Scribd
Upload
24 views

Luby Codes - Differential Equations: 28 de Outubro de 2011

The document discusses Luby codes and differential equations related to analyzing their performance under erasures. It contains the following key points: 1) It presents differential equations to model the evolution of left and right node degrees over time as edges are removed. This allows calculating the expected fraction of nodes with each degree. 2) It proves that if the fraction of surviving right-degree-one nodes stays positive, the decoding process will continue successfully. 3) The main theorem states that if the fraction of erased coordinates is below a threshold, the decoding algorithm will terminate successfully with high probability using O(k) steps, where k is the number of left nodes.

Uploaded by

Pedro Oliveira
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
24 views

Luby Codes - Differential Equations: 28 de Outubro de 2011

The document discusses Luby codes and differential equations related to analyzing their performance under erasures. It contains the following key points: 1) It presents differential equations to model the evolution of left and right node degrees over time as edges are removed. This allows calculating the expected fraction of nodes with each degree. 2) It proves that if the fraction of surviving right-degree-one nodes stays positive, the decoding process will continue successfully. 3) The main theorem states that if the fraction of erased coordinates is below a threshold, the decoding algorithm will terminate successfully with high probability using O(k) steps, where k is the number of left nodes.

Uploaded by

Pedro Oliveira
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 18

Luby Codes - Differential Equations

28 de outubro de 2011

()

28 de outubro de 2011

1 / 17

Left-hand Evolution - Differential Equation


(i)

Lt : number of edges of left-degree i at time t


lt (i) :=

(i)

Lt
E

After confirm that the conditions of the THM are attained we have to
solve the diff. eq.:
dli ( )
d

ili ( )
, i = 1, . . . , d max deg on the left
e( )

Here, to ease notation we hold degree i fix and drop out this from li ( ):
Z
Z
dl( )
dl(s)
ds
d
= i

= i
ds
l( )
l( )
a l(s)
a e(s)
Z
R

ds
i ds ds

lnl(s) a = i
ds l( ) = l(a)e a e(s)
a e(s)
()

28 de outubro de 2011

2 / 17

Now defining x( ) := e

ds
a e(s) ds

l( ) = l(a)x( )i

() we obtain

l(x) = l(a)x i function of x

Notice that from (), x(0) = 1, and that for li ( = a = 0) = i , therefore


li (x) = i x i
Finally notice that li (x) goes to zero as goes to /al , x runs over the
interval (0, 1]

()

28 de outubro de 2011

3 / 17

Right-hand Evolution

(i)

Rt : number of edges of right degree i at time t


(i)

(i)

rt := REt
P (i)
et = i rt : fraction of the edges remaing at time t
d: maximum degree on the left

()

28 de outubro de 2011

4 / 17

After check the conditions of


 the THM are satisfied, we have a.s.:
(i)
Rt = Eri (t/E ) + O E 5/6 for E t ( + ) E and ri ( )
Where the ri ( ) are the solutions to

()

dri ( )
dt

= (ri+1 ( ) ri ( ))

dr1 ( )
dt

= (r2 ( ) r1 ( ))

i (a( ) 1)
, for
e( )

i >1

i (a( ) 1)
1
e( )

28 de outubro de 2011

(1)
(2)

5 / 17

Progression of r1 ( )

If r1 ( ) > 0 there is a node of degree one on the right


and the process continues. Otherwise, r1 ( ) = 0
process stops!
Hence we would like r1 ( ) > 0 until all nodes on the
left are deleted and the process terminates successfully.

()

28 de outubro de 2011

6 / 17

Determination of rj (1)

Recall that rj (1) is the expected fraction of edges of


right-degree j at time 0
Prior time t = 0 each node on the left is deleted
randomly with probability 1
The resulting graph is a random graph over those with
the given degree sequence, ({m })
To the nodes on the right it is as though each edge is
deleted with probability 1

()

28 de outubro de 2011

7 / 17

Hence, an edge whose right incident node had degree


j before remains in the graph and has degree i after
with probability


j 1 i
(1 )j1
i 1
Thus
rj (1) =

X
mj

()


m1 j
(1 )mj
j 1

28 de outubro de 2011

8 / 17

Appendix B

dri ( )
dt
dr1 ( )
dt

i (a( ) 1)
, for i > 1
e( )
i (a( ) 1)
= (r2 ( ) r1 ( ))
1 (E 4)
e( )


X
m1 j
rj (1) =
m
(1 )mj initial condition
j 1
= (ri+1 ( ) ri ( ))

(3)
(4)
(5)

mj

Proposition 1: For the solution to the system of differential equations


is given by
r1 (x) = (x) [x 1 + (1 (x))]
(6)
where x is defined via

()

dx
d

x
= e(
)

28 de outubro de 2011

9 / 17

Proposition 2

Let B(, ) be a bipartite random graph with k message bits Let be


fixed so that
(1 (x)) > 1 x, for x (0, 1].
For all > 0 there is some k0 such that for all k k0 , if the message bits
of C (B) are erased i.i.d. with prob. , then with probability at least
1 k 2/3 e

3
k
2

the recovery algorithm terminates with at most k message bits erased

()

28 de outubro de 2011

10 / 17

Proposition 2 - Proof

Let E = kal . Recall that al is the average degree of the nodes in the
left side of the graph.
al is a constant for fixed and , given by al = P1i .
i
.
Let = al . By ( 6) and the preceding discussions, with probability at

least 1 k 2/3 e k/2 the number of nodes of degree one on the right
is
r1 (x) = (x) [x 1 + (1 (x))] + O(k 5/6 )

R
for x ( , 1], where = exp 0 ds/e(s) .

