P. 1
markov-chains

# markov-chains

|Views: 2,308|Likes:

### Availability:

See more
See less

08/17/2013

pdf

text

original

4.3 Electrical networks

151

An electrical network has a countable set I of nodes, each node i having a
capacity 1ri > O. Some nodes are joined by wires, the wire between i and
j having conductivity aij = aji 2:: o. Where no wire joins i to j we take
aij = O. In practice, each 'wire' contains a resistor, which determines the
conductivity as the reciprocal of its resistance. Each node i holds a certain
charge Xi, which determines its potential ¢i by

A current or flow of charge is any matrix (T'ij : i, j E I) with T'ij = -T'ji.
Physically it is found that the current T'ij from i to j obeys Ohm's law:

Thus charge flows from nodes of high potential to nodes of low potential.
The first problem in electrical networks is to determine equilibrium flows
and potentials, subject to given external conditions. The nodes are parti-
tioned into two sets D and aD. External connections are made at the nodes
in aD and possibly at some of the nodes in D. These have the effect that
each node i E aD is held at a given potential Ii, and that a given current 9i
enters the network at each node i ED. The case where gi = 0 corresponds
to a node with no external connection. In equilibrium, current may also
enter or leave the network through aD , but here it is not the current but
the potential which is determined externally.

Given a flow (T'ij : i, j E I) we shall write T'i for the total flow from i to

the network:

'Yi = 2:'Yij'

JEI

In equilibrium the charge at each node is constant, so

T'i = 9i for i E D.

Therefore, by Ohm's law, any equilibrium potential ¢ = (¢i : i E I) must
satisfy

{

LjEIaij(¢i - ¢j) = 9i,

¢i = fi'

for i E D

(4.8)

There is a simple correspondence between electrical networks and reversible
Markov chains in continuous-time, given by

for i =I j.

152

4. Further theory

We shall assume that the total conductivity at each node is finite:

ai = Laij < 00.

j#i

Then ai = 7riqi = -7riqii. The capacities 7ri are the components of an
invariant measure, and the symmetry of aij corresponds to the detailed
balance equations. The equations for an equilibrium potential may now be
written in a form familiar from the preceding section:

{

-Q¢ = c in D

(4.9)

(4.10)

where Ci = 9i/7ri. It is natural that c appears here and not 9, because ct
and f have the same physical dimensions. We know that these equations
may fail to have a unique solution, indicating the interesting possibility
that there may be more than one equilibrium potential. However, to keep
matters simple here, we shall assume that I is finite, that the network is
connected, and that aD is non-empty. This is enough to ensure uniqueness
of potentials. Then, by Theorem 4.2.4, the equilibrium potential is given
by

¢i = Ei (iTc(Xt)dt+f(XT))

where T is the hitting time of aD.

In fact, the case where aD is empty may be dealt with as follows: we

must have

L9i=O

iEI

or there is no possibility of equilibrium; pick one node k, set aD = {k}, and
replace the condition ~ k = 9k by ¢i = O. The new problem is equivalent to
the old, but now aD is non-empty.

A

B

c

2

1

2

1

2

1

2

D

E

F

4.3 Electrical networks

153

Example 4.3.1

Determine the equilibrium current in the network shown on the preceding
page when unit current enters at A and leaves at F. The conductivities are
shown on the diagram. Let us set cPA = 1 and cPF = O. This will result in
some flow from A to F, which we can scale to get a unit flow. By symmetry,

cPE = 1-cPB and cPD = 1-cPe. Then, by Ohm's law, since the total current
leaving Band C must vanish

(cPB - cPA) + (cPB - cPE) + 2(cPB - cPe) = 0,

2(cPe - cPF) + 2(cPe - cPB) = O.

Hence, cPB = 1/2 and cPe = 1/4, and the associated flow is given by ~ A B =
1/2, ~ B e = 1/2, ~ e F = 1/2, ~ B E = O. In fact, we were lucky - no scaling
was necessary.

Note that the node capacities do not affect the problem we considered.
Let us arbitrarily assign to each node a capacity 1. Then there is an asso-
ciated Markov chain and, according to (4.10), the equilibrium potential is
given by

cPi = Ei(lxT=A) = JP>i(XT = A)

where T is the hitting time of {A, F}. Different node capacities result
in different Markov chains, but the same jump chain and hence the same
hitting probabilities.
Here is a general result expressing equilibrium potentials, flows and
charges in terms of the associated Markov chain.

