a(x)u = f (x)
(8)
1
 d d 1
 d
− dx (p(x) du
 
L= w(x) − dx (p(x) dx ) + q(x) for equation w(x) dx ) + q(x)u = f (x)

∂ ∂2
(9) L= ∂t − ∂x2
for equation ut − uxx = f
and so on.
Then: if u1 solves Lu1 = f1 (x), and u2 solves Lu2 = f2 (x), then a
linear combination u = c1 u1 + c2 u2 solves Lu = c1 f1 + c2 f2 (because L is
linear). Similarly, we can superimpose P any number of solutions:
P if uk solves
Luk = fk (x) for k = 1, . . . , n then u = k ck uk solves Lu = k ck fk .
The same is true if we also have some homogeneous conditions, initial
(IC) or boundary (BC). A condition is called homogeneous when a linear
functional of u is required to be zero. For example, we could consider the
problem Lu = f with
-L as in (7) and the (IC) u(x0 ) = 0
-L as in (8) and the (IC) u(x0 ) = 0, u0 (x0 ) = 0
-L as in (8) and the (BC) u(x0 ) = 0, u(x1 ) − 3u0 (x1 ) = 0
-L as in (9) and the (IC) u(t0 , x) = 0 and (BC) u(x0 ) = 0, u(x1 ) = 0
and so on.
For this non-homogeneous problems with homogeneous conditions, again,
a superposition of forcing terms is solved by a superposition of solutions.

Why not consider a ”continuous superposition”? After all, f can be

written as a ”superposition of impulses”:
Z ∞
f (x) = f (t)δ(x − t) dt
−∞
Then solve, for each parameter t, the equation LG = δ(x − t), denote its
solution by G = G(x, t), and then superimpose:
Z ∞
(10) u(x) = f (t)G(x, t) dt
−∞
is expected to solve Lu = f . We can similarly solve homogeneous (IC) or
(BC) problems.
G(x, t) is called the Green’s function of the problem.

The tasks ahead are then: to find ways of calculating the Green’s function
of given problems, to show that the superposition (10) is indeed a solution,
10 RODICA D. COSTIN

and to see what information about solutions can be read directly in the
Green’s function.
4.3. First examples.
4.3.1. The simplest example, C ∞ case. Consider L as in (7) and the homo-
geneous initial value problem
du
(11) dx + a(x)u = f (x), u(x0 ) = 0, for x ≥ x0
Let us find its Green’s function: solve
dG
(12) + a(x)G = δ(x − t), G(x0 , t) = 0 for all t
dx
R
The integrating factor is exp( a(x) dx).
Assume a ∈ C ∞ . Then so is the integrating factor, and then we can
multiply the equation by it (since δ can be multiplied by C ∞ functions) and
we get
d  Rxx a(s) ds  Rx
a(s) ds
e 0 G = e x0 δ(x − t)
dx
hence
d  Rxx a(s) ds  Rt
a(s) ds
(13) e 0 G = e x0 δ(x − t)
dx
Recalling that
dH d
= δ(x), and similarly H(x − t) = δ(x − t)
dx dx
integrating (13)2
Rx Rt
a(s) ds a(s) ds
e x0 G(x, t) = e x0 H(x − t) + C(t)
therefore
Rx Rx
− a(s) ds
G(x, t) = e− t a(s) ds
H(x − t) + e x0 C(t)
| {z } | {z }
particular sol. general sol. homog. eq.
R x0
Imposing the initial condition we obtain C(t) = −e− t a(s) ds H(x 0 − t)
hence
Rx
(14) G(x, t) = e− t a(s) ds
χ[x0 ,x] (t)
where 
1 if t ∈ [x0 , x]
χ[x0 ,x] (t) = H(x − t) − H(x0 − t) =
0 otherwise
is a step function.
The solution of (11) is
Z ∞ Z x Rx
u(x) = f (t)G(x, t) dt = f (t)e− t a
dt
x0 x0
2We need to use here the fact that the only distributions whose derivatives are 0 are
the constants, which requires a proof (not included here).
 DISTRIBUTIONS 11

4.3.2. The simplest example, non-smooth coefficients case. Consider again

the problem (11), only now with a(x) not C ∞ (for example, we could have
a(x) = |x|). Then we do not know that we can multiply (meaningfully) the
distribution δ by the integrating factor.
Then the equation is solved as follows: we solve for x < t, and separately
for x > t and we match the two pieces. There are two main facts here.

Fact 1. We can rely on the intuition that δ(x) = 0 for x < 0 and also
δ(x) = 0 for x > 0. (This can be rigorously stated, but we will not do that.)
Therefore δ(x − t) = 0 for x < t and also for x > t.

Fact 2. Examining (12) we see that G (as a function of x) has a jump

discontinuity of magnitude 1 at x = t. This condition will match the two
solutions on the two intervals.
dG
Now let us solve (12). For x < t Rthe equation is dx + a(x)G = 0 with
x
− a
the general solution G(x, t) = A(t)e . Imposing the initial condition we
x0

must have A(t) = 0, hence G(x, t) = 0 for x < t.

