Lecture 16: The adiabatic approximation:

introduction (11/10/2005)
Last week we considered the sudden approximation in which the Hamiltonian is changing abruptly.
Today, we will on the contrary consider the case in which the Hamiltonian is changing very slowly.
Additional reading if you wish: Griffiths 10.1
We will see that the n-th lowest energy eigenstate simply evolves into the n-the lowest energy
eigenstate of the new Hamiltonian that may be continuously transformed from the initial one. We
will use both time-independent as well as time-dependent perturbation theory to derive the result.

Slowly changing systems: Adiabatic theorem

Recall that last time we derived wavefunctions in the sudden approximation. It was valid when


where is the time interval in which the Hamiltonian changes from H0 to H00 and E is the typical
spread of the energies of the physical system. We will use the term external time for the
changes of the definition of the physical system and the term internal time for h
/E, the
typical periodicity of the internal processes. Recall that if |i can be written in terms of the
eigenstates |ni of H0 as well as the eigenstates |n0 i of H00 , then
|i =

X
n

cn |ni,

| 0 i =

X
n0

dn0 |n0 i

d n0 =

X
n

cn hn0 |ni

Once again, this result is valid if the external time is much shorter than the internal time. What
about the opposite limit? In this limit, the following adiabatic theorem is true:
d n 0 = cn

for n = n0

where the eigenstates are ordered so that they can be continuously connected if you continuously
change time. In other words, the amplitude for the system to be in the n-th lowest-lying eigenstate
of the future Hamiltonian equals the initial amplitude for the system to be in the n-th lowest lying
eigenstate of the past Hamiltonian.

Proving the adiabatic theorem

Because the evolution is a linear operator on the Hilbert space, it is sufficient to prove the adiabatic
theorem for the basis of eigenstates of the original Hamiltonian. An initial eigenstate |ni of H0 must
evolve into |n0 i, an eigenstate of H00 = H0 + H 0 , as long as H 0 is turned on very slowly.
Let us consider a special case in which the Hamiltonian changes along a straight line in the
space of operators which means that
H 0 = V f (t),

f (0) = 0,

f (T ) = 1

which means that the Hamiltonian gradually changes between H0 and H0 + V as t goes from 0 to
T . The assumption that the line is straight is not too constraining; you can always approximate a
curved line by a sequence of straight lines.
Draw a graph of eigenvalues of H(t) to show an example.
1

First-order calculations
Before we start to compute time-dependent perturbation, it is useful to recall that the eigenvalues
of H0 + V may be computed from first-order time-independent perturbation theory:
|n0 i = |ni +

Vkn
|ki,
k6=n En Ek
X

Vkn hk|V |ni

Lets now start with the time-dependent perturbation theory. We start with the state |ni at t = 0
which means that
hk|(0)i = 0 for k 6= n
where hk| is an H0 eigenstate and we want to prove that it will also be true that
hk 0 |(T )i = 0 for k 0 6= n0
where n0 = n refers to the same ordering and hk 0 | is an eigenstate of H0 + V . Note that the absolute
value of the normalization of |i is conserved by the unitarity of the Schrodinger equation (the
conservation of probabilities) but the phase is undetermined and we wont determine it either.
Now we will use first-order perturbation theory to determine |(t)i. The general formula we
derived previously says that
X
ck (t)|kieEk t/ih .
|(t)i =
k

where
cn (t) = 1 +
and, for k 6= n,

Vnn Z t
f (t0 )dt0
i
h 0

Vkn Z t
0
f (t0 )e(En Ek )t /ih dt0
i
h 0
This was discussed in lecture 10. A nice way to evaluate these integrals is per partes (by parts).
Note that
i
h
d h (En Ek )t0 /ih i
0
e(En Ek )t /ih =
e
En Ek dt0
which may seem as an unnecessary complication but it will be helpful because we also obtained the
denominator (En Ek ) in the time-independent perturbation theory and they may cancel. At any
rate, the formula above allows us to integrate by parts now:
ck (t) =

i
t
d h
Vkn
0
f (t0 ) 0 e(En Ek )t /ih dt0
ck (t) =
En Ek 0
dt
"
#
Z t
0
df
(t
)
Vkn
0
0
f (t)e(En Ek )t /ih
e(En Ek )t /ih dt0
=
En Ek
dt0
0
Z

