115C (QM III): The Double Well

Or, How I Learned to Stop Worrying and Love Approximations

David Grabovsky
April 28, 2021

Contents
1 Introduction 2
(a) Qualitative Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
(b) Expansion about a Minimum . . . . . . . . . . . . . . . . . . . . . . . . . . 3
(c) Zeroth-Order Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Perturbation Theory 4
(a) Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
(b) First-Order Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
(c) First-Order Wave Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
(d) Second-Order Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
(e) A Visual Demonstration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
(f) Convergence Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3 Variational Techniques 11
(a) One Gaussian: Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
(b) One Gaussian: Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
(c) Two Gaussians: Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . 13
(d) Two Gaussians: Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
(e) First Excited State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4 The WKB Method 16

(a) Tunneling at Zero Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
(b) The Quantization Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
(c) Preparation I: Large-a Quantization . . . . . . . . . . . . . . . . . . . . . . . 18
(d) Preparation II: Harmonic Approximation . . . . . . . . . . . . . . . . . . . . 19
(e) The Energy Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
(f) Tunneling and Imaginary Time . . . . . . . . . . . . . . . . . . . . . . . . . 20
(g) Instantons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

A References 23

1
 1 INTRODUCTION

1 Introduction
In this sequence of exercises, we consider the quartic double-well Hamiltonian,
p2
2
H= + V (x), V (x) = mλ x2 − a2 , (1.1)
2m
where m > 0 is the mass of the particle, a > 0 is a constant with units of distance, and
λ > 0 is a constant whose sole purpose is to make sure that V has units of energy. (Do NOT
think of λ as a small parameter for use in perturbation theory!)

This Hamiltonian is a rich and interesting source of intuition on the standard approx-
imation techniques of quantum mechanics. It is also a rare example of a system where
the calculations are nontrivial enough to pose a challenge, but manageable enough to done
explicitly and come away (hopefully) with insight instead of pain.

High-level summary. Let us summarize where we are headed. We begin in this section
with preparatory work: we will qualitatively understand the behavior of low-energy states
in this potential, bemoan the lack of an exact solution, zoom into one of the two wells and
approximate it by a harmonic potential, and then write down the exact solution to this
(approximate) problem. This is a zeroth-order attempt at a solution, since it only captures
the behavior of the ground state in one of the wells, and is insensitive to tunneling.
We therefore turn to perturbation theory in an attempt to do better, treating the non-
harmonic parts of the Hamiltonian as a “small” perturbation. After obtaining the first-order
corrections to the ground state energy and wave function, we will face the harsh truth:
the perturbation isn’t really small, and perturbation theory fails to converge or display
sensible tunneling behavior across both wells. We will then explore a variational approach
to the problem, where by the use of a rather ad hoc trial wave function, we will successfully
model the ground state. Finally, we will pass to a WKB analysis of the problem. We will
examine tunneling and construct a proof that the energy splitting between low-lying states is
nonperturbatively small. We will conclude by introducing imaginary time and its attendant
reformulation of tunneling in terms of the classical motion of instantons.

(a) Qualitative Analysis

Graph the potential V (x). Then, graph the approximate shape of the ground state wave
function ψ0 (x) and the first excited state wave function ψ1 (x). Label each one.
Then, try your best to solve the potential exactly, i.e. to obtain the eigenstates and
energies of the Hamiltonian. You will fail: the exact solutions cannot be written down in
closed form using elementary functions. Such is life sometimes.

Solution. As expected, the potential V (x) has a double well. It is graphed below (blue),
together with the ground state (orange) and first excited state (green) wave functions. The
precise shape of the wave functions should not be taken literally: their only salient features
are that they have bumps localized on the wells, and that the ground state is symmetric
while the first excited state is antisymmetric.

2
(b) Expansion about a Minimum 1 INTRODUCTION

Figure 1: The double well and its low-energy states (m = λ = ~ = 1, a = 2).

Looking up the double well on Wikipedia will immediately throw the reader into a
firestorm of special functions, series expansions, and so on. More directly, the Schrödinger
equation takes the schematic form
−ψ 00 + (c1 x2 + c2 x4 )ψ ∼ Eψ. (1.2)
This is simply a rather nasty ODE whose solutions certainly exist (and some of which look like
the wave functions above), but which cannot be expressed in terms of elementary functions.

(b) Expansion about a Minimum

Expand the potential about its minimum at x = a. That is, calculate the Taylor series of
V (x) at x = a to lowest nontrivial order. You should obtain a quadratic function of x; call
it V (2) (x). Thus rewrite the full potential in the form
V (x) = V (2) (x) + V 0 (x) = A(x − B)2 + V 0 (x), (1.3)
where A and B are constants that you should determine, and V 0 (x) is the non-quadratic
part of the potential, which you should also determine.

Solution. The quadratic approximation to V (x) is given by taking derivatives:

V (2) (x) = 4mλa2 (x − a)2 . (1.4)
In the notation of the problem, this gives A = 4mλa2 and B = a. As for the remaining bit
of the potential, there are various ways to write it down, none of which are too enlightening:
V 0 (x) = mλ(x − a)3 (x + 3a) = mλ(x − a)3 (x − a) + 4a =

= mλx4 − 6mλa2 x2 + 8mλa3 x − 3mλa4 . (1.5)

It may, at first sight, be disturbing that V 0 (x) contains a quadratic term. This is to be
expected, though; we expanded only around one of the two quadratic-looking minima, leaving
the other one intact. So its quadratic imprint survives in V 0 .

3
(c) Zeroth-Order Solution 2 PERTURBATION THEORY

(c) Zeroth-Order Solution

p 2
Consider the quadratic Hamiltonian H (0) = 2m + V (2) (x). By an appropriate change of
coordinates from x to a new coordinate ξ, show that H (0) takes the same form as the
p2
Hamiltonian for a harmonic oscillator, HQHO = 2m + 12 mω 2 ξ 2 , and find the frequency ω
of this oscillator in terms of the parameters a and λ. Write down the ground state wave
function and the ground state energy of H (0) .

Solution. Consider the change of coordinates given by a shift of x to the location of the
minimum at a: x −→ ξ = x − a. Then the Hamiltonian is

(0) p2
H = + 4mλa2 ξ 2 . (1.6)
2m
This is the QHO Hamiltonian, with the role of 21 mω 2 (the usual prefactor at the quadratic
term in the QHO) played by 4mλa2 . In other words, we have the identification
1 √ √
4mλa2 = mω 2 =⇒ ω = 8λa2 = 2a 2λ. (1.7)
2
This makes sense: the location a of the well also controls the steepness of its walls, which is
precisely what the frequency ω usually measures.
The ground state wave function is given by its usual expression:

 mω 1/4 √ !1/4 √ !
(0) − mωξ
2
2ma 2λ ma 2λ
ψ0 (x) = e 2~ = exp − (x − a)2 . (1.8)
π~ π~ ~

This ground state is a Gaussian bump in the well centered at x = a, and is an excellent
description of half of the true ground state of the system. But approximating V (x) by V (2)
ignores the well at x = −a, and the Gaussian wave function above doesn’t know about it.
The ground state energy, as usual, is

(0) ~ω √
E0 = = ~a 2λ. (1.9)
2
Notice that if we had expanded around x = −a instead, we would have obtained a similar-
looking Gaussian ground state and, importantly, the same ground state energy.