Theorem 4.3.2. Consider a finite network with external connections at
two nodes A and B, and the associated Markov chain ( X t ) t ~ o .
(a) The unique equilibrium potential cP with cPA = 1 and cPB = 0 is given
by

where TA and TB are the hitting times ofA and B.
(b) The unique equilibrium flow ~ with ~ A = 1 and ~ B = -1 is given by

where rij is the number of times that ( X t ) t ~ O jumps from i to j before
hitting B.
(c) The charge X associated with ~ , subject to XB = 0, is given by

154

4. Further theory

Proof. The formula for ¢ is a special case of (4.10), where c = 0 and f =
l{A}. We shall prove (b) and (c) together. Observe that if Xo = A then

if i = A

if i ~ {A,B}

if i = B

so if ~ i j = EA(rij - rji) then ~ is a unit flow from A to B. We have

00
roo = '""'1{,-," _0,-,"

0

N}

'1,) L-J ~ n - ' I , , ~ n + l = ) , n < B

n=O

where NB is the hitting time of B for the jump chain ( Y n ) n ~ O . So, by the
Markov property of the jump chain

00

lEA(rij) = LIPA(Yn = i, Yn+! = j,n < NB)
n=O

00

= LIPA(Yn = i,n < NB)1rij.
n=O

Set

fTB

Xi = lEA 10 l{Xt=i}dt

and consider the associated potential'l/Ji = Xi/1ri. Then

00

Xiqij = XiQi1rij = LIPA(Yn = i,n < NB)1rij = lEA(rij)
n=O

so

('l/Ji -'l/Jj)aij = Xiqij - Xjqij = ~ i j ·

Hence'l/J = ¢, ~ is the equilibrium unit flow and X the associated charge, as
required. D

The interpretation of potential theory in terms of electrical networks
makes it natural to consider notions of energy. We define for a potential
¢ = (¢i : i E I) and a flow ~ = ( ~ i j : i, j E I)

E( i,jEI

I(r) = l L ,fj aijl.

i,jEI

4.3 Electrical networks

155

The 1/2 means that each wire is counted once. When ¢ and ~ are related
by Ohm's law we have

E(¢) = l L (¢i - ¢j)rij = I(r)

i,jEI

and E(¢) is found physically to give the rate ofdissipation ofenergy, as heat,
by the network. Moreover, we shall see that certain equilibrium potentials
and flows determined by Ohm's law minimize these energy functions. This
characteristic of energy minimization can indeed replace Ohm's law as the
fundamental physical principle.

Theorem 4.3.3. The equilibrium potential and flow may be determined
as follows.

(a) The equilibrium potential ¢ = (¢i : i E I) with boundary values
¢i = Ii for i E 8D and no current sources in D is the unique solution
to

minimize E (¢)

subject to ¢i = fi for i E aD.

(b) The equilibrium flow r = (rij : i, j E I) with current sources ri = 9i
for i E D and boundary potential zero is the unique solution to

minimize I (~ )

subject to ~ i = 9i for i E D.

Proof. For any potential ¢ = (¢i : i E I) and any flow ~ = ( ~ i j : i,j E I)
we have

L (¢i - ¢j)rij = 2L ¢i'Yi.

i,jEI

iEI

(a) Denote by ¢ = (¢i : i E I) and by ~ = ( ~ i j : i,j E I) the equilibrium
potential and flow. We have ~ i = 0 for i E D. We can write any potential
in the minimization problem in the form ¢ +c, where c = (ci : i E I) with

Ci = 0 for i E aD. Then

L (Ei - Ej)(¢i - ¢j)aij = L (Ei - Ej)rij = 2LEi'Yi = 0
i,jEI

i,jEI

iEI

so

E(¢ +c) = E(¢) +E(c) ~ E(¢)

with equality only if c = o.

156

4. Further theory

(b) Denote by cP = (cPi : i E I) and by ~ = ( ~ i j : i, j E I) the equilibrium
potential and flow. We have cPi = 0 for i E aD. We can write any flow in
the minimization problem in the form ~ +8, where 8 = (8ij : i,j E I) is a
flow with 8i = 0 for i E D. Then

L 'Yij8ijai/ = L (¢i - ¢j)8ij = 2L ¢i8i = 0,

i,jEI

i,jEI

iEI

so

I ( ~ +8) = I ( ~ ) +1(8) ~ 1(8)

with equality only if 8 = o. D

The following reformulation ofpart (a) ofthe preceding result states that
harmonic functions minimize energy.

Corollary 4.3.4. Suppose that ¢ = (¢i : i E I) satisfies

{

Q¢ = 0 in D

Then cP is the unique solution to

minimize E (¢)

subject to ¢ = f in aD.