For x > t we Rhave again dG dx + a(x)G = 0, with the general solution
− tx a
G(x, t) = B(t)e . There are no other conditions on this interval, and
B(t) is determined from the jump condition at x = t: we must have
G(x, t) x=t+
− G(x, t) x=t−
=1
hence B(t) = 1, giving for G the formula (14).
12 RODICA D. COSTIN

d
4.4. Autonomous equations. If the coefficients of L = L( dx ) do not de-
pend on x, the equation Lu = f (x) is called autonomous. For example, (6),
(9) are autonomous equations.
Autonomous equations are translation-invariant, and we want to take
advantage of that. More precisely, if u(x) solves Lu = 0 then u(x − t) solves
the same (homogeneous) equation for any t. And is u(x) solves Lu = f (x)
then also u(x − t) solves L[u(x − t)] = f (x − t) for any t. So instead of
searching for a Green’s function G(x, t) by solving LG = δ(x − t) with
boundary conditions, we could first solve for LF = δ, then add to F (x − t)
the general solution of the homogeneous equation, and then impose the
boundary conditions, the result being the Green’s function. The procedure
is illustrated below. But first, a definition.
d
Definition 5. Let L = L( dx ) be an autonomous differential operator.
A solution F (x) of LF = δ is called a fundamental solution of the differ-
ential operator L.
R
Then upart (x) = (f ∗ F )(x) = f (t)F (x − t)dt is a particular solution of
Lu = f . The general solution of Lu = f is upart + uhomog where uhomog is
the general solution of Lu = 0.
4.4.1. A simple example. Consider the autonomous problem (6) (which, of
course, is a particular case of (7), for a(x) ≡ α). Its fundamental solution
is found as in §4.3.1 (only calculating for t = 0): we find a solution of
dF
(15) dx + αF = δ(x)
(no conditions imposed). Multiplying be the integrating factor eαx and
integrating we obtain
F (x) = e−αx [H(x) + C]
Any particular value for C gives a fundamental solution of L. For exam-
ple, for C = − 12 we obtain the symmetric fundamental solution F (x) =
1 −αx .
2 sign(x) e
To find the Green’s function of L with condition u(x0 ) = 0, for x ≥ x0
we look for it in the form F (x − t) + C(t)e−αx (note that the free con-
stant in the solution of the homogeneous equation is allowed to depend on
t). We determine C(t) by imposing the boundary condition: G(x0 , t) =
e−α(x0 −t) H(x − t) − C(t)e−α(x0 −t) = 0 and solving for C(t) we get
G(x, t) = e−α(x−t) (H(x − t) − H(x0 − t))
R∞
The solution of the initial value problem is then given by u(x) = x0 G(x, t)f (t) dt
Rx
(which for x > x0 equals the familiar u(x) = e−αx x0 eαt f (t) dt).
Alternatively, the solution can be found as follows. A particular solution
of du
dx + αu = f is then (for, say, C = 0)
Z ∞
upart (x) = (f ∗ F )(x) = f (t)e−α(x−t) H(x − t) dt
−∞
 DISTRIBUTIONS 13

The general solution is obtained by adding the general solution of the

homogeneous equation:
ugen (x) = upart (x) + Ae−αx
where the constant A is determined by imposing the initial conditions.

4.4.2. A second order example.

d2
(i) Find the fundamental solutions of L = dx 2 , and
(ii) the Green’s function of −L with boundary conditions u(0) = 0, u(π) =
0 for x ∈ [0, π].
We need to solve: F 00 = δ.
(i) Solution I. By direct integration, recalling Rthat H 0 = δ we obtain
0 x
F (x) = H(x) + C, and integrating again F (x) = −∞ H(t)dt + Cx + D =
xH(x) + Cx + D.
Solution II. We solve the equation for x < 0, giving F (x) = ax + b, and
for x > 0, giving F (x) = cx + d, and we match the two solutions at x = 0
using the following.
1) −F 0 has a jump discontinuity of magnitude 1 at x = 0 (the graph of F
has an ”angle” at x = 0): F 0 (0+) − F 0 (0−) = −1, which gives c − a = −1,
and
2) F must be continuous at x = 0: F (0+) = F 0 (0−), giving b = d.
Therefore f (x) = cx + d for x > 0 and F (x) = (c + 1)x + d for x < 0,
which is, of course F (x) = cx + d + xH(x), the same we had before.
(ii) A fundamental solution of −L is −xH(x) hence we look for G(x, t) in
the form −(x − t)H(x − t) + a(t)x + b(t). Imposing the boundary conditions
G(0, t) = 0, G(π, t) = 0 we obtain G(x, t) = −(x − t)H(x − t) + (1 − πt )x.

