Spatial Random Processes

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 โ€“ 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‘ โˆˆ [0,1]
Such a graph is referred to as ๐บ(๐‘›, ๐‘).
Question: What is the size of the largest component? Denote it as ๐ถ๐‘š๐‘Ž๐‘ฅ of ๐บ(๐‘›, ๐‘).
๐œ†
๐‘›
We define: ๐‘ = , ๐œ† > 0
Exercise: Prove that ๐ธ[๐‘‘๐‘’๐‘”๐‘Ÿ๐‘’๐‘’(๐‘ฃ)] =
๐‘›โˆ’1
๐œ†
๐‘›
We will see that:
๐‘›โ†’โˆž
If ๐œ† < 1, than โˆƒ๐ถ๐œ† > 0 ๐‘ . ๐‘ก. โˆ€๐œ– > 0. ๐‘ƒ[(๐ถ๐œ† โˆ’ ๐œ–) log ๐‘› โ‰ค |๐ถ๐‘š๐‘Ž๐‘ฅ | โ‰ค (๐ถ๐œ† + ๐œ–) log ๐‘›] โ†’
๐‘›โ†’โˆž
If ๐œ† > 1, then โˆƒ๐ถโ€ฒ๐œ† > 0 ๐‘ . ๐‘ก. โˆ€๐œ– > 0 ๐‘ƒ[(๐‘ โ€ฒ๐œ† โˆ’ ๐œ–)๐‘›] โ‰ค |๐ถ๐‘š๐‘Ž๐‘ฅ | โ‰ค (๐ถ โ€ฒ๐œ† + ๐œ–)๐‘›] โ†’
1
1
Fix vertex v. Do breadth first exploration of component ๐ถ(๐‘ฃ) of ๐บ(๐‘›, ๐‘) containing v vertices.
Can be either N (neutral), A(Active) or I(Inactive).
At time ๐‘ก = 0, ๐‘ฃ โˆˆ ๐ด and all other vertices are in N.
At time ๐‘ก โ‰ฅ 1, if ๐ด = ๐œ™, terminate.
Otherwise, choose ๐‘ค โˆˆ ๐ด.
Put all N neighbors of ๐‘Š in ๐ด. Put ๐‘Š in ๐ผ.
Note: Algorithm terminates after |๐ถ(๐‘ฃ)| steps.
Suppose at time ๐‘ก,
|๐‘| = ๐‘› โˆ’ ๐‘˜ vertices are left and |๐ด| โ‰ฅ 1.
The number of vertices that go from N to A in stem ๐‘ก has the distribution:
๐œ†
๐œ†
๐ต๐‘–๐‘›(๐‘› โˆ’ ๐‘˜, ๐‘) = ๐ต๐‘–๐‘› (๐‘› โˆ’ ๐‘˜, ) โ‰ˆ ๐ต๐‘–๐‘›(๐‘›, )
๐‘›
๐‘›
๐‘›
Recall that ๐‘‹~๐ต(๐‘›, ๐‘), ๐‘ โˆˆ [0,1] โ†” ๐‘ƒ[๐‘‹ = ๐‘˜] = ( ) ๐‘๐‘˜ (1 โˆ’ ๐‘)๐‘›โˆ’๐‘˜ , 0 โ‰ค ๐‘˜ โ‰ค ๐‘›
๐‘˜
Expect similarity with a branching process!
Branching Process
t=0
t=1
๐œ†
๐ต๐‘–๐‘›(๐‘›, )
๐‘›
๐œ†
๐ต๐‘–๐‘›(๐‘›, )
๐‘›
t=2
๐‘ƒ๐‘œ๐‘–(๐œ†)
๐‘Œ~๐‘ƒ๐‘œ๐‘–(๐œ†), ๐œ† โˆˆ (0, โˆž)๐‘–๐‘“ ๐‘ƒ[๐‘Œ = ๐‘˜] =
๐œ†๐‘˜ ----โˆ’๐œ†
๐‘’
, ๐‘˜โˆˆโ„ค
๐‘˜!
