Professional Documents
Culture Documents
Finite graph:
𝑛% ≔ |𝑉% | < +∞ and 𝑚% ≔ |𝐸% | < +∞
Notation:
What is (lazy)
random walk?
Move to neighbor with prob. 1/2deg(𝑥)
What is (lazy)
random walk?
Time 0
What is (lazy)
random walk?
Time 1 (prob. 1/6)
What is (lazy)
random walk?
Time 2 (prob. 1/2)
What is (lazy)
random walk?
Time 3 (prob. 1/10)
What is (lazy)
random walk?
Time 4 (prob. 1/2)
What is (lazy)
random walk?
Time 5 (prob. 1/2)
What is (lazy)
random walk?
Time 6 (prob. 1/6)
What is (lazy)
random walk?
Time 7 (prob. 1/2)
What is (lazy)
random walk?
Notation:
Given 𝐺 = (𝑉, 𝐸) with 𝑉 ⊂ ℕ and 𝑖 ∈ ℕ.
(K)
𝑋J := 𝑖-th random walker on 𝐺 at time 𝑡
Multiple lazy (K) (K)
random walks deg J ≔ degree of 𝑋J
We write:
An estimator
Y (K) (K)
𝐸𝑆𝑇T,U ≔ 𝐸𝑆𝑇(𝑋J , deg J : 𝑖 ≤ 𝐾; 𝑡 ≤ 𝑇)
A good Want:
estimator
For all 𝐺 ∈ 𝒢 there exist 𝐾O 𝐺 and
𝑇O (𝐺) such that ∀𝑥 ∈ 𝑉% , 𝐾 ≥ 𝐾O , 𝑇 ≥ 𝑇O ,
2
ℙ%N Y T,U − 𝛾 𝐺
𝐸𝑆𝑇 ≤ 𝜖𝛾 𝐺 ≥ .
3
Time complexity:
𝑲𝟎 𝑮 𝑻𝟎 𝑮 = min. total # of RW steps
(want this as small as possible)
Time Want:
complexity
For all 𝐺 ∈ 𝒢 there exist 𝐾O 𝐺 and
𝑇O (𝐺) such that ∀𝑥 ∈ 𝑉% , 𝐾 ≥ 𝐾O , 𝑇 ≥ 𝑇O ,
2
ℙ%N Y T,U − 𝛾 𝐺
𝐸𝑆𝑇 ≤ 𝜖𝛾 𝐺 ≥ .
3
Given what my group of 𝐾
walkers has seen so far,
Best guess our best guess is that 𝐺
given time has 2 billion vertices.
complexity
budget
1. Do sublinear estimators always exist?
∑N,€∈|×| • 1[𝑥 ∼% 𝑦]
ℎ% ≔ min
The case of |⊂./ , 𝑑|𝑆|
O} | ~j/ /0
expanders
Thm (Pinsker): When 𝑛 ≫ 1, nearly all 𝑑-
regular graphs 𝐺 on 𝑛 vertices have
ℎ% ≥ ℎ 𝑑 > 0.
Theorem:
𝒢…,† ≔ {𝑑 − 𝑟𝑒𝑔. 𝐺 𝑤𝑖𝑡ℎ ℎ% ≥ ℎ}
Why no
Can get ℎ% ≥ ℎ > 0
sublinear
before & after
self-stopping
estimator?
Huge 6-reg
𝑛 − 𝑐 vertices
A lower bound
for expanders (...) (...)
(...)
A lower bound
for expanders (...) (...)
Coupling:
𝑃% 𝑥, 𝑦 : = ℙ%N (𝑋• = 𝑦)
Some theory
for LRW •
𝑖𝑓 𝑥 = 𝑦;
0
= •
𝑖𝑓 𝑥 ∼ 𝑦;
0 ••‘ N
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒.
Transition matrix:
Some theory %
𝐼 𝐴
𝑃 = +
for LRW 2 2𝑑
ℙ%N 𝑋J = 𝑦 = 𝑃% J
𝑥, 𝑦 ,
Matrix powers
so the spectrum of 𝑃% is important.
%
is well defined, and 𝑡˜™š ≤ 2ℎ%›0 .
(Cheeger’s inequality)
Theorem:
If 𝐺 is connected, then as 𝑡 → +∞,
J
deg 𝑦
𝑃 𝑥, 𝑦 → 𝜋 𝑦 ≔ .