5. Green’s function of self-adjoint Sturm-Liouville problems

5.1. Liouville transformation. Linear, second order, differential equa-
2
tions can be brought to the simple form ddzy2 − g(z)y = 0 by a change of
both dependent and independent variables. Moreover, the transformation
preserves norms and eigenvalues. This is the Liouville transformation, ex-
plained in the present section.
d
Given an eigenvalue equation L(x, dx )u = λu where
d d 2 d
L(x, dx ) = −P (x) dx 2 − Q(x) dx − R(x)

we saw that if
Z x 
Q 1 R
p(x) = exp , w= p, q = − p
x0 P P P
then substituting
p p0 q
P = , Q= , R=−
w w w
14 RODICA D. COSTIN

the operator L is presented in a self-adjoint form

   
d 1 d d
(16) L(x, dx ) = − p(x) + q(x)
w(x) dx dx
which is formal self-adjoint in a domain in L2 (I, w(x)dx) (where I is an
interval and some homogeneous boundary conditions are assumed).
We can simplify further the equation/operator, bringing it to a form with
p = 1 and w = 1, using the Liouville transformation as follows.
I. There is a change of the independent variable z = z(x) so that in the
new variable z the weight w and p become 1. To find this substitution, using
d2 02 d2
d
the fact that dx = z 0 dz
d
, and dx 00 d
2 = z dz 2 + z dz , L defined by (16)

d2
 
1 d
L= −p 2 − p0 +q u
w dx dx
becomes
d2
 
1 d
L= −pz 02 2 − (pz 00 + p0 z 0 ) + q u
w dz dz
pz 02 d2 pz 00 + p0 z 0 d
 
q
= − 2− +
w dz pz 02 dz pz 02
R x q w(t)
We want pz 02 /w = 1, therefore z(x) = a p(t) dt where a can be chosen
at will. With this choice L is
 00
d2 p0

z d q
L=− 2 − + +
dz z 02 pz 0 dz w
II. The first derivative term can be eliminated (for any linear second order
equation) by a substitution of the dependent variable of the form u = φy
where the function φ is suitably chosen. In our case, this substitution in the
eigenvalue equation Lu = λu gives
 00
p0
 00
d2 y d2 φ p0
 
dφ dy z dy z dφ q
−φ 2 −2 −y 2 − 02 + 0 φ − 02 + 0 y + φy = λφy
dz dz dz dz z pz dz z pz dz w
hence, and dividing by φ
(17) 
z 00 p0 dy
 00
d2 y 1 d2 φ p0
   
1 dφ z 1 dφ q
− 2 + + + − − + + y = λy
dz 2 φ dz z 02 pz 0 dz φ dz 2 z 02 pz 0 φ dz w
dy
The condition that the coefficient of dz vanishes is that
1 dφ z 00 p0
(18) 2 + 02 + 0 = 0
φ dz z pz
or
φ0 z 00 p0
2 + 0 + = 0 therefore φ2 pz 0 = C and choose φ = (pw)−1/4
φ z p
Furthermore, this change of coordinates is unitary:
 DISTRIBUTIONS 15

Theorem 6. Liouville transformation

Assume p(x), w(x) > 0 on [a, b], and p, p0 , q, w continuous on [a, b].
Then the change of variables (x, u) 7→ (z, y) given by
Z xs
w(t)
(19) z(x) = dt, u = φy, φ = (pw)−1/4
a p(t)
transforms the eigenvalue equation Lu = λu with L given by (16) into
2
(20) − ddzy2 + g(z)y = λy
where
d

1 dφ

(pw)1/4
 h i0 0
−1/4
(21) g(z) = q − φ =q− p (pw)
dz φ2 dz w
R b qw
Moreover, the change of variable preserves the L2 norm: if c = a p
then Z b Z c
|u(x)|2 w(x)dx = |y(z)|2 dz
a 0
and it is an isomorphism between the Hilbert spaces L2 ([a, b], w(x)dx) and
L2 ([0, c], dz).
Remark. A linear change of variables in x can be used to replace the in-
terval [0, c] by any other interval: the linear operator T defined as (T f )(t) =
f (ct/α) is an isomorphism between L2 ([0, c], dz) and L2 ([0, α], αc dt). Replac-
ing the constant weight αc by 1 represents just a multiplication of all functions
by the same factor, and does not modify the eigenvalues and eigenspaces of
operators.

Proof of Theorem 6. The formula for z(x) and φ have been found before.
To deduce the formula for g(z) we only need to simplify the coefficient of y
in (17); using (18) this coefficient becomes
1 d2 φ 1 dφ 2
   
q d 1 dφ q
− 2
+2 + =φ 2
+
φ dz φ dz w dz φ dz w
giving the first equality in (21); the second equality follows immediately
using the formulas for z 0 and for φ.
To show that the Liouville transformation preserves the norm we simply
change the variable of integration:
Z b Z c Z c √
2 2 2 1 1 2 p
|u(x)| w(x)dx = φ |y| w 0 dz = 1/2
|y| w √ dz
a 0 z 0 (pw) w
2

As a consequence, all the theorems we prove for (20) can be transcribed

for (16), and the first form is certainly much simpler (it is called a normal
16 RODICA D. COSTIN

