Thresholds for Some Basic Properties Eight Lectures on Random Graphs: Lecture 2
Oleg Pikhurko Carnegie Mellon University
Basic Thresholds – p.1 Gn,p = random order-n graph with edge probability p. Whp = with high probability (approaching 1 as n → ∞). Markov's Inequality: for a random variable X ≥ 0 and a real a > 0 E [ X ] Pr [ X ≥ a ] ≤ . a
Graphs and Properties
Graph property = a collection of graphs. Monotone = adding edges cannot violate it.
Basic Thresholds – p.2 Whp = with high probability (approaching 1 as n → ∞). Markov's Inequality: for a random variable X ≥ 0 and a real a > 0 E [ X ] Pr [ X ≥ a ] ≤ . a
Graphs and Properties
Graph property = a collection of graphs. Monotone = adding edges cannot violate it.
Gn,p = random order-n graph with edge probability p.
Basic Thresholds – p.2 Graphs and Properties
Graph property = a collection of graphs. Monotone = adding edges cannot violate it.
Gn,p = random order-n graph with edge probability p. Whp = with high probability (approaching 1 as n → ∞). Markov's Inequality: for a random variable X ≥ 0 and a real a > 0 E [ X ] Pr [ X ≥ a ] ≤ . a
Basic Thresholds – p.2 Proof Define p0 ∈ [0, 1] by p1 + (1 − p1) p0 = p2.
Let G1 ∈ Gn,p1 and G0 ∈ Gn,p0 . Then G1 ∪ G0 ∼ Gn,p2 .
Pr [ G1 ∈ A ] ≤ Pr [ G1 ∪ G0 ∈ A ] .
Monotone Properties
Theorem For any monotone A and p1 ≤ p2
Pr [ Gn,p1 ∈ A ] ≤ Pr [ Gn,p2 ∈ A ] .
Basic Thresholds – p.3 Monotone Properties
Theorem For any monotone A and p1 ≤ p2
Pr [ Gn,p1 ∈ A ] ≤ Pr [ Gn,p2 ∈ A ] .
Proof Define p0 ∈ [0, 1] by
p1 + (1 − p1) p0 = p2.
Let G1 ∈ Gn,p1 and G0 ∈ Gn,p0 . Then G1 ∪ G0 ∼ Gn,p2 .
Pr [ G1 ∈ A ] ≤ Pr [ G1 ∪ G0 ∈ A ] .
Basic Thresholds – p.3 1 Example p0 = n is a threshold for having a cycle. Indeed, if p = o(1/n), then n n (i − 1)! Pr [ ∃ cycle ] ≤ E [ #cycles ] = pi ≤ (np)i → 0. i 2 Xi≥3 Xi≥3
2+ε n n−1 if p > n , then E [ e(G) ] = p 2 > (2 + ε) 2 . By Chernoff’s bound, whp e(G) ≥ n.
Thresholds
p0 = p0(n) is a threshold for a monotone property A if ∀p(n)
0, if p/p0 → 0, Pr [ Gn,p ∈ A ] → ( 1, if p/p0 → ∞.
Basic Thresholds – p.4 2+ε n n−1 if p > n , then E [ e(G) ] = p 2 > (2 + ε) 2 . By Chernoff’s bound, whp e(G) ≥ n.
Thresholds
p0 = p0(n) is a threshold for a monotone property A if ∀p(n)
0, if p/p0 → 0, Pr [ Gn,p ∈ A ] → ( 1, if p/p0 → ∞.
1 Example p0 = n is a threshold for having a cycle. Indeed, if p = o(1/n), then
n n (i − 1)! Pr [ ∃ cycle ] ≤ E [ #cycles ] = pi ≤ (np)i → 0. i 2 Xi≥3 Xi≥3
Basic Thresholds – p.4 Thresholds
p0 = p0(n) is a threshold for a monotone property A if ∀p(n)
0, if p/p0 → 0, Pr [ Gn,p ∈ A ] → ( 1, if p/p0 → ∞.
1 Example p0 = n is a threshold for having a cycle. Indeed, if p = o(1/n), then
n n (i − 1)! Pr [ ∃ cycle ] ≤ E [ #cycles ] = pi ≤ (np)i → 0. i 2 Xi≥3 Xi≥3 2+ε n n−1 if p > n , then E [ e(G) ] = p 2 > (2 + ε) 2 . By Chernoff’s bound, whp e(G) ≥ n.
Basic Thresholds – p.4 Proof Choose p0 = p(1/2), i.e.
Pr [ Gn,p0 ∈ A ] = 1/2.
p0 exists as f(p) = Pr [ Gn,p ∈ A ] is a polynomial with f(0) = 0 and f(1) = 1.
