Spatial Random Processes Introduction Literature - R. van der Hofstad: Random graphs and complex nerworks. www.win.tue.nl/~hofstad/NotesRGCN2009.pdf B.Bollobas: Random graphs For probability background – R.Durrett, probability theory and examples (2005) Main subject We will look at the Erdos-Renyi random graph (1960) ๐ Consider n vertices, let each one of the ( ) possible edges appear independently with 2 possibility ๐ ∈ [0,1] Such a graph is referred to as ๐บ(๐, ๐). Question: What is the size of the largest component? Denote it as ๐ถ๐๐๐ฅ of ๐บ(๐, ๐). ๐ ๐ We define: ๐ = , ๐ > 0 Exercise: Prove that ๐ธ[๐๐๐๐๐๐(๐ฃ)] = ๐−1 ๐ ๐ We will see that: ๐→∞ If ๐ < 1, than ∃๐ถ๐ > 0 ๐ . ๐ก. ∀๐ > 0. ๐[(๐ถ๐ − ๐) log ๐ ≤ |๐ถ๐๐๐ฅ | ≤ (๐ถ๐ + ๐) log ๐] → ๐→∞ If ๐ > 1, then ∃๐ถ′๐ > 0 ๐ . ๐ก. ∀๐ > 0 ๐[(๐ ′๐ − ๐)๐] ≤ |๐ถ๐๐๐ฅ | ≤ (๐ถ ′๐ + ๐)๐] → 1 1 Fix vertex v. Do breadth first exploration of component ๐ถ(๐ฃ) of ๐บ(๐, ๐) containing v vertices. Can be either N (neutral), A(Active) or I(Inactive). At time ๐ก = 0, ๐ฃ ∈ ๐ด and all other vertices are in N. At time ๐ก ≥ 1, if ๐ด = ๐, terminate. Otherwise, choose ๐ค ∈ ๐ด. Put all N neighbors of ๐ in ๐ด. Put ๐ in ๐ผ. Note: Algorithm terminates after |๐ถ(๐ฃ)| steps. Suppose at time ๐ก, |๐| = ๐ − ๐ vertices are left and |๐ด| ≥ 1. The number of vertices that go from N to A in stem ๐ก has the distribution: ๐ ๐ ๐ต๐๐(๐ − ๐, ๐) = ๐ต๐๐ (๐ − ๐, ) ≈ ๐ต๐๐(๐, ) ๐ ๐ ๐ Recall that ๐~๐ต(๐, ๐), ๐ ∈ [0,1] ↔ ๐[๐ = ๐] = ( ) ๐๐ (1 − ๐)๐−๐ , 0 ≤ ๐ ≤ ๐ ๐ Expect similarity with a branching process! Branching Process t=0 t=1 ๐ ๐ต๐๐(๐, ) ๐ ๐ ๐ต๐๐(๐, ) ๐ t=2 ๐๐๐(๐) ๐~๐๐๐(๐), ๐ ∈ (0, ∞)๐๐ ๐[๐ = ๐] = ๐๐ ----−๐ ๐ , ๐∈โค ๐! ๐→∞ ๐๐ ๐ −๐ , ๐! ๐ Exercise: If ๐~๐ต๐๐(๐, ๐), then ๐[๐ = ๐] → ๐∈โค We will see that: If ๐ < 1, then branching process (๐๐๐ (๐)) dies out almost surely. If ๐ > 1, then it lives forever with probability >0. Further Questions 2 - ๐ = 1 → |๐ถ๐๐๐ฅ | ≈ ๐3 - If ๐ = 1 + ๐ฟ๐ , ๐ฟ๐ → 0 Other properties of ๐บ(๐, ๐) Add geometry to the graph Fix a large graph ๐บ๐ Retain (delete) every edge independently with probability ๐. If ๐บ๐ = ๐๐ (the full graph), we get ๐บ(๐, 1 − ๐) Consider a random walk on ๐บ๐ , study the components ๐บ๐ \ {๐๐ก1 , … , ๐๐ก๐ }, ๐ก๐ ≥ 0 Study random graphs s.t. ๐[๐๐๐๐๐๐(๐ฃ) ≥ ๐] ≈ ๐ −๐ , ๐ > 0 - ๐→∞ Branching Processes Time ๐ = 0, 1, … ๐1,1 t=0 ๐2,1 ๐2,๐1,1 t=1 t=2 In generation ๐ ≥ 0, individual ๐ gives birth to ๐๐+1,๐ children, then dies. Formally, we define {๐๐,๐ |๐ ≥ 1, ๐ ≥ 1} as independent, identically distributed (iid) random variables with values in โค > 0 Define ๐๐ , (๐ ≥ 0) recursively by ๐0 = 1 โฎ ๐๐=∑๐๐ −1 ๐ ๐=1 ๐,๐ ๐0 = to the total number of individuals in generation 1. Notation ๐ = ๐1,1 (since all distribute identically) ๐๐ = ๐[๐ = ๐] =probability that individual has ๐ children. ๐ = ๐ [โ{๐๐ = 0}] = ๐[๐๐๐๐๐๐ ๐ ๐๐๐๐ ๐๐ข๐ก] ๐≥1 For ๐ ∈ [0,1] ๐บ๐ฅ (๐ ) = ๐ธ[๐ ๐ฅ ] “offspring generation function” Theorem 1.1: - If ๐ธ[๐ฅ] < 1, ๐กโ๐๐ ๐ = 1 - If ๐ธ[๐ฅ] > 1, ๐กโ๐๐ ๐ < 1 - If ๐ธ[๐ฅ] = 1, ๐๐๐ ๐[๐ = 1] < 1, ๐กโ๐๐ ๐ = 1 Moreover, ๐ is the smallest solution in [0,1] of ๐ = ๐บ๐ฅ (๐) Exercise 1.1: Prove ๐ธ[๐๐ ] = ๐๐ , where ๐ = ๐ธ[๐] Exercise 1.2: For ๐ < 1, prove ๐ธ[๐] = --------End of lesson 1 1 1−๐ where ๐ = ∑∞ ๐=0 ๐๐ Reminder - {๐ฅ๐,๐ }๐,๐≥1 ๐๐๐, โค ≥ 0 − ๐ฃ๐๐๐ข๐๐ ๐๐−1 - ๐0 = 1, ๐๐ = ∑๐=1 ๐๐,๐ - ๐ = ๐1,1 - ๐๐ = ๐[๐ = ๐], ๐ = ๐๐ โค ≥ 0 - ๐ = ๐[โ๐≥0{๐๐ = 0}] - ๐บ๐ฅ (๐ ) = ๐ธ[๐ ๐ฅ ], ๐ ∈ [0,1] Theorem 1.1 (survival vs. extinction): if ๐ธ[๐] < 1 then ๐ = 1, If ๐ธ[๐] > 1 then ๐ < 1, If ๐ธ[๐] = 1 and ๐[๐ = 1] < 1 then ๐ = 1 ๐ is the smallest solution in [0,1] of ๐ = ๐บ๐ฅ (๐) Example: ๐~๐๐๐(๐), ๐ > 0 ∞ ๐ ๐บ๐ฅ (๐ ) = ∑∞ ๐=0 ๐ ๐๐ = ∑๐=0 ๐ ๐ ๐๐ −๐ ๐ ๐! = ๐ ๐(๐ −1) Proof: First prove that ๐ = ๐บ๐ฅ (๐) Set ๐๐ = ๐ [ ๐๐ = 0 ] โ ๐๐๐๐๐๐๐ ๐๐๐ ๐๐ ๐ By ๐ additivity, ๐๐ ↑ ๐ as ๐ → ∞. Let ๐บ๐ (๐ ) = ๐ธ[๐ ๐๐ ], ๐ ∈ [0,1] (↑ means that ๐๐ ≤ ๐๐+1 , ๐๐ → ๐) Note that is we set ๐บ๐ (0) = ๐๐ ๐ ๐บ๐ (0) = ∑∞ ๐=1 0 ๐[๐๐ = ๐] = ๐๐ + 0 + 0 + โฏ + 0 = ๐๐ . ๐๐ ๐บ๐ (๐ ) = ∑∞ ๐=0 ๐ธ[๐ 1{๐๐ =๐} ] , where 1๐ด (๐ค) = { TODO: Draw the full drawing ๐ก=1 ๐ก=๐ 1 ๐๐ ๐ค ∈ ๐ด 0 ๐๐ ๐ค ∈ Ω\๐ด (๐1 ) (1) ๐๐ = ๐๐ + โฏ + ๐๐ (๐) where ๐๐ Is the number of generation n-individuals descending from (๐) (๐) generation 1-individual ๐ and ๐๐ = ๐๐−1 (๐) (๐๐ ) ๐≥0 ๐ ๐ ๐๐−1 ] ๐๐−1 ] are independent of ๐1 hence ๐บ๐ (๐ ) = ∑∞ ๐๐ = ∑∞ ๐๐ = ๐=0 ๐ธ[๐ ๐=0 ๐ธ[๐ ๐ ∑∞ ๐=0 ๐บ๐−1 (๐ ) ๐๐ = ๐บ๐ฅ (๐บ๐−1 (๐ )). = ๐บ1 (๐บ๐−1 (๐ )) ๐ = 0: ๐๐ = ๐บ1 (๐๐−1 ), ๐ → ∞ ๐ → ๐บ1 (๐) By the dominated convergence theorem. ๐บ1 (๐๐−1 ) = ๐ธ[(๐๐−1 )๐1 ] ๐→∞ (๐๐−1 ) ๐ง1 → ๐ ๐1 Let ๐ ∈ [0,1], ๐ = ๐บ๐ฅ (๐) Now we need to show that ๐ ≤ ๐, or equivalently, ๐๐ ≤ ๐ ∀๐. Indeed: ๐0 = 0 ≤ ๐ ∈ [0,1], if ๐๐−1 ≤ ๐, then ๐๐ = ๐บ๐ฅ (๐๐−1 ) ≤ ๐บ๐ฅ (๐) ๐๐ฆ ๐๐ ๐ ๐ข๐๐๐ก๐๐๐ = ๐. Case ๐[๐ ∈ {0,1}] = 1 ๐[๐ = 0] = ๐ > 0 ๐→∞ ๐๐ = 1 − ๐[๐๐ > 0] = 1 − (1 − ๐)๐ → 1. Now assume ๐[๐ฅ ≤ 1] < 1. ๐บ๐ฅ′ (๐ ) = ๐ธ[๐๐ ๐−1 ] > 0 ๐บ๐ฅ′′ (๐ ) = ๐ธ[๐(๐ − 1)๐ ๐−2 ] > 0, ๐ ∈ (0,1). (Ex: Prove this by using the mean value theorem and the dominant convergence theorem.) ๐บ๐ฅ (1) = 1 TODO: Draw the graph of the function ๐บ๐ฅ (๐ ) Either ๐ = ๐บ๐ฅ (๐ ) has one or two solutions in [0,1]. One solution ↔ ๐บ๐ฅ′ (1) ≤ 1 ๐บ๐ฅ′ (1) = ๐ธ[๐] Exercise 1.4: Compute ๐ If ๐0 = 1 − ๐, ๐2 = ๐, ๐ ∈ [0,1] “Binary branching”. Exercise 1.5: Prove ๐[๐๐ > 0] ≤ ๐๐ , ๐ = ๐ธ[๐]. Hint: use the Chebishev inequality and the previous exercise. Dominated Convergence Theorem ๐→∞ Let (๐๐ )๐≥1 be a sequence of random variables, ๐๐ → ๐ a.s. Assume that ∃ a random variable ๐ ≥ 0 almost surely ๐ธ[๐] < ∞ such that |๐๐ | ≤ ๐ ∀๐ almost ๐→∞ surely, Then ๐ธ[๐๐ ] → ๐ธ[๐]. Exercise 1.6: Use the Dominated Convergence Theorem to fill in the details of the proof of theorem 1.1. Change of Notation (random walk perspective) TODO: Draw the drawing ๐ก=0 ๐ก =n Redenote ๐๐,๐ as they appear in breadth-first search exploration of the tree. ๐. ๐. (๐1,1 , ๐2,1 , … , ๐2,2 , ๐3,1 , … ) =: (๐1 , ๐2 , ๐3 , … ) Define the random variables ๐๐ , ๐ผ๐ , ๐ ≥ 0 such that ๐๐ = ๐(๐๐ ,๐ผ๐) Observation: (๐๐ , ๐ผ๐ ) depend only on (๐1 , … , ๐๐−1 ) Lemma 1.7: (๐๐ )๐≥1 are iid with the same distribution as ๐. Proof: Let ๐1 , … , ๐๐ ∈ โค ≥ 0 ๐[๐1 = ๐ฅ1, … , ๐๐ = ๐ฅ๐ ] = ∑ ∑ ๐ [(๐1 = ๐ฅ1 , … , ๐๐−1 = ๐ฅ๐−1 , ๐๐ = ๐, ๐ผ๐ = ๐ ) ๐๐,๐ = ๐ฅ๐ ] ๐≥1 ๐≥1 ๐−1 ∞ ๐ด ๐−1 ๐ด depends only on ๐ ({๐๐,๐ }๐=1,๐=1 {๐๐,๐ }๐=1 ) ∑๐ ∑๐ ๐[๐ด]๐[๐ = ๐ฅ๐ ] = ๐[๐1 = ๐ฅ1 , … , ๐๐−1 = ๐ฅ๐−1 ]๐[๐ = ๐ฅ๐ ] ๐[๐ = ๐ฅ๐ ] ๐๐๐๐ข๐๐ก๐๐๐ = ๐[๐ = ๐ฅ1 ] โ … โ Independence of ๐๐,๐ Notation (ctd): (๐๐ )๐≥0 equals to the number of active individuals at stage ๐. ๐0 = 1 ๐๐ = ๐๐−1 + (๐๐ − 1), ๐ ≥ 1 = (๐1 , … , ๐๐ ) − (๐ − 1) = number of active individuals. ๐ = (∑๐≥0 ๐๐ ) = inf{๐ ≥ 0 โถ ๐๐ = 0} ๐0 = 1 ๐1 = 1 + (2 − 1) ๐2 = 2 + (2 − 1) ๐3 = 3 + (0 − 1) ๐4 = 2 + (0 − 1) = 1 ๐ 5 = 1 + (0 − 1) = 0 ------- end of lesson 2 In the proof of theorem 1.1 we showed: ๐บ๐ (๐ ) = ๐ธ[๐๐๐ ] = ๐บ๐ (๐บ๐−1 (๐ )), ๐ ∈ [0,1] Exercise 1.7: Set ๐บ๐ (๐ ) = ๐ธ[๐ ๐ ], ๐ ∈ [0,1] Prove that ๐บ๐ (๐ ) = ๐ โ ๐บ๐ (๐บ๐ (๐ )), where ๐บ๐ (๐ ) = ๐ธ[๐ ๐ ] Exercise 1.8: Prove the following : If ๐ = ๐ธ[๐] = 1, then ๐ธ[๐] = ∞ (hint: Prove first that ๐ธ[๐] = 1 + ๐ธ[๐]) Exercize 1.9: Recall ๐ธ[๐๐ ] = ๐๐ - ๐ Prove that (๐๐ = ๐๐๐ ) ๐≥0 is a martingale (w.r.t. natural filtration). Show that: lim ๐๐ = ๐∞ exists and 0 ≤ ๐๐ ≤ ∞ almost surely. ๐→∞ - For ๐ > 1and ๐ธ[๐ 2 ] < ∞ show that (๐๐ )๐≥0 is bounded in ๐ฟ2 . Use theorem 1.1 and onvergence theorem (๐ฟ2 ) to prove that ๐[๐∞ = 0] = ๐ = ๐[๐ < ∞]. Deduce that {๐∞ = 0} = {๐ < ∞} almost surely. (⇒ ๐๐ ~(๐∞ โ ๐๐ )๐๐ {๐=∞} ) Theorem 1.10: (Extinction with large total progeny) ๐ −๐ผ๐ If ๐ = ๐ธ[๐] > 1, then ๐[๐ ≤ ๐ < ∞] ≤ 1−๐ ๐ผ where ๐ผ = sup(๐ก − log ๐ธ[๐ ๐ก๐ ]) > 0 ๐ Remark: In ๐บ (๐, ๐) , ๐ > 1, we will see that ∃๐ > 0 ๐ . ๐ก. there are no components of size 1 between ๐ log ๐ and ๐ โ ๐, with high probability as ๐ → ∞. Lemma 1.11: (Cramer/Chernoff bound): Let {๐๐ }∞ ๐=1 be iid. 1. For any ๐ ∈ (๐ธ[๐1 ], ∞), ๐[∑๐๐=1 ๐๐ ≥ ๐๐] ≤ ๐ −๐๐ผ(๐) 2. For any ๐ ∈ (−∞, ๐ธ[๐1 ]), ๐[∑๐๐=1 ๐๐ ≤ ๐๐] ≤ ๐ −๐๐ผ(๐) Where ๐ผ(๐) = sup(๐ก๐ − log ๐ธ[๐ ๐ก๐1 ]) in (1) t≥0 and ๐ผ(๐) = sup(๐ก๐ − log ๐ธ[๐ ๐ก๐1 ]) in (2) t≤0 Proof of Lemma 