221A Lecture Notes

Variational Method 1 Introduction

Most of the problems in physics cannot be solved exactly, and hence need to be dealt with approximately. There are two common methods used in quantum mechanics: the perturbation theory and the variational method. The perturbation theory is useful when there is a small dimensionless parameter in the problem, and the system is exactly solvable when the small parameter is sent to zero. The system is then studied in power series expansion in the small parameter. For instance, the quantum electrodynamics is a perfect example where the perturbation theory is useful. The small parameter is the ne-structure constant = e2 /(4 0 hc) 1/137 1. In the limit where 0, the photons and electrons are free particles and there is no interaction; hence the system is exactly solvable. Physical quantities are calculated in power series in . The variational method is useful to study the ground state, but not very useful for the study of excited states. On the other hand, it is not required that the system has a small parameter, nor that the system is exactly solvable in a certain limit. Therefore it has been useful in studying strongly correlated systems, such as the fractional Quantum Hall eect. The study of multi-electron atoms relies on the HartreeFock theory based on the variational method. Certain systems also exhibit drastic dierence between the perturbative ground state and true ground state, such as the superconductivity. The BardeenCooperSchrieer (BCS) theory of supercoductivity is based on the variational method. On the other hand, the success of the variational method depends on the initial guess what the ground-state wave function looks like, and an excellent physical intuition is required for a successful application.

Fundamentals

By denition, the ground state has the lowest energy eigenvalue of a given system, H|0 = E0 |0 , (1) 1

while any other states have higher energy eigenvalues, H|n = En |0 , En > E0 . (2) Suppose we have a Hamiltonian we cannot solve for eigenstates and eigenvalues exactly. Based on a physical intuition, we try to come up with a guess what the ground state looks like: | . Many prefer to call it an ansatz or 0 a trial wave function instead of a guess to make it sound scientic, but a guess is nonetheless a guess. If the ansatz is not exactly right, it is a linear combination of dierent Hamiltonian eigenstates,

|0 =
n=0

|n n|0 =
n=0

cn |n .

(3)

The ansatz is exact if c0 = 1 and cn = 0 for all n = 0. The proper normalization requires |c0 |2 + n=0 |cn |2 = 1. The expectation value of the energy with the ansatz is |cn |2 E0 = E0 . (4) |cn |2 En |c0 |2 E0 + 0|H|0 = |c0 |2 E0 +
n=0 n=0

Namely the expectation value must be always greater than or equal to the ground state energy. This way, one can obtain an upper limit on the groundstate energy eigenvalue by employing an arbitrary trial ket. Clearly a better upper limit is obtained by a better trial ket, namely a better overlap with the true ground state (e.g., c0 as close to unity as possible). The converse is also true: a state with a lower expectation value of the Hamiltonian should have a better overlap with the true ground state. Given this observation, one can look for a ground state wave function by introducing parameters to the trial ket | 1 , 2 , ) , calculating the energy 0( expectation value E(1 , 2 , ) = 1 , 2 , )|H| 1 , 2 , ) , 0( 0( (5) and looking for as low expectation value as possible, namely minimizing it: E E = = = 0. (6) 1 2 The success of this method still depends on having a good trial ket with the good set of parameters. This is what is called the variational method: you vary the parameters within the ansatz and nd the parameter set that minimizes the energy expectation value. If the ansatz is close enough to the true ground state, so is the minimum energy expectation value. If the ansatz has the correct functional form, it leads to the exact ground state. 2

Examples

If the functional form of the ground-state wave function is guessed correctly, then the variational method gives the true ground-state wave function. Sakurai discusses the example of the hydrogen atom, where the ground state wave function is 1 Z 3/2 Zr/a0 , 2e = 4 a0 (7)

where a0 = h2 /(me e2 ) is the Bohr radius (in Gaussian unit). The energy eigenvalue is Ze2 Z 2 e4 me E0 = = . (8) 2a0 2 2 h See Appendix A.6 of Sakurai. Suppose you didnt know this exact form, but guessed that the groundstate wave function is spherical and exponentially damping, er/d . The properly normalized wave function is then = 0 1 r/d e . d3 (9)