Monotone ⇒ ∃ Threshold
Theorem (Bollobás-Thomason’87) Every non-trivial monotone property A has a threshold.
Basic Thresholds – p.5 Monotone ⇒ ∃ Threshold
Theorem (Bollobás-Thomason’87) Every non-trivial monotone property A has a threshold.
Proof Choose p0 = p(1/2), i.e.
Pr [ Gn,p0 ∈ A ] = 1/2.
p0 exists as f(p) = Pr [ Gn,p ∈ A ] is a polynomial with f(0) = 0 and f(1) = 1.
Basic Thresholds – p.5 m As 1 − (1 − p) ≤ pm ≤ p0,
1 m ≤ Pr [ H 6∈ A ] ≤ Pr [ ∀i G 6∈ A ] = 1 − Pr [ G ∈ A ] . 2 i n,p ⇒ Pr [ Gn,p ∈ A ] < ε.
Other direction: take Gn,p0 ∪ · · · ∪ Gn,p0 .
p0 = p(1/2) is a threshold
m Given ε > 0, let (1 − ε) < 1/2. Let p < p0/m. Let G1, . . . , Gm ∈ Gn,p and
H = G1 ∪ · · · ∪ Gm ∼ Gn,1−(1−p)m .
Basic Thresholds – p.6 Other direction: take Gn,p0 ∪ · · · ∪ Gn,p0 .
p0 = p(1/2) is a threshold
m Given ε > 0, let (1 − ε) < 1/2. Let p < p0/m. Let G1, . . . , Gm ∈ Gn,p and
H = G1 ∪ · · · ∪ Gm ∼ Gn,1−(1−p)m .
m As 1 − (1 − p) ≤ pm ≤ p0,
1 m ≤ Pr [ H 6∈ A ] ≤ Pr [ ∀i G 6∈ A ] = 1 − Pr [ G ∈ A ] . 2 i n,p ⇒ Pr [ Gn,p ∈ A ] < ε.
Basic Thresholds – p.6 p0 = p(1/2) is a threshold
m Given ε > 0, let (1 − ε) < 1/2. Let p < p0/m. Let G1, . . . , Gm ∈ Gn,p and
H = G1 ∪ · · · ∪ Gm ∼ Gn,1−(1−p)m .
m As 1 − (1 − p) ≤ pm ≤ p0,
1 m ≤ Pr [ H 6∈ A ] ≤ Pr [ ∀i G 6∈ A ] = 1 − Pr [ G ∈ A ] . 2 i n,p ⇒ Pr [ Gn,p ∈ A ] < ε.
Other direction: take Gn,p0 ∪ · · · ∪ Gn,p0 .
Basic Thresholds – p.6 E [ # spanning trees ] → 0 ⇒ G 6∈ C. E [ # spanning trees ] → ∞ 6⇒ G ∈ C.
Connectivity Property C
Idea 1: Connectivity = ∃ spanning tree
E [ # spanning trees ] = nn−2 · pn−1.
The “window” is 1 p = (1 + o(1)) . n
Basic Thresholds – p.7 6⇒ G ∈ C.
Connectivity Property C
Idea 1: Connectivity = ∃ spanning tree
E [ # spanning trees ] = nn−2 · pn−1.
The “window” is 1 p = (1 + o(1)) . n
E [ # spanning trees ] → 0 ⇒ G 6∈ C. E [ # spanning trees ] → ∞
Basic Thresholds – p.7 Connectivity Property C
Idea 1: Connectivity = ∃ spanning tree
E [ # spanning trees ] = nn−2 · pn−1.
The “window” is 1 p = (1 + o(1)) . n
E [ # spanning trees ] → 0 ⇒ G 6∈ C. E [ # spanning trees ] → ∞ 6⇒ G ∈ C.
Basic Thresholds – p.7 Cuts or Isolated Components?
Ck = # k-components
Observation: G ∈ C iff Ck = 0 ∀k ∈ [1, n/2].
Idea 2: Look at Cuts
Cut:
Basic Thresholds – p.8 Ck = # k-components
Observation: G ∈ C iff Ck = 0 ∀k ∈ [1, n/2].
Idea 2: Look at Cuts
Cut:
Cuts or Isolated Components?
Basic Thresholds – p.8 Idea 2: Look at Cuts
Cut:
Cuts or Isolated Components?
Ck = # k-components
Observation: G ∈ C iff Ck = 0 ∀k ∈ [1, n/2].
Basic Thresholds – p.8 Connectivity Threshold
Theorem (Erdos-Ren˝ yi’60) Let
log n c p = + . n n
−c e−e , |c| = O(1), Then Pr [ Gn,p ∈ C ] → 0, c → −∞, 1, c → +∞.
logn In particular, p0(n) = n is a threshold for connectivity.