1.11: ๐ ๐ ๐ ๐ [∑ ๐๐ ≥ ๐๐] = ๐ [exp (−๐ก๐๐ + ๐ก ∑ ๐๐ ) ≥ 1] = ๐ธ[1๐ด ] ≤ ๐ธ 1๐ด โ exp (−๐ก๐๐ + ๐ก ∑ ๐๐ ) โ ๐=1 ๐=1 ๐=1 ∀๐ก≥0 [ ] ≥1 ๐ ๐ ≤ ๐ธ exp (−๐ก๐๐ + ๐ก ∑ ๐๐ ) = ๐ โ ๐=1 [ ] ≥0 ๐ก๐๐ ]) = exp(−๐(๐ก๐ − log ๐ธ[๐ ) Now optimize over ๐ก ≥ 0 ⇒(1). −๐ก๐๐ ๐ธ [∏ ๐ ๐ ๐ก๐๐ ]=๐ −๐ก๐๐ ๐=1 ∏ ๐ธ[๐ ๐ก๐๐ ] ๐=1 (2) is left as an exercise! Proof of theorem 1.10: ∞ ∞ ∞ ๐[๐ ≤ ๐ < ∞] = ∑ ๐ [ ๐ โ= ๐ ] ≤ ∑ ๐[๐1 + โฏ + ๐๐ = ๐ − 1] = ∑ ๐[๐1 + โฏ + ๐๐ ≤ ๐] ๐=๐ ⊆{๐๐ =0} ๐=๐ ๐=๐ ๐๐ = (๐1 + โฏ + ๐๐ ) − (๐ − 1) ๐๐๐๐๐ 1.11,(2),๐=1 ๐[๐ ≤ ๐ < ∞] ≤ ∞ ∑ ๐ −๐๐ผ(1) ๐=๐ Define – ๐(๐ก) = ๐ก − log ๐ธ[๐ ๐ก๐ ] , ๐ก ≤ 0 ๐(0) = 0 − log 1 = 0 1 ๐ ′ (๐ก) | = 1 − โ ๐ธ[๐๐ 0๐ ] = 1 − ๐ธ[๐] < 0 0๐ ] ๐ธ[๐ ๐ก=0 Since the derivative is negative, there is some negative value where ๐(๐ก) approaches from above. So ∃๐ก < 0: ๐(๐ก) > 0 ⇒ ๐ผ(1) > 0 Duality Principle Assume ๐ ∈ (0,1) (๐ = ๐[๐ < ∞]) Condition the branching process to die out. (only consider the probability ๐[๐ด|๐ < ∞] = ๐[๐ด∩{๐<∞}] ๐[๐<∞] (A event) Question: How does this change the branching process? Definition 1.12: (Conjugate Distribution) Let ๐๐ฅ = ๐[๐ = ๐ฅ], ๐ฅ ∈ โค≥0 and assume ๐0 > 0 (⇔ ๐ > 0). Then the conjugate distribution (๐๐ฅ′ )๐ฅ∈โค≥0 associated to ๐ is defined as ๐๐ฅ′ = ๐ ๐ฅ−1 ๐๐ฅ , ๐ฅ ∈ โค≥0 Lemma 1.13: ∑ ๐๐ฅ′ = 1 ๐ฅ≥0 Proof: ∑ ๐๐ฅ′ = ∑ ๐ ๐ฅ−1 ๐๐ฅ = ๐ฅ≥0 ๐ฅ≥0 1 1 1 ∑ ๐ ๐ฅ ๐๐ฅ = ๐บ๐ (๐) = โ ๐ = 1 ๐ ๐ ๐ ๐ฅ≥0 Denote the branching process with offspring distribution ๐′ (๐′ branching process) by ′ {๐๐,๐ }๐≥1,๐≥1 Lemma 1.14 (๐ > 0) ๐[๐ ′ < ∞] = 1 Proof: Let ๐ ∈ [0,1], assume that ๐บ๐ ′ (๐ ) = ๐ . Want to show ⇒ ๐ = 1, use theorem 1.1. ∞ ∞ 1 ๐บ๐ ′ (๐ ) = ∑ ๐ ๐[๐ = ๐] = ∑ ๐ ๐ (๐ ๐−1 ๐๐ ) = ๐บ๐ (๐ ๐) = ๐ ๐ ๐ ′ ๐=0 ๐=0 So ๐ ๐≥๐ ๐บ๐ (๐ ๐) = ๐ ๐ ⇒ โ ------ End of lesson 3 ๐ ๐ ≥ ๐ ⇒ ๐ = 1 Duality Principle Assume that ๐๐[๐ < ∞] ∈ (0,1) Given ๐๐ฅ = ๐[๐ = ๐ฅ], ๐ฅ ≥ 0, ๐0 > 0 Define conjugate distribution ๐′ by ๐๐ฅ′ = ๐ ๐ฅ−1 ๐๐ฅ , ๐ฅ ∈ โค≥0 ′ {๐๐,๐ }๐≥1,๐≥1 ๐′ - A branching process (๐ ′ , ๐๐′ , ๐1′ , ๐2′ … ) then ๐[๐ ′ < ∞] = 1 Theorem 1.15: (duality princinple) Let ๐ = (๐๐ฅ )๐ฅ∈โค≥0 be a distribution on โค≥0 with ๐0 > 0 and let ๐′ be its conjugate distribution. Let (๐๐ )๐≥1 be a branching process. Let (๐๐′ )๐≥1 be a branching process Then ๐[(๐1 , … , ๐ ๐ ′ ) = (๐ฅ1 , … , ๐ฅ๐ก )|๐ < ∞] = ๐[(๐1′ , … , ๐๐′ ) = (๐ฅ1 , … , ๐ฅ๐ก )|๐ < ∞] For ๐ฅ1 , … , ๐ฅ๐ ∈ โค≥0 . In words, ๐-branching process conditioned to die out is distributed as a ๐′ branching process. Proof: Denote ๐ป = (๐1 , … , ๐๐ ), ๐ป ′ = (๐1′ , … , ๐๐′ ′ ), โ = (๐ฅ1 , … , ๐ฅ๐ก ) Recall: ๐๐ = (๐1 + โฏ + ๐๐ ) − (๐ − 1) ๐ = inf{๐ ≥ 1|๐๐ = 0} ๐ ๐ = (๐ฅ1 , … , ๐ฅ๐ ) − (๐ − 1) Assume that ๐๐ > 0 for 1 ≤ ๐ < ๐ก and ๐๐ก = 0. Otherwise (1) otherwise the equation is true trivially (0 = 0) ๐ก ๐ก ๐=1 ๐=1 ๐[๐ป = โ, ๐ < ∞] ๐ต๐ฆ ๐๐๐๐๐ 1.7 1 1 ๐[๐ป = โ|๐ < ∞] = = ∏ ๐๐ฅ๐ = ∏ ๐๐ฅ′ ๐ ๐1−๐ฅ๐ ๐[๐ < ∞] ๐ ๐ ๐ก = 1 ๐ก−∑๐ก ๐ฅ โ ๐ ๐=1 ๐ ∏ ๐๐ฅ′ ๐ ๐ โ ๐=1 ๐[๐ป ′ =โ] ∑๐ก๐=1 ๐ฅ๐ But ๐ก − = ๐ก − (๐ ๐ก + (๐ก − 1)) = ๐ก − ๐ก + 1 = 1 ′ So it equals: ๐[๐ป = โ] Which is what we needed to prove. Exercise 1.16: Let ๐ be ๐๐๐(๐) - distributed, ๐ > 0. Recall from Lemma 1.11 ๐ผ(๐) = sup(๐ก๐ − log ๐ธ[๐ ๐ก๐ฅ ]) , ๐ > 0 (๐ก ≥ 0 if ๐ > ๐) (๐ก ≤ 0 if ๐ < ๐) Show that in both cases (๐ > ๐, ๐ < ๐) ๐ ๐ผ(๐) = sup (๐ก๐ − log ๐ธ[๐ ๐ก๐ฅ ]) = ๐ − ๐ + ๐ log ( ) ๐ t∈Rโ Exercise 1.17 (coupling) Let (๐๐ฅ )๐ฅ∈โค , (๐๐ฅ )๐ฅ∈โค be different probability distributions on โค. We want a random variable (๐, ๐), โค2 valued, such that ๐[๐ = ๐ฅ] = ๐๐ฅ , ๐ฅ ∈ โค≥0 ๐[๐ = ๐ฅ] = ๐๐ฅ , ๐ฅ ∈ โค≥0 ๐[๐ ≠ ๐] as small as possible. Prove that: 1 2 (๐๐ฅ −min{๐๐ฅ ,๐๐ฅ })(๐๐ฅ −min{๐๐ฅ ,๐๐ฅ }) ๐๐๐ (๐,๐) (i) ∑๐ฅ∈โค(๐๐ฅ − min{๐๐ฅ , ๐๐ฅ }) = ∑๐ฅ∈โค(๐๐ฅ − min{๐๐ฅ , ๐๐ฅ }) = ∑๐ฅ∈โค|๐๐ฅ − ๐๐ฅ | = ๐๐๐ (๐, ๐) (ii) ๐๐ฅ,๐ฆ = min{๐๐ฅ , ๐๐ฅ } if ๐ฅ = ๐ฆ, and (iii) otherwise. Is a distribution on โค2 With ๐[(๐, ๐) = (๐ฅ, ๐ฆ)] = ๐๐ฅ,๐ฆ , (๐ฅ, ๐ฆ) ∈ โค2 get ๐[๐ ≠ ๐] = ๐๐๐ (๐, ๐) and this is the best. if ๐[(๐, ๐) = (๐ฅ, ๐ฆ)] = ๐′๐ฅ,๐ฆ , ๐[๐ = ๐ฅ] = ๐๐ฅ , ๐[๐ = ๐ฅ] = ๐๐ฅ ⇒ ๐[๐ ≠ ๐] ≥ ๐๐๐ (๐, ๐) Possible for any random variable (๐ ′ , ๐′) with values in โค2 , with ๐[๐ ′ = ๐ฅ] = ๐๐ฅ , ๐[๐ ′ = ๐ฅ] = ๐๐ฅ ⇒ ๐[๐ ′ ≠ ๐ ′ ] > ๐๐๐ (๐, ๐) Poisson branching processes Notation: Use superscript “*” for Poisson branching processes. (๐๐∗ , ๐๐∗ , ๐ ∗ , … ) Recall: ๐๐ −๐ ๐ , ๐! - ๐[๐ ∗ = ๐] = - ๐ธ[๐ ∗ ] = ๐ ๐บ๐ ∗ (๐ ) = ๐ ๐(๐ −1) , ๐ ≥ 0 ๐ ∈ โค≥0 Definition 2.1: Call 0 < ๐ < 1 < ๐ a conjugate pair if ๐๐ −๐ = ๐๐ −๐ Note: ๐: ๐ฅ → ๐ฅ๐ −๐ฅ is strictly increasing on [0,1] It is strictly decreasing on [1, ∞] ๐(0) = 0, ๐(∞) = 0 TODO: Draw function… So for any ๐ > 1, there is a unique conjugate ๐ ∈ [0,1) and vice-versa. Theorem 2.2: (Poisson duality principle): Let ๐ < 1 < ๐ be a conjugate pair The ๐๐๐(๐) - a branching process conditioned on extinction is distributed as ๐๐๐(๐) - a branching process. Proof: By theorem 1.15, The conditioned process is a branching process with offspring distribution ๐๐ฅ′ = ๐๐๐ฅ−1 ๐๐ฅ −๐ ๐ , ๐ฅ! ๐ฅ > 0 where ๐๐ = extinction probability for ๐๐๐(๐)- branching process. ๐๐ ∈ (0,1) i.e. (๐๐๐ )๐ฅ −๐ 1 ๐๐ฅ′ = ๐ ( ) ๐ฅ! ๐๐ By theorem 1.1: ๐๐ = ๐บ๐ ∗ (๐๐ ) = ๐ ๐(๐๐ −1) So ๐๐ฅ′ = (๐๐๐ )๐ฅ ๐ฅ! ๐ −๐ โ 1 ๐ ๐(๐๐ −1) = (๐๐๐ )๐ฅ ๐ฅ! ๐ −๐๐๐ , ๐ฅ ≥ 0 Hence, ๐′ = ๐๐๐(๐๐๐ ) ? ๐๐๐ ๐ −๐๐๐ = ๐๐ −๐ Have to prove ๐๐๐ = ๐ ๐๐๐ ๐ −๐๐๐ = ๐๐ ๐(๐๐ −1) ๐ −๐๐๐ = ๐๐ −๐ So, ๐๐๐ ≠ ๐ (since ๐ > ๐ (theorem 1.1) ⇒ ๐ธ[๐ ∗ ] > 1 ⇒ ๐๐ > 1) And (๐๐๐ )๐ −(๐๐๐ ) = ๐๐ −๐ Since ๐ฅ๐ −๐ฅ ↑ on [0,1], ↓ on [1, ∞] There is at most 1 ๐ฅ diferent from ๐ such that ๐ฅ๐ −๐ฅ = ๐๐ −๐ = ๐ Hence, ๐๐๐ = ๐ โ Theorem (Cayley’s formula): The number of labeled trees on ๐ vertices is equal to ๐๐−2 TODO: Draw a tree Theorem 3.15: In reference ---- End of lesson 4 ∞ ∗ ๐ = ∑ ๐๐∗ = inf{๐ ≥ 0: (๐1 + โฏ + ๐๐ ) − (๐ − 1) = 0} ๐=1 Theorem 2.3 (Total Progeny of Poisson Branching Process): For ๐๐๐(๐) - branching process. ๐ > 0 (๐๐ )๐−1 −๐๐ ๐[๐ ∗ = ๐] = ๐ , ๐! ๐≥1 Proof: By Induction on ๐. ∗ Check for ๐ = 1: ๐[๐ ∗ = 1] = ๐[๐1,1 = 0] = ๐ −๐ which fully agrees with the formula. ๐ ≥ 2, assume the result holds for all ๐′ < ๐. ๐−1 ∗ ๐[๐ = ๐] = ∑ ๐[๐ ∗ = ๐, ๐1,1 = ๐] ∗ ๐=1 If we observe the children that “sprung” from all children of the first one, the sum of all vertices is their sum + 1 (for the root) equals the sum of all vertices in the graph. ∗ ๐ ∗ is distributed as 1 + ๐ (1) + โฏ + ๐ (๐1,1 ) , where {๐ (2) }๐≥1 are iid with the same distribution as ๐∗ ๐−1 ∗ ๐[๐ = ๐] = ∑ ๐[๐ (1) + โฏ + ๐ (๐) = ๐ − 1, ๐1,1 = ๐] ∗ ๐=1 But these random variables are independent (by our assumption) ๐−1 =∑ ∗ ๐[๐ (1) = ๐1 , … , ๐ (๐) = ๐๐ , ๐1,1 = ๐] = ∑ ๐=1 (๐1 ,…,๐๐ ) ๐๐ ≥1, ∑๐ ๐๐ =๐−1 ๐−1 ∑ ∑ ∗ ๐[๐ (1) = ๐1 ] โ … โ ๐[๐ (๐) = ๐๐ ] โ ๐[๐1,1 = ๐] ๐=1 (๐1 ,…,๐๐ ) ๐๐ ≥1, ∑๐ ๐๐ =๐−1 Now we can use our induction hypothesis! Each of the ๐๐ ’s is smaller than ๐. ๐−1 ∑ ๐ ∑ ๐−1−๐ −๐(๐−1) ๐ ๐ โ ∏( ๐=1 (๐1 ,…,๐๐ ) ๐๐ ≥1, ∑๐ ๐๐ =๐−1 So, so far we have shown that: ๐=1 (๐๐ )๐๐−1 ๐๐ −๐ )โ ๐ ๐๐ ! ๐! ๐−1 ∗ ๐−1 −๐๐ ๐[๐ = ๐] = ๐ ๐ 1 ∑ ๐! ๐=1 โ ๐ ∑ ∏( (๐1 ,…,๐๐ ) ๐=1 ๐๐ ≥1 ∑๐ ๐๐ =๐−1 (๐๐ )๐๐ −1 ) = ๐๐−1 ๐ −๐๐ โ ๐๐ ๐๐ ! =:๐๐ Lemma 2.4: Let ๐ฟ๐ be the number of labeled trees on ๐ vertices. Then, ๐ฟ๐ = ๐๐−2 = (๐ − 1)! ๐๐ ๐๐−2 With the lemma 2.4, we get ๐[๐ ∗ = ๐] = ๐๐−1 ๐ −๐๐ โ (๐−1)! = (๐๐)๐−1 ๐! ๐ −๐๐ โ (theorem 2.3). How do we prove lemma 2.4? We use lemma 2.5. Lemma 2.5 (Cayley’s Formula): ๐ฟ๐ = ๐๐−2 Example: ๐=5 TODO: Draw graph Proof that lemma 2.5 ⇒ Lemma 2.4: For ๐ ≥ 1, ๐1 , … , ๐๐ , ๐๐ ≥ 1, ∑๐ ๐๐ = ๐ − 1, define ๐ก(๐1 ,…,๐๐) =number of labeled trees of ๐ vertices such that ๐ฃ1 has ๐ neighbors, (tree\{๐ฃ1 }) – has components with ๐1 , … , ๐๐ vertices. To choose such a tree: 1. Split {๐ฃ2 , … , ๐ฃ๐ } into sets of size ๐1 , … , ๐๐ 2. Choose ๐ labeled trees of size ๐1 , … , ๐๐ 3. Choose a vertex in each of the ๐ trees, and connect it to ๐ฃ1 ๐ ๐ก(๐1 ,…,๐๐) (๐๐ )๐๐−1 ๐−1 ๐ −2 ๐ −2 =( ) โ (๐1 1 โ … โ ๐๐ ๐ ) โ (๐1 โ … โ ๐๐ ) = (๐ − 1)! ∏ ( ) ๐1 , … , ๐๐ ๐๐ ! ๐=1 Hence, ๐−1 ๐ฟ๐ = ∑ ๐−1 ∑ ๐๐ ≥1 ๐=1 ∑๐ ๐๐ =๐−1 (๐๐๐๐๐๐๐๐๐!) Now we’re done! ⇒ ๐ฟ๐ = (๐ − 1)! โ ๐๐ ๐ก(๐1 ,…,๐๐) = ∑ ๐=1 1 ๐! ∑ ๐๐ ≥1 ∑๐ ๐๐ =๐−1 ๐๐๐๐๐๐๐! ๐ก(๐1 ,…,๐๐) Lemma 2.5 (๐ฟ๐ = ๐๐−2 , ๐ ≥ 1): Definition: A rooted tree is a tree with one distinguished vertex