form). In concrete problems, however, if we need specific formulas, we do

not want to be bothered by sometimes involved changes of variables, or, we
are interested in some w(x) which vanishes at a or at b. For this reasons we
will keep the form (16) for most times.
5.2. Existence and uniqueness of the Greens’s function for non-
homogeneous problems are not guaranteed. Not all non-homogeneous
Sturm-Liouville problems have solutions. And when they exist, solutions
may not be unique.
To see an example, solve
d2 u
(22) = f (x), u0 (0) = 0, u0 (1) = 0
dx2
We attempt to find the Green’s function of the problem. For this, we
need to solve
G00 = δ(x − t); G0 (0, t) = G0 (1, t) = 0 (∗)
One integration of (*) gives
G0 (x, t) = H(x − t) + C(t); H(−t) + C(t) = H(1 − t) + C(t) = 0
This is already impossible! No point in trying to integrate further.
What is happening?
If we had such a Green’s function, then (22) would have a solution for
any say continuous f , and the solution would be
Z ∞
u(t) = G(x, t)f (t)dt
−∞
Note
P −1 the similarity with the linear algebra problem Ax = b ⇒ x = A−1 b; xi =
Aij bj .
Let’s try to solve one of the simplest instances of (22), when f = 1.
The solution of this inhomogeneous equation is, as we know, any par-
ticular solution u0 plus the general solution of the homogeneous equation
u00h = 0. The latter is u0 = ax + b. For the former we can choose u0 = x2 /2.
Now we impose the boundary conditions:
(x2 /2 + ax + b)0 |x=0 = a = 0; (x2 /2 + ax + b)0 |x=1 = 1 + a = 0
giving −1 = 0. What does this mean? Remember from linear algebra
that Ax = B where A is a matrix is guaranteed to have a solution if A
is invertible. Likewise, an equation of the form Lf = g where L is some
differential operator is guaranteed to have a solution if L is invertible. The
operator d2 /dx2 defined on {f ∈ C 2 [0, 1]|f 0 (0) = f 0 (1) = 0} is simply not
invertible!
It is desirable to have a theorem which gives conditions guaranteeing ex-
istence and uniqueness of the Green’s function. This is done in the following
section.
 DISTRIBUTIONS 17

6. Conditions for existence and uniqueness of the solution to

real-valued, non-homogeneous Sturm-Liouville problems
In the present section we assume problems with real coefficients. Equa-
tions with complex coefficients can be split into the real and the imaginary
parts.

6.1. Existence of eigenvalues for real-valued Sturm-Liouville prob-

lems. Recall the following facts:
Theorem 7. Consider the linear second order differential operator in self-
adjoint form (16), with p(x), w(x) > 0 for x ∈ (a, b)3, and p, p0 , w, q contin-
uous on [a, b] and real-valued.
Denote H = L2 ((a, b), w(x)dx). Let α, α0 ∈ R, not both zero, and β, β 0 ∈
R, not both zero. Denote
D(L) = {u ∈ H | u0 , u00 ∈ H, p(αu + α0 u0 ) x=a
= 0, p(βu + β 0 u0 ) x=b
= 0}
Then D(L) is dense in H, and L is self-adjoint on the domain D(L).
Therefore:
(i) all the eigenvalues of L : D(L) → H are real, and eigenfunctions are
(can be chosen) real-valued,
(ii) eigenfunctions corresponding to distinct eigenvalues are orthogonal in
H,
(iii) the eigenvalues are simple,
(iv) there is a lowest eigenvalue, and the eigenvalues form an increasing
sequence λ1 < λ2 < . . . < λn < . . .
(v) limn→∞ λn = ∞, and the corresponding eigenfunctions form an or-
thogonal basis for H.
Notes.
1. If p(a) = 0, then no condition is needed at x = a in D(L), and similarly
if p(b) = 0, no condition is needed at x = b.
2. The derivatives of functions in L2 (but not necessarily differentiable as
functions) are considered in the sense of distributions–when these derivatives
are functions. For example |x| ∈ L2 (−1, 1) has its derivative |x|0 ∈ L2 (−1, 1)
(see Exercise 1. of §2.3).
Furthermore: For boundary conditions which are not regular, but mixed
(including the periodic case):
αu(a) + α0 u0 (a) + α2 u(b) + α20 u0 (b) = 0
βu(b) + β 0 u0 (b) + β2 u(a) + β20 u0 (a) = 0
results similar to those of Theorem 7 hold, except that the eigenvalues may
not be simple (they are at most double, since this is a second order differ-
ential equations).

3Note that p(x), w(x) may vanish at the endpoints x = a, x = b.

18 RODICA D. COSTIN

7. Green’s function for self-adjoint Sturm-Liouville problems

7.1. When the Green’s function of Sturm-Liouville problems exists
and is unique. We will find conditions which ensure existence and unique-
ness of the Green’s function of given self-adjoint Sturm-Liouville problems
and of the solution to non-nomogeneous problems. Theorems 8 and 9 em-
body the Fredholm alternative for self-adjoint Sturm-Liouville problems:
either the homogeneous problem has a nontrivial solution, or any non-
homogeneous problem has a unique solution. This theorem is true in more
general contexts: for compact operators, as are, for example, the integral
operators which give solutions to second order problems.
Remark. As before, L denotes a differential operator, and D(L) its do-
main. The initial and boundary condition of the problem are specified by
conditions on the functions in D(L). We choose D(L) to consist of square
integrable functions so that we can take advantage of the powerful Hilbert
spaces techniques and insights. On the other hand, the Green’s function
of a given problem is defined without reference to a precise space of func-
tions: one only needs to specify a differential operator, the domain of the
independent variable and boundary conditions; all calculations are done in
distribution sense. For example, instead of spaces of square integrable func-
tions, we could consider spaces of C ∞ functions (with the same boundary
conditions), or Lp functions.
Recall that λ = 0 is an eigenvalue of (L, D(L)) if and only if the ho-
mogeneous problem has a nontrivial (i.e. nonzero) solution (i.e. there is
u ∈ D(L), u 6≡ 0 so that Lu = 0).
Theorem 8 below states that the Green’s function is guaranteed to ex-
ist, and be unique, if there are no nontrivial solutions of the homogeneous
equation. Recall the example in §5.2 where the Green’s function either did
not exist, or was not unique, depending on a condition on the forcing term;
it has a nontrivial solution of the homogeneous problem, namely u(x) = 1.
Furthermore, Theorem 8 also states that in this case, then there is a
unique solution of the non-homogeneous problem, for any forcing term (in
the appropriate L2 space).
Theorem 8. Let (L, D(L)) be a regular Sturm-Liouville problem, as in The-
orem 7 with p > 0 on [a, b].
Denote by H = L2 ((a, b), w(x)dx) and h·, ·iw its inner product.
Assume λ = 0 is not an eigenvalue. Then:
1. The problem has a Green’s function.
2. The Green’s function is unique.
3. Denoting by u1 , u2 , . . . , un , . . . an orthonormal basis of H formed by
eigenfunctions then
∞
X 1
(23) G(x, t) = w(t) un (x) un (t)
λn
n=1
where the series converges in H for each t.
 DISTRIBUTIONS 19

