THE MYSTERY OF STOCHASTIC MECHANICS

Edward Nelson
Department of Mathematics
Princeton University
1
Classical Hamilton-Jacobi theory
N particles of various masses on a Euclidean
space.
Incorporate the masses in the at Riemannian
metric m
ij
, the mass tensor. Then if v
i
is a veloc-
ity, v
i
= m
ij
v
j
is a momentum.
Kinetic energy:
1
2
v
i
v
i
.
Potential energy: V .
Lagrangian: L =
1
2
v
i
v
i
V .
Position at time t of the conguration: (t).
Initial time: t.
Final time: t
1
.
Hamiltons principal function:
S(x, t) =
_
t
1
t
L
_
(s)
_
ds.
Substantial derivative (derivative along trajec-
tories):
D =

t
+v
i

i
.
Then DS = L.
2
Vector eld: v with
d
dt
= v.
Principle of least action in Hamilton-Jacobi
theory: v is a critical point for S, for unconstrained
variations.
That is, let v

be another vector eld, let

v = v

v, and denote by a prime quantities with

v replaced by v

. Then
D(S

S) = D

DS + (D D

)S

= L

L v
i

i
S

= L

L v
i

i
S + o(v).
Now
L

L = v
i
v
i
+ o(v),
S

S =
_
t
1
t
(v
i

i
S)v
i
ds + o(v).
Since this is true for all variations, we have the
Hamilton-Jacobi condition:
v
i
=
i
S.
3
Together with DS = L; i.e.,
s
t
+v
i

i
S =
1
2
v
i
v
i
V,
this gives the Hamilton-Jacobi equation
S
t
+
1
2

i
S +V = 0.
If we take the gradient we obtain Newtons
equation F = ma.
4
Stochastic Hamilton-Jacobi theory
Following Guerra and Morato, construct a
conservative Markovian dynamics for a Markov
process .
Wiener process (Brownian motion) w with
Edw
i
(t)dw
i
(t) = hdt.
Kinematics:
d(t) = b
_
(t), t
_
dt +dw(t).
Dynamics: E
_
Ldt = 0 (heuristically).
The trajectories are non-dierentiable, so what
is the meaning of
1
2
d
i
dt
d
i
dt
in the Lagrangian L?
5
Let dt > 0 and d(t) = (t + dt) (t), so
d
dt
is a quotient, not a derivative.
Compute E
1
2
d
i
dt
d
i
dt
up to o(1).
d
i
=
_
t+dt
t
b
i
_
(r), r
_
dr +dw
i
.
Note: dw is of order dt
1/2
.
d
i
=
_
t+dt
t
b
i
_
(t) +
_
r
t
b
_
(s), s)ds +w(r) w(t), r
_
dr
+dw
i
= b
i
dt +
k
b
i
W
k
+dw
i
+ O(dt
2
)
where
W
k
=
_
t+dt
t
[w
k
(r) w
k
(t)] dr.
Therefore
6
1
2
d
i
d
i
=
1
2
b
i
b
i
dt +b
i
dw
i
dt +
k
b
i
W
k
dw
i
+
1
2
dw
i
dw
i
+ o(dt
2
).
Now
EW
k
dw
i
= h
k
i
_
t+dt
t
(r t)dt =
h
2

k
i
,
so
k
b
i
W
k
=
h
2

i
b
i
.
The term b
i
dw
i
dt is singular, of order dt
3/2
,
but its expectation is 0.
Finally, Edw
i
dw
i
=
h
2
ndt. Hence we have the
sought-for result:
E
1
2
d
i
dt
d
i
dt
=
1
2
b
i
b
i
+
h
2

i
b
i
+
h
2
n
dt
+ o(1).
7
The singular term
h
2
n
dt
is a constant, not de-
pending on the trajectory, and it drops out of the
variation.
Let
L
+
=
1
2
b
i
b
i
+
h
2

i
b
i
V
and
I = E
_
t
1
t
L
+
_
(s)
_
ds.
Let b be a vector eld, b

= b + b, and as
before denote by primes quantities with b replaced
by b

. Let

satisfy
d

(t) = b

(t), t
_
dt +dw(t)
with

(t) = (t) for the initial time t.