An important feature of electrical networks is that networks with a small
number of external connections look like networks with a small number
of nodes altogether. In fact, given any network, there is always another
network of wires joining the externally connected nodes alone, equivalent
in its response to external flows and potentials.
Let J ~ I. We say that a = (aij : i,j E J) is an effective conductivity on
J if, for all potentials I = (Ii :i E J), the external currents into J when J
is held at potential I are the same for (J, a) as for (I, a). We know that I
determines an equilibrium potential ¢ = (cPi : i E I) by

{

L.jEI(¢i - ¢j)aij : 0 for ~ ~ J

cPi - Ii for 'I, E J.

Then ais an effective conductivity if, for all I, for i E J we have
L(¢i - ¢j)aij = :E(Ji - Ji) aij'

JEI

jEJ

For a conductivity matrix a on J, for a potential I = (Ii : i E J) and a flow
8 = (8ij : i, j E J) we set

-

1"

2

E(/) = 2 L.J (Ii - Ij) aij,

i,jEJ

1(8) = ~ L 8fjui?·

i,jEJ

4.3 Electrical networks

Theorem 4.3.5. There is a unique effective conductivityagiven by

aij = aij + L aik

k ~ J

where for each j E J, q) = ( ¢ ~ :i E I) is the potential defined by

157

{

~ k E I ( < p 1 - < p { ) a i ~ = 0

¢ ~ = 8ij

for i tf- J

for i E J.

(4.11)

Moreover, ais characterized by the Dirichlet variational principle

E(/) = inf E(¢),

and also by the Thompson variational principle

inf 1(8) =

inf

I ( ~ ) .

Oi==gi onJ {J
gi on

"Yi==

ooff J

Proof. Given I = (Ii: i E J), define ¢ = (¢i : i E I) by

jEJ

then ¢ is the equilibrium potential given by

{

~ j E I aij(¢i - ¢j) = 0

¢i = Ii

and, by Corollary 4.3.4, ¢ solves

for i fj. J

for i E J,

minimize E(¢)

subject to ¢i = Ii for i E J.

We have, for i E J

Z:=aij

JEI

jEJ

k ~ J j E J

jEJ

In particular, taking I == 1 we obtain

Laij = Laijo

JEI jEJ

158

4. Further theory

Hence we have equality of external currents:

2)cPi - cPj)aij = 'r)1i - Ii )Uij ·

JEI

jEJ

Moreover, we also have equality of energies:

L (cPi - cPj)2aij = 2L cPi L(cPi - cPj)aij

i,jEI

iEI JEI

= 2 L Ii LUi -!i)Uij = L Ui _!i)2Uij.
iEJ jEJ

i,jEJ

Finally, if gij = (Ii - Ij )aij and ~ i j = (¢i - ¢j)aij, then

L "Yljai/ = L (cPi - cPj)2aij

i,jEI

i,jEI

= L (Ii - Ii)2Uij = L gfjUi/,

i,jEJ

i,jEJ

so, by Theorem 4.3.3, for any flow 8 = (8ij : i,j E I) with 8i = gi for i E J
and bi = 0 for i ¢. J, we have

L ~ 2 -1 > L 2 --1

v··a··

g··a··

~ J ~ J -

~ J ~ J

i,jEI

i,jEJ

o

Effective conductivity is also related to the associated Markov chain

( X t ) t ~ O in an interesting way. Define the time spent in J

At =itl{Xs EJ}ds

and a time-changed process (Xt ) t ~ O by

X t = Xr(t)

where

r(t) = inf{s ~ 0 : As > t}.

We obtain ( X t ) t ~ O by observing ( X t ) t ~ O whilst in J, and stopping the clock
whilst ( X t ) t ~ O makes excursions outside J. This is really a transformation
of the jump chain. By applying the strong Markov property to the jump
chain we find that (Xt ) t ~ O is itself a Markov chain, with jump matrix IT
given by

1'fij = 'lrij + L'lrik¢k

k ~ J

for i,j E J,

where

4.4 Brownian motion

159

¢{ = JP>k(XT = j)

and T denotes the hitting time of J. See Example 1.4.4. Hence ( X t ) t ~ O
has Q-matrix given by

Qij = qij +L qik¢{

k ~ J

Since ¢-i = (¢{ : k E I) is the unique solution to (4.11), this shows that

so ( X t ) t ~ O is the Markov chain on J associated with the effective conduc-

tivitya.

There is much more that one can say, for example in tying up the non-
equilibrium behaviour of Markov chains and electrical networks. More-
over, methods coming from one theory one provide insights into the other.
For an entertaining and illuminating account of the subject, you should
see Random Walks and Electrical Networks by P. G. Doyle and J. L. Snell
(earus Mathematical Monographs 22, Mathematical Association of Amer-
ica, 1984).

scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->