2𝑚
Ergodicity and
the uniform Moreover, there is a 𝑡•jKž = 𝑂 𝑡˜™š log 𝑛
mixing time such that for all 𝑡 ≥ 𝑡•jKž and 𝑥, 𝑦 ∈ 𝑉% :
3 𝑃J 𝑥, 𝑦 5
≤ ≤
4 𝜋(𝑦) 4
Leskovec, Lang, Dasgupta & Mahoney
(Internet Mathematics 2009):
Real-life social
networks
evidence of polylog relaxation time
in real-life social networks.
The case of regular
graphs
Collisions vs. intersections
Corollary:
Regular J
1
𝑃 𝑥, 𝑦 → .
graphs 𝑛
J
1
𝑡 ≫ 𝑡•jKž ⇒ 𝑃 𝑥, 𝑦 ∼ .
𝑛
Idea # 1: collision counts
(•)
Take i.i.d. LRWs 𝑋J , … , 𝑋J T .
Estimating the
§
number of 𝐶JT ≔ ¦ 1 𝑋J K = 𝑋J .
vertices K,§~T
K §
𝐶JT ≔ ¦ 1 𝑋J = 𝑋J .
K,§~T
Estimating the
number of
vertices For 𝑡 ≫ 𝑡•jKž , 𝐾 ≫ 𝑛:
T«
𝔼%N 𝐶JT ≈ and
j
T-
𝑉𝑎𝑟N% 𝐶JT ≪ « .
j
Idea # 1: collision counts
(•) T
LRWs 𝑋 ,…,𝑋 with 𝐾 = 2𝑘. Set:
Estimating the
number of (K) 0K›• 0K
𝐼J ≔ ¦ 1 𝑋±² = 𝑋±« ;
vertices O~±² ,±« ~J›•
1 (K)
𝐼J ≔ ¦ 𝐼J
𝑘
•~K~³
Theorem (Ben-Hamou, O., Peres):
𝑛% = Θ 𝑎𝑘
%
𝑡
Stars˜™š =Θ
form an 𝑘 0
Lower bounds 3-regular expander
for “all” 𝐭𝐢𝐦𝐞𝑎 or
with = 2𝑎
𝛀 𝒌 𝟐
vertices.
𝒂
𝟑 𝟏
relaxation = 𝛀( 𝐤 𝟐 𝟒 𝒌𝒂 ) 𝟐
times
No rw is able to distinguish Gk,¸ and
𝑮𝒂 or 𝑮𝟐𝒂 : edges become 𝑘-paths w/ handles 3
0 Ô
(slows RW down by 𝑘 factor) ¸2 k & tu
log
Sketch of
first moment Click here .
upper bounds
Collisions give suboptimal bounds
(also require adding more RWs over time).
Takeaways
from the
regular case Intersections use whole paths and are
optimal for “all” values of the relaxation
time.
Missing:
Results for
nonregular graphs
(•) T
LRWs 𝑋 ,…,𝑋 with 𝐾 = 2𝑘. Set:
Estimating the
0K›• 0K
number of (K)
1 𝑋±² = 𝑋±«
edges ℐJ ≔ ¦ 0K
;
O~±² ,±« ~J›• deg % (𝑋±« )
1 (K)
ℐJ ≔ ¦ ℐJ
𝑘
•~K~³
Theorem (Ben-Hamou, O., Peres):
𝑡0 2
ℙ%N 1−𝜖 𝑚 ≤ ≤ 1+𝜖 𝑚 ≥ .
2ℐJ 3
Corollary: (loosely stated)
𝑛 𝜋(𝑦)
What about =¦
2𝑚 deg(𝑦)
the number of €
J×ØÙÚ Ì
Additional steps.