Denition: is critical for L in case
I

I = o(b).
8
Stochastic Hamiltons principal function:
S(x, t) = E
x,t
_
t
1
t
L
+
_
(s), s
_
ds.
Then
DS = L
+
where now D is the mean forward derivative:
Df = lim
dt0+
E
t
_
df(t)
dt
_
.
As before,
D(S

S) = D

DS + (D D

)S

= L

+
L
+
b
i

i
S

= L

+
L
+
b
i

i
S + o(b).
Now
L

+
L
+
= b
i
b
i
+
h
2

i
b
i
+ o(b)
and
()
I

I = E
_
t
1
t
_
b
i

i
S +
h
2

i
_
b
i
ds + o(b).
9
Digression on Markovian kinematics:
Markov process: given the present, past and
future are independent.
The two directions of time are on an equal
footing.
In addition to the forward drift b there is the
backward drift b

.
v =
b +b

2
current velocity,
u =
b b

2
osmotic velocity.
Let be the probability density of . Then
v
t
= (v) current equation,
u
t
=
h
2

osmotic equation.
End of digression.
10
The equation (*), namely
I

I = E
_
t
1
t
_
b
i

i
S +
h
2

i
_
b
i
ds + o(b),
is awkward because it involves
i
b
i
. Integrate by
parts:
E
h
2

i
b
i
_
(s), s
_
=
_
R
n
h
2
_

i
b
i
_
dx
=
_
R
n
b
i
u
i
dx.
Since b u = v,
I

I = E
_
t
1
t
(v
i

i
S)b
i
ds + o(b).
This is true for all variations, so is critical
for L if and only if the stochastic Hamilton-Jacobi
condition holds:
v
i
=
i
S.
11
Let
R =
h
2
log ,
so

i
R = u
i
.
If we write out DS = L
+
and express every-
thing in terms of R and S, we obtain the stochastic
Hamilton-Jacobi equation:
(1)
S
t
+
1
2

i
S
i
S +V
1
2

i
R
i
R
h
2
R = 0.
Expressing the current equation
v
t
= (v)
in terms of R and S we obtain:
(2)
R
t
+
i
R
i
S +
h
2
S = 0.
These two equations are a system of coupled
nonlinear partial dierential equations expressing
necessary and sucient conditions for a Markov
process to be critical.
How can we solve them?
12
Let
= e
1
h
(R+iS)
.
Then the system is equivalent to the Schrodinger
equation

t
=
i
h
_

h
2
+V
_
.
13
magnetic elds
Riemannian manifolds
spin
Bose-Einstein and Fermi-Dirac statistics
existence of nite-energy Markov processes
momentum
interference
14
Non-locality
Consider two particles on R
1
with initial wave
function
(0) =
1

2
e

1
4

1
(0)xx
where

1
(0) =
_
1 1
1 2
_
and
x =
_
x
1
x
2
_
.
The x
1
and x
2
axes are unrelated; the par-
ticles may be separated by an arbitrarily large
amount a.
At time 0, turn on a linear restoring force
(harmonic oscillator with circular frequency ) for
the second particle. Then the particles are dynami-
cally uncoupled.
15
Since the particles are very widely separated
and dynamically uncoupled, we should expect that
E
1
(t)
1
(0)
does not depend on .
In fact it does not to fourth order in t, but
nevertheless the trajectory of the rst particle is
immediately aected by the choice of in a far
distant place.
For me this is unphysical, especially since the
eect does not depend on the separation a.
16
A wrong prediction
Consider two harmonic oscillators, about two
widely separated points a
1
and a
2
, with circular
frequency 1, and let X
i
be the Heisenberg position
operator, in units of distance from a
i
, with Heisen-
berg momentum operator P
i
.
Then
X
i
(t) = cos t X
i
(0) + sin t P
i
(0)
P
i
(t) = sin t X
i
(0) + cos t P
i
(0)
for i = 1, 2.
Let the Heisenberg state vector
0
be a real
Gaussian centered at (0, 0), and write A = (
0
, A
0
).
Then X
i
(t) = P
i
(t) = 0 since this is true for
t = 0. The operators X
1
(t) and X
2
(s) commute.
Choose
0
so that the correlation X
1
(0)X
2
(0)
is 90%. Thus the oscillators are entangled but dy-
namically uncoupled. The quantum mechanical
correlation function X
1
(t)X
2
(0) is periodic of pe-
riod 2 since X
1
(t) is.
Hence X
1
(t)X
2
(0) = .9 whenever t is a mul-
tiple of 2.
17
But let (
1
,
2
) be the corresponding Markov
process of stochastic mechanics.
This is a diusion process and it eventually
loses all memory of where it started.
Thus
lim
n
E
_

1
(2n)
2
(0)
_
= 0
whereas
X
1
(2n)X
2
(0) = .9.
Here we have an empirical dierence between
the predictions of quantum mechanics and sto-
chastic mechanics. Measurements of the position
of the rst particle at time t and of the second par-
ticle at time 0 do not interfere with each other, and
the two theories predict totally dierent statistics.
Does anyone doubt that quantum mechanics is
right and stochastic mechanics is wrong?
18