2 Perturbation Theory
In the last section we saw that the quadratic approximation of V fails to predict the correct
shape of the ground state wave function. In this section, we will attempt to fix this problem
by treating V 0 as a small perturbation on top of the unperturbed, exactly solvable Hamilto-
nian H (0) . As we will soon see, this is an approximation doomed to fail. However, that will
not stop us from first getting some useful results.

4
(a) Preparation 2 PERTURBATION THEORY

(a) Preparation
Recall that you found V 0 (x) in part 1(c). Copy it down again, but this time express it in
terms of ξ instead of in terms of x. Now look at the unperturbed Hamiltonian H (0) : are
there any degeneracies? If not, we are clear to use non-degenerate perturbation theory to
analyze it. If so, find the “good” states in preparation for degenerate perturbation theory.

Solution. For reference, the perturbation is

V 0 (x) = mλ(x − a)3 (x − a) + 4a = mλξ 3 (ξ + 4a) = mλξ 4 + 4mλaξ 3 .

(2.1)

The Hamiltonian H (0) is just a harmonic oscillator, so its energy levels are indexed by the
(0)
integers and are nondegenerate. (This means that the eigenstates ψn of H (0) are already
“good” states.) Whew!

(b) First-Order Energy

Calculate the first-order correction to the ground state energy due to the perturbation V 0 (x).

Solution. In nondegenerate perturbation theory, the first-order correction to the ground

state energy is given by the expectation value of the perturbation:
(1) (0) (0)
E0 = hψ0 |V 0 |ψ0 i (2.2)

There are several ways to go from here. Most naı̈vely, one could recall the explicit expressions
(0)
(1.5) for V 0 (x) and (1.8) for ψ0 (x), and then set about to calculate a dizzying number of
integrals. Things look much nicer in the ξ coordinates: here V 0 only has two terms, and
(0)
ψ0 (ξ) is precisely the wave function of the ground state |0i of the harmonic oscillator.
The calculation simplifies dramatically:
(1)
E0 = h0|mλξ 4 + 4mλaξ 3 |0i = mλ h0|ξ 4 |0i + 4mλa h0|ξ 3 |0i . (2.3)

We now argue that the ξ 3 term vanishes. This can be seen by writing down the integral
that defines this matrix element: the integrand is ξ 3 times a Gaussian, which is an odd
function being integrated over the symmetric interval (−∞, ∞), so the integral vanishes.
Equivalently, one can expand ξ ∼ a+ + a− in terms of the ladder operators. Observe that
no term in the expansion of ξ 3 will have an equal number of raising and lowering operators.
Thus the resulting states will always be orthogonal, and the inner product will vanish.
As for the ξ 4 term, the ladder operator method proves useful:
r  2
~ (1) ~
ξ= (a− + a+ ) =⇒ E0 = mλ h0| (a− + a+ )4 |0i . (2.4)
2mω 2mω
Now we need to expand the fourth power of the sum above, taking care to preserve the order
of all 16 resulting terms. Fortunately, not all of these terms contribute to the result: any
term with an unequal number of a+ and a− operators (e.g. a+ a− a2+ ) will act on |0i by raising
more times than it lowers (or vice versa), leading to a “mismatched” inner product like h0|2i.

5
(c) First-Order Wave Function 2 PERTURBATION THEORY

These terms, in addition to any term where a− acts directly on |0i, all vanish. Only 6 terms
that have two a+ and two a− operators, and of those, four immediately annihilate |0i:

h0| (a− + a+ )4 |0i = h0| (a+(a(

+ a(
− a− + (
(( a (
+ a(
− a(
+ a− + (
(( a (
+ a(
− a(
− a+ +
((
†
a† a |0i =


+ a− a− a+ a+ + a− a+ a− a+ +  aa
= h0| (a− a− a+ a+ + a− a+ a− a+ ) |0i = 2 h0|0i + 1 h0|0i = 3. (2.5)

All together, then, we have

2 2
3~2
 
~ ~
En(1) = 3mλ = 3mλ √ = . (2.6)
2mω 4ma 2λ 32ma2
To first order, then, the ground state has energy

(0) (1)
√ 3~2
E0 ≈ E0 + E0 = ~a 2λ + . (2.7)
32ma2
This makes sense: each next term enters with a higher power of ~.

(c) First-Order Wave Function

Calculate the first-order correction to the ground state wave function due to V 0 (x). You may
(0)
leave your answer in terms of the unperturbed harmonic oscillator eigenstates ψn —no need
to actually compute the explicit spatial wave functions.

Solution. The first-order correction to the ground state is given by

(1)
X hψn(0) |V 0 |ψ (0) i
|ψ0 i = (0)
0
(0)
ψn(0) , V 0 = mλξ 4 + 4mλaξ 3 . (2.8)
n6=0 E0 − En

The energy denominator is pretty quick to calculate:

√
 
(0) (0) ~ω 1
E0 − En = − ~ω n + = −n~ω = −2n~a 2λ. (2.9)
2 2
The main problem here is to calculate the matrix element
0
Vn0 = hn|V 0 |0i = hn|mλξ 4 + 4mλaξ 3 |0i =
 2  3/2
~ 4 ~
= mλ hn|(a− + a+ ) |0i + 4mλa hn|(a− + a+ )3 |0i . (2.10)
2mω 2mω
(The ladder operator method is the only viable approach this time, since writing down the
(0)
wave function ψm (ξ) explicitly in terms of Hermite functions will get us nowhere.) This
calculation is, to say the least, not entirely pleasant. The key is to verify that
√ √
(a− + a+ )4 |0i = 3 |0i + 6 2 |2i + 2 6 |4i ,
3
 √  √
(a− + a+ ) |0i = 2 + 2 |1i + 6 |3i . (2.11)

6
(d) Second-Order Energy 2 PERTURBATION THEORY

It remains to take the inner product of these results with all possible bras hn| except for the
ground state, since n 6= 0. The infinite sum above reduces to a finite number of terms:
2 X 3/2 X
hn|(a− + a+ )4 |0i (0) hn|(a− + a+ )3 |0i (0)
 