4. For any f ∈ H the solution u ∈ D(L) of Lu = f is given by

Z b
(24) u(x) = G(x, t) f (t) dt := (Gf )(x)
a
and its expansion in terms of eigenfunctions is
∞
X 1
(25) u(x) = hun , f iw un (x)
λn
n=1

5. Furthermore,
∞
X
(26) w(t) un (t) un (x) = δ(x − t)
n=1

Remarks. Before looking at the proof of the Theorem, let us examine

its implications.
1. The existence of a Green’s function G(x, t) so that solutions u ∈ D(L)
of Lu = f are given by (24) can be stated as the existence of the inverse of
the operator L (with given conditions); the inverse in an integral operator
G, with integral kernel G(x, t), and whose domain needs to be specified.
Theorem 8 gives a class of problems for which the inverse is defined on the
whole Hilbert space: if 0 is not an eigenvalue, Theoren 8 point 4. states that
the operator G : H → D(L) is the inverse of the operator L : D(L) → H.
Otherwise, Theorem 9 establishes the domain of the inverse as the orthogonal
complement of the eigenfunction corresponding to the 0 eigenvalue.
2. Moreover, the operator G is bounded: from (25) and by Theorem 7
(iv), it follows that kuk ≤ M kf k where M = max 1/λ2n .
3. We see from (23) that
(27) w(x)G(x, t) = w(t)G(t, x)

Proof of Theorem 8. (To make it fully rigorous, we would need to be more

careful with some steps...) 1. We construct the Green’s function: Equation
Ly = δ(x − t) for x ∈ [a, b] can be written as
p0 (x) q(x) w(x)
−Gxx − p(x) Gx + p(x) G= p(x) δ(x − t)
hence
p0 (x) q(x) w(t)
−Gxx − p(x) Gx + p(x) G= p(t) δ(x − t)
Since the right hand side is zero for x < t, and also for x > t, solving on
these two intervals with their respective boundary condition we get

C(t) ya (x) for x < t
G(x, t) =
D(t) yb (x) for x > t
where ya , yb are nonzero solutions of Ly = 0, satisfying αya (a)+α0 ya0 (a) = 0,
respectively βyb (b) + β 0 yb0 (b) = 0 (since p 6= 0).
20 RODICA D. COSTIN

The two branches of G(x, t) must be matched at x = t. As in §4.4.2, we

must have G(x, t) continuous in x at x = t, and its x-derivative must have
a jump discontinuity of magnitude w(t)
p(t) at x = t:

C(t)ya (t) = D(t)yb (t)

(28) w(t)
D(t)yb0 (t) − C(t)ya0 (t) = p(t)

If ya and yb are linearly dependent, then yb =const ya then ya satisfies

the boundary condition at x = b as well, and since Lya = 0, then ya is an
eigenfunction to the eigenvalue 0, which contradicts the assumption that 0
is not an eigenvalue.
Therefore ya , yb are linearly independent, hence their Wronskian W :=
ya yb0 − ya0 yb 6= 0. This means that the linear system (28) has a unique
solution:
w(t)yb (t) w(t)ya (t)
C(t) = , D(t) =
W (t)p(t) W (t)p(t)
On the other hand we can calculate the Wronskian W (t) from the differ-
ential equation that ya,b satisfy, −py 00 − p0 y 0 + qy = 0, or, in matrix form
0 " #
0 1
 