called “the root” and the edges are all oriented, and oriented away from the root. Example: TODO: Draw an oriented tree It’s not possible that a vertex will have 2 incoming edges. So a vertex has 1 incoming degree. For labeled vertices ๐ฃ1 , … , ๐ฃ๐ , ๐ธ = {〈๐ฃ๐ , ๐ฃ๐ 〉, ๐ ≠ ๐} ๐ ๐ = |{(๐1 , … , ๐๐ ) ∈ ๐ธ ๐−1 |({๐ฃ1 , … , ๐ฃ๐ }, {๐1 , … , ๐๐−1 }) − ๐๐ ๐ ๐๐๐๐ก๐๐ ๐ก๐๐๐}| Example: TODO: Draw vertices example To choose such a sequence as (was supposed to be drawn) above: 1. Choose a labeled tree on {๐ฃ1 , … , ๐ฃ๐ } 2. Choose a root 3. Choose order in which to add the ๐ − 1 edges. Hence, ๐ ๐ = ๐ฟ๐ โ ๐ โ (๐ − 1)! = ๐ฟ๐ โ ๐! Alternatively, ๐ ๐ = ๐ ๐′ , where ๐๐′ is the number of rooted trees after you remove the last ๐ edges. ๐ ๐′ = |{(๐1 , … , ๐๐ ) ∈ ๐ธ ๐−1 |({๐ฃ1 , … , ๐ฃ๐ }, {๐1 , … , ๐๐−1−๐ }) − ๐๐ ๐ ๐๐๐๐ก๐๐ ๐๐๐๐๐ ๐ก ๐ค๐๐กโ ๐ + 1 ๐๐๐๐๐๐๐๐๐ก๐ ∀๐ = 0, … , ๐ − 1}| ๐ ๐ ≤ ๐ ๐′ obviously (every sequence in a set of ๐ ๐ can generate a sequence in ๐ ๐′ ) But also ๐ ๐′ ≤ ๐ ๐ , since for ๐ = 1, every forest is a rooted tree. Suppose edges ๐1 , … , ๐๐−1−๐ have been added. Number of choices for ๐๐−๐ = (๐ข, ๐ฃ) โ ๐ โ โ ๐ ๐ข ๐๐๐๐๐ก๐๐๐๐ฆ ๐ฃ ๐๐๐ฆ ๐๐๐๐ก ๐๐กโ๐๐ ๐กโ๐๐ ๐กโ๐ ๐๐๐ ๐๐ ๐ถ(๐ข) Hence ๐−1 ๐๐ = ๐ ๐′ = ∏(๐๐) = ๐๐−1 โ (๐ − 1)! = ๐๐−2 ๐! = ๐ฟ๐ (๐!) ๐=1 --- end of lesson 5 Theorem 2.3: For ๐๐๐(๐) = ๐ต. ๐. ๐[๐ ∗ = ๐] = (๐ > 0), (๐๐ )๐−1 −๐๐ ๐ , ๐! ๐≥1 Proof of Lemma 2.4: (Correction): ๐ ≥ 1, ๐1 , … , ๐๐ , ๐๐ ≥ 1, ∑ ๐๐ = ๐ − 1 ๐ TODO: Draw correction drawing ๐ก(๐1 , … , ๐๐ ) = number of labeled trees on ๐ vertices {๐ฃ1 , … , ๐ฃ๐ }, where ๐ฃ1 belongs to labeled edges 1, … , ๐. And ๐ฃ1 has ๐๐ descendants attached to edge ๐, ๐ ∈ {1, … , ๐} Reminder of what we did last time: ๐−1 ( ) โ๐1 , … , ๐๐ ๐ก(๐1 , … , ๐๐ ) = ๐๐ ๐ค๐๐ฆ๐ ๐ก๐ ๐๐๐๐ก๐๐ก๐๐๐ {๐ฃ2 ,…,๐ฃ๐ } ๐๐๐ก๐ ๐๐๐๐ข๐๐ ๐๐ ๐ ๐๐ง๐ ๐1 ,…,๐๐ ๐1 −2 ๐ −2 = ๐ × … × ๐๐ ๐ โ1 ๐๐ ๐โ๐๐๐๐๐ ๐๐ ๐๐๐๐๐๐๐ ๐ ๐ข๐๐ก๐๐๐๐ × (๐1 × … × ๐๐ ) โ ๐๐ ๐โ๐๐๐๐๐ ๐๐ ๐๐๐๐โ๐๐๐๐ ๐๐ ๐ฃ1 ๐ Remark: (๐ , … , ๐ ) - is the number of ways to partition ๐ objects into ๐ labeled bins of sizes 1 ๐ ๐1 , … , ๐๐ And the number of labeled trees on ๐ vertices equals ๐−1 ∑ ๐=1 1 ๐! ∑ ๐ก(๐1 ,…,๐๐) โ (๐1 ,…,๐๐ ) โ๐๐≥1,∑๐ ๐−1 # ๐๐ ๐ก๐๐๐๐ ๐ค๐๐กโ deg(๐ฃ1 )=๐ Corollary 2.6 (differentiability of the extinction probability): Let ๐(๐) = ๐๐ [๐ ∗ < ∞], ๐ > 1 And let 0 < ๐ < 1 be the conjugate of ๐ (cf. definition 2.1) So ๐๐ −๐ = ๐๐ −๐ Then −๐′ (๐) = ๐(๐)(๐ − ๐) <∞ ๐(1 − ๐) Proof: Assume ๐ > 1 + ๐, ๐ > 0 By theorem 2.3: ∞ ๐(๐) = ∑ ๐=1 (๐๐)๐−1 −๐๐ ๐ ๐! By using the mean-value theorem, and the dominant convergence theorem, we can check that we can differentiate term-by-term. So: ∞ ๐ ′ (๐) =∑ ๐=1 (๐๐)๐−1 −๐๐ (๐ − 1)๐๐−2 ๐๐−1 −๐๐ ๐๐ − ๐ = ๐! ๐! ๐๐ฆ ๐กโ๐๐๐๐๐ 1.15 (๐๐ข๐๐๐๐ก๐ฆ) 1 1 1 1 ๐ธ๐ [๐ 1{๐ ∗ <∞} ] (1 − ) + ๐(๐) = ๐ธ๐ [๐ ∗ |๐ ∗ < ∞] โ ๐(๐) (1 − ) + ๐(๐) = ๐ ๐ ๐ ๐ 1 1 ๐(๐) ๐๐ฆ ๐ธ๐ฅ.1.3(?) ๐(๐) (1 − ๐) ๐(๐) ๐(๐)(๐ − 1) + ๐(๐)(1 − ๐) ∗ ๐ธ๐ [๐ ] โ ๐(๐) (1 − ) + = + = ๐ ๐ 1−๐ ๐ ๐(1 − ๐) ๐(๐)(๐ − ๐) = โ ๐(1 − ๐) ∗ Recall exercise 1.17: ๐, ๐ be different distributions on โค, then there exists a distribution ๐ on โค2 such that if (๐) (๐) (๐) (๐, ๐) = ๐, then ๐ = ๐, ๐ = ๐ and ๐[๐ ≠ ๐] = ๐ ๐๐ (๐, ๐) = 1 − ∑๐ฅ∈โค min{๐๐ฅ , ๐๐ฅ } Exercise 2.7: Let (๐๐ )๐๐=1 be iid ๐ฅ๐ ~๐ต๐(๐), i.e. ๐[๐๐ = 1] = ๐, ๐[๐๐ = 0] = 1 − ๐, ๐ ∈ [0,1], then show that ∑๐๐=1 ๐ฅ๐ ~๐ต๐๐(๐, ๐) Exercise 2.8: Let (๐๐ )๐๐=1 be independent, ๐ฅ๐ ~๐๐๐ (๐๐ ), ๐๐ > 0 Show that ∑๐๐=1 ๐๐ ~๐๐๐(∑๐๐=1 ๐๐ ) Theorem 2.9: (coupling of Binomial and Poisson random variables) For ๐ > 0, there exists a โค2 values random variable (๐, ๐) such that: ๐ - ๐~๐ต๐๐ (๐, ๐) - ๐~๐๐๐(๐) - ๐[๐ ≠ ๐] ≤ ๐2 ๐ Proof: By Ex.1.17, there are iid random variables {(๐ผ๐ , ๐ฝ๐ )}๐๐=1 such that: ๐ - ๐ผ๐ ~๐ต๐ (๐) - ๐ฝ๐ ~๐๐๐ (๐) - ๐[๐ผ๐ = ๐ฝ๐ ] = ∑๐ฅ∈โค≥0 min{๐[๐ผ1 = ๐ฅ], ๐[๐ฝ1 = ๐ฅ]} = ๐ ๐ ๐ ๐ ๐ 1−๐ฅ≤๐ −๐ฅ ๐ ๐ ๐ min {1 − , ๐ −๐ } + min { , ๐ −๐ } = 1 − + ๐ −๐ ๐ ๐ ๐ ๐ ๐ Define: ๐ ๐ (๐, ๐) = ∑(๐ผ๐ , ๐ฝ๐ ) ๐=1 ๐ By exercise 2.7, ๐~๐ต๐๐ (๐, ๐) By exercise 2.8: ๐~๐๐๐(๐) Finally: ๐ ๐ ๐ ๐ ๐ ๐[๐ ≠ ๐] ≤ ∑ ๐[๐ผ๐ ≠ ๐ฝ๐ ] = ๐ โ ( − ๐ −๐ ) ≤ ๐ โ ๐ ๐ ๐ ๐=1 ๐ ๐ Exercise 2.10: Let (๐๐ )๐≥1 , ๐๐ ~๐ต๐๐ (๐, ) , ๐~๐๐๐(๐), ๐ ≥ 0. (๐) ๐→∞ Prove that ๐๐ → ๐, meaning: For any ๐ด ⊆ โค, ๐[๐๐ ∈ ๐ด] → ๐[๐ ∈ ๐ด] Theorem 2.11 (Poisson and Binomial branching processes): ๐ ๐ Let ๐ and ๐ ∗ be the total progenies of a ๐ต๐๐ (๐, ) - branching process ๐ and of a ๐๐๐(๐) branching process ๐ ∗, ๐ > 0. Then for ๐ ≥ 1 ๐[๐ ≥ ๐] = ๐[๐ ∗ ≥ ๐] + ๐(๐, ๐) Where |๐(๐, ๐)| ≤ 2๐2 ๐−1 ∑๐ =1 ๐[๐ ∗ ๐ > ๐ ] ≤ 2๐2 ๐ ๐ Proof: By theorem 2.9, we can define the branching processes: (๐1 , ๐2 , … ) (Bin) (๐1∗ , ๐2∗ , … ) (Poi) (random walk perspective) Such that ๐[๐๐ ≠ ๐๐∗ ] ≤ ๐2 , ๐ ๐≥1 ๐ ≥ 1: {๐ ≥ ๐พ} depends only on {๐1 , … , ๐๐−1 } - A key observation! Now: ๐−1 ๐[๐ ≥ ๐, ๐ < ๐] ≤ ∑ ๐[๐๐ = ๐๐∗ ๐๐๐ 1 ≤ ๐ ≤ ๐ − 1, ๐๐ ≠ ๐๐ ∗ ๐ ≥ ๐ ≥ ๐ ] ∗ ๐ =1 We can deduce from ๐ ≥ ๐ ≥ ๐ that ๐ ∗ ≥ ๐ ! ๐−1 ≤ ∑ ๐[๐๐ ≠ ๐๐ ∗ , ๐ ∗ ≥ ๐ ] ๐ =1 But these two events are independent ๐−1 = ∑ ๐[๐๐ ≠ ๐ =1 Similarly: ๐−1 ๐๐ ∗ ]๐[๐ ∗ ๐2 ≥ ๐ ] ≤ ∑ ๐[๐ ∗ ≥ ๐ ] ๐ ๐ =1 ๐−1 ๐[๐ ≥ ๐, ๐ < ๐] ≤ ∑ ๐[๐๐ = ๐๐∗ , ๐๐๐ ๐ ≤ ๐ − 1, ๐๐ ≠ ๐๐∗ , ๐ ∗ ≥ ๐ ≥ ๐ ] ≤ ∗ ๐ =1 ๐−1 ∑โ ๐[๐๐ ≠ ๐๐ ∗ ] ๐[๐ ∗ ≥ ๐ ] ๐ =1 ≤ ๐2 ๐ We are interested in: |๐[๐ ≥ ๐] − ๐[๐ ∗ ≥ ๐]| ≤ |๐[๐ ≥ ๐, ๐ ∗ ≥ ๐] + ๐[๐ ≥ ๐, ๐ ∗ < ๐] − ๐[๐ ∗ ≥ ๐, ๐ ≥ ๐] − ๐[๐ ∗ ≥ ๐, ๐ < ๐]| ≤ ๐−1 2๐2 ∑ ๐[๐ ∗ ≥ ๐ ] โ ๐ ๐ =1 --- end of lesson 6 Exercise 1.17 ๐, ๐ different distributions on โค (๐๐ฅ − min{๐๐ฅ , ๐๐ฅ })(๐๐ฆ − min{๐๐ฆ , ๐๐ฆ }) ๐๐ ๐ฅ = ๐ฆ ๐๐ฅ,๐ฆ = { ๐ ๐๐ (๐, ๐) min{๐๐ฅ , ๐๐ฅ } ๐๐ ๐ฅ = ๐ฆ (๐ฅ, ๐ฆ) ∈ โค2 1 1 ๐๐๐ (๐, ๐) = ∑|๐๐ฅ − ๐๐ฅ | = ( ∑ (๐๐ฅ − ๐๐ฅ ) + ∑ (๐๐ฅ − ๐๐ฅ )) = 2 2 ๐ฅ∈โค ๐ฅ:๐๐ฅ ≥๐๐ฅ ๐ฅ:๐๐ฅ <๐๐ฅ 1 (∑ ๐๐ฅ − ∑ ๐๐ฅ − ∑ ๐๐ฅ + ∑ ๐๐ฅ − ∑ ๐๐ฅ ) = 2 ๐ฅ∈โค ๐ฅ:๐๐ฅ <๐๐ฅ ๐ฅ:๐๐ฅ ≥๐๐ฅ ๐ฅ:๐๐ฅ <๐๐ฅ ๐ฅ:๐๐ฅ <๐๐ฅ 1 (1 − ∑ ๐๐ฅ − ∑ ๐๐ฅ + ∑ ๐๐ฅ − ∑ ๐๐ฅ − ∑ ๐๐ฅ ) = 2 ๐ฅ:๐๐ฅ <๐๐ฅ ๐ฅ:๐๐ฅ ≥๐๐ฅ ๐ฅ∈โค ๐ฅ:๐๐ฅ ≤๐๐ฅ ๐ฅ:๐๐ฅ <๐๐ฅ 1 (2 − 2 ∑ min{๐๐ฅ − ๐๐ฅ }) = 1 − ∑ min{๐๐ฅ , ๐๐ฅ } 2 ๐ฅ∈โค ๐ฅ∈โค ๐ is a distribution on โค2 : ∑ ∑ ๐๐ฅ,๐ฆ = ∑ ๐๐ง,๐ง + ∑ ∑ ๐๐ฅ,๐ฆ = ∑ min{๐๐ง , ๐๐ง } + ∑(๐๐ฅ − min{๐๐ฅ , ๐๐ฅ }) = 1 ๐ฅ ๐ฆ ๐ง∈โค ๐[๐ = ๐ฅ] = ∑ ๐๐ฅ,๐ฆ ๐ฆ∈โค ๐ฅ∈โค ๐ฆ≠๐ฅ ๐ง∈โค ๐ฅ∈โค (๐, ๐)~๐ (๐๐ฅ − min{๐๐ฅ , ๐๐ฅ })๐๐๐ (๐, ๐) = min{๐๐ฅ , ๐๐ฅ } + = ๐๐ฅ ๐๐๐ (๐, ๐) Similarly, ๐[๐ = ๐ฅ] = ๐๐ฅ ๐[๐ ≠ ๐] = 1 − ๐[๐ = ๐] = 1 − ∑ ๐๐ฅ,๐ฅ = 1 − ∑ min{๐๐ฅ , ๐๐ฅ } = ๐๐๐ (๐, ๐) ๐ฅ∈โค ๐ฅ∈โค Cannot do better! Assume (๐ ′ , ๐ ′ ): ๐ ′ ~๐, ๐ ′ ~๐ ๐[๐ ′ = ๐ ′ ] = ∑ ๐[๐ ′ = ๐ ′ = ๐ฅ] ๐ฅ∈โค But ๐[๐ ′ = ๐ ′ = ๐ฅ] ≤ ๐[๐ ′ = ๐ฅ] = ๐๐ฅ And also ๐[๐ ′ = ๐ ′ = ๐ฅ] ≤ ๐[๐ ′ = ๐ฅ] = ๐๐ฅ And both are smaller or equal to min{๐๐ฅ , ๐๐ฅ } ⇒ ๐[๐ ′ ≠ ๐ ′ ] = 1 − ๐[๐ ′ = ๐ ′ ] ≥ 1 − ∑ min{๐๐ฅ , ๐๐ฅ } = ๐๐๐ (๐, ๐) ๐ฅ∈โค 3. The Eedos-Renyi random graph Definition 3.1: E.R. random graphs, ๐บ(๐, ๐) for ๐ ≥ 1, ๐ ∈ [0,1]. Let (๐ผ{๐,๐} )1≤๐<๐≤๐ be iid random variables with distribution ๐ต๐(๐). Then ๐บ(๐, ๐) = ([๐], ๐ธ) Where [๐] = {1, … , ๐}, ๐ธ = {{๐, ๐} ≤ [๐]: ๐ ≠ ๐, ๐ผ{๐,๐} = 1} Exercise 3.2: Let ๐ โ number of triangles in ๐บ(๐, ๐) ๐ ๐ธ[๐] = ( ) ๐3 3 1 Let ๐ = ๐๐ผ , ๐ผ > 0 Show that: - ๐→∞ For ๐ผ > 1, ๐[๐ > 0] → ๐→∞ For ๐ผ < 1, ๐[๐ > 0] → ๐ผ = 1? What about ๐พ๐ , ๐ ≥ 4? (hint, prove that ๐ฃ๐๐(๐) ๐→∞ → 0 ๐ธ[๐]2 0 (can be proven using a chebishev inequality) 1 (should use a second moment method! – the variant of ๐) and use chebishev). Notation: - For ๐ข, ๐ฃ ∈ [๐], "๐ข ↔ ๐ฃ” means “There is a nearest-neighbor path from ๐ข to ๐ฃ in ๐บ(๐, ๐)” - For ๐ฃ ∈ [๐], ๐ถ(๐ฃ) = {๐ค ∈ [๐]: ๐ฃ ↔ ๐ค} Consider the following exploration algorithm for ๐ถ(1) set ๐ด0 = {1}, ๐1 = [๐]\{1} - For ๐ก > 0, if ๐ด๐ก−1 = ∅, then ๐ด๐ก = ∅, ๐๐ก = ∅ (terminate) If ๐ด๐ก−1 ≠ ∅, let ๐ค๐ก−1 be the smallest vertex in ๐ด๐ก−1 and define: ๐ฃ๐ก = {๐ค ∈ ๐๐ก−1 |{๐ค๐ก−1 , ๐ค} ∈ ๐ธ} ๐ด๐ก โ (๐ด๐ก−1 \{๐ค๐ก−1 }) ∪ ๐๐ก ๐๐ก โ ๐๐ก−1 \๐๐ก Example: Draw graphical example. Note that: - |๐ด0 | = 1, |๐ด๐ก | = |๐ด๐ก−1 | + |๐๐ก | − 1 for ๐ก ≥ 1 - |๐๐ก | = ๐ − |๐ด๐ก | − ๐ก, ๐ก ≥ 0 - |๐ถ(1)| = inf{๐ก ≥ 1 | |๐ด๐ก | = 0} Lemma 3.3: For ๐ฅ1 , … , ๐ฅ๐ ∈ โค≥0 , let: - ๐ 0 = 1, ๐ ๐ก = ๐ ๐ก−1 + ๐ฅ๐ก − 1 for ๐ก ≥ 1 - ๐๐ก = ๐ − ๐ ๐ก − ๐ก If ๐ 1 , ๐ 2 , … , ๐ ๐ ≥ 1, ๐0 , ๐1 , … , ๐๐−1 ≥ 0 Then ๐ ๐[|๐1 | = ๐ฅ1 , … , |๐๐ | = ๐ฅ๐ ] = ∏(๐(๐๐ก−1 , ๐, ๐ฅ๐ก )) ๐ Where ๐(๐, ๐, ๐) = ( ) ๐๐ (1 − ๐)๐−๐ ๐ ๐ก=1 Proof: By