(1) ~ ~
|ψ0 i = −mλ |ψn i − 4mλa |ψn i =
2mω n6=0 n~ω 2mω n6=0
n~ω
 2 " √ √ #  3/2 " √ √ #
~ 6 2 (0) 2 6 (0) ~ 2 + 2 (0) 6 (0)
= −mλ |ψ2 i + |ψ4 i − 4mλa |ψ1 i + |ψ i =
2mω 2~ω 4~ω 2mω ~ω 3~ω 3
" √ #  1/2 " √ #
~ √ (0) 6 (0) ~ √ 
(0) 6 (0)
=− √ 3 2 |ψ2 i + |ψ4 i − √ 2 + 2 |ψ1 i + |ψ3 i .
3
64ma 2λ 2 3
64ma 2λ 3
(2.12)
This is a bit of a mess, but we can clean
√ it up a bit using dimensional analysis. Observe
that the quantities mλa4 and ~ω ∝ ~a λ both have units of energy. This means that
√
   
4 3 ~ ~
[mλa ] = [~a λ] =⇒ [ma ] = √ =⇒ √ = [1]. (2.13)
λ ma3 λ
Thus the prefactor that appears twice in (2.12) is indeed dimensionless, and we can give it
√
a name: let γ ≡ ~/ 64ma3 2λ . Then to first order, the full wave function is
√ √
(0) (1)
 √ √ √ 6√ 6
|ψ0 i ≈ |ψ0 i + |ψ0 i = |0i − 2 + 2 γ |1i − 3 2 γ |2i − γ |3i − γ |4i . (2.14)
3 2

(d) Second-Order Energy

Calculate the second-order correction to the ground state energy due to V 0 (x).

Solution. The second-order correction to the ground state energy is

D E2
(0) (0)
X ψn V 0 ψ0
(2)
E0 = (0) (0)
, V 0 = mλξ 4 + 4mλaξ 3 . (2.15)
n6=0 E 0 − E n
√
We already know that the energy denominator is −n~ω = −2n~a 2λ, and we calculated all
of the necessary matrix elements above. It remains to sum their squares:
2 2
hn|(a− + a+ )4 |0i hn|(a− + a+ )3 |0i
 4 X  3 X
(2) 2 ~ 2 ~
E0 = −(mλ) − (4mλa) =
2mω n6=0 n~ω 2mω n6=0 n~ω
 4    3 " √
2 #
~ 36 · 2 4 · 6 ~ 2 + 2 6
= −(mλ)2 + − (4mλa)2 + =
2mω 2~ω 4~ω 2mω ~ω 3~ω
" 4 3 #" √ #
~2 (1 + γ)  √ 
 
~ ~ 50 + 4 2
= − (mλ)2 + (4mλa)2 = 50 + 4 2 .
2mω 2mω ~ω 32ma2
(2.16)

7
(e) A Visual Demonstration 2 PERTURBATION THEORY

This is slightly worrying: the second-order energy correction looks larger than the first-
order correction. All together, to second order we have

(0) (1) (2)

√ ~2 h  √ i
E0 ≈ E0 + E0 + E0 = ~a 2λ + 3 + (1 + γ) 50 + 4 2 . (2.17)
32ma2

(e) A Visual Demonstration

Using a table of the first few stationary states of the harmonic oscillator and a graphing
software, plot both the unperturbed and first-order-corrected wave functions on the same
graph. What happens at x = −a? Does this trouble you?

Solution. See the attached Mathematica notebook for the computation. We pretty much
(0) (4)
just look up and hard-code the√ first four stationary QHO wave functions ψ0 (x), ..., ψ4 (x),
being careful to substitute 2a 2λ for ω and x − a for ξ wherever they appear.

Figure 2: PT wave function, m = λ = ~ = 1, a = 1.75.

(0) (0) (1)

The differences between ψ0 (orange) and ψ0 ≈ ψ0 + ψ0 (green) are minuscule. The
corrected wave function looks like it is shifted to the left. More precisely, what happens is
that the well at x = a is slightly steeper than its zeroth-order harmonic approximation for
(0)
x & a, and slightly shallower for x . a. The perturbative corrections to ψ0 respond to
these effects by making the wave function decay more steeply for x > a and less steeply for
x < a. This seems somewhat promising: the true ground state has a bump in both wells, so
the increased amplitude of the wave function for x < a indicates that perturbation theory
is at least trying to capture the effects of tunneling through the central barrier. Alas, no
bump appears at x = −a, and the corrected wave function seems to pretend that the second
well—which is certainly part of the perturbing potential—simply isn’t there.
Could this forebode a failure of perturbation theory?

8
(f) Convergence Issues 2 PERTURBATION THEORY

(f ) Convergence Issues
To summarize, we are given the Hamiltonian H = H (0) + V 0 (ξ), where H (0) is the harmonic
oscillator Hamiltonian in the coordinate ξ, and V 0 (ξ) is viewed as a small perturbation.
First-order perturbation theory for this problem seems to be tractible. However: Argue,
using any means you like, that perturbation theory does not converge for this system. To be
precise, this means that the all-orders perturbative ground state energy and wave function,
(∞) (0) (1) (2) (∞) (0) (1) (2)
E0 = E0 + E0 + E0 + · · · , ψ0 = ψ0 + ψ0 + ψ0 + · · · , (2.18)

are not equal to the true ground state energy E0 and wave function ψ0 . You do not need to
be rigorous by any means, but you do need to provide a convincing physical argument.

Solution. There are many different ways to argue here. We provide three, but these are
far from the only valid reasons for the failure of the convergence of perturbation theory.

Argument 1: V 0 as a perturbation. The “perturbation” V 0 (x) is not small! It is

certainly true that when x is near a, the full potential V (x) is well approximated by V (2) (x),
and so V 0 (x) is small there. But this approximation is only good locally, and as we move
away from x = a the difference V 0 (x) = V (x) − V (2) (x) grows large. In particular,

V 0 (−a) = mλa4 − 6mλa4 − 8mλa4 − 3mλa4 = −16mλa4 (2.19)

is huge in the deep-well regime where a is large. So it is not really reasonable to use V 0 (x)
as a perturbation to begin with. Also, V 0 (x) contains no explicit small parameters that
distinguish its size from that of V (x) or V (2) (x), so there is no reason to even suspect that
it should be small to begin with.

Argument 2: wave functions and analyticity. The content of this argument is really
only precise in the context of the WKB approximation and will be more fully fleshed out
by the end of the document. For now, we keep things vague. The basic idea is that several
quantities, like the WKB tunneling probability and the true energy splitting between the
ground and first excited states, have expressions whose functional form includes the factor
e−1/~ . The function f (x) = e−1/|x| is well-defined and smooth for all x 6= 0, and has a
removable singularity at x = 0 that can be “plugged” by defining f (0) = 0 to yield a
globally smooth function. And yet, it is a curious fact that the Taylor series of f at x = 0
is identically zero: every single term in the Taylor series vanishes, because the function is
simply too flat at the origin. It is a weird example of a smooth but non-analytic function.
Physically, this means that any attempt to capture effects that look like e−1/~ in a power
series is doomed to fail. We therefore say that such effects are nonperturbative, because they
are too small to be noticed by perturbation theory.