Now we calculate the expectation value of the Hamiltonian H= The kinetic energy term gives 2 d 1 2 1 3 h2 d2 h + er/d = 3 d = . 2 2 dr r dr d 2m d 2md2 (11) The potential energy term, on the other hand, gives 4r2 drer/d Ze2 1 = 3 (Ze2 ) r d 4rdre
2r/d

Ze2 p2 . 2m r

(10)

p2 1 2 h = 3 2m d 2m

1 d = 3 (Ze2 )4 d 2

Ze2 . (12) d

By minimizing the expectation value, E(d) = d d h2 Ze2 2md2 d 3 = 2 h2 Ze2 + 2 = 0, 2md3 d (13)

we nd d = h2 /(Zme e2 ), the exact result. Note that this is the theoretically justied version of the HW #3, problem 2. Suppose, however, you didnt guess it correctly and instead tried a Gaussian, 2 3/2 r2 /d2 e . (14) = d2 The expectation value of the kinetic energy is p2 2m = = 2 d2 2 d2
3/2

2 h 2m 2 h 2m

4r2 drer

2 /d2

d2 2 d 2 2 er /d + dr2 r dr 4r2 6 2 4 d d (15)

3/2

4r2 dre2r

2 /d2

3 2 h = . 2md2 That of the potential energy is Ze2 r 2 d2 2 Ze 2 2 . = d = Ze2

4r2 dre2r

2 /d2

1 r (16)

The minimization of the expectation value yields E 3 2 h Ze2 2 2 3 2 h Ze2 2 2 = = 3 + 2 = 0, d d 2md2 d md d and hence 3 h2 d= , 2 2 Zme e2 4 Z 2 e4 me = 0.85E0 . Emin = 3 h2

(17)

(18)

Even though the functional form is quite wrong, we still got a good answer within 15%! Indeed the expectation value is larger in the absolute value (less in magnitude) than the true ground-state energy. Actually, with some thinking, we would not choose a Gaussian. It is useful to think about the asymptotic behavior of the wave function. Because the Coulomb potential goes to zero at the innity, the bound state wave function satises (asymptotically) h2 d2 = |E| 2m dr2 4 (19)

whose solution is er , (20) where = 2m|E|/ . Therefore, we know that Gaussian was not a very h good guess, but nonetheless it did quite well. On the other hand, this type of consideration does suggest an exponential form, which turns out to be the correct ansatz, and then the variational method xes the exponent. Sakurai discusses a particle in the box problem. Instead of the exact ground-state wave function x 1 , = cos a 2a he tries = |a| |x| , (22) and nds the result Emin = 1.00298E0 , a remarkable agreement. In coming up with this guess, the main argument was to make sure the boundary condition, (a) = 0, is satised. Again, having some thoughts about the boundary conditions turn out to be useful. (21)

Excited States

In general, the application of the variational method is dicult for excited states, but it is not impossible. One such way is to identify a quantum number that distinguishes the excited state from the ground state; then one can look for the lowest energy state with that quantum number. Because of the dierent quantum numbers, the trial wave function is guaranteed to be orthogonal to the ground state. The variational method gives a rigorous upper limit on the lowest energy state with a given quantum number. For instance, one can look for the 2p state of the hydrogen atom because it is the lowest energy state with l = 1. Using a trial wave function that involves Y1m , it is guaranteed to be orthogonal to the ground-state wave function due to its distinct angular momentum quantum numbers. The true wave function is Z 3/2 Zr Zr/2a0 m e Y1 , (23) = 2a0 3a0 with the energy eigenvalue Ze2 . (24) E1 = 8a0 5

On general grounds, the radial wave functions start with the power rl h2 l(l+1) because of the centrifugal barrier 2mr2 . Therefore, we can guess = 4 3d5
1/2

rer/d Y1m .