Basic Thresholds – p.9 n = E [ C ] ≤ (1 − p)k(n−k)kk−2pk−1 k k X n X ≤ (en/k)k & (1 − x) ≤ e−x k h k i ≤ n O(log n) e−np+kp → 0. X Thus whp C2 = · · · = Cbn/2c = 0.
log n O(1) p = n + n
bn/2c Let := k=2 . P P Pr Ck ≥ 1 ≤ E Ck h X i h X i
Basic Thresholds – p.10 n ≤ (en/k)k & (1 − x) ≤ e−x k h k i ≤ n O(log n) e−np+kp → 0. X Thus whp C2 = · · · = Cbn/2c = 0.
log n O(1) p = n + n
bn/2c Let := k=2 . P P Pr Ck ≥ 1 ≤ E Ck h X i hnX i = E [ C ] ≤ (1 − p)k(n−k)kk−2pk−1 k k X X
Basic Thresholds – p.10 k ≤ n O(log n) e−np+kp → 0. X Thus whp C2 = · · · = Cbn/2c = 0.
log n O(1) p = n + n
bn/2c Let := k=2 . P P Pr Ck ≥ 1 ≤ E Ck h X i hnX i = E [ C ] ≤ (1 − p)k(n−k)kk−2pk−1 k k X n X ≤ (en/k)k & (1 − x) ≤ e−x k h i
Basic Thresholds – p.10 Thus whp C2 = · · · = Cbn/2c = 0.
log n O(1) p = n + n
bn/2c Let := k=2 . P P Pr Ck ≥ 1 ≤ E Ck h X i hnX i = E [ C ] ≤ (1 − p)k(n−k)kk−2pk−1 k k X n X ≤ (en/k)k & (1 − x) ≤ e−x k h k i ≤ n O(log n) e−np+kp → 0. X
Basic Thresholds – p.10 log n O(1) p = n + n
bn/2c Let := k=2 . P P Pr Ck ≥ 1 ≤ E Ck h X i hnX i = E [ C ] ≤ (1 − p)k(n−k)kk−2pk−1 k k X n X ≤ (en/k)k & (1 − x) ≤ e−x k h k i ≤ n O(log n) e−np+kp → 0. X Thus whp C2 = · · · = Cbn/2c = 0.
Basic Thresholds – p.10 − −e c It is enough to prove Pr [ C1 = 0 ] → e because
0 ≤ Pr [ C1 = 0 ] − Pr [ C ∈ C ] ≤ Pr [ ∃i ∈ [2, n/2] Ci > 0 ] → 0.
log n O(1) p = n + n (cont.)
Thus, whp Gn,p ∈ C iff C1 = 0 (i.e. no isolated vertices).
Basic Thresholds – p.11 log n O(1) p = n + n (cont.)
Thus, whp Gn,p ∈ C iff C1 = 0 (i.e. no isolated vertices).
− −e c It is enough to prove Pr [ C1 = 0 ] → e because
0 ≤ Pr [ C1 = 0 ] − Pr [ C ∈ C ] ≤ Pr [ ∃i ∈ [2, n/2] Ci > 0 ] → 0.
Basic Thresholds – p.11 µ E [ # successes ] = n × = µ n n µi e−µ Pr [ i successes ] = pi (1 − p)n−i → . i i!
Poisson Distribution with Mean µ
µ n independent trials, Pr [ success ] = n , constant µ. Poisson(µ) = # successes as n → ∞.
Basic Thresholds – p.12 n µi e−µ Pr [ i successes ] = pi (1 − p)n−i → . i i!
Poisson Distribution with Mean µ
µ n independent trials, Pr [ success ] = n , constant µ. Poisson(µ) = # successes as n → ∞. µ E [ # successes ] = n × = µ n
Basic Thresholds – p.12 Poisson Distribution with Mean µ
µ n independent trials, Pr [ success ] = n , constant µ. Poisson(µ) = # successes as n → ∞. µ E [ # successes ] = n × = µ n
n µi e−µ Pr [ i successes ] = pi (1 − p)n−i → . i i!
Basic Thresholds – p.12 The k-th Factorial Moment:
Mk[X] = E [ (X)k ] = E [ X(X − 1) . . . (X − k + 1) ] .
For fixed k k(n−1)−(k) −c k −c Mk[C1] = (n)k (1 − p) 2 → (e ) = Mk[Poisson(e )].
−c This is known to imply that C1 → Poisson(e ). In particular,
− −e c Pr [ C1 = 0 ] → e .
Isolated Vertices
n−1 −c E [ C1 ] = n(1 − p) → e .
