Monte Carlo Variance Reduction

Antithetic Variate
antithetic variate technique to reduce Monte Carlo estimator
Review : Use Monte Carlo to estimate Elg(x)
X (g(x) <
E(g(x)
1 . Simulate X for B simulations w/ iid U2 unit 10, 1)
2 .
Pass thr inverse adf Ex" (n) to get X
3 Pass thru function
.

a
calculate estimated value of g(x)
4
. from Mc simulation

+
Fx (u)

&
U, S
XI < g(x,) =
9
,
(g(xz)
U2
X2 E(g(x)) +... +
=
92 = 9, +
92 g
: : B
up) XB (g(XB) =
9B
Isquare root of variance) o

standard error : sample population's standard error SE

=
n

Ba (B 2) = variability multiple samples of a

across

SECE(g(x)))
.

=
population ; estimated

"to make E(g(X)) more precise SECE(g(x))) by

,
↓ ↑ simulations
"
can become inefficient :
to half SECE(g(X)) BX4 ,

Antithetic : reduces 9 + 92 +... + g by introducing some opposite

behavior btun adjoining values of 9
"
instead of B indep replications , pair values St they have If depen-
.

dency to each other work w/ replications/pairs

(f)
dependency introduces covariate terms
each pair has comes from same dist) each pair has
If
-1 correlation/covariance
+ P correlation
Fx (u)
3 < g(x,) =
9
,

&
U, X
1 -
u, =
42
3
X2 (g(x2) =
92 E(g(x)) = 91 +

92
+
93 +
94 +... +
9

Uz -
X3 (g(Xz) =
93
-
1 -

my
=
14 X4 (g(x4) =
94

: :
VB-1
~ X
1 -

V+ =
4B B (g(X() =
9B

SECE(g(X)))Av Ug 2POgB
= log
NgH
+ + =

"remember who (p) is negative ,

so SE indeed smaller
 Control Variate
T= g(X) w/ unknown
expectation ECY] by control variate
:
assumes can approx system
-

we

*
simplified model z =
g (x) w/ known expectation ECZ] (constant)
Yand z both dependent on same random variable X

sample diff btwn Y and

f =
Y z
=
+ E[z] at same rv input X
.

-
mean of Y ECO] ECY] ECE] + ELE] ELY] <My Mo
+ Elz]
Mix - z)
= -
= = =

=var(Y-z]
variance of Y var(M + ] =
var (Mc z1] =
N
=
ELLY-z-ELY-z))2]
n

E(z]) ]
<
=
E[LY- E(Y] -

(z-
N

=
E[(Y-E[Y])2 -

2(Y- ECT])(z-E(z]) + 12 -
ECE])3]
N

=
var[i] + var[z]-zcor[Y z] ,

original Mc variance : var (My) var(i)

=

want coV(Y, z]4 that var [u + ] ↓: control variate z correlated w/Y

so
strongly
St 2 coV[i,
.
z] > var[E]

MC in Statistical Inference
-

confidence interval : statistical tests performed visually

outside of it occurs < 5% of time
-
95 % Cl covers 95 % of bootstrapped means (anything
3 (significantly different
< 0 05
p-value of anything outside of CI .
 stratified sampling
g(X) g and random variable input
-

output variable T= w/ function X

-

for MC Simulation ,
run prob density function fx (knowns <probl, that X sampled from
fx falls into Xi can be precisely calculated
of values at value most of the time
majority Ymax'thus sampling same

1 .
Split range of input random variable X into (Strata : similar groups (value groups Xi , i = 1, . .

.,
4)
and Y sampled w/in each
group (also divides ↑into corresponding strata Yi)

samples x from X: generates samples Y =

g(x) from is
4 Yi variance small

2
. free to choose # of samples we generate win each group

need more samples in 1st group be spread of y variables is much greater ; and group needs

only few samples to calculate mean

3
.
separately estimate expectation/mean value of output random variable win each group
,
then
combine to get ECY] =
My
ECY] WiECYi]
=

Since ECTi] unknown for any i, ECT] =

My = Wil
,

var [My] = wi var(ii]

ni

A A then 1 4 W2 ↓
↑ , ,

Ymax- g(x) fx
(X

-X
XI Xz fy
w
,
= 0 .
3 W2 = 0
.
7

max
 Markov Chains
ex1 :
Assume there's a restaurant that serves 3 types of food : 1) hamburger ,
2) pizza 3) hot dog.
,

on any given day, they serve only litem and it depends on what they served yesterday .

Laka predict what will serve tomorrow if what they're serving today.
we can
they we know

0 2
transition ~
P(X4 hotdog) Xz pizza) 0 7
.

= = =
.

diagram :
Hamburger
7 r
0 3 0 5
after Steps P(ham) =
0 3519
.
.

a .

,
6 0 2
L
0 . .

J P(p) = 0 .
2124
Pizza > Hotdog 30 .
5
P(HO) =
0 4356
.

0 .
7

random walk of 10 days

Ham-p-Ham-Hot >
- Ham -Hot-Hot >
-
Hot - >
Ham -p

i. e
. prob corresponding to each food item (prob dist of states)
# of occurrences PCHaM) = P(P) =
To P(H) =
# of days

Markov chain sequence of events where future probs ONLY depend on present
-

1) current state , future state (future state only depends on current state
P(Xn + 1 =
x(Xn = x)
2) sum of weights of outgoing arrows from any state =
transition : each arrow from one state to another

stationary dist/equilibrium state

:
prob dist does not change w/time for

Markov chain ; output cow rector =

input row rector ITA =
I

It is left eigenvector w/ eigenvalue

-

=
1 Av a =

elements of TT =
1 (i . e. 21 +
122] + +[3] =
1)
I w/ eigenvalue =
>
stationar state if there exists / eigenvectors
-

>

represent w/ transition matrix where elements are edge weight connecting 2

corresponding vertices crow--column)

goal find prob of each state
:

row rector it :
elements represent states' probs ; prob dist of states
 ham Piz not

ham 0 2 .
0. 6 0 2 .

A= 0 3 g 0 7
Pit .
.

not 0 5 .
8 0 S .

If on Pizzaday To = [0 1 0]
1 .
,

ToA =
[0 1 0] 0 2
.
0. 6 0 2 .
= [0 . 300 .
7]
0 3 7 future probs
.
g 0 .

corresponding
0 5 .
8 0 S . to pizza state

2 . H, A =
20 .
3 0 0 .
73 0 2 .
0. 6 0 2 .
=
[0 410 18
. .
0 .
41]
0 3 .
g 0 .
7

0 5 .
8 0 S .

3 .
T2A =
[0 41 .
0 . 18 0 .
41] 0 2 .
0. 6 0 2 .
20 340
=

. .
250 .
41]
0 3 .
g 0 .
7

0 5 8 0 S
· .
.

=
[]
exc : 0 .

75G Bull "Bear 20. S

,
=
PSo start at initial state then
,
-
Sz =
PS, =
Paso transition
0 .
4 multiply by n

Sn =
"So matrices pr
Freediagram :
0 . 75 Bull transition matrix : column-row
,

0 . 75 Bu A B
;

&
Bear
Bull 255

(
0

=A
.

P
04 Bull
>
0 .
259 Bear
0 .
63 Bear