๐‘›โ†’โˆž ๐œ†๐‘˜
๐‘’ โˆ’๐œ† ,
๐‘˜!
๐œ†
Exercise: If ๐‘‹~๐ต๐‘–๐‘›(๐‘›, ๐‘›), then ๐‘ƒ[๐‘‹ = ๐‘˜] โ†’
๐‘˜โˆˆโ„ค
We will see that:
If ๐œ† < 1, then branching process (๐‘ƒ๐‘œ๐‘ (๐œ†)) dies out almost surely.
If ๐œ† > 1, then it lives forever with probability >0.
Further Questions
2
-
๐œ† = 1 โ†’ |๐ถ๐‘š๐‘Ž๐‘ฅ | โ‰ˆ ๐‘›3
-
If ๐œ† = 1 + ๐›ฟ๐‘› , ๐›ฟ๐‘› โ†’ 0
Other properties of ๐บ(๐‘›, ๐‘)
Add geometry to the graph
Fix a large graph ๐บ๐‘›
Retain (delete) every edge independently with probability ๐‘.
If ๐บ๐‘› = ๐‘˜๐‘› (the full graph), we get ๐บ(๐‘›, 1 โˆ’ ๐‘)
Consider a random walk on ๐บ๐‘› , study the components ๐บ๐‘› \ {๐‘‹๐‘ก1 , โ€ฆ , ๐‘‹๐‘ก๐‘› }, ๐‘ก๐‘› โ‰ฅ 0
Study random graphs s.t. ๐‘ƒ[๐‘‘๐‘’๐‘”๐‘Ÿ๐‘’๐‘’(๐‘ฃ) โ‰ฅ ๐‘˜] โ‰ˆ ๐‘˜ โˆ’๐œ , ๐œ > 0
-
๐‘›โ†’โˆž
Branching Processes
Time ๐‘› = 0, 1, โ€ฆ
๐‘‹1,1
t=0
๐‘‹2,1
๐‘‹2,๐‘‹1,1
t=1
t=2
In generation ๐‘› โ‰ฅ 0, individual ๐‘– gives birth to ๐‘‹๐‘›+1,๐‘– children, then dies.
Formally, we define {๐‘‹๐‘›,๐‘– |๐‘› โ‰ฅ 1, ๐‘– โ‰ฅ 1} as independent, identically distributed (iid) random
variables with values in โ„ค > 0
Define ๐‘๐‘› , (๐‘› โ‰ฅ 0) recursively by
๐‘0 = 1
โ‹ฎ
๐‘๐‘›=โˆ‘๐‘๐‘› โˆ’1 ๐‘‹
๐‘–=1
๐‘›,๐‘–
๐‘0 = to the total number of individuals in generation 1.
Notation ๐‘‹ = ๐‘‹1,1 (since all distribute identically)
๐‘ƒ๐‘– = ๐‘ƒ[๐‘‹ = ๐‘–] =probability that individual has ๐‘– children.