y y
= q p 0
y0 p −p
y0
R
and, recalling that W =const exp[ Tr(M atrix)], it follows that W =const/p,
or W (x)p(x) =const≡ κ1 . Therefore C(t) = κw(t)yb (t), D(t) = κw(t)ya (t),
giving 
κ w(t) yb (t) ya (x) for x < t
G(x, t) =
κ w(t) ya (t) yb (x) for x > t
2. Suppose G1 and G2 are two Green’s function of the problem. Then
G1 − G2 belongs to D(L) and satisfies L(G1 − G2 ) = 0. Since 0 is not an
eigenvalue, we must have G1 − G2 = 0 hence G1 = G2 .
3. We know that the eigenfunctions of (L, D(L)) form an orthonormal
basis of H (by Theorem 7). Since G(·, t) ∈ H for each t (because it is
continuous on [a, b]), then
∞
X
(29) G(x, t) = cn (t) un (x) for some cn (t) with {cn (t)}n=1,2,... ∈ `2
n=1
and where
cn (t) = hun , G(·, t)iw , n = 1, 2, . . .
Now apply L to both sides of (29). We need to be careful: L is not bounded,
i.e. it is not continuous, therefore we cannot say that L(limN →∞ N
P
PN n=1 ∗) =
limN →∞ L( n=1 ∗). It turns out to be true in distribution sense. Here is a
justification.
We know that Lun = λn un for all n. Therefore
(30) λn cn (t) = λn hun , G(·, t)iw = hλn un , G(·, t)iw = hLun , G(·, t)iw
 DISTRIBUTIONS 21

Now,
Z b
(−pu0n )0 + qun G(x, t) dx = (un , wLG(·, t))

hLun , G(·, t)iw =
a
(we define, in the sense of distributions, as usual, (u, Lφ) = (Lu, φ)) hence
(un , wLG(·, t)) = (un , δ(x − t) w(t)) = un (t)w(t)
Question: Which steps above are not quite correctly justified?
Therefore
un (t)w(t)
cn (t) =
λn
which used in (29) gives (23).
4. Formula (24) for the solution of a second order linear equation follows
from the classical theory of linear differential equations (or can be checked
directly). Using (23) in (24) gives (25).
5. Besides the expansion (25), we can expand any u ∈ H as
∞
X X∞ Z b
u(x) = hun , uiw un (x) = un (t)u(t)w(t)dt un (x)
n=1 n=1 a
∞
Z b "X #
= [exchange order in distribution sense] un (t)un (x) u(t)w(t) dt
a n=1
therefore
∞
X
w(t) un (t)un (x) = δ(x − t)
n=1
2
22 RODICA D. COSTIN

7.2. When the Green’s function for self-adjoint Sturm-Liouville

problems exists and it is not unique.
Note: If 0 is an eigenvalue of a self-adjoint Sturm-Liouville problem, the
dimension of its eigenspace must be one (any eigenspace is at most two-
dimensional).
Theorem 9 shows that if there are nontrivial solutions of the homoge-
neous problem (i.e. 0 is an eigenvalue), then a non-homogeneous equation
is solvable only when the forcing term is orthogonal to the kernel of the
homogeneous problem; the solution is not unique because we can add to it
any solution of the homogeneous problem (i.e. in its kernel).

Theorem 9. Let (L, D(L)) be as in Theorem 7 with p > 0 on [a, b].

If 0 is an eigenvalue of L, then an equation Lu = f has solutions u ∈ D(L)
if and only if

hu0 , f iw = 0 for u0 ∈ D(L) with Lu0 = 0

When the solution exists, it is not unique, but there is a one-parameter

family of solutions, namely
∞
X 1
(31) u(x) = Cu0 (x) + hun , f iw un (x)
λn
n; λn 6=0

Proof.
We only need to retrace the proof of Theorem 8 and see what happens if
0 is an eigenvalue.
We first see that we cannot find a Green’s function. Then a direct verifi-
cation show that (31) satisfies Lu = f . Are there other solutions? Suppose
that u2 also satisfies Lu2 = f . Then u − u2 satisfies L(u − u2 ) = 0, hence
u − u2 is a solution of the homogeneous equation, hence u2 = u+ [a solution
in D(L) of the homogeneous equation].
If u2 is linearly dependent of u0 , then u2 has the same formula as u, only
with a different constant C.
If u2 and u0 are linearly independent, then there are two independent
solutions of the homogeneous equation satisfying the boundary conditions,
hence any solutions satisfies the boundary conditions, but this is not possible
unless α = α0 = β = β 0 = 0, contradiction! 2

7.3. The resolvent. We can use a line of reasoning similar to that of §7.1
and §7.2 to find the formula for the resolvent of L. Let z ∈ C, not an
eigenvalue of L. This means that 0 is not an eigenvalue of L − zI. Using
Theorems 8 and 9 for the operator L − z instead of L we find

Theorem 10. Let (L, D(L)) be a regular Sturm-Liouville problem, as in

Theorem 7 with p > 0 on [a, b].
(i) Let z ∈ C be not an eigenvalue of L.
 DISTRIBUTIONS 23

Then Green’s function G(x, t, z) of L − z has the expansion

∞
X 1
(32) G(x, t, z) = w(t) un (x) un (t)
λn − z
n=1
and the solution of Lu − zu = f , u ∈ D(L) is
∞
X 1
(33) u(x) = (L − z)−1 f = hun , f iw un (x)
λn − z
n=1
(ii) If z = λk is an eigenvalue of L then Lu − zu = f , u ∈ D(L) has
solutions if and only if huk , f iw = 0, and in this case there is a one-parameter
family of solutions, namely
∞
X 1
u(x) = Cuk + hun , f iw un (x)
λn − λk
n=1, n6=k

Theorem 10 can be formulated as an instance of the Fredholm’s alterna-

tive:
Given (L, D(L)):
either (L − z)u = f has a unique solution for every f ,
or there is a nonzero solution of (L − z)u = 0.

