Numerical Path Integrals: Austin Liu

Numerical path integrals
Austin Liu
Bernardi Group
Department of Applied Physics and Materials Science
California Institute of Technology
[email protected]
July 18, 2016

Overview
1 Review of Quantum Mechanics

Propagators using Path Integrals
2 Numerical Integration
Importance Sampling
Metropolis Algorithm
Review of Quantum Mechanics
Notation and background
For simplicity, first consider a time-independent Hamiltonian.

d
i~ |i = H |i
dt
|(t)i = e i H(tt0 )/~ |(t0 )i
K (x, t; x0 , t0 ) = hx|U(t, t0 )|x0 i
(x, t) = hx|(t)i = hx|U(t, t0 )|(t0 )i
Z
= dx0 hx|U(t, t0 )|x0 i hx0 |(t0 )i
Propagator is the kernel of the integral transform.

Time-sliced path integral
Following the approach by Zee1 - consider the propagator:
hxN , tN |x0 , t0 i
= hxN | e iH(tN t0 )/~ |x0 i
Z
= dx1 dx2 . . . dxN1 hxN |e iHt/~ |xN1 i hxN1 |e iHt/~ |xN2 i
. . . hx1 |e iHt/~ |x0 i
1
A. Zee QFT in a nutshell
Each transition amplitude is:

i p2 i
hxj+1 |e iHt/~ |xj i = hxj+1 |e ~ 2m t e ~ V (xj )t |xj i
Z
i i p2
= e ~ V (xj )t dp hxj+1 |e ~ 2m t |pi hp|xj i
Z
i i p2
= e ~ V (xj )t dp e ~ 2m t hxj+1 |pi hp|xj i
e ipx/~
The convention here is hx|pi =
2~
,
Z
iHt/~ ~i V (xj )t dp i p2 t+ ip(xj+1 xj )
hxj+1 |e |xj i = e e ~ 2m ~
2~
i r 2
e ~ V (xj )t 2m~ 2m~(xj+1 x j)
= e 4it~ 2
2~ it
x x 2
j+1 j
r i PN1
m ~ j=0 t t
V (xj )
= e
2~it
2 +bx b2 p
dx e ax = e 4a a . Hence:
R
Note
hxN |e iH(tN t0 )/~ |x0 i

2
m N/2 N1 YZ i PN1 m xj+1 xj
V (xj )
~ j=0 t 2 t
= dxj e
2~it
j=1
hxN , tN |x0 , t0 i = hxN |e iH(tN t0 )/~ |x0 i

2
m N/2 N1 YZ i PN1 m xj+1 xj
V (xj )
~ j=0 t 2 t
= dxj e
2~it
j=1
Z xN
i
= D[x(t)] exp S[x(t)]
x0 ~
where
Z tN
S[x(t)] = dt L(x, x)
t0
1
L(x, x) = mx 2 V (x)
2
Imaginary time
Let t i :
2
x 1 x
LE x, = m V (x)
2
so:
Z xN
1
hxN , N |x0 , 0 i = D[x( )] exp S[x( )]
x0 ~
where
Z N
S[x( )] = d LE (x, x)
0
Numerical Integration
Importance Sampling
Importance Sampling
hxN , tN |x0 , t0 i
2
m N/2 N1 YZ ~1 N1
P
m xj+1 xj
+V (xj )
j=0 2
= dxj e
2~
j=1
We want to sample more points in the regions which contribute

more to the integral.
Importance Sampling
Implementation in 1-D
In the 1-D case2 , want to replace original integral

Z b Z b Z 1
f (x(y ))
I = dx f (x) = dx p(x) = dy J(y )f (x(y ))
a a p(x) 0
dx
where J(y (x)) = dy = 1/p(x) is the Jacobian of the
transformation. p(x) is interpreted as a probability density which
we are choosing points from.
1
G. P. Lepage VEGAS Documentation.
http://pythonhosted.org/vegas/background.html
Importance Sampling
Now estimate the integral:

1 X
I S (1) = J(yi )f (x(yi )) = E(f (x(y ))J(y )) = E(f (x)p(x))
M
i
where we select M points yi from U(0, 1).

Importance Sampling
Since:
Z 1
1
I2 = 2 2
dy J (y )f (x(y )) I 2
= Var(f (x(y ))J(y ))
M 0
Z b
1 2 2
= dx J(y (x))f (x) I
M a
where we select M points yi from U(0, 1). So the variance

vanishes if we can choose:
Rb
f (x)dx
J(y (x)) = a
f (x)
Importance Sampling
The grid gives the transformation

Desired grid function:
x(y = i/N) = xi
x0 = a
for i = 0, 1, . . . , N with Jacobian:
x1 = x0 + x0
.. dx xi
. J(y ) = Ji = = Nxi
dy 1/N
xN = xN1 + xN1 = b
for i/N < y < (i + 1)/N
Importance Sampling
The variance of the estimate is:
Z b
2 1 2 2
I = J(y (x))f (x)dx I
M a
Z xi+1 !
1 X
= Ji f 2 (x)dx I 2
M xi i
so as
X Z xi+1 X Z xi+1
2
Ji f (x)dx = Nxi f 2 (x)dx I 2
i xi i xi
the variance vanishes. Therefore adjust the grid such that xi is

small when |f (x)| is large.
Importance Sampling
The VEGAS routine assumes that a separable Jacobian can be

used in the multidimensional case (to save memory, so the number
of points to store goes as Nd and not N d ).
Suppose we want a mapping from x 1 , x 2 to y 1 , y 2 , then we can
approximate:

1 2
(x1 , x2 ) dx1 dx2
J(y , y ) =
=
(y1 , y2 ) dy1 dy2
Similar conditions on the grid in each direction can be obtained.

Importance Sampling
Example in 2-D
Sum of two Gaussians that are diagonally offset:
Importance Sampling
Example in 2-D
Final grid:
Importance Sampling
Stratified sampling
The wasted evaluations along the second diagonal may be reduced

by stratified sampling.
Divide integration region into hypercubes
Perform MC integral in each region
Obtain variances and define new grid, keeping number of
hypercubes constant, repeat until contributions to 2 from
each are close enough.
Importance Sampling
Evaluation of the propagator
We can evaluate the integral for the propagator:

2
N1 Z i PN1 m x j+1 xj
m N/2 Y ~ j=0 t 2 t
V (xj )
dxj e
2~it
j=1
Importance Sampling
Evaluation of the propagator
With V (xj ) = 12 m 2 xj2 :

Want to compute averages of the form:

R
f (x)p(x)dx
hf i = R
p(x)dx
For example, the quantity:
Dx x(2 )x(1 )e S[x]

R
hx(2 )x(1 )i = R
Dx e S[x]
(set ~ = 1 for notational ease)

note that hx|e H |xi = Dx e S[x] , so rewrite the numerator

R
(sum over the repeated index for the energy eigenstates,

considering paths where xf = xi = x) as:
Z
dx hx|e H(f 2 ) xe H(2 1 ) xe H(2 i ) |xi
Z
= dx hx|En i hEn |e H(f 2 ) xe H(2 1 ) xe H(2 i ) |En i hEn |xi
Z
= dx hx|En i hEn |xe H(2 1 ) x|En i hEn |xi e En (f i )
= hxe H(2 1 ) x|En i e En (f i )

The denominator becomes

Z Z
H(f i )
dx hx|e |xi = dx hx|En i hEn |e H(f i ) En i hEn |xi
= e En (f i )
so we have (with T = f i , = 2 1 )
hxe H x|En i e En T
P
hx(2 )x(1 )i G ( ) = n P En T
ne
hE0 |xe (HE0 ) x|E0 i
| hE0 |x|E1 i |2 e (HE0 )
q
at large times, note that x = ~
2m (a + a)
Finally obtain
log(G ( )/G ( + )) (E1 E0 )

Algorithm
(i) (i)
First generate Ncf paths x (i) = {x0 , xN1 }. Beginning with an
arbitrary path, randomize xj at the j th site using the
Metropolis-Hastings algorithm as follows:
Generate random number in uniformly in the range (, ).
Compute S, the change in the action caused by replacing
xj xj + .
If S < 0, accept the move.
If S > 0, accept the move with probability e S
Algorithm
The overall algorithm is:

Initialize all points xji to 0, for i {0, , Ncf 1},
j {0, , N 1}.
Update the path 5Ncor times.
Update the path Ncor times, noting the acceptance ratio, once
this is satisfactory, run the simulation and compute the
desired function for a path G (t).
Average over Ncf to get hG (t)i.
The relevant measurement for us is:

1 X
G (t) = hx(tj + t)x(tj )i
N
j
1 X
= hx(j+n) mod N xj i
N
j
Want to calculate the variance of the quantity hG (t)i.

Estimate the variance by generating bootstrap copies of G .
Given {G (i) , i = 1, Ncf } Monte Carlo measurements,
randomly sample Ncf times to get a copy of G .
Obtain new estimate of E = (E1 E0 ).
Multiple such estimates allow us to estimate the variance of
E .
t Delta E(t) Error

------------
0 1.02946 0.0462194
0.5 1.01308 0.0655298
1 0.933914 0.119393
1.5 0.92962 0.183209
2 0.993295 0.273887
Future directions
Compare convergence of multidimensional integration routine with
X
G (x, t; 0, 0) = m (x)m (0)e iEm t
m
for simple cases.

Look at convergence of:
X m (x)m (x 0 ) 0 (x)m0 (x 0 )
(x, x 0 , ) = m
0
(E m Em0 ) i
m,m
References
G. P. Lepage (2005)
Lattice QCD for Novices. arXiv:hep-lat/0506036
G. P. Lepage (1978)
A new algorithm for adaptive multidimensional integration
Journal of Computational Physics 27(2), 192 203
G. P. Lepage (2014)
vegas 3.0 Documentation http://pythonhosted.org/vegas/
R. Shankar (1994)
Principles of Quantum Mechanics
A. Zee (2003)
Quantum Field Theory in a Nutshell
H. Gould, J. Tobochnik, W. Christian (2011)

An Introduction to Computer Simulation Methods, 367 401.
The End

Numerical Path Integrals: Austin Liu

Uploaded by

Copyright:

Available Formats

Numerical Path Integrals: Austin Liu

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Numerical Path Integrals: Austin Liu

Uploaded by

Copyright:

Available Formats

Numerical path integrals

Numerical path integrals

July 18, 2016

1 Review of Quantum Mechanics

Notation and background

For simplicity, first consider a time-independent Hamiltonian.

Propagator is the kernel of the integral transform.

Time-sliced path integral

Following the approach by Zee1 - consider the propagator:

. . . hx1 |e iHt/~ |x0 i

Time-sliced path integral

Each transition amplitude is:

Time-sliced path integral

hxN |e iH(tN t0 )/~ |x0 i

Time-sliced path integral

hxN , tN |x0 , t0 i = hxN |e iH(tN t0 )/~ |x0 i

We want to sample more points in the regions which contribute

In the 1-D case2 , want to replace original integral

Now estimate the integral:

where we select M points yi from U(0, 1).

where we select M points yi from U(0, 1). So the variance

The grid gives the transformation

the variance vanishes. Therefore adjust the grid such that xi is

The VEGAS routine assumes that a separable Jacobian can be

Similar conditions on the grid in each direction can be obtained.

The wasted evaluations along the second diagonal may be reduced

Evaluation of the propagator

We can evaluate the integral for the propagator:

Evaluation of the propagator

With V (xj ) = 12 m 2 xj2 :

Want to compute averages of the form:

For example, the quantity:

Dx x(2 )x(1 )e S[x]

(set ~ = 1 for notational ease)

note that hx|e H |xi = Dx e S[x] , so rewrite the numerator

(sum over the repeated index for the energy eigenstates,

= hxe H(2 1 ) x|En i e En (f i )

The denominator becomes

log(G ( )/G ( + )) (E1 E0 )

The overall algorithm is:

The relevant measurement for us is:

Want to calculate the variance of the quantity hG (t)i.

t Delta E(t) Error

Compare convergence of multidimensional integration routine with

for simple cases.

H. Gould, J. Tobochnik, W. Christian (2011)

You might also like