Argument 3: “formal” “proof.” Even in the absence of an explicit small parameter,

we can regard this brand of perturbation theory as a series expansion in powers of ~. If you
don’t like this idea, we could also—for purely pedagogical purposes—modify the problem
by hand and consider the perturbation εV 0 (x), where ε  1 is a small parameter. (Thus in

9
(f) Convergence Issues 2 PERTURBATION THEORY

(∞) (∞)
the following, think of ε as ~.1 ) In either case, we shall see presently that E0 and ψ0
diverge. The key will be to remember what this means, so a brief reminder follows.
Definition 2.1 (Convergence, in physics language). A power series
N
X
F (ε) = ak εk = a0 + a1 ε + a2 ε2 + · · · + aN εN (2.20)
k=0

is said to converge at ε = ε∗ if, upon fixing ε = ε∗ , increasing the number N of terms in the
series to infinity causes it to approach a finite limit. The set of ε∗ for which this is true is
called the domain of convergence D of the series, and the largest open interval (ε0 − r, ε0 + r)
contained in D defines the radius of convergence of the series.
(∞) (∞)
In the language developed above, both E0 and ψ0√ are power series in ε ∼ ~ around
(0) (0) 2
ε = 0. In particular, setting ε = 0 recovers E0 = ~a 2λ and ψ0 ∼ e−ξ . We will show
(∞)
that E0 has zero radius of convergence, which means that its power series converges for
ε∗ = 0 but for no other nearby value. Suppose, towards a contradiction, that the series
does converge for some small, positive value ε∗ of the perturbation strength. Thus the series
(∞)
E0 (ε) is convergent at ε = 0, with radius of convergence at least r = ε∗ . This implies that
(∞) (∞)
E0 (−ε∗ ) must converge just as well as E0 (ε∗ ). So far, nothing special about E (∞) has
(∞)
been used. But attempting to compute E0 (−ε∗ ) entails doing perturbation theory with
−εV 0 instead of +εV 0 , which flips the perturbing potential upside down. No matter how
small |ε∗ | is, the inverted potential −V 0 is unbounded below and leads to certain disaster. To
wit, it causes the amplitude of the wave function to diverge when |x|  a, rendering it (and
its energy) ill-defined. One sometimes describes this situation by saying that the particle will
spontaneously tunnel through the inverted potential barrier into a region of negative-infinite
potential. (See also the phrase “spontaneous vacuum decay.”) In any case, the point is that
(∞) (∞)
E0 (−ε∗ ) cannot possibly converge, so neither can E0 (ε∗ ) for any positive value of ε∗ .
(∞) (∞)
Thus E0 has zero radius of convergence, and the same argument also works for ψ0 .

Addendum. This argument was first given by Dyson in the context of the renormalized
perturbation theory commonly used in quantum field theory, but it applies equally well in
nonrelativistic single-particle quantum mechanics.
(1)
One may be extremely confused at this point: how come we got sensible results for E0
(1)
and ψ0 earlier?! The answer is rather subtle and depends on the order in which limits are
taken. If ε is held fixed and the number N of terms in the series is increased, the series
initially approaches the correct value, reaches its best approximation at some finite N = N∗ ,
and then begins to diverge from the right answer for N > N∗ . On the other hand, if we fix
N and then send ε −→ 0, the resulting N -term sum will approach its unperturbed (ε = 0)
value. This holds for any fixed N , so even if we choose N  N∗ (where the series is a bad
approximation for any finite ε), cranking down ε −→ 0 will make the approximation better
and better. For this reason, we call such series perturbative or asymptotic series.
More precisely, ε = ~1/2 . This is due to the annoying technical detail that V 0 consists of ξ 3 and
1
4
ξ terms, and the two terms produce energy and wave function corrections that come with half-integer
and integer powers of ~, respectively. This means that the ~-expansions of E0 and ψ0 both look like
a0 + a1/2 ~1/2 + a1 ~ + a3/2 ~3/2 + · · · . This is a minor detail, however, and doesn’t affect the argument.

10
 3 VARIATIONAL TECHNIQUES

3 Variational Techniques
Having understood what perturbation theory can and cannot do for us, we turn in this
section to the variational principle. We will choose variational wave functions of several
different forms, and examine the bounds they give on the ground state energy.

(a) One Gaussian: Minimization

2
Consider the trial wave function ψβ (x) = Ae−β(x−a) , where β > 0 is the variational param-
eter. Normalize ψβ by finding A, and use the variational principle to determine the value
of β that gives the lowest upper bound on the ground state energy of the system. Call this
value β∗ . Be warned: the result will not be pretty.

Solution. We begin by normalizing the wave function:

Z Z ∞
∗ 2 2
hψβ |ψβ i = dx ψβ (x)ψβ (x) = |A|2
dx e−2(x−a) /β =
R −∞
r  1/4  1/4
π ! 2β 2β 2
= |A| 2
= 1 =⇒ A = =⇒ ψβ (x) = e−βξ . (3.1)
2β π π
Next we need to use the variational principle, which states that

E0 ≤ hHiψβ = hψβ |H|ψβ i ≡ Ev . (3.2)

Now there are a number of ways to calculate the matrix element. Here is one:
 2 
p (2) 0
Ev = hψβ | + V + V |ψβ i =
2m
1
= hψβ |p2 |ψβ i + 4mλa2 hψβ |ξ 2 |ψβ i + 4mλa hψβ |ξ 3 |ψβ i + mλ hψβ |ξ 4 |ψβ i . (3.3)
2m
The first two terms are the kinetic energy and harmonic potential, and the second two terms
are the “perturbing” terms. We dutifully calculate each one in turn:
Z ∞
1 ~2 d2
~2 β
hT iβ = 2
hψβ |p |ψβ i = − dx ψβ∗ (x) 2 ψβ (x) = ,
2m 2m −∞ dx 2m
Z ∞
(2) 2 2 2 mλa2
hV iβ = 4mλa hψβ |ξ |ψβ i = 4mλa |ψβ (x)|2 (x − a)2 = ,
−∞ β
Z ∞
(3) 3
hV iβ = 4mλa hψβ |ξ |ψβ i = 4mλa |ψβ (x)|2 (x − a)3 = 0,
Z ∞ −∞
3mλ
hV (4) iβ = mλ hψβ |ξ 4 |ψβ i = mλ |ψβ (x)|2 (x − a)4 = . (3.4)
−∞ 16β 2
Putting all of these together, we find a β-dependent upper bound on E0 :
~2 β mλa2 3mλ
E0 ≤ Ev (β) = + + . (3.5)
2m β 16β 2

11
(b) One Gaussian: Expansion 3 VARIATIONAL TECHNIQUES

Now we seek the value of β that minimizes Ev (β). To find it, we compute the derivative
∂Ev
∂β
, set it to zero, and solve for the minimizing value of β:

∂Ev ~2 mλa2 3mλ !