๐œ‚ = ๐‘ƒ [โ‹ƒ{๐‘๐‘› = 0}] = ๐‘ƒ[๐‘๐‘Ÿ๐‘œ๐‘๐‘’๐‘ ๐‘  ๐‘‘๐‘–๐‘’๐‘  ๐‘œ๐‘ข๐‘ก]
๐‘›โ‰ฅ1
For ๐‘  โˆˆ [0,1]
๐บ๐‘ฅ (๐‘ ) = ๐ธ[๐‘† ๐‘ฅ ]
โ€œoffspring generation functionโ€
Theorem 1.1:
- If ๐ธ[๐‘ฅ] < 1, ๐‘กโ„Ž๐‘’๐‘› ๐œ‚ = 1
- If ๐ธ[๐‘ฅ] > 1, ๐‘กโ„Ž๐‘’๐‘› ๐œ‚ < 1
- If ๐ธ[๐‘ฅ] = 1, ๐‘Ž๐‘›๐‘‘ ๐‘ƒ[๐‘‹ = 1] < 1, ๐‘กโ„Ž๐‘’๐‘› ๐œ‚ = 1
Moreover, ๐œ‚ is the smallest solution in [0,1] of ๐œ‚ = ๐บ๐‘ฅ (๐œ‚)
Exercise 1.1: Prove ๐ธ[๐‘๐‘› ] = ๐œ‡๐‘› , where ๐œ‡ = ๐ธ[๐‘‹]
Exercise 1.2: For ๐œ‡ < 1, prove ๐ธ[๐‘‡] =
--------End of lesson 1
1
1โˆ’๐œ‡
where ๐‘‡ = โˆ‘โˆž
๐‘›=0 ๐‘๐‘›
Reminder
-
{๐‘ฅ๐‘›,๐‘– }๐‘›,๐‘–โ‰ฅ1 ๐‘–๐‘–๐‘‘, โ„ค โ‰ฅ 0 โˆ’ ๐‘ฃ๐‘Ž๐‘™๐‘ข๐‘’๐‘‘
๐‘๐‘›โˆ’1
- ๐‘0 = 1, ๐‘๐‘› = โˆ‘๐‘–=1
๐‘‹๐‘›,๐‘–
- ๐‘‹ = ๐‘‹1,1
- ๐‘ƒ๐‘– = ๐‘ƒ[๐‘‹ = ๐‘–], ๐‘– = ๐‘–๐‘› โ„ค โ‰ฅ 0
- ๐œ‚ = ๐‘ƒ[โ‹ƒ๐‘›โ‰ฅ0{๐‘๐‘› = 0}]
- ๐บ๐‘ฅ (๐‘ ) = ๐ธ[๐‘  ๐‘ฅ ], ๐‘  โˆˆ [0,1]
Theorem 1.1 (survival vs. extinction):
if ๐ธ[๐‘‹] < 1 then ๐œ‚ = 1,
If ๐ธ[๐‘‹] > 1 then ๐œ‚ < 1,
If ๐ธ[๐‘‹] = 1 and ๐‘ƒ[๐‘‹ = 1] < 1 then ๐œ‚ = 1
๐œ‚ is the smallest solution in [0,1] of ๐œ‚ = ๐บ๐‘ฅ (๐œ‚)
Example:
๐‘‹~๐‘ƒ๐‘œ๐‘–(๐œ†), ๐œ† > 0
โˆž
๐‘–
๐บ๐‘ฅ (๐‘ ) = โˆ‘โˆž
๐‘–=0 ๐‘  ๐‘๐‘– = โˆ‘๐‘–=0
๐‘ ๐‘– ๐œ†๐‘– โˆ’๐œ†
๐‘’
๐‘–!
= ๐‘’ ๐œ†(๐‘ โˆ’1)
Proof: First prove that ๐œ‚ = ๐บ๐‘ฅ (๐œ‚)
Set ๐œ‚๐‘› = ๐‘ƒ [
๐‘๐‘› = 0 ]
โŸ
๐‘–๐‘›๐‘๐‘Ÿ๐‘’๐‘Ž๐‘ ๐‘–๐‘›๐‘” ๐‘–๐‘› ๐‘›
By ๐œŽ additivity, ๐œ‚๐‘› โ†‘ ๐œ‚ as ๐‘› โ†’ โˆž.
Let ๐บ๐‘› (๐‘ ) = ๐ธ[๐‘  ๐‘๐‘› ], ๐‘  โˆˆ [0,1]
(โ†‘ means that ๐œ‚๐‘› โ‰ค ๐œ‚๐‘›+1 , ๐œ‚๐‘› โ†’ ๐œ‚)
Note that is we set ๐บ๐‘› (0) = ๐œ‚๐‘›
๐‘–
๐บ๐‘› (0) = โˆ‘โˆž
๐‘–=1 0 ๐‘ƒ[๐‘๐‘› = ๐‘–] = ๐œ‚๐‘› + 0 + 0 + โ‹ฏ + 0 = ๐œ‚๐‘› .
