Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu...
Transcript of Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu...
![Page 1: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/1.jpg)
Analysis of random walks using tabu lists
K. Altisen, S. Devismes, A. Gerbaud and P. Lafourcade
Laboratoire Verimag
Université Joseph Fourier, Grenoble 1
1 / 18
![Page 2: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/2.jpg)
1 Random walk with tabu list
2 Termination
3 Optimal update rules for m-free graphs
2 / 18
![Page 3: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/3.jpg)
1 Random walk with tabu list
2 Termination
3 Optimal update rules for m-free graphs
3 / 18
![Page 4: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/4.jpg)
Simple random walk routing
0
1
2
3
4
5
6
7
8
9
1/4
1/4
1/4
1/4
source
destination
All graphs are �nite, simple,
connected, with at least two vertices.
4 / 18
![Page 5: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/5.jpg)
Simple random walk routing
0
1
2
3
4
5
6
7
8
9
1/4
1/4
1/4
1/4
source
destination
At each step, the walker moves to a
neighbor chosen uniformly at random.
The walker represents a packet
forwarded from node to node.
4 / 18
![Page 6: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/6.jpg)
Simple random walk routing
0
1
2
3
4
5
6
7
8
9
1/4
1/4
1/4
1/4
source
destination
Advantages:
X no routing table;
X local computations;
X scalable;
X robust; and
X load-balanced.
Drawback : slow.
4 / 18
![Page 7: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/7.jpg)
Mean hitting times
Hitting time: number of steps to
deliver the packet to its �nal
destination.
Motivation: reducing the mean
hitting times.
5 / 18
![Page 8: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/8.jpg)
Mean hitting times
Let (Xn)n>0 be a random walk on a
graph G . For every two vertices x
and y of G :
The hitting time of y is the
random variable
Hy (G ) = inf{n > 0 : Xn = y} .
The mean hitting time of y
starting from x is the mean
value ExHy (G ).
5 / 18
![Page 9: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/9.jpg)
Mean hitting times
Let (Xn)n>0 be a random walk on a
graph G . For every two vertices x
and y of G :
The hitting time of y is the
random variable
Hy (G ) = inf{n > 0 : Xn = y} .
The mean hitting time of y
starting from x is the mean
value ExHy (G ).
0
1
2
3
4
5
E1H0(K6) = 5
5 / 18
![Page 10: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/10.jpg)
Mean hitting times
Let (Xn)n>0 be a random walk on a
graph G . For every two vertices x
and y of G :
The hitting time of y is the
random variable
Hy (G ) = inf{n > 0 : Xn = y} .
The mean hitting time of y
starting from x is the mean
value ExHy (G ).
0
1
2
3
4
5
E1H0(S6) = 2× 5
5 / 18
![Page 11: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/11.jpg)
Mean hitting times
Let (Xn)n>0 be a random walk on a
graph G . For every two vertices x
and y of G :
The hitting time of y is the
random variable
Hy (G ) = inf{n > 0 : Xn = y} .
The mean hitting time of y
starting from x is the mean
value ExHy (G ).
0
1
2
3
4
5
6
7
8
9
E1H0(G ) = 39
5 / 18
![Page 12: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/12.jpg)
Random walk with tabu list
The walker has a �nite memory in which it stores occurrences of vertices
already visited, called tabu list. The occurrences are sorted in chronological
order. The number of blocks is the length of the tabu list. The tabu list is
full if all blocks are allocated. The policy to insert or remove occurrences of
vertices in the tabu list is called the update rule.
31
lastoccurenceinserted
(a) The tabu listhas length 3 with2 allocated blocks.
25124
lastoccurenceinserted
(b) The tabu listhas length 5 and isfull.
6 / 18
![Page 13: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/13.jpg)
Random walk with tabu list
If the tabu list is not full, then three cases may occur:
x t1x
t3
x
t1x
t3
(c) The current vertex isinserted.
x t1x
t3
x
t1t3
(d) The current vertex isinserted and an older oc-currence of the currentvertex is discarded.
x t1x
t3
t1x
t3
(e) The current vertex isnot inserted.