The first term, without the integral sign, was evaluated using f (0) = 0. Can we get rid of the
second term, the integral? Yes, we can, assuming that the oscillations will be much faster than the
time-dependence of df /dt0 which means if
1 df
|En Ek |

f dt
h

If this condition is satisfied, there will be massive cancellations from the phase that rotates around
in the complex plane and that is only slowly modulated by df /dt. We divided the inequality by
2

f because we are really comparing the integral to the first term that is proportional to f . The
condition is nothing else than the adiabatic condition. It requires that the external changes
occuring with angular frequency (1/f )(df /dt) are much slower than the internal motion whose
angular frequency is |En Ek |/
h. Finally, when you drop the integral, you see that the formula
for ck (t) exactly agrees with the formula for the first-order time-independent perturbation theory
calculating |n0 i, the eigenstate of H0 + V . In this sense, the change of basis induced by the timeindependent perturbation theory exactly cancels or exactly agrees with the change of the actual
state induced by the time-dependent perturbation theory.
If you want to use the word cancel, it is useful to think about the original quantity hk 0 |(T )i
that had to cancel for k 0 6= n0 . It does, to the leading order, if you realize that a conjugation (plus
kn-interchange) of the first formula in this subsection yields
hk 0 | = hk|

Vkn
hn|,
E

E
n
k
n6=k
X

Vkn hk|V |ni,

k = k0

The scalar product hk 0 |(T )i is then

X
Vkn
Vkn

Ek T /ih = O(V 2 )
hk|
h
n| |nieEn T /ih +
e(En Ek )T /ih |kie
E

E
E

k
k
=n n
n
6=k
k6
X

which cancels to the leading order in V . The only two contributions come from k = k and n
= n.
= h
They are proportional to hk|ki
n|ni = 1 and they cancel.
The first-order approximation is valid if V is small. If it is not small, we can divide V the total
change of the Hamiltonian to N pieces whose size is V /N. Since each of these transitions
induces changes of the wavefunction of second order, i.e. O(V 2 /N 2 ), the total change will be
V2
O N 2
N

= O(V 2 /N) 0,

This implies that if the changes of the Hamiltonian are (very) slow, the adiabatic theorem is (very)
accurate for arbitrarily large changes of the Hamiltonian.

Example 1: particle in a box

Consider a particle in the infinite well potential, i.e. a one-dimensional box. We will suddenly change
the size (or length) of the box from
a 2a

i.e. from [0, a] to [0, 2a]. What happens? First, ask the question in classical physics. If you change
the size of the box abruptly, the particle will typically be away from the wall and the instantly
changed size of the box wont change the kinetic energy of the particle. On the other hand, if you
increase the size slowly, the particle will be losing its kinetic energy because whenever it hits the wall,
the wall is escaping and the particle is reflected with a smaller velocity. No heat is being transferred
in this process which is why it is called an adiabatic process. How much will the kinetic energy
change in this adiabatic process? Recall that
pV = const,

+1

where is one half of the number of degrees of freedom; = 3/2 for monotonic gas or 5/2 for
diatomic gas or 3 for complex molecules. In our case, = 1/2 and thus = 3. Doubling the volume
3

(length) V adiabatically therefore implies changing the pressure by the factor of eight. This means
that the velocity goes to one half of the initial value: the momentum of the particle gets to one half
and the time between two reflections from the wall gets multiplied by four (because the box is twice
as big and the particle is twice as slow). In other words, the kinetic energy should decrease to one
quarter.
How does it look like in quantum mechanics? The initial (ground state) wavefunction is
1 (x) =

2
x
sin
,
a
a


Ei = E1 =

2h
2
2ma2

Let us first use the sudden approximation. The box suddenly doubles and the final wavefunction is
f (x) =

cn nf inal (x)

n=1

where the coefficients are the simple inner products with

q the big-box wavefunctions in which some
f inal
factors of two are changed namely with n (x) = a1 sin(xn/2a) i.e. the integrals

Z




n=2:
nx
x
2 a
sin
sin
dx =
cn =
a 0
2a
a
n=
6 2:

2
2
2
a

sin(n/2)
2(n/2a/a)

sin(n/2+)
2(n/2a+/a)

These is a rather obscure result, but when we square it, it becomes better:
2

|cn | =

32
(n2 4)2 2
1
2

n = 1, 3, 5 . . .
n=2
n = 4, 6, 8 . . .

The final average energy is computed as a weighted average of the eigenvalues Enf
hEf i =

n=1

|cn |

Enf

= Ei ,

Enf

1
=
4

n2 2 h
2
2ma2

and it is of course conserved because you can express it using the same integral. This statement
may sound as a complicated nontrivial assertions about some nontrivial sums involving functions of
n, but it must right and therefore you can use such statements to calculate these sums.
On the other hand, if we increase the size of the box slowly, the particle will remain in the ground
state as dictated by the adiabatic theorem. Because the size of the box doubled and E 1/a2 , the
energy will go to one quarter of the initial value exactly as in our result from classical physics!

Example 2: electron in magnetic field

Imagine that the field B is in the positive z-direction, and so is the spin: |+i. The sign in the vector
indicates the spin relatively to the direction of the magnetic field. The states |+i and |i are energy
eigenstates if the magnetic field is turned on.
If you suddenly revert the magnetic field, the absolute spin wont be changed, but the relative
spin with respect to the magnetic field will be reverted:
|+if = |ii ,

|if = |+ii .