๐‘๐‘›
๐บ๐‘› (๐‘ ) = โˆ‘โˆž
๐‘–=0 ๐ธ[๐‘  1{๐‘๐‘› =๐‘–} ] , where 1๐ด (๐‘ค) = {
TODO: Draw the full drawing
๐‘ก=1
๐‘ก=๐‘›
1 ๐‘–๐‘“ ๐‘ค โˆˆ ๐ด
0 ๐‘–๐‘“ ๐‘ค โˆˆ ฮฉ\๐ด
(๐‘1 )
(1)
๐‘๐‘› = ๐‘๐‘› + โ‹ฏ + ๐‘๐‘›
(๐‘–)
where ๐‘๐‘› Is the number of generation n-individuals descending from
(๐‘–) (๐‘‘)
generation 1-individual ๐‘– and ๐‘๐‘› = ๐‘๐‘›โˆ’1
(๐‘–)
(๐‘๐‘› )
๐‘–โ‰ฅ0
๐‘–
๐‘–
๐‘๐‘›โˆ’1 ]
๐‘๐‘›โˆ’1 ]
are independent of ๐‘1 hence ๐บ๐‘› (๐‘ ) = โˆ‘โˆž
๐‘๐‘– = โˆ‘โˆž
๐‘๐‘– =
๐‘–=0 ๐ธ[๐‘ 
๐‘–=0 ๐ธ[๐‘ 
๐‘–
โˆ‘โˆž
๐‘–=0 ๐บ๐‘›โˆ’1 (๐‘ ) ๐‘๐‘– = ๐บ๐‘ฅ (๐บ๐‘›โˆ’1 (๐‘ )).
= ๐บ1 (๐บ๐‘›โˆ’1 (๐‘ ))
๐‘  = 0: ๐œ‚๐‘› = ๐บ1 (๐œ‚๐‘›โˆ’1 ), ๐‘› โ†’ โˆž
๐œ‚ โ†’ ๐บ1 (๐œ‚)
By the dominated convergence theorem.
๐บ1 (๐œ‚๐‘›โˆ’1 ) = ๐ธ[(๐œ‚๐‘›โˆ’1 )๐‘1 ]
๐‘›โ†’โˆž
(๐œ‚๐‘›โˆ’1 ) ๐‘ง1 โ†’
๐œ‚ ๐‘1
Let ๐œ“ โˆˆ [0,1], ๐œ“ = ๐บ๐‘ฅ (๐œ“)
Now we need to show that ๐œ‚ โ‰ค ๐œ“, or equivalently, ๐œ‚๐‘› โ‰ค ๐œ“ โˆ€๐‘›.
Indeed: ๐œ‚0 = 0 โ‰ค ๐œ“ โˆˆ [0,1], if ๐œ‚๐‘›โˆ’1 โ‰ค ๐œ“, then ๐œ‚๐‘› = ๐บ๐‘ฅ (๐œ‚๐‘›โˆ’1 ) โ‰ค ๐บ๐‘ฅ (๐œ“)
๐‘๐‘ฆ ๐‘Ž๐‘ ๐‘ ๐‘ข๐‘š๐‘๐‘ก๐‘–๐‘œ๐‘›
=
๐œ“.
Case ๐‘ƒ[๐‘‹ โˆˆ {0,1}] = 1
๐‘ƒ[๐‘‹ = 0] = ๐‘ > 0
๐‘›โ†’โˆž
๐œ‚๐‘› = 1 โˆ’ ๐‘ƒ[๐‘๐‘› > 0] = 1 โˆ’ (1 โˆ’ ๐‘)๐‘› โ†’
1.
Now assume ๐‘ƒ[๐‘ฅ โ‰ค 1] < 1.
๐บ๐‘ฅโ€ฒ (๐‘ ) = ๐ธ[๐‘‹๐‘  ๐‘‹โˆ’1 ] > 0
๐บ๐‘ฅโ€ฒโ€ฒ (๐‘ ) = ๐ธ[๐‘‹(๐‘‹ โˆ’ 1)๐‘  ๐‘‹โˆ’2 ] > 0, ๐‘  โˆˆ (0,1).
(Ex: Prove this by using the mean value theorem and the dominant convergence theorem.)