= − − = 0. (3.6)
∂β β=β∗ 2m β∗2 8β∗3

This is a cubic equation for β∗ , but luckily it can be solved exactly. There are two imaginary
roots, corresponding to critical points of Ev in the unphysical region β < 0, and one real
and positive root. The answer isn’t pretty:

4m2 λa2 h 2  √ i−1/3

β∗ = 3m λ 27~ + 729~2 − 1536m2 λa6 +
~
1 m2 λ 

√ 1/3
+ 27~ + 729~2 − 1536m2 λa6 =
2~ 9

4m2 λa2 −1/3 1  α 1/3 2

 √ 
= (3α) + , 2 2 6
α ≡ m λ 27~ + 729~ − 1536m λa . (3.7)
~ 2~ 9

(b) One Gaussian: Expansion

We are interested in the regime where a is very large. Expand β∗ to leading nontrivial order
in a in the large-a limit, and use this approximate value of β∗ to find the lowest upper bound
on the ground state energy. Compare your results to perturbation theory.

Solution. It is actually possible to plug β∗ directly into Ev and obtain an exact result for
the lower bound on the energy. But its form is rather unenlightening. We’ll make things
easier on ourselves by taking a to be large, in which case the expression for α above is
dominated by the a6 term inside the square root. We therefore obtain
1/3 √
4m2 λa2 h 2 √ 1 m2 λ √
i−1/3 
6 6
ma 2λ
β∗ ≈ 3m λ −1536λa + −1536λa = . (3.8)
~ 2~ 9 ~

The last equality is the result of drastic and somewhat miraculous simplifications that save
us at the last second from what looks like certain doom at the hands of imaginary numbers.
We’re in good shape: this expression looks very similar to the frequency ω we found earlier.
In fact, this is exactly the value of β∗ we would find if the last term in (3.5), which is the
energy contribution from the ξ 4 part of the potential, wasn’t there at all. So it is reasonable
to suspect that we have landed back in harmonic oscillator territory, and the following
calculation will confirm that suspicion.
Plugging this value of β into Ev gives the least upper bound

~2 β∗ mλa2 3mλ √ 3~2

E0 ≤ Emin = Ev (β∗ ) ≈ + + = ~a 2λ + , (3.9)
2m β∗ 16β∗2 32ma2

which is exactly the first-order-corrected energy we found above!

12
(c) Two Gaussians: Minimization 3 VARIATIONAL TECHNIQUES

(c) Two Gaussians: Minimization

By now, we are desperate to see results which mimic the behavior of the true ground state.
Consider, therefore, a trial wave function with a bump in each well:
A h −β(x+a)2 −β(x−a)2
i
ψβ (x) = √ e +e . (3.10)
2
As before, normalize ψβ by finding A, and try to determine the value of β (call it β∗ ) that
gives the tightest upper bound on E0 . You will not be able to solve for β∗ analytically in
closed form. Instead, argue that it is a reasonable approximation to use the value of β∗
obtained for the one-Gaussian trial wave function from above.

Solution. Here we repeat the same steps as in the two parts above. Most of the algebra
was done in Mathematica, so we omit most of the intermediate steps here.
We begin by normalizing the wave function:
r  1/4  −1/2
π −2a2 β

! π 2
hψβ |ψβ i = |A|2
1+e = 1 =⇒ A = 1 + e−2a β =⇒
2β 2β
 1/4  −1/2 h
1 π 2 2 2
i
ψβ (x) = √ 1 + e−2a β e−β(x+a) + e−β(x−a) . (3.11)
2 2β
Next, we perform about a zillion integrals in order to compute

E0 ≤ Ev = hψβ |H|ψβ i = hT iβ + hV (2) iβ + hV (3) iβ + hV (4) iβ . (3.12)

The final result contains an energy contribution from each bump of the wave function in
response to each of the terms in the potential. All together, we have

3mλ mλa2 mλa4 ~2 β ~2 a2 β 2

Ev = + + + − +
16β 2 4β 2 2m m
3mλa2 mλa4 ~2 a2 β 2
 
tanh a2 β .


+ − + (3.13)
4β 2 m

At this point, the natural thing to do is to differentiate Ev with respect to β, set the result
equal to zero, and attempt to solve for β. But as threatened, ∂E v
∂β β=β∗
= 0 has no closed-
form solution for β∗ in this case. Womp womp. Notice, however, that for large a, the
tanh factor above rapidly approaches 1 from below. If we just replaced it by 1, then we
would obtain exactly the one-Gaussian bound (3.5) on the ground state energy. So the two
functions to be minimized differ by an exponentially small quantity. In particular, their
minimizing values β∗ are also exponentially close. But now comes the kicker: since both
the one-Gaussian and two-Gaussian functions Ev are locally flat at their respective minima,
using the one-Gaussian value of β∗ in the two-Gaussian energy function can incur at most
the square of an exponentially small error. This is because near the minimum, we have
Ev (β) ∼ Ev (β∗ ) + Ev00 (β∗ )(β − β∗ )2 , and β − β∗ is exponentially small. This means that using
the wrong value β∗ cannot harm the sensitivity of Ev to exponentially small contributions.

13
(d) Two Gaussians: Expansion 3 VARIATIONAL TECHNIQUES

(d) Two Gaussians: Expansion

Now use the approximate (i.e. large-a) value of β∗ from part (b) to find the “best” upper
bound on E0 . Compare this upper bound to the one obtained in part (b): did we do better
this time? If not, why not? If so, by how much did we beat the previous bound?