Remark: formula (32) shows that the Green’s function of the resolvent
(L−z)−1 of (L, D(L)) is analytic in the parameter z, except for z equal to the
eigenvalues of (L, D(L)), where G has poles of order one. This is intuitively
expected, by functional calculus. In infinite dimensions this happens only
for special operators, though.
24 RODICA D. COSTIN

8. Continuous spectrum
Eigenvalue problems on infinite intervals, or when coefficients are not
smooth enough, may have a continuum of ”eigenvalues”.
For example, consider the problem
(34) u00 + λu = 0, for x ∈ (−∞, ∞)
with the boundary condition
(35) u is bounded at ± ∞
Then√ any λ ≥ √0 is an ”eigenvalue”, corresponding to the ”eigenfunctions”
sin( λx), cos( λx). I used quotation makes because we can only use these
words in a specified a Hilbert space, and we specified none. However, the
problem (34), (35) has a valid physical significance, and we need a solid
mathematical foundation for its treatment.
To have an intuitive picture on what is going on with the problem (34),
(35), we should consider the even simpler operator
d
(36) L = −i dx
with the boundary conditions (35), that is, on
(37) D(L) = {u ∈ L2 (R) | u0 ∈ L2 (R), u is bounded at ± ∞}
We have
d iξx
−i dx e = ξ eiξx
so eiξx behaves like an eigenfunction corresponding to the eigenvalue ξ, for
any ξ ∈ R. Since eiξx does not belong to the Hilbert space L2 (R), we will call
it a generalized eigenfunction, and ξ will be called a generalized eigenvalue.
Moreover, these generalized eigenfunctions are ”complete” in L2 (R) in
the sense that any f in the Hilbert space can be expanded in terms of the
generalized eigenfunctions—only not as a series, but as an integral: any
f ∈ L2 (R) can be written as
Z ∞
f (x) = eixξ g(ξ) dξ
−∞

where, of course, g(ξ) = √12π fˆ(ξ) ∈ L2 (R) is the continuous analogue of the
coefficients of the expansion of f .
The Fourier transform is the unitary transformation of L2 (R) which di-
d
agonalizes the operator dx . Indeed, since
Z ∞
d
−i dx d
f (x) = −i dx F −1 fˆ(ξ) = −i dx
d
eixξ √12π fˆ(ξ) dξ
−∞
Z ∞
= eixξ √12π ξ fˆ(ξ) dξ = F −1 [ξ fˆ(ξ)]
−∞
we see that
−1
d
F −i dx F =ξ
 DISTRIBUTIONS 25

it is the operator of multiplication of fˆ(ξ) by ξ. These calculations are for-

mal, more work would be needed to justify differentiation under the integral
operator etc. But the result above, carefully stated is correct Given a func-
tion g, the multiplication operator Mg defined as Mg f = gf is the continuum
analogue of a diagonal operator.
d d
Notes: Recall that dx is skew-symmetric in L2 (R), hence ±i dx is sym-
d
metric, justifying the choice of the factor i in front of dx . We choose the
minus sign so that the eigenvalue is ξ (rather than −ξ for the other choice).
In general, in the case of continuous spectrum, we have generalized eigen-
functions, and, when they form a complete set, functions can be expanded
as integrals, rather than series.
The formal definition of the spectrum of an operator is as follows.
Definition 11. Consider an operator L, D(L). The resolvent set of L con-
sists of all the numbers z ∈ C for which (L − z)−1 exists and it is bounded.
The spectrum of L, σ(L), consists of all complex numbers which are not
in the resolvent set.
Note that λ ∈ σ(L) is an eigenvalue means that L − z is not one-to-one.
But there are other reasons why some points belong to the spectrum of L:
these are the generalized eigenvalues.
d
8.0.1. A first order example: L = −i dx on R. Consider the problem (36), (37),
with spectrum σ(L) = R, and which has no proper eigenfunctions (but
plenty of generalized ones). Take z in its resolvent set, that is, z ∈ C \ R.
This means can invert L−z and (L−z)−1 is a bounded operator. Let us find
the resolvent (L − z)−1 . A simple way to solve (L − z)u = f for f ∈ L2 (R)
(or any differential equations with constant coefficients defined on the whole
line) is to take the Fourier transform in the equation, which gives
fˆ
ξ û − z û = fˆ so that û =
ξ−z
therefore
fˆ(ξ)
Z ∞
(38) u(x) = √2π 1
eixξ dξ
−∞ ξ−z
which is a bounded operator, since, using the Cauchy-Schwartz inequality,
and then Parseval’s identity
Z ∞
dξ π
(39) 2
kuk ≤ 2
kfˆk2 = kf k2
−∞ |ξ − z| |=z|
To exhibit the Greens’s function of the problem, rewrite (38) as

fˆ(ξ)
Z ∞ Z ∞ ixξ Z ∞
e
(40) u(x) = √12π eixξ 1
dξ = 2π f (y)e−iξy dydξ
−∞ ξ − z −∞ ξ − z −∞
Z ∞ Z ∞ i(x−y)ξ
1 e
= f (y) dξ
2π −∞ −∞ ξ − z
26 RODICA D. COSTIN

hence
∞
eiξ(x−y)
Z
1
(41) G(x, y, z) = dξ
2π −∞ ξ−z

We can further simplify (41): for =z > 0 we have G(x, t) = ieizt H(t), which
is seen by deformation of the path of integration upwards in the complex ξ
plane, towards ∞ and collecting the residue at ξ = z if t > 0, and downwards
if t < 0, yielding zero. Similarly, for =z < 0, G(x, t) = ieizt H(−t).
We see that u(x) in (38) is not defined if z ∈ R. This alone does not imply
that the equation (L − z)u = f has no solution, for now it only implies that
a particular formula for a possible solution does not apply.
Let’s however try to solve