6 / 18
![Page 14: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/14.jpg)
Random walk with tabu list
If the tabu list is full, then three cases may occur:
x t1x
t3t4t5
x
t1x
t3t5
(f) The current vertex isinserted and an occur-rence of another vertex isdiscarded.
x t1x
t3t4t5
x
t1t3t4t5
(g) The current vertex isinserted and an older oc-currence of the currentvertex is discarded.
x t1x
t3t4t5
t1x
t3t4t5
(h) The current vertex isnot inserted.
6 / 18
![Page 15: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/15.jpg)
Random walk with tabu list
At each step, the walker moves to a neighbor chosen uniformly at random
among those without occurrence in the tabu list. If all neighbors have an
occurrence in the tabu list, the walker moves to a neighbor chosen
uniformly at random.
0
1
2
3
4
5
6
7
8
9
1/2
1/2
2527
0
1
2
3
4
5
6
7
8
9
1/4
1/4
1/4
1/4
28676
6 / 18
![Page 16: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/16.jpg)
First In First Out
The FIFOm update rules, m > 0.
x t1t2
almost surely
x
t1t2
x t1t2t3t4t5
almost surely
x
t1t2t3t4
Here, m = 5. The current vertex is inserted almost surely. If the tabu list is full,
then the occurence of the vertex stored in the last block is discarded, almost
surely.
7 / 18
![Page 17: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/17.jpg)
First In First Out
The FIFO40 update rule.
7 / 18
![Page 18: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/18.jpg)
First In First Out
Contrarily to the snake, the walker never dies.
7 / 18
![Page 19: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/19.jpg)
A non deterministic update rule
x t1t2
almost surely
x
t1t2
x t1t2t3t4t5
probability 1/5
x
t1t2t4t5
The current vertex is inserted almost surely. If the tabu list is full, then an
occurence in the tabu list, chosen uniformly at random, is discarded. Here, t3 is
discarded.
8 / 18
![Page 20: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/20.jpg)
Hitting time in lollilop graph
For the simple random walk, G. Brightwell and P. Winkler showed in 1990:
maxG :|V|=n
maxx ,y∈V
ExHy (G ) = E1H0(Lollilopn) =4n3
27(1+ o(1)) .
For the random walk with tabu list and update rule FIFO1,
E1H0(Lollilopn) =4n2
9(1+ o(1)) .
1
23
4
5
6
7 8
9
10 11 12 13 14 0
clique with 2n/3 vertices path with n/3 vertices
9 / 18
![Page 21: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/21.jpg)
1 Random walk with tabu list
2 Termination
3 Optimal update rules for m-free graphs
10 / 18
![Page 22: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/22.jpg)
Trap
0 1 2
3
4
X0 X1
X2
X3
X4
X5
X6
X7
X0
1X1
2
1
X2
3
2
1
X3
4
3
1
X4
2
4
1
X5
3
2
1
X6
4
3
1
X7
2
4
1
In some cases, there is a positive probability that a vertex is never reached.11 / 18
![Page 23: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/23.jpg)
Finite mean hitting times
An update rule is called trivial if the current vertex is never inserted in the
tabu list when the tabu list contains no element.
The unique update rule with zero length is trivial.
Consider a trivial update rule. If the tabu list is initially empty, then it
remains empty forever and the walker performs a simple random walk.
Consequently, all mean hitting times are �nite for all graphs.
12 / 18
![Page 24: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/24.jpg)
Finite mean hitting times
Theorem
Consider a non trivial update rule.
All mean hitting times are �nite for
all connected graphs, if and only if,
when the current vertex has no
occurrence in the tabu list:
If the tabu list is not full, then
the probability to insert the
current vertex is positive.
If the tabu list is full, then the
probability to remove the last
element is positive.
x t1t2
probability > 0
x
t1t2
x t1t2t3t4t5
probability > 0
x
t1t2t3t4
12 / 18
![Page 25: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/25.jpg)
Finite mean hitting times
Sketch of proof.
If one of the two conditions is not
full�lled, then we exhibit a graph such
that, with positive probability, the
walker never reaches a given vertex.