However, if you now rotate the direction of the magnetic field slowly, the adiabatic derivation above
assures us that the electronic spin will follow it. The higher energy eigenstate will evolve into a new
4

higher energy eigenstate, and so forth; the relative orientation of the spin and the magnetic field
wont change.
Whats remarkable is that even a (very) weak magnetic field is able to lead the electronic spin
in this way; however, it must be rotated (very) slowly. How fast can the field change to be able to
drag the spin of the electron with it? The internal motion has frequency
~
E
2~ B
e
=
= B
h

m
Typical magnets in your lab can have 1 Tesla, and therefore
=

1.602 1019 C
(1 Tesla) = 1.76 1011 s1
9.109 1031 kg

The units of the results are really radians per second and therefore the period is 36 picoseconds. Its
enough to be patient for 36 picoseconds in reverting the field, and the electron will follow it! Very
good. We can afford a weaker field. Take the Earths magnetic field 50 microtesla, and replace the
electron by a nucleon which is harder to rotate because its magnetic moment is smaller, roughly
n 11, 000e . Even with these numbers you obtain 1.8 milliseconds. These are very short time
intervals.
But you should realize one thing: these changes of the magnetic field cant align the spins. For
example, if the spin is perpendicular to the magnetic field (for example spin in the x-direction and
the magnetic field in the z-direction), we have seen that the spin will precess around the z-axis.
If you now slowly change the direction of the magnetic field, the spin will remain perpendicular
to the magnetic field and it will precess about its magnetic axis. Even without the adiabatic
approximations you know that the direction of the precession will be reverted if you change the
magnetic field to the opposite one either quickly or slowly, it does not matter.
Note that the fact that the relative orientation between the magnetic field and the spin wants to
be conserved in the adiabatic process morally follows from the requirement that the energy wants
to be preserved as much as possible.

Level crossing
Reference: Phys. Rev. A 23, 3107 (1981).
So far we assumed that the energy eigenstates can be nicely ordered and they are never equal.
But what happens if you imagine that two eigenvalues cross? Draw a graph with a big X-cross, and
label the upper vertices as |2i, |10 i and the lower vertices as |1i, |20i. It is a graph of energy as a
function of q (imagine q is time). If you follow the straight lines, you might think that |1i will evolve
into |10 i and |2i will evolve into |20 i.
What does the adiabatic approximation say about it? It does not say anything because it requires
the inequality
En Ek
1 dq

q dt
h

which is inevitably violated when two relevant eigenvalues agree.

Fine. The original adiabatic approximation is not the only method in the world and we would
like to answer the question anyway. What is the final state originating from |1i?
The answer is that if dq/dt is fast in a new sense to be defined soon then |1i evolves to |10 i
along the straight lines. If dq/dt is slow, then the adiabatic crossing evolves |1i to |20 i. Draw the
previous graph but replace the cross in the middle by a hyperbola oriented in such a way to have
horizontal lines in the middle.
5

Level crossing: calculation

Start with the eigenstates
H0 (q)|1i = E1 (q)|1i,

H0 (q)|2i = E2 (q)|2i.

You can always conjugate H0 in such a way that the states |1i, |2i are q-independent. Assume they
are. Now imagine that the full Hamiltonian has a small, q-independent perturbation V :
H = H0 + V,

V11 = V22 = 0;

The Hamiltonian is then

H=

V12 = h1|V |2i =

6 0

E1 (q) V12
V12
E2 (q)

and its eigenvalues are (a is lower, b is higher)

E1 + E2

Ea (q) =
2

(E1 E2 )2
+ |V12 |2 ,
4

E1 + E2
Eb (q) =
+
2

(E1 E2 )2
+ |V12 |2 .
4

Check that their sum is the trace of the Hamiltonian matrix while their product is the determinant.
A nice way to parameterize the eigenstates is in terms of the angle defined by
tan (q) =

2V12
,
E2 E1

where is equal to /2 at the crossing. The eigenstates are

|ai = + cos(/2)|1i + sin(/2)|2i,
|bi = sin(/2)|1i + cos(/2)|2i.
The minimal separation between Ea and Eb is at the crossing, (Eb Ea )qc = 2V12 . The adiabatic
condition that will evolve |1i into |20 i and vice versa is therefore (assume that V12 is real positive)
1 dq
2V12

.
q dt
h

The equations can actually be solved in closed form even if the process is a bit faster and the result
is
!
2V12
P120 = 1 exp
h
(1/q)(dq/dt)
The fact that it is not quite equal to one for faster processes is used to revert the populations in
atomic physics (negative temperature). In some sense, even the solar neutrino oscillations may be
interpreted as an example of this formula and V12 plays the role of the off-diagonal matrix elements
of the (squared) mass operator.