(42) (L − z)u = f ; f ∈ L2 (R), z ∈ R

We can proceed by variation of parameters. Let’s take the particular ex-

ample z = 0, as for general zR we can do a similar analysis by variation of
x 2
parameters. We get f (x) = i −∞ g(s)ds + C. Take g(x) = e−x ∈ L2 . Since
R∞ √
we want f ∈ L2 we must have C = 0. But then , since −∞ g(s)ds = π,
√
limx→∞ f (x) = i π and f is not in L2 , thus L is not invertible in L2 , where
L−1 was supposed to be defined.
Note that the generalized eigenvalues (z ∈ R) are still singular points of
the Green’s function, only in this case they are points of jump discontinu-
ities. But in its integral representation (41) they are poles of the integral
kernel. Also, the norm of the resolvent goes to infinity when z approaches
the spectrum of L, as seen in (39).
Of course, (38) coincides with the expression obtained by solving (L −
z)u = f by multiplying with an integrating factor: imposing the boundary
R x izy
condition (37) to the general solution u(x) = e−izx (C + i e f (y)dy) we
R∞
obtain u(x) = i 0 eizt f (x − t)dt for =z > 0 and the symmetric formula for
=z < 0.

2
8.0.2. A second order example: L = − xd2 on R. Let
d 2
(43) L = − dx 2

with the boundary conditions (35), that is, on

(44) D(L) = {u ∈ L2 (R) | u0 , u00 ∈ L2 (R), u is bounded at ± ∞}

Since (43) is the square of (36) we expect these two operators to have a
d2 ±ixξ
common set of generalized eigenvectors, and indeed, − dx 2e = ξ 2 e±ixξ so
that the spectrum of (43) , (44) is [0, +∞).

8.1. Example: Bessel operator.

 DISTRIBUTIONS 27

8.1.1. Bessel functions. Consider the Bessel equation

(45) x2 u00 + xu0 + (x2 − n2 )u = 0
For n real, we can assume n ≥ 0.
Solutions of Bessel’s equation are called Bessel functions. The point x = 0
is a singular point of the equation. It is not hard to see that there are
two independent solutions of the form xn , respectively x−n , multiplying a
convergent power series having an infinite radius of convergence. The first
solution is the Bessel function J:
1
Jn (x) = xn n 1 + O(x2 )


2
This is the only solution of (45) which is bounded at x = 0 (if n > 0).
It can be shown that all solutions of (45) are bounded for x → ∞ (they
oscillate and decay as x−1/2 ).

8.1.2. An eigenvalue problem for the Bessel operator. Consider the eigen-
value problem
1  2

(xu0 )0 + λ − nx2 u = 0, for x ∈ [0, ∞)
x
with boundary conditions
u is finite for x → 0 and for x → ∞
√
For any λ > 0 the Bessel functions Jn ( λx) are generalized eigenfunctions
and it can be shown that any f ∈ L2 ([0, ∞), xdx) can be expanded in terms
of the eigenfunctions: there exists a function F so that
Z ∞ √ Z ∞
f (x) = Jn ( λx)F (λ) dλ = Jn (tx)F (t2 ) 2t dt
0 0
A general representation theorem of this nature exists for self-adjoint
operators.
d
8.2. A cautionary tale: −i dx on L2 ([0, ∞) Formally self-adjoint is not
d
the same as self-adjoint! Consider the operator L = −i dx on D(L) =
2 0 2
{f ∈ L ([0, ∞), f ∈ L ([0, ∞), f (0) = 0}. Check first that the operator is
symmetric.
Now let’s look at (L + i) and see whether it is invertible.
For that, we have to solve the equation
−iu0 + iu = f ; f ∈ L2 ([0, ∞)
Equivalently,
u0 − u = if ; f ∈ L2 ([0, ∞)
By variation of parameters, and imposing the condition u(0) = 0 we get
Z x
u(x) = ex e−s f (s)ds
0
28 RODICA D. COSTIN
R∞
Take any f > 0 in L2 . Check that e−s f ∈ L2 and 0 e−s f (s)ds = Cf > 0.
It then follows that
u(x)e−x → Cf as x → ∞
and thus u 6∈ L2 .
In fact you can similarly check that for any z = x + iy with y < 0 (L − z)
is not invertible. Thus the spectrum of L contains the whole lower half plane
in z, far from being purely real! That’s not what happens with self-adjoint
matrices. In fact : −i dx d
on L2 ([0, ∞) on D(L) = {f ∈ L2 ([0, ∞), f 0 ∈
L2 ([0, ∞), f (0) = 0} is not self-adjoint!