For example, in the following
situation, the vertex 1 is never
removed from the tabu list. Thus,
the vertex 0 is never reached.
0 1
5
4
3
2
6
X0
12 / 18
![Page 26: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/26.jpg)
Finite mean hitting times
Sketch of proof.
If one of the two conditions is not
full�lled, then we exhibit a graph such
that, with positive probability, the
walker never reaches a given vertex.
For example, in the following
situation, the vertex 1 is never
removed from the tabu list. Thus,
the vertex 0 is never reached.
0 1
5
4
3
2
6
4
3
2
1X4
12 / 18
![Page 27: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/27.jpg)
Finite mean hitting times
Sketch of proof.
If one of the two conditions is not
full�lled, then we exhibit a graph such
that, with positive probability, the
walker never reaches a given vertex.
For example, in the following
situation, the vertex 1 is never
removed from the tabu list. Thus,
the vertex 0 is never reached.
0 1
5
4
3
2
6
5
3
2
1
X5
12 / 18
![Page 28: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/28.jpg)
Finite mean hitting times
Sketch of proof.
If one of the two conditions is not
full�lled, then we exhibit a graph such
that, with positive probability, the
walker never reaches a given vertex.
For example, in the following
situation, the vertex 1 is never
removed from the tabu list. Thus,
the vertex 0 is never reached.
0 1
5
4
3
2
6
6
5
2
1
X6
12 / 18
![Page 29: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/29.jpg)
Finite mean hitting times
Sketch of proof.
If one of the two conditions is not
full�lled, then we exhibit a graph such
that, with positive probability, the
walker never reaches a given vertex.
For example, in the following
situation, the vertex 1 is never
removed from the tabu list. Thus,
the vertex 0 is never reached.
The reciprocal assertion follows from
the theory of Markov chains on �nite
space.
0 1
5
4
3
2
6
6
5
2
1
X6
12 / 18
![Page 30: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/30.jpg)
1 Random walk with tabu list
2 Termination
3 Optimal update rules for m-free graphs
13 / 18
![Page 31: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/31.jpg)
Optimal update rules for m-free graphs
A graph is m-free if a tabu random
walk with FIFOm update rule cannot
visit a vertex in its current tabu list,
unless the current vertex has degree
one.
All graphs are 1-free.
A graph is 2-free if and only if it
does not contain any triangle
with a degree two vertex.
Every (m + 1)-regular graph is
m-free.
A graph with girth greater than
or equal to m + 2 is m-free.
If 1 6 k < m, then all m-free
graphs are k-free.
14 / 18
![Page 32: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/32.jpg)
Optimal update rules for m-free graphs
A graph is m-free if a tabu random
walk with FIFOm update rule cannot
visit a vertex in its current tabu list,
unless the current vertex has degree
one.
All graphs are 1-free.
A graph is 2-free if and only if it
does not contain any triangle
with a degree two vertex.
Every (m + 1)-regular graph is
m-free.
A graph with girth greater than
or equal to m + 2 is m-free.
If 1 6 k < m, then all m-free
graphs are k-free.
14 / 18
![Page 33: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/33.jpg)
Optimal update rules for m-free graphs
A graph is m-free if a tabu random
walk with FIFOm update rule cannot
visit a vertex in its current tabu list,
unless the current vertex has degree
one.
All graphs are 1-free.
A graph is 2-free if and only if it
does not contain any triangle
with a degree two vertex.
Every (m + 1)-regular graph is
m-free.
A graph with girth greater than
or equal to m + 2 is m-free.
If 1 6 k < m, then all m-free
graphs are k-free.
0
1
2
3
4
5
The clique with 6 elements is 5-regular,
hence 4-free.
14 / 18
![Page 34: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/34.jpg)
Optimal update rules for m-free graphs
A graph is m-free if a tabu random
walk with FIFOm update rule cannot
visit a vertex in its current tabu list,
unless the current vertex has degree
one.
All graphs are 1-free.
A graph is 2-free if and only if it
does not contain any triangle
with a degree two vertex.
Every (m + 1)-regular graph is
m-free.
A graph with girth greater than
or equal to m + 2 is m-free.