Basic Thresholds – p.13 For fixed k k(n−1)−(k) −c k −c Mk[C1] = (n)k (1 − p) 2 → (e ) = Mk[Poisson(e )].
−c This is known to imply that C1 → Poisson(e ). In particular,
− −e c Pr [ C1 = 0 ] → e .
Isolated Vertices
n−1 −c E [ C1 ] = n(1 − p) → e .
The k-th Factorial Moment:
Mk[X] = E [ (X)k ] = E [ X(X − 1) . . . (X − k + 1) ] .
Basic Thresholds – p.13 −c This is known to imply that C1 → Poisson(e ). In particular,
− −e c Pr [ C1 = 0 ] → e .
Isolated Vertices
n−1 −c E [ C1 ] = n(1 − p) → e .
The k-th Factorial Moment:
Mk[X] = E [ (X)k ] = E [ X(X − 1) . . . (X − k + 1) ] .
For fixed k
k(n−1)−(k) −c k −c Mk[C1] = (n)k (1 − p) 2 → (e ) = Mk[Poisson(e )].
Basic Thresholds – p.13 Isolated Vertices
n−1 −c E [ C1 ] = n(1 − p) → e .
The k-th Factorial Moment:
Mk[X] = E [ (X)k ] = E [ X(X − 1) . . . (X − k + 1) ] .
For fixed k
k(n−1)−(k) −c k −c Mk[C1] = (n)k (1 − p) 2 → (e ) = Mk[Poisson(e )].
−c This is known to imply that C1 → Poisson(e ). In particular,
− −e c Pr [ C1 = 0 ] → e .
Basic Thresholds – p.13 Examples, connectivity: sharp, having a triangle: not sharp, having a cycle: ‘one-sided sharp’.
Sharp Threshold
Connectivity Threshold
p0 is a sharp threshold for a monotone A if ∀ ε > 0 whp
Gn,(1−ε)p0 6∈ A and Gn,(1+ε)p0 ∈ A.
Basic Thresholds – p.14 Sharp Threshold
Connectivity Threshold p0 is a sharp threshold for a monotone A if ∀ ε > 0 whp
Gn,(1−ε)p0 6∈ A and Gn,(1+ε)p0 ∈ A. Examples, connectivity: sharp, having a triangle: not sharp, having a cycle: ‘one-sided sharp’.
Basic Thresholds – p.14 Theorem (Friedgut’99) If for a monotone A
∂ 1 Pr [ G ∈ A ] = O , ∂p n,p p then ∀ ε > 0 there is a finite family F of graphs such that ∀ n, p Pr [ Gn,p ∈ A 4 {an F-subgraph} ] ≤ ε.
Friedgut's Theorem
∂ 1 Note: Sharp threshold ⇒ ∂pPr [ Gn,p ∈ A ] =6 O p .
Basic Thresholds – p.15 Friedgut's Theorem
∂ 1 Note: Sharp threshold ⇒ ∂pPr [ Gn,p ∈ A ] =6 O p . Theorem (Friedgut’99) If for a monotone A
∂ 1 Pr [ G ∈ A ] = O , ∂p n,p p then ∀ ε > 0 there is a finite family F of graphs such that ∀ n, p Pr [ Gn,p ∈ A 4 {an F-subgraph} ] ≤ ε.
Basic Thresholds – p.15 Theorem (Achlioptas-Friedgut’99) For fixed k ≥ 3 k-colorability has a sharp threshold.
Applying Friedgut's Theorem
Difficult to apply: the type of A ∪ B depends on which one appears ‘earlier’.
Basic Thresholds – p.16 Applying Friedgut's Theorem
Difficult to apply: the type of A ∪ B depends on which one appears ‘earlier’.
Theorem (Achlioptas-Friedgut’99) For fixed k ≥ 3 k-colorability has a sharp threshold.
Basic Thresholds – p.16 Gn,p: edge probability p.
Theorem For a monotone A,
Pr Gn,M ∈ A is a non-decreasing function of M.
Model Gn,M
Gn,M : random M edges.
Basic Thresholds – p.17 Theorem For a monotone A,
Pr Gn,M ∈ A is a non-decreasing function of M.
Model Gn,M
Gn,M : random M edges.
Gn,p: edge probability p.
Basic Thresholds – p.17 Model Gn,M
Gn,M : random M edges.
Gn,p: edge probability p.
Theorem For a monotone A,
Pr Gn,M ∈ A is a non-decreasing function of M.
Basic Thresholds – p.17 Theorem For any non-trivial monotone A, there is a threshold M0, that is,
0, if M/M0 → 0, Pr Gn,M ∈ A → ( 1, if M/M0 → ∞. log n+c n Theorem Let M = n 2 . Then