Solution.
√ The approximate value of β∗ obtained for the one-Gaussian wave function was
β∗ ≈ ma 2λ/~. If we plug this value into the exact expression for Ev , we find
√
3mλa4 5~a λ 3~2
Ev (β∗ ) ≈ − + √ + +
2 4 2 32ma2
√ ! √ √ !
3mλa4 ma3 2λ 3~a λ ma3 2λ
+ tanh + √ tanh . (3.14)
2 ~ 4 2 ~

For large a, the tanh terms above rapidly approach 1 from below. Replacing them both by 1
would yield the one-Gaussian bound (3.9) on E0 . So the two results differ by an exponentially
small quantity, and moreover the two-Gaussian energy is just slightly below the one-Gaussian
energy. Let us make this a bit more quantitative:
√ !" √ !#
(1G) (2G) 3 ~a 2λ 4 ma3 2λ
∆Ev = Ev − Ev = + mλa 1 − tanh ≈
2 4 ~
√ ! " √ #
~a 2λ 2ma3 2λ √ 3
≈3 4
+ mλa exp − ∼ 3a4 e−2 2a . (3.15)
4 ~

Here we have used the identity 1−tanh(x) = 2/(1 + e2x ) and dropped the 1 in the denomina-
tor in the large-a limit. We can see clearly that the difference between the perturbative result
(1G) (2G)
Ev and the improved variational result Ev is exponentially small in both the limits of
large a and small ~. The ~-dependence also assumes the form e−1/~ that was threatened at
the end of the perturbation theory section. Although the result has the advantage of being
morally correct, it is rather ad hoc, since we had to postulate a two-Gaussian trial wave
function in order to see the nonperturbative tunneling effects. Before we do things right
with WKB, we turn to one last—and even more ad hoc—variational trick.

(e) First Excited State

Now consider and normalize the antisymmetric two-Gaussian trial wave function:
− A h −β(x+a)2 −β(x−a)2
i
ψβ (x) = √ e −e . (3.16)
2
Intuitively, ψβ− should be an excellent proxy for the first excited state at large a. Even
though the variational principle only applies to the ground state, we can throw ψβ− at it
anyway and obtain a reasonable prediction for the energy of the first excited state. Using
the same approximate value β = β∗ found in part (c), plug ψβ−∗ into the variational principle
to obtain a “bound” on the energy of the first excited state. Compare this to your result
from part (d) for E0 , and calculate the energy splitting between the two lowest states.

14
(e) First Excited State 3 VARIATIONAL TECHNIQUES

Solution. The calculations here are almost identical to the ones in the previous part, so
we skip the algebra and report the results. We begin with normalization:
 1/4  −1/2
π 2β
ψβ− ψβ− = 1 =⇒ A = 1 − e−2a =⇒
2β
 1/4  −1/2 h
1 π 2 2 2
i
ψβ− (x) =√ 1 − e−2a β e−β(x+a) − e−β(x−a) . (3.17)
2 2β
The expectation value of H in this state is given by

3mλ mλa2 mλa4 ~2 β ~2 a2 β 2

Ev− = + + + − +
16β 2 4β 2 2m m
3mλa2 mλa4 ~2 a2 β 2
 
coth a2 β .


+ − + (3.18)
4β 2 m
√
Next, we plug in β∗ ≈ ma 2λ/~ and obtain the energy of the first excited state:
√
3mλa4 5~a λ 3~2
Ev (β∗ ) ≈ − + √ + +
2 4 2 32ma2
√ ! √ √ !
3mλa4 ma3 2λ 3~a λ ma3 2λ
+ coth + √ coth . (3.19)
2 ~ 4 2 ~

These results are identical to the ones obtained in the symmetric two-Gaussian state, except
with tanh replaced by coth everywhere. This is a good sign: tanh(x) approaches 1 from
below, while coth(x) approaches 1 from above, so the symmetric state has lower energy than
the antisymmetric state being analyzed presently. Finally, we compute the energy splitting
between the ground and first excited states:
√ ! 3
√ !
~a 2λ 2ma 2λ
∆Ev = Ev− − Ev+ = 3 mλa4 + csch ≈ (3.20)
4 ~
√ ! "
3
√ #
~a 2λ 2ma 2λ √ 3
≈ 6 mλa4 + exp − ∼ 6a4 e−2 2a . (3.21)
4 ~

Here we have used csch(x) = 2/(ex − e−x ) and dropped the e−x in the denominator in the
large-a limit. At long last, we have hacked together a “derivation” that the energy gap of
the double well potential is nonperturbatively small, both in a1 and in ~.

15
 4 THE WKB METHOD

4 The WKB Method

We conclude our tour de force of the double well with its study through the unreasonably
powerful WKB approximation. In particular, we will use the WKB method to calculate
tunneling probabilities, wave functions, and energy splittings in the double well.

(a) Tunneling at Zero Energy

Consider a particle of negligible energy E ≈ 0 stuck in the right well at x = a. Calculate the
probability T for the particle to tunnel through the central barrier to x = −a. Compare your
(0)
answer to the square amplitude of the Gaussian wave function, ψ0 , evaluated at x = −a.

Solution. The tunneling probability is expressed in terms of the extinction coefficient γ,

which is in turn calculated from the classical momentum of the particle:
1 a
Z p
−2γ
T =e , γ= dx |p(x)|, p(x) = 2m(E − V (x)). (4.1)
~ −a

We have an explicit expression for V (x), and we assume E = 0. Therefore:

p p
|p(x)| = 2m(E − V (x)) ≈ −2mV (x) =
q √
= −2m · mλ(x2 − a2 )2 = m 2λ x2 − a2 . (4.2)

The extinction coefficient can then be calculated exactly:

√ Z √
m 2λ a 4ma 3
2λ
γ= dx x2 − a2 = . (4.3)
~ −a 3~

The tunneling probability is therefore assembled:

√ !
8ma3 2λ
T = e−2γ = exp − . (4.4)
3~

As expected, T decays extremely rapidly with increasing a.

Meanwhile, the probability computed from the Gaussian is
√ !1/4 √ !
(0) 2ma 2λ ma 2λ 2
ψ0 (x) = exp − (x − a) =⇒
π~ ~
s √ 3
√ !
(0)
2 2ma 2λ 8ma 2λ
ψ0 (−a) = exp − . (4.5)
π~ ~

Evidently they have the same decay exponent! Physics works. :)

16
(b) The Quantization Condition 4 THE WKB METHOD

(b) The Quantization Condition

Now suppose that the energy E is small but nonzero. We will build up to a derivation of
the quantization condition and the energy splitting in the double well potential. To begin,
define the following related quantities:
1 x2 1 x1
Z Z Z x1
θ= dx p(x), φ= dx |p(x)|, S(x) = dx0 |p(x0 )|, (4.6)
~ x1 ~ −x1 x

where ±x1 and ±x2 are the locations of the inner and outer turning points of the double
well, respectively. Intuitively, θ and φ measure oscillation and extinction in their respective
regions. The interpretation of S(x) will remain mysterious for now.
It can be shown via the connection formulæ (but don’t bother proving this) that in the
region [0, x1 ], the WKB wave function takes the form
C± 
2 cos θeS(x)/~ + sin θe−S(x)/~ ,

ψWKB (x) = p (4.7)
|p(x)|
It turns out to be impossible to calculate θ, φ, and S(x) in closed form for the double
well potential V (x), but we can still say a few things on general grounds. Recall that the
symmetry of the potential V (x) forces ψWKB to be either symmetric (even) or antisymmetric
0
(odd). In the odd case, ψWKB (0) = 0, while in the even case, ψWKB (x) = 0. Impose these
boundary conditions to derive the quantization condition tan θ = ±2eφ , where the plus sign
is taken for even ψWKB , and the minus sign for odd ψWKB .
Hint: for the odd case, first prove that d|p|
dx x=0
= 0.