j
5/6 Ô
Time tunif n is not enoug
𝑛% = Θ 𝑘𝑞
𝑚% = Θ 𝑘 0
%
𝑡•jKž = Θ 𝑞0
Kk
Extra time is 𝐭𝐢𝐦𝐞 = 𝛀 𝒒𝒌
necessary for 𝒕𝒖𝒏𝒊𝒇 𝒎
= 𝛀( )
vertices 𝒏
(walk q on path)
However, once a good esti
the
𝑮𝒒 or 𝑮𝟐𝒒 : clique of sizemean degree,
𝑘 with paths which ca
of length
𝑞 or 2𝑞 attached to each vertex (𝑘 ≫ 𝑞)
Corollary: (loosely stated)
J 0
𝑃 𝑥, 𝑦
𝑑N0 𝑡 ≔¦ −1 𝜋(𝑦)
𝜋 𝑦
Estimating €
𝑑N0 𝑡 𝑃0J 𝑥, 𝑥 1
≔ −
2𝑚 deg 𝑥 2𝑚
(can be related to intersections)
Conclusion
and open problems
Can estimate # of vertices, edges and
mixing time via multiple RW:
Summary of
results
1. no self-stopping in general
2. sometimes (but not always) sublinear
What about other parameters?
Main open
problems What other models of local access to the
graph? (eg. “cluster growth”)
(jump to end)
Sketch of first moment
intersection bound
for regular graphs
Recall:
(•) T
𝑋 ,…,𝑋 on regular graph with 𝐾 = 2𝑘.
K 0K›• 0K
Sketch of first 𝔼%N 𝐼J ≔ ¦ ℙ%N 𝑋±² = 𝑋±« ;
moment O~±² ,±« ~J›•
bound
Claim:
𝑡0 % K 𝑡0 –/0
≤ 𝔼N 𝐼J ≤ + 𝐶𝑡˜™š
𝑛 𝑛
Intersection and return probabilities:
• 0
ℙ%N 𝑋±² = 𝑋±«
Sketch of first
moment = ∑€ 𝑃±² 𝑥, 𝑦 𝑃±« (𝑥, 𝑦) (meet at some 𝑦)
bound
= ∑€ 𝑃±² 𝑥, 𝑦 𝑃±« (𝑦, 𝑥) (symmetry of 𝑃)
•
⇒ 𝑃± 𝑥, 𝑥 − = ∑Kß0 𝜆K± 𝜓K0 𝑥 ≥ 0.
j
Intersection and return probabilities:
% • • 0
𝔼N [𝐼J ]≔ ¦ ℙ%N 𝑋±² = 𝑋±«
O~±² ,±« ~J›•
Sketch of first
moment J« ±² ݱ« •
= + ∑±² ,±« ~J›• 𝑃 𝑥, 𝑥 −
bound j j
𝑡0 ±
1
≤ + ¦(𝑠 + 1) 𝑃 𝑥, 𝑥 −
𝑛 𝑛
±ßO
Back to the spectral representation
•
Sketch of first ∑±ßO 𝑠 + 1 𝑃± 𝑥, 𝑥 − =
j
moment
bound = ∑±ßO (𝑠 + 1) ∑Kß0 𝜆K± 𝜓K0 𝑥
›•
with all 0 ≤ 𝜆K ≤ 𝜆0 = 1 − 𝑡˜™š
± • à
𝑃 𝑥, 𝑥 − ≤ (Aldous-Fill’94)
j ±Ý•
Sketch of first 1
±
moment ⇒ ¦ 𝑠+1 𝑃 𝑥, 𝑥 −
𝑛
bound O~±~𝒕𝒓𝒆𝒍
–
0
≤ ¦ 𝐶 𝑠+1≤ 𝐶𝑡˜™š .
O~±~𝒕𝒓𝒆𝒍
End of sketch:
𝑡0 1
𝔼%N 𝐼J • ±
= + ¦(𝑠 + 1) 𝑃 𝑥, 𝑥 −
𝑛 𝑛
Sketch of first ±ßO
moment
𝑡0 ±
1
bound ≤ + 𝐶 ¦ (𝑠 + 1) 𝑃 𝑥, 𝑥 −
𝑛 𝑛
±~Jáâã
–
𝑡0 0
≤ + 𝐶𝑡˜™š .
𝑛
Weighted intersections:
For min degree 𝑑:
0K›• 0K
(K)
1 𝑋±² = 𝑋±«
Non-regular ℐJ ≔ ¦ 0K
;
graphs O~±² ,±« ~J›• deg % (𝑋±« )
(back)
± •O••‘ N
𝑃 𝑥, 𝑥 − 𝜋(𝑥) ≤
† ±Ý•
(Peres-O.’19)
Thank you!