picture. TODO: Draw picture proof. ๐[|๐1 | = ๐ฅ1 , … , |๐๐ | = ๐ฅ๐ ] = ∑ ∑ ๐ [|๐1 | = ๐ฅ1 , … , |๐๐−1 | = ๐ฅ๐−1 , ๐ด๐−1 = ๐ด, ๐๐−1 = ๐, ∑ ๐ผ{๐,๐ค} = ๐ฅ๐ ] ๐ด⊆[๐] ๐⊆[๐] ๐ด≠∅ ๐ค∈๐ ๐ is the smallest vertex in ๐ด. {|๐1 } = ๐ฅ1 , … , ๐๐ก−1 = ๐} ∈ ๐(๐ผ๐ : ๐ ๐๐๐ก ๐๐ ๐๐๐๐ ๐๐๐๐ ๐ ๐ก๐ ๐) { ∑ ๐ผ{๐,๐ค} = ๐ฅ๐ } ∈ ๐(๐ผ๐ : ๐ ๐๐๐๐ ๐๐๐๐ {๐} ๐ก๐ ๐) ๐ค∈๐ So these two events must be independent! So: ๐[|๐1 | = ๐ฅ1 , … , |๐๐ | = ๐ฅ๐ ] = ∑ ∑ ๐[|๐1 | = ๐ฅ1 , … , ๐๐−1 = ๐] โ ๐(๐๐−1 , ๐, ๐ฅ๐ ) If the probability of ๐[|๐1 | = ๐ฅ1 , … , ๐๐−1 = ๐] is positive then necessarily |๐| = |๐๐−1 | = ๐๐−1 ⇒ ๐[|๐1 | = ๐ฅ1 , … , |๐๐ | = ๐ฅ๐ ] = ๐[๐1 = ๐ฅ! , … , |๐๐−1 | = ๐ฅ๐−1 ] × ๐(๐๐−1 , ๐, ๐ฅ๐ ) So now we can use induction to complete the proof. --- end of lesson Definition: 3.4: ๐ ๐ก−1 Let (๐ผ๐ก,๐ )๐ก,๐≥1 be iid ๐ต๐(๐). Define ๐0 = ๐ − 1, for ๐ก ≥ 1 : ∑๐=1 ๐ผ๐ก,๐ , ๐๐ก = ๐๐ก − ๐๐ก , ๐๐ก = (๐1 + โฏ + ๐๐ก ) − (๐ก − 1) ๐๐ก ~|๐๐ก |, ๐๐ก ~|๐๐ก |, ๐๐ก ~|๐ด๐ก | Lemma 3.5 For ๐ฅ๐ , ๐ ๐ , ๐๐ as in lemma 3.3, ๐ ∏ ๐(๐๐ก−1 , ๐, ๐ฅ๐ก ) , ๐[๐1 = ๐ฅ1 , … , ๐๐ = ๐ฅ๐ ] = { ๐๐ ๐1 , … , ๐๐ก−1 ≥ 0 ๐ก=0 0, ๐๐กโ๐๐๐ค๐๐ ๐ Proof: {๐1 = ๐ฅ1 , … , ๐๐ = ๐ฅ๐ } ⊆ {๐1 = ๐1 , … , ๐๐−1 = ๐๐−1 } So if ๐๐ < 0, then ๐[… ] = 0. Otherwise: ๐[๐1 = ๐ฅ1 , … , ๐๐−1 = ๐ฅ๐−1 , ๐๐ = ๐ฅ๐ ] = ๐[๐1 = ๐ฅ1 , … , ๐๐−1 = ๐ฅ๐−1 ] โ ๐(๐๐−1 , ๐, ๐ฅ๐ ) ๐๐ฆ ๐๐๐ ๐๐−1 Because ๐๐ = ∑๐=1 ๐ผ๐,๐ We can continue by induction. And that finishes the proof. โ Corollary 3.6: For ๐ ≥ 1, and any ๐: โค๐≥0 → โ ๐ธ[๐(|๐1 |, … , |๐๐ |)1{|๐ด1 |>0,…,|๐ด๐−1 |>0} ] = ๐ธ[๐(๐1 , … , ๐๐ )1{๐1 >0,…,๐๐−1 >0} ] Proof: Combine Lemma 3.3 and Lemma 3.5. โ Theorem 3.7: Let ๐ − and ๐ + be the total progenies of a ๐ต๐๐(๐ − ๐, ๐) and ๐ต๐๐(๐, ๐) branching processes. Where 1 ≤ ๐ ≤ ๐ and ๐ ∈ [0,1] then ๐[๐ − ≥ ๐] ≤ ๐[|๐ถ(1)| ≥ ๐] ≤ ๐[๐ + ≥ ๐]. +/− Proof: For (๐ผ๐ก,๐ ) as in Definition 3.4, define ๐๐ก +/− Then (๐๐ก ) are iid~๐ต๐๐(๐/๐ − ๐, ๐). ๐ก≥1 +/− +/− +/− ๐๐ก : = (๐1 + โฏ + ๐๐ก ) − (๐ก − 1), (๐) +/− Then ๐ +/− = inf {๐ก ≥ 1|๐๐ก = 0} ๐/๐−๐ = ∑๐=1 ๐ก ≥ 1, ๐ผ๐ก,๐ , ๐ก ≥ 1. +/− ๐0 =1 ๐[|๐ถ(1)| ≥ ๐] = ๐[|๐ด1 | > 0, … , |๐ด๐−1 | > 0] = ๐[๐1 > 0, … , ๐๐−1 > 0] ≤ + ๐[๐1+ > 0, … , ๐๐−1 > 0] = ๐[๐ + ≥ ๐] (by cor 3.6 if ๐ ≡ 1) ๐๐ก ≤ ๐๐ก+ , ๐ก ≤ ๐ − 1 On the other hand, ๐[|๐ถ(1)| ≤ ๐] = ๐[โ๐๐ก=1{|๐๐ก | = 0, |๐ด๐ก−1 | = 1, |๐ด๐ก−2 | > 0, … , |๐ด1 | > 0}] (exploration stops at ๐ก) ๐ = ∑ ๐[๐๐ก = 0, ๐๐ก−1 = 1, ๐๐ก−2 > 0, … , ๐1 > 0] ๐ก=1 (by Cor 3.6) = ๐[๐ ≤ ๐] where ๐ = inf{๐ก ≥ 1|๐๐ก = 0} Claim: If {๐ ≤ ๐} ⊆ {๐ − ≤ ๐} Indeed, if ๐ ≤ ๐ ๐๐ = ๐ − 1 − ๐1 − โฏ − ๐๐ = ๐ − ๐ − ๐๐ ≥ ๐ − ๐ So ๐1 ≥ ๐2 ≥ โฏ ≥ ๐๐ ≥ ๐ − ๐ Hence, ๐๐ก− ≤ ๐๐ก , ๐ก ≤ ๐ ๐๐ก− ≤ ๐๐ก , ๐ก ≤ ๐ ⇒ ๐− ≤ ๐ Which proves this claim. โ Exercise 3.8: Consider random variables ๐, ๐ such that ๐~๐ต๐๐(๐, ๐) and conditionally on ๐, ๐~๐ต๐๐(๐, ๐), ๐ ≥ 1, ๐, ๐ ∈ [0,1] ๐ I.e. ๐[๐ = ๐|๐ = ๐] = ( ) ๐ ๐ (1 − ๐)๐−๐ ๐ (0 ≤ ๐ ≤ ๐ ≤ ๐) Prove that ๐~๐ต๐๐(๐, ๐๐) It would be probable to assume: |๐๐ก |~๐ต๐๐(๐ − 1, (1 − ๐)๐ก ) But this is not correct… Proposition 3.9: For 0 ≤ ๐ก ≤ ๐, ๐๐ก ~๐ต๐๐(๐ − 1, (1 − ๐)๐ก ) Proof: By induction. ๐0 = ๐ − 1. This is true. Assume that ๐๐ก ~๐ต๐๐(๐ − 1, (1 − ๐)๐ก ) Need to check that conditionally on ๐๐ก , ๐๐ก+1 ~๐ต๐๐(๐๐ก , 1 − ๐) (By Ex 3.6) 0 ≤ ๐ ≤ ๐ ≤ ๐ − 1 ๐[๐๐ก+1 = ๐. ๐๐ก = ๐], ๐๐ก+1 = ๐๐ก − ๐๐ก+1 = ๐[ ๐ โ ๐ก+1 = ๐ − ๐, ๐๐ก = ๐] ๐๐๐๐๐๐๐๐๐๐๐๐ = ๐(๐, ๐, ๐ − ๐)๐[๐๐ก = ๐] = =∑๐ ๐=1 ๐ผ๐ก+1,๐ ๐(๐, (1 − ๐), ๐)๐[๐๐ก = ๐] โ ๐ ( ) ๐๐−๐ (1 − ๐)๐ ๐−๐ ๐ผ๐ = ๐ − 1 − log ๐ , (see Ex. 1.16, ๐ผ(๐)) ๐>0 ๐ Lemma 3.8: For ๐ = ๐ , ๐ ∈ (0,1) ๐ ≥ 1, ๐[|๐ถ(1)| > ๐] ≤ ๐ −๐ผ๐ ๐ Proof: Let ๐ > 0, ๐ = − log ๐ > 0 ๐[|๐ถ(1)| > ๐] = ๐[|๐ด1 | > 0, … , |๐ด๐ | > 0] ๐ถ๐๐ 3.6 = ≤ ๐[๐1 > 0, … , ๐๐ > 0] ≤ ๐[๐๐ > 0] = ๐ ๐ โ๐ + ๐ − 1 ≥ ๐ ๐๐๐ข๐๐ฃ๐๐๐๐๐ก ๐ก๐: [ ] ๐๐ + ๐ − 1 = ๐1 + โฏ + ๐๐ = ๐ − 1 − ๐๐ ~๐ต๐๐(๐ − 1,1 − (1 − ๐)๐ ) ๐ ๐(๐๐ +๐−1) ≥๐ ๐๐ So ๐[|๐ถ(1)| > ๐] ≤ ๐[๐ ๐(๐๐ +๐−1) ≥ ๐ ๐๐ ] ๐๐ฆ ๐๐๐๐๐๐ฃ ≤ ๐ −๐๐ ๐ธ[๐ ๐(๐๐ +๐−1) ] ๐−1 ๐ธ[๐ ๐(๐๐ +๐−1) ] = ((1 − ๐) + ๐๐ ๐ ) (๐) ๐๐ + ๐ − 1 = ๐1 + โฏ + ๐๐−1 , ๐๐ iid ~๐ต๐(๐) ๐−1 ⇒ ๐ธ[๐ ๐(๐๐ +๐−1) ] = ๐ธ [๐ ๐(∑๐−1 ๐=1 ๐๐ ) ๐−1 ] = ∏(๐ธ[๐ ๐๐๐ ]) = (๐ธ[๐ ๐๐1 ]) ๐−1 = ((1 − ๐) + ๐๐ ๐ ) ๐=1 So ๐[|๐ถ(1)| > ๐] ≤ ๐ −๐๐ ๐ ๐−1 ((1 − ๐) + ๐๐ ) (1−๐ฅ)≤๐ −๐ฅ ≤ ๐ exp(−๐๐ + ๐(๐ − 1)(๐ − 1)) ≤ exp(−๐๐ + ๐(๐ ๐ − 1)๐) ๐ = 1 − (1 − ๐)๐ ≤ ๐๐ (∗) (comment: Seems like ๐ก turned into ๐) ⇒ ๐[|๐ถ(1)| > ๐] ≤ exp(−๐๐ + ๐(๐ ๐ − 1)๐) ≤ exp(−๐๐ + ๐๐(๐ ๐ − 1)๐) = ๐ ๐= ๐ ๐ exp (−๐(๐ − ๐๐(๐ − 1))) = exp (−๐(๐ − ๐(๐ ๐ − 1))) If we try to minimize the expression, we get that ๐ = − log ๐ So let’s just set it in: ๐ −๐๐ผ๐ Why is (*) true? Show that at zero they are equal, and the derivative of the right hand side is always larger then the derivative of the left hand side. --- end of lesson Theorem 3.9 (Upper bound on the largest sub-critical component): ๐ 1 Fix ๐ ∈ (0,1), ๐ = ๐ , ๐ > ๐ผ ๐ Then there is a ๐ฟ = ๐ฟ(๐, ๐) such that ๐[|๐ถ๐๐๐ฅ | > ๐ log ๐] < ๐−๐ฟ Proof: ๐[|๐ถ๐๐๐ฅ | > ๐ log ๐] = ๐ [ โ {|๐ถ(๐ฃ)| > ๐ log ๐}] ≤ ∑ ๐[|๐ถ(๐ฃ)| > ๐ log ๐] = ๐ฃ∈[๐] ๐๐ฆ ๐ฟ๐๐๐๐ 3.8 ๐๐[|๐ถ(1)| > ๐ log ๐] ≤ ๐ฃ∈[๐] ๐ โ ๐−๐ผ๐๐ , ๐ผ๐ ๐ > 1 = โ Next, lower bound: Lemma 3.10: Let ๐≥๐ = ∑๐๐ฃ=1 1{|๐ถ(๐ฃ)|≥๐} ๐≥๐ = ๐ธ[|๐ถ(1)|1{|๐ถ(1)|≥๐} ] Then ๐ฃ๐๐(๐≥๐ ) ≤ ๐๐≥๐ 2 2 Proof: ๐ฃ๐๐(๐≥๐ ) = ๐ธ [(∑๐๐ฃ=1 1{|๐ถ(๐ฃ)|≥๐} ) ] − ๐ธ[∑๐๐ฃ=1 1{|๐ถ(๐ฃ)|≥๐} ] = ๐ ๐ ∑ ∑ ๐[|๐ถ(๐)| ≥ ๐, |๐ถ(๐)| ≥ ๐] − ๐[|๐ถ(๐)| ≥ ๐]๐[|๐ถ(๐)| ≥ ๐] ๐=1 ๐=1 Split according to ๐ ↔ ๐ and ๐ ↔ ๐ ๐ connected to ๐ and ๐ not connected to ๐ ๐ ๐ ๐[|๐ถ(๐)| ≥ ๐, |๐ถ(๐)| ≥ ๐, ๐ ↔ ๐] = ∑ ∑ ๐[|๐ถ(๐)| = ๐, ๐ ↔ ๐, |๐ถ(๐)| = ๐] = ๐=๐ ๐=๐ ๐ ∑ ๐[|๐ถ(๐) = ๐|, ๐ ↔ ๐] โ ๐ [|๐ถ(๐)| = ๐] = ∑ ๐[|๐ถ(๐)| = ๐, ๐ ↔ ๐] โ ๐ [|๐ถ(๐)| ≥ ๐] ≤ (๐−๐) ๐,๐ ๐ (๐−๐) ๐=๐ ๐ ∑ ๐[|๐ถ(๐)| = ๐, ๐ ↔ ๐] โ ๐ [|๐ถ(๐)| ≥ ๐] = ∑ ๐[|๐ถ(๐)| = ๐] โ ๐[|๐ถ(๐)| ≥ ๐] (๐−๐) ๐=๐ ๐=๐ (we can strike these out since it only increases the probability of the event) So let’s go back to the entire variance: The term we just calculated cancels the second term! So we are only left with the case where they are connected: ๐ ๐ ๐ฃ๐๐(๐≥๐ ) ≤ ∑ ∑ ๐[|๐ถ(๐)| ≥ ๐, ๐ถ(๐) ≥ ๐, ๐ ↔ ๐] ๐=1 ๐=1 ๐↔๐⇒ ๐โ๐๐ ๐ ๐ ๐๐๐๐๐๐ฆ ๐ ๐กโ๐ ๐ ๐๐๐ ๐๐ฃ๐๐๐ก = ๐ ∑ ∑ ๐[|๐ถ(๐)| ≥ ๐] = ๐=1 ๐=1 ๐ ๐ ∑ ๐ธ 1{|๐ถ(๐)|≥๐} โ ∑ 1{๐∈๐ถ(๐)} = ๐๐≥๐ ๐=1 ๐=1 โ [ ] =|๐ถ(๐)| Theorem 3.11 (lower bound on largest subcritical component): ๐ Fix ๐ ∈ (0,1), ๐ = ๐ 1 Then for every ๐ < ๐ผ there exists ๐ฟ = ๐ฟ(๐, ๐), and ๐ = ๐(๐, ๐) such that ๐ ๐[|๐ถ๐๐๐ฅ | ≤ ๐ log ๐] ≤ ๐๐−๐ฟ Proof: Let ๐ = ⌊๐ log ๐⌋, ๐ denotes a constant depending only on ๐, ๐ changing from place to place. Claim 1: For any 0 < ๐ผ < 1 − ๐ผ๐ ๐ ๐ธ[๐≥๐ ] ≥ ๐๐ผ , for ๐ ≥ ๐ Claim 2: ๐ฃ๐๐(๐≥๐ ) ≤ ๐๐๐1−๐๐ผ๐ From these claims, it follows that we can prove this theorem. Why? ๐[|๐ถ๐๐๐ฅ | ≤ ๐ log ๐] = ๐[๐≥๐ = 0] ≤ ๐[|๐๐≥0 − ๐ธ[๐≥๐ ]| ≥ ๐ธ[๐≥๐ ]] = ๐๐ฆ ๐กโ๐ ๐[|๐๐≥0 − ๐ธ[๐≥๐ ]|2 ≥ ๐ธ[๐≥๐ ]2 ] ๐ฃ๐๐(๐≥๐ ) ๐๐๐๐๐๐ ๐๐๐1−๐๐ผ๐ ≤ ≤ ๐ธ[๐≥๐ ]2 ๐2๐ผ 2 (For instance, if ๐ผ = 3 (1 − ๐๐ผ๐ ) then it’s ≤ ๐ log ๐ ๐ (1−๐๐ผ๐ ) 3 ) Let’s prove claim 1: ๐ Let ๐, ๐ ∗ be the total progenies of ๐ต๐๐ (๐ − ๐, ๐) branching process, and ๐๐๐(๐๐ ) branching process wher ๐๐ = ๐(๐−๐) ๐ By theorem 3.7, ๐[|๐ถ(1)| ≥ ๐] ≥ ๐[๐ ≥ ๐] ๐๐ฆ ๐กโ๐๐๐๐๐ 2.11 ≥ ๐ 1 ๐(๐ − ๐) = ( ) ๐ ๐−๐โ ๐ ๐๐ ∗ ∗ ๐[๐ ≥ ๐] ≥ ๐[๐ = ๐] ๐๐ฆ ๐กโ๐๐๐๐๐ 2.3 = We can use sterling’s formula: (๐๐ ๐)๐−1 −๐ ๐ ๐ ๐ ๐! ๐[๐ ∗ ≥ ๐] − 2๐2๐ ๐ ๐ ๐ ๐ ๐! = ( ) โ √2๐๐(1 + ๐(1)) ๐ And then: ≥ ๐๐๐ ๐ −๐๐ ๐ ๐ ๐ (1 + ๐(1)) ≥ ๐ −๐ผ๐๐ ๐(1+๐) ๐๐๐ ๐ ๐ข๐๐๐๐๐๐๐๐ก๐๐ฆ ๐๐๐๐๐ ๐ ๐√2๐๐๐๐ (๐ผ๐ = ๐ − 1 − log ๐) = ๐ −๐ผ๐ ๐(1+๐) For ๐ ≥ 0: 1 − ๐๐ผ๐ (1 + ๐) > ๐ผ for ๐ ≥ ๐(๐, ๐, ๐) >๐ผ ∗ ๐ธ[๐≥๐ ] = ๐๐[|๐ถ(1)| ≥ ๐] ≥ ๐๐[๐ ≥ ๐] − ๐๐ ๐ ≥ ๐ (๐ −๐ผ๐ ๐(1+๐) − ๐๐ ) ๐ =๐ โ 1−๐ผ๐ ๐(1+๐) − ๐ log ๐ ≥ ๐๐ผ for ๐ ≥ ๐. (For ๐ chosen sufficiently small) Now let’s prove claim 2: Proof of claim 2: |๐ถ(1)| We can write |๐ถ(1)| = ∑๐=1 1{|1|>๐} = ∑๐๐=1 1{|1|>๐} So: ๐ ๐ ๐≥๐ = ∑ ๐[|๐ถ(1)| ≥ 1, |๐ถ(1)| ≥ ๐] = ๐๐[|๐ถ(1)| ≥ ๐] + ∑ ๐[|๐ถ(1)| ≥ ๐] ๐=1 ๐=๐+1 ∞ ๐๐ −๐ผ๐ (๐−1) + ∑ ๐ −๐ผ๐ (๐−1) ≤ ๐๐๐−๐๐ผ๐ + ๐๐ โ −๐ผ๐ ๐ ≤ ๐๐๐−๐๐ผ๐ ≤๐−๐๐ผ๐ ๐=๐+1 So by lemma 3.10 get claim 2. โ Exercise 3.12 (second moment method): Let ๐ be a random variable such that 0 ≤ ๐ธ[๐] < ∞ and 0 < ๐ธ[๐ 2 ] < ∞ Let 0 ≤ ๐ < 1. ๐ธ[๐ฅ]2 Prove that: ๐[๐ > ๐๐ธ[๐]] ≥ (1 − ๐)2 ๐ธ[๐ฅ 2 ] In our proof, we used it with ๐ = 0. Hint: Prove first that (1 − ๐)๐ธ[๐] ≤ ๐ธ[๐ โ 1{๐>๐๐ธ[๐]} ] ๐๐ฆ ๐๐๐๐๐ 3.8 ≤