Solution. The even case is easier: setting ψWKB (0) = 0 yields

C±  !
2 cos θeS(0)/~ + sin θe−S(0)/~ = 0 =⇒

ψWKB (0) = p
|p(0)|
2 cos θe S(0)/~
= − sin θe−S(0)/~ =⇒ tan θ = −2e2S(0)/~ . (4.8)

We’re almost there! It remains to observe that since the potential is even, we have
Z x1
1 x1
Z
0 0 ~
S(0) = dx |p(x )| = dx |p(x)| = φ, (4.9)
0 2 −x1 2
where the first equality follows because p(x) does not change sign on [0, x1 ]. Therefore:

tan θ = −2e2φ~/2~ = −2eφ . (4.10)

The even case takes a bit more work, since we need to compute the derivative of the
WKB wave function. Using the chain rule and the fundamental theorem of calculus, we find

0 C±d|p(x)|  S(x)/~ −S(x)/~


ψWKB (x) = 2 cos θe + sin θe +
|p(x)|3/2 dx
    
C± S(x)/~ |p(x)| −S(x)/~ |p(x)|
+p 2 cos θe + sin θe − . (4.11)
|p(x)| ~ ~

17
(c) Preparation I: Large-a Quantization 4 THE WKB METHOD

Now we set x = 0. We recall that S(0) = ~2 φ, and set p0 = |p(0)| > 0 for shorthand. We’ll
also need to show that dp/dx vanishes at zero:

dp d hp i dV dp dV
= 2m[E − V (x)] ∝ =⇒ ∝ = 0. (4.12)
dx dx dx dx x=0 dx x=0

0
Putting these pieces together, we find that the entire first term in ψWKB (0) above vanishes:

0 C± h p 
0
 p i
0 !
ψWKB (0) = √ 2 cos θeφ + sin θe−φ − = 0 =⇒
p0 ~ ~
2 cos θeφ = sin θe−φ =⇒ tan θ = +2eφ . (4.13)

(c) Preparation I: Large-a Quantization

In the WKB limit, the two wells are deep and widely separated. Irrespective of the specific
form of V (x), this makes φ grows large. This means that tan θ = ±2eφ is exponentially large,
which occurs only when θ is close to a half-integer multiple of π. (This is why the formula
above really was a quantization condition.) Introduce a small parameter ε and write
 
1
θ = n+ π + ε. (4.14)
2

The small deviation ε should be related to how large φ is. Upon imposing tan θ = ±2eφ ,
show that ε ≈ ∓ 12 e−φ (to first order in ε), and therefore show that
 
1 1
θ = n+ π ∓ e−φ . (4.15)
2 2

Notice that this means that there are two WKB wave functions for each value of n, differing
in the region [0, x1 ] from each other by a tiny amount parametrized by ∆θ ∼ e−φ .

Solution. We begin by using an angle addition formula to isolate ε in tan θ:

sin n + 12 π + ε


(−1)n cos ε
  
1 1
tan θ = tan n + π+ε =  1

 = n+1
=− . (4.16)
2 cos n + 2 π + ε (−1) sin ε tan ε

Then, imposing the quantization condition, we have

1 1
tan θ = − = ±2eφ =⇒ tan ε ≈ ε = ∓ e−φ . (4.17)
tan ε 2
This completes the proof, since we can now plug ε directly into the quantization condition:
   
1 1 1
θ = n+ π+ε= n+ π ∓ e−φ . (4.18)
2 2 2

18
(d) Preparation II: Harmonic Approximation 4 THE WKB METHOD

(d) Preparation II: Harmonic Approximation

Find the exact locations of the four classical turning points in the double well potential V (x).
Call the inner ones ±x1 and the outer ones ±x2 . Expand x1,2 to leading order in large a, and
check that they match the locations of the turning points in the harmonic potential V (2) (x).
Then, calculate θ and φ in terms of the energy E for the harmonic potential V (2) (x).

Solution. The four turning points ±x1,2 of V (x) are defined by V (±x1,2 ) = E:
s r r !
2 E 1 E
E = mλ x2 − a2 =⇒ ±x1,2 = ± a2 ±


≈± a± . (4.19)
mλ 2a mλ

The outer turning points take the plus sign inside the square root, while the inner turning
points take the minus sign. The overall sign tells us whether we are on the left or the right.
Meanwhile, the turning points of the harmonic potential are given by
r
(2) 2 2 2 E 1 E
E = V (x) = 4mλa (x − a) =⇒ (x − a) = =⇒ x 1,2 = a ± , (4.20)
4mλa2 2a mλ
which matches the results above.
Now we attack the integration of the momentum, which is given for V (2) by
q p
p(x) = 2m[E − V (2) (x)] = 2m[E − 4mλa2 (x − a)2 ]. (4.21)

The integrals defining θ and φ are tricky but doable; see Griffiths, problem 9.17. The θ
integral can be computed exactly:
πE
θ = (math) = √ . (4.22)
2a 2λ~
The φ integral can also be done exactly, but the answer is not as pretty:
 q   r
E mλ
q
2
φ = (math) = √ z0 z02 − 1 − ln z0 + z02 − 1 , z0 = 2a . (4.23)
a 2λ~ E

In the WKB limit E  V (2) (x), this expression collapses to

√
E  2  Ez02 4ma3 2λ
φ≈ √ z0 − ln(2z0 ) ≈ √ = . (!) (4.24)
a 2λ~ a 2λ~ ~

(e) The Energy Splitting

Combine your results from parts (c) and (d) to obtain a formula for the energies En± of the
symmetric (+) and antisymmetric (−) WKB wave functions. (You do not need to show that
the WKB wave functions satisfy the Schrödinger equation and therefore have well-defined
energies. It is true but unenlightening to prove.) Calculate, at long last, the energy splitting
∆E0 = E0+ − E0− between the two lowest states in the double well potential.