If 1 6 k < m, then all m-free
graphs are k-free.
0
1
2
3
4
5
6
7
8
9
X0
X1
X2
85
That graph is not 2-free because X3
belongs to {5, 8} almost surely.
14 / 18
![Page 35: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/35.jpg)
Optimal update rules for m-free graphs
A graph is m-free if a tabu random
walk with FIFOm update rule cannot
visit a vertex in its current tabu list,
unless the current vertex has degree
one.
All graphs are 1-free.
A graph is 2-free if and only if it
does not contain any triangle
with a degree two vertex.
Every (m + 1)-regular graph is
m-free.
A graph with girth greater than
or equal to m + 2 is m-free.
If 1 6 k < m, then all m-free
graphs are k-free.
0
1
2
3
4
5
6
7
8
9
X0
X1
X2
X3
85
That graph is 2-free because it does not
contain any triangle with a degree two
vertex.
14 / 18
![Page 36: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/36.jpg)
Theorem
For every nonnegative integer m and for every m-free graph, FIFOm yields
minimal mean hitting times among all update rules with length less than or
equal to m.
15 / 18
![Page 37: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/37.jpg)
Theorem
For every nonnegative integer m and for every m-free graph, FIFOm yields
minimal mean hitting times among all update rules with length less than or
equal to m.
Sketch of proof.
Let m be a positive integer. A walker (Xn)n>0 backtracks at step n if Xnbelongs to {Xmax{n−2,0}, . . . ,Xmax{n−m−1,0}} while there exists a neighbor
of Xn−1 which does not have any occurrence in the tabu list.
On one hand, we show that any tabu random walk with FIFOm update rule
does not backtrack on any m-free graph.
On the other hand, starting from any tabu random walk, we remove all
backtracks to get a stochastic process that follows the same law than a
tabu random walk with FIFOm update rule.
15 / 18
![Page 38: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/38.jpg)
Theorem
For every nonnegative integer m and for every m-free graph, FIFOm yields
minimal mean hitting times among all update rules with length less than or
equal to m.
Sketch of proof.
0
1
2
3
4
5
6
7
8
9
X0
X1
X2 = X5
X3
X4
X6
(a) The walker backtracks atstep 5.
0
1
2
3
4
5
6
7
8
9
(b) We remove the back-tracks. 15 / 18
![Page 39: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/39.jpg)
Since all graphs are 1-free:
Corollary
For every update rule R of length 0 or 1, the mean hitting times associated
to R are greater than or equal to the corresponding mean hitting times
associated to FIFO1.
Your mean hitting times are always lower if someone follows you. It does not hold
on all graphs for two people or more.
16 / 18
![Page 40: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/40.jpg)
Better sometimes to forget
0 1 2 3
X0
0 1 2 3
X11
0 1 2 3
X22
1
0 1 2 3
X33
2
1
17 / 18
![Page 41: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/41.jpg)
Comparisons between FIFOm update rules, m > 0
Kn: clique with n vertices.
Ln: path with n vertices.
Fn: �ower graph with n petals.
0
1
2
3
45
6
7
89
(a) F4
0
1
2
3
4567
8
9
10
11
12
13
1415
(b) F7
18 / 18
![Page 42: Analysis of random walks using tabu lists · 2013-12-18 · Analysis of random walks using tabu lists K. Altisen, S. Devismes, A. Gerbaud and .P Lafourcade Laboratoire Verimag Université](https://reader033.fdocuments.in/reader033/viewer/2022050609/5fb0c45a2db39f047f2237e2/html5/thumbnails/42.jpg)
Comparisons between FIFOm update rules, m > 0
Kn: clique with n vertices.
Ln: path with n vertices.
Fn: �ower graph with n petals.
0 1 2 3 4 m > 5
0 × K3 K3 K3 K3 K3
1 × × K3 K3 K3 K3
2 F7 F4 × K4 K4 K4
3 L4 L4 L4 × K5 K5
4 L4 L4 L4 L5 × K6
k > 5 L4 L4 L4 L5 L6
Lm+2 if m < k ,
× if m = k ,
Kk+2 if m > k .
18 / 18