By assumptions this number is positive (for large enough k), which


proves the proposition. 

()

28 de outubro de 2011

11 / 17

Decoding processing termination

The foregoing proposition does not prove that the decoding process
terminates successfully
Combinatorial arguments are needed in order to demonstrate that any
small subset of left nodes has many right neighbors

Lemma Let B(, ) a bipartite random graph with


k left nodes, such that 1 = 2 = 0. Then there is
some > 0, such that, with probability 1 O(k 3/2 ),
the recovery process restricted to the subgraph
induced by any fraction of the left nodes
terminates successfully

()

28 de outubro de 2011

12 / 17

Lemma 1 (Proof)

Let S be any set of nodes on the left of size |S| k,


where will be chosen later
Let a be the average number of these nodes
If the number of nodes on the right that are neighbors
of S is greater than a|S|/2, then one of these nodes
has only one neighbor in S (Why?) and so, the
process continue.
Thus, this only is need to show that the initial graph
is a good expander on small sets. (what is that?)

()

28 de outubro de 2011

13 / 17

Let Es denote the event that a subset of size s of the


nodes on the left has at most as/2 neighbors.
Firstly Pr(Es ) will be bounded and then sum Pr(Es )
over all values of s no larger than k.
For this, fix any subset S of the left nodes of size s,
and any subset T of the right nodes of size as/2.


k
ways of
There are ks ways of choosing S, and as/2
choosing T
The probability that 
T contains all as neighbors of the
as
Hence, one has
vertices in S is as/2
k
as
   
k
k
as
Pr (Es )
as
2k
s
s
()

28 de outubro de 2011

14 / 17

Note that

n
k

(ne/k)k , hence one has


Pr (Es )

 s (a/21)s
k

sc 2
k

s/2

where c is a constant depending on and a.


Since the graph does not have nodes of degreee one or two, one has
that Pr(E1 ) = Pr(E2 ) = 0.
Choosing 1/(2c 2 ) yields
k
X
s=1

Pr(Es )

k  2 s/2
X
sc
s=3

= O

k k

X 1
3c 2
+
k k s=4 2s

which shows that, with high probability, the original graph is an


expander on small subsets. 
()

28 de outubro de 2011

15 / 17

Theorem 2 (Main)

Let k be an integer, and suppose that C = C (B1 , . . . , Bm , C ) is a cascade


of described bipartite graphs, where B1 has k L-nodes. Suppose that each
Bi is chosen at random with edge degree specified by (x) and (x), such
that
1 = 2 = 0,
and suppose that is such that
(1 (x)) > 1 x, 0 < x 1

(E 7)

Then, if at most a fracion of the coordinates of an encoded word in C


are erased i.i.d, the algorithm A1 terminates successufully with probability
1 O(k 3/4 ) in O(k) steps

()

28 de outubro de 2011

16 / 17

Proof of the Main Theorem

Part (A) At each level of the cascade, the nr of edges is


1
E = kl R 1
0 (t)dt = cte
that is (nr of left nodes) (aver. degree), so the total nr of edges in the cascade = nr of steps to recovery is O(k)

Part (B) (Chernoff bound (ref.19, Prob. 4.7(c), pp. 98). Let Ej the event: There is some j such that the fraction of left

.
8
nodes lost on the left side of the graph Bj is larger than = + 1/ k
The adapted Chernoff bound reads: for any j, the probability that there are more erasures than

Pr Ej

k
2j

k
2j

3/4

4
2j e2 k

1/4

Remark: = j + k j3/4
2
2

()

28 de outubro de 2011

17 / 17

Proof of the Main Theorem (cont. Part B)

Union bound: required probability is at most equal to the sum of these probabilities which is log(k)
log(k)
that there are at most 2 such j s.

exp(
2

4
k)

. Note

Recall Proposition 2: Let B(, ) with k msg bits. Let so that


(1 (x)) > 1 x
For all > 0 there is k0 such that for k k0 , if the message bits of C(B) are erased i.i.d. with prob. , then with
prob. at least 1 k 2/3 e

3
k/2

the algorithm A1 terminates with at most k message bits erased


For large enough k, (E7) is satisfied for instead of for any > 0 and any of the graphs Bj the algorithm A1 stops
with less than k/2j nodes uncorrected with prob. 1 O (exp (k )) for some > 0.
Recall Lemma 1: Let B(, ) with 1 = 2 = 0. Then there is some > 0, such that with prob.1 O(k 3/2 ), the
recovery process restricted to the subgraph induced by any fraction of the left nodes terminates successfully

3/2 
Lemma 1 applies for small , the recovery process ends successfully with prob. 1 O
.
2j /k
The algorithm fail probability is thus at most
completes the proof. 

()

P  j 3/2
,
j 2 /k



j = 0 . . . log(k)/2. This equals O k 3/4 , which

28 de outubro de 2011

18 / 17

You might also like