๐บ๐‘ฅ (1) = 1
TODO: Draw the graph of the function ๐บ๐‘ฅ (๐‘ )
Either ๐‘  = ๐บ๐‘ฅ (๐‘ ) has one or two solutions in [0,1].
One solution โ†” ๐บ๐‘ฅโ€ฒ (1) โ‰ค 1
๐บ๐‘ฅโ€ฒ (1) = ๐ธ[๐‘‹]
Exercise 1.4:
Compute ๐œ‚
If ๐‘0 = 1 โˆ’ ๐‘, ๐‘2 = ๐‘, ๐‘ โˆˆ [0,1]
โ€œBinary branchingโ€.
Exercise 1.5: Prove ๐‘ƒ[๐‘๐‘› > 0] โ‰ค ๐œ‡๐‘› , ๐œ‡ = ๐ธ[๐‘‹].
Hint: use the Chebishev inequality and the previous exercise.
Dominated Convergence Theorem
๐‘›โ†’โˆž
Let (๐‘‹๐‘› )๐‘›โ‰ฅ1 be a sequence of random variables, ๐‘‹๐‘› โ†’ ๐‘‹ a.s.
Assume that โˆƒ a random variable ๐‘Œ โ‰ฅ 0 almost surely ๐ธ[๐‘Œ] < โˆž such that |๐‘‹๐‘› | โ‰ค ๐‘Œ โˆ€๐‘› almost
๐‘›โ†’โˆž
surely, Then ๐ธ[๐‘‹๐‘› ] โ†’
๐ธ[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
๐‘ก=0
๐‘ก =n
Redenote ๐‘‹๐‘›,๐‘– as they appear in breadth-first search exploration of the tree.
๐‘–. ๐‘’. (๐‘‹1,1 , ๐‘‹2,1 , โ€ฆ , ๐‘‹2,2 , ๐‘‹3,1 , โ€ฆ ) =: (๐‘‹1 , ๐‘‹2 , ๐‘‹3 , โ€ฆ )
Define the random variables ๐‘๐‘˜ , ๐ผ๐‘˜ , ๐‘˜ โ‰ฅ 0 such that ๐‘‹๐‘˜ = ๐‘‹(๐‘๐‘˜ ,๐ผ๐‘˜)
Observation: (๐‘๐‘˜ , ๐ผ๐‘˜ ) depend only on (๐‘‹1 , โ€ฆ , ๐‘‹๐‘˜โˆ’1 )
Lemma 1.7:
(๐‘‹๐‘˜ )๐‘˜โ‰ฅ1 are iid with the same distribution as ๐‘‹.
Proof: Let ๐‘‹1 , โ€ฆ , ๐‘‹๐‘˜ โˆˆ โ„ค โ‰ฅ 0
๐‘ƒ[๐‘‹1 = ๐‘ฅ1, โ€ฆ , ๐‘‹๐‘˜ = ๐‘ฅ๐‘˜ ] = โˆ‘ โˆ‘ ๐‘ƒ [(๐‘‹1 = ๐‘ฅ1 , โ€ฆ , ๐‘‹๐‘˜โˆ’1 = ๐‘ฅ๐‘˜โˆ’1 , ๐‘๐‘˜ = ๐‘›, ๐ผ๐‘˜ = ๐‘– ) ๐‘‹๐‘›,๐‘– = ๐‘ฅ๐‘˜ ]
๐‘›โ‰ฅ1 ๐‘–โ‰ฅ1
๐‘›โˆ’1 โˆž
๐ด
๐‘–โˆ’1
๐ด depends only on ๐œŽ ({๐‘‹๐‘š,๐‘— }๐‘š=1,๐‘–=1 {๐‘‹๐‘›,๐‘— }๐‘—=1 )
โˆ‘๐‘› โˆ‘๐‘– ๐‘ƒ[๐ด]๐‘ƒ[๐‘‹ = ๐‘ฅ๐‘˜ ] = ๐‘ƒ[๐‘‹1 = ๐‘ฅ1 , โ€ฆ , ๐‘‹๐‘˜โˆ’1 = ๐‘ฅ๐‘˜โˆ’1 ]๐‘ƒ[๐‘‹ = ๐‘ฅ๐‘˜ ]
๐‘ƒ[๐‘‹ = ๐‘ฅ๐‘˜ ]
๐‘–๐‘›๐‘‘๐‘ข๐‘๐‘ก๐‘–๐‘œ๐‘›
=
๐‘ƒ[๐‘‹ = ๐‘ฅ1 ] โˆ™ โ€ฆ โˆ™
Independence of ๐‘‹๐‘›,๐‘–
Notation (ctd): (๐‘†๐‘– )๐‘–โ‰ฅ0 equals to the number of active individuals at stage ๐‘–.