19
(f) Tunneling and Imaginary Time 4 THE WKB METHOD

Solution. Let’s begin by recalling and gathering our results so far:

√
πEn± 4ma3 2λ
 
1 1 −φ
θ = n+ π∓ e , θ= √ , φ≈ . (4.25)
2 2 2a 2λ~ ~
This allows us to immediately solve for the energies:
√ √
√ 2λ~ −4ma3 √2λ/~
    
2a 2λ~ 1 1 1 a
En± = n+ π ∓ e−φ = 2a 2λ~ n + ∓ e . (4.26)
π 2 2 2 π
The energy splitting between any pair of symmetric and antisymmetric states, including the
gap between the ground and first excited state, is
√
2a 2λ~ −4ma3 √2λ/~ ~ω −mωa2 /~
∆En = e = e . (4.27)
π π

(f ) Tunneling and Imaginary Time

p
When a particle tunnels through a potential barrier, its momentum p = 2m[E − V (x)]
becomes imaginary. Classically, this is clearly nonsense... or is it? If we think of momentum
as p = mv, then an imaginary p implies that v = ẋ is imaginary: we can view this as the
velocity of motion that happens over an imaginary time interval.
Consider, now, the classical action for a point particle in an arbitrary potential V (x),
Z T Z T "  2 #
1 dx
S= dt L = dt m − V (x) . (4.28)
0 0 2 dt

Make the transformation t −→ τ = it with β = iT , and express the transformed (Euclidean)

action in terms of τ . This change of coordinates is called a Wick rotation, and τ is called
the imaginary or Euclidean time. You should find that
Z β "  2 #
1 dx
S −→ i dτ m + V (x) ≡ SE . (4.29)
0 2 dτ

Obtain the Euler-Lagrange equations that follow from S 0 . Compare them to the usual Euler-
Lagrange equations: what happened to the potential? Then explain the following cryptic
claim: “Quantum tunneling is classical motion in imaginary time.”

Solution. The transformation t −→ τ = it induces the following changes:

d d d d d
dt −→ dτ = i dt ⇐⇒ dt = −i dτ, −→ = −i ⇐⇒ =i . (4.30)
dt dτ dt dt dτ
Substituting these changes, along with t = T =⇒ iT = β into the original action yields
Z β "  2 # Z β "  2 #
1 dx 1 dx
S −→ SE = (−i dτ ) m i − V (x) = i dτ m + V (x) . (4.31)
0 2 dτ 0 2 dτ

20
(g) Instantons 4 THE WKB METHOD

As for the Euler-Lagrange equations, their derivation is completely identical to the usual
case. The factor of i in front of the action does not appear in the equations of motion, and
the only thing that changes is the sign of V . To wit:
d2 x
 
∂ ∂L ∂L dV
δSE = 0 =⇒ = =⇒ m 2 = + . (4.32)
∂τ ∂(dx/dτ ) ∂x dτ dx
The potential evidently gets flipped upside down. This is unsurprising, since the move to
imaginary time has the effect of flipping the sign of the kinetic term relative to the potential,
so the Euler-Lagrange equations reflect motion in the inverted potential.
We conclude that the action for a particle that tunnels over a potential barrier, which is
purely imaginary, can be viewed as i times the action for a particle that rolls in the upside-
down version of the same potential. And as we saw above, such an inverted potential is
obtained by putting the system in imaginary time.

(g) Instantons
Using the explicit form of the double well potential V (x), solve the Euler-Lagrange equations
to find the classical trajectory x(τ ) of the imaginary-time system, subject to the constraints
that the total energy is zero and x(0) = x. This solution is called an instanton: why? (Hint:
consider large a.) Next, evaluate SE on this trajectory by plugging the solution x(τ ) back
into SE and performing the integral dτ explicitly on the interval [−β, β].
Finally, take the limit of large a. If the real-time action S transforms to SE , what does
the quantity eiS/~ transform to? Does this remind you of any of the previous results? It
turns out that the phase factor eiS/~ is a crucial ingredient in the theory of Feynman path
integrals. This suggests that path integrals in imaginary time have a good deal to say about
the WKB approximation (indeed, at a deep level they are identical!) and about tunneling.

Solution. Our equations of motion, using overdots to denote τ -derivatives, are

dV
mẍ = = 4mλx(x2 − a2 ) =⇒ ẍ = 4λx(x2 − a2 ). (4.33)
dx
This looks hard to solve; it is better to take advantage of the hint that the total energy is
conserved and vanishes. In imaginary time, the Lagrangian is L = T − (−V ) = T + V (i.e.
it is the energy of the real-time system), so the imaginary-time Hamiltonian is
1
2
H = T + (−V ) = mẋ2 − mλ x2 − a2 = 0. (4.34)
2
This has the effect of “automatically” integrating the equation of motion partway. The
resulting equation is separable and can be solved exactly:
√
 2
1 dx
2 dx
= λ x2 − a2 =⇒ = ± 2λ x2 − a2 =⇒


2 dτ dτ
1 −1 x
  √  √ 
− tanh = ± 2λ(τ − τ0 ) =⇒ x(τ ) = ±a tanh a 2λτ . (4.35)
a a
(We have set τ0 = 0 to ensure x(0) = 0.)

21
(g) Instantons 4 THE WKB METHOD

This solution is called an instanton because at large a, tanh approaches an abrupt step
function. The particle sits, almost stationary, on one maximum (x = −a) of the inverted
potential for all t < 0. Then, right around t = 0, it quickly rolls across to the other maximum
(x = a), where it stays for all t > 0. The tunneling is localized in an instant, and is described
by the motion of a particle (an “-on” in particle parlance), hence the name instanton.
It remains to plug this into the action and evaluate it. This is most easily done in
Mathematica, which gives
 √  1  dx(τ ) 2
4 4
V (x(τ )) = mλa sech a 2λ τ = m = T. (4.36)
2 dτ

Therefore the Lagrangian is obtained and integrated:

 √ 
L = T + V = 2T = 2V = 2mλa4 sech4 a 2λ τ =⇒
Z β √  √  √
4ma3 2λ  1 2

SE = i dτ (T + V ) = tanh a 2λ β 1 + sech a 2λ β i. (4.37)
−β 3 2

Now at large a, sech(ax) approaches zero, and tanh(ax) approaches 1, both exponentially
rapidly. Therefore in the crude large-a approximation, we have
√ " √ #
4ma3 2λ t→it 4ma 3
2λ
SE ≈ =⇒ eiS/~ −→ exp − , (4.38)
3 3~

which reminds us strongly of just about every result we’ve obtained so far using WKB.
A beautiful story begins here: topology, path integrals, instantons, semiclassical approx-
imations, and even quantum gravity all converge here. But this is a story for another time.

22
 A REFERENCES

A References
• https://journals.aps.org/pr/pdf/10.1103/PhysRev.85.631

• http://hitoshi.berkeley.edu/221A-F04/asymptotic.pdf

• https://wwwphy.princeton.edu// verlinde/PHY305/wkb-path.pdf

• Griffiths (QM 3ed), chapter 9; specifically, problem 9.17.

23