(25)

The prefactor is determined to make it properly normalized. The kinetic energy term has the expectation value h2 d l(l + 1) p2 2 d h2 = + = , 2m 2m dr2 r dr r2 2md2 while the potential energy term gives Ze2 Ze2 = . r 2d (27) (26)

The minimization of the expectation value E(d) = d d gives d= and the energy Ze2 me . (30) 8 2 h This is the exact result as expected, because we had the correct functional form. It is interesting to see that the variational method may also give the correct result for excited states even when there is no distinct quantum number. By extending the trial function for l = 0 to include a polynomial, Emin = = N (1 + kr)er/d Y00 , where N 2 = 1 r2 dr(1 + kr)2 e2r/d = d3 (1 + 3dk + 3d2 k 2 ), 4 (31) h2 2Ze2 me (29) Ze2 h2 2md2 2d h2 Ze2 = 3 + 2 =0 md 2d (28)

(32)

we compute the expectation value of the energy. We nd E(d, k) = H = h2 (1 + dk + d2 k 2 ) Ze2 (2 + 4dk + 3d2 k 2 ) . 2md2 2d (33) By minimizing it, Mathematica nds two real solutions 1 1 + 3dk + 3d2 k 2 (d, k) = ( with energies Z 2 e4 me Z 2 e4 me E= , , (35) 2 2 h 8 2 h respectively. Indeed, the second solution is precisely that for the excited n = 2 state, Z 3/2 Zr Zr/2a0 0 = 2 e Y0 , (36) 2a0 a0 with the energy eigenvalue Ze2 . (37) E1 = 8a0 What is going on here is clear: the excited state does not quite minimize the expectation value, but it gives a local minimum for the function E(d, k). Of course, this example is a cheat, because we used the exact form. In general, the variational method can give a good approximation to the groundstate wave function, and then one can try to nd another wave function carefully chosen to be orthogonal to the (approximate) ground-state wave function. The better approximation the groundstate is, namely that it is high in the true ground-state content, whatever that is orthogonal to it is low in the ground-state content, and has a good chance of being an excited state. If the groundstate had been obtained poorly, so is the excited state, presumably even worse. There is no guarantee that this method works in general, but this example shows it is worth a shot. h2 2 2 h Ze2 me ) , 0), ( 2 , Ze2 me Ze me 2 2 h (34)

Caveat

One word of caution about the variational method is that there is no way to judge how close your result is to the true result. The only thing you can 7

do is to try out many Anstze and compare them. For example, the Laugha lins paper on the Fractional Quantum Hall Eect, R.B. Laughlin, Phys. Rev. Lett. 50, 13951398 (1983); http://link.aps.org/abstract/PRL/ v50/p1395, proposed a trial wave function that beat other wave functions that had been proposed earlier, such as Wigner crystal. (Of course there were many other reasons why his wave function was right to deserve the Nobel Prize.) Once your wave function gives a lower energy than your rivals, you won the race (for a moment, at least).

Why Not Full Variation?

One of you were smart to ask: why dont we do the full arbitrary variation without taking a specic ansatz? Unfortunately, you will be brought back to where you started, namely the time-independent Schrdinger equation. o One can set up an analog of the action, which is the expectation value of the Hamitonian, E[ , ] = H = dx H. (38) We can regard E as a functional of the wave function , and try to minimize it. Because is complex, we can regard and as independent. We can take an arbitrary variation without respect to and require that the expectation value is stationary. At the rst sight, it leads to a wrong result: 0 = E = and hence H = 0. (40) We got this wrong result because the expectation value must be varied subject to the requirement that the wave function stays properly normalized. One way to enforce this requirement is to use the Lagrange multiplier, E[ , , ] = dx H dx 1 . (41) dx H, (39)

We regard this expression as a functional of , , and a function of the Lagrange multiplier . The stationary condition with respect to obviously

requires the terms in the parentheses vanish; hence the properly normalized wave function. There is an important additional eect: the stationary condition with respect to also changes, E = dx H dx = 0. (42)

Now that we can treat arbitrary, we nd H = 0. (43)

In other words, the Lagrange multiplier, introduced as a mathematical tool to enforce the proper normalization of the wave function, ends up being the energy eigenvalue! This is the correct result. On the other hand, we are back to square one. The reason we chose the variational method was because we couldnt solve the Schrdinger equation exactly. We unfortunately didnt o gain anything. By the way, the use of Lagrange multiplier that ends up being the energy comes back prominently when we discuss the HartreeFock method of multielectron atoms in 221B.