๐‘†0 = 1
๐‘†๐‘– = ๐‘†๐‘–โˆ’1 + (๐‘‹๐‘– โˆ’ 1), ๐‘– โ‰ฅ 1 = (๐‘‹1 , โ€ฆ , ๐‘‹๐‘– ) โˆ’ (๐‘– โˆ’ 1) = number of active individuals.
๐‘‡ = (โˆ‘๐‘›โ‰ฅ0 ๐‘๐‘› ) = inf{๐‘– โ‰ฅ 0 โˆถ ๐‘†๐‘– = 0}
๐‘†0 = 1
๐‘†1 = 1 + (2 โˆ’ 1)
๐‘†2 = 2 + (2 โˆ’ 1)
๐‘†3 = 3 + (0 โˆ’ 1)
๐‘†4 = 2 + (0 โˆ’ 1) = 1
๐‘ 5 = 1 + (0 โˆ’ 1) = 0
------- end of lesson 2
In the proof of theorem 1.1 we showed: ๐บ๐‘› (๐‘ ) = ๐ธ[๐‘†๐‘๐‘› ] = ๐บ๐‘‹ (๐บ๐‘›โˆ’1 (๐‘ )), ๐‘  โˆˆ [0,1]
Exercise 1.7: Set ๐บ๐‘‡ (๐‘ ) = ๐ธ[๐‘† ๐‘‡ ], ๐‘  โˆˆ [0,1]
Prove that ๐บ๐‘‡ (๐‘ ) = ๐‘  โˆ™ ๐บ๐‘‹ (๐บ๐‘‡ (๐‘ )), where ๐บ๐‘‹ (๐‘ ) = ๐ธ[๐‘† ๐‘‹ ]
Exercise 1.8: Prove the following :
If ๐œ‡ = ๐ธ[๐‘‹] = 1, then ๐ธ[๐‘‡] = โˆž
(hint: Prove first that ๐ธ[๐‘‡] = 1 + ๐ธ[๐‘‡])
Exercize 1.9: Recall ๐ธ[๐‘๐‘› ] = ๐œ‡๐‘›
-
๐‘
Prove that (๐‘€๐‘› = ๐œ‡๐‘›๐‘› )
๐‘›โ‰ฅ0
is a martingale (w.r.t. natural filtration). Show that:
lim ๐‘€๐‘› = ๐‘€โˆž exists and 0 โ‰ค ๐‘€๐‘› โ‰ค โˆž almost surely.
๐‘›โ†’โˆž
-
For ๐œ‡ > 1and ๐ธ[๐‘‹ 2 ] < โˆž show that (๐‘€๐‘› )๐‘›โ‰ฅ0 is bounded in ๐ฟ2 . Use theorem 1.1 and
onvergence theorem (๐ฟ2 ) to prove that ๐‘ƒ[๐‘€โˆž = 0] = ๐œ‚ = ๐‘ƒ[๐‘‡ < โˆž]. Deduce that
{๐‘€โˆž = 0} = {๐‘‡ < โˆž} almost surely.
(โ‡’ ๐‘๐‘› ~(๐‘€โˆž โˆ™ ๐œ‡๐‘› )๐‘œ๐‘› {๐‘‡=โˆž} )
Theorem 1.10: (Extinction with large total progeny)
๐‘’ โˆ’๐ผ๐‘˜
If ๐œ‡ = ๐ธ[๐‘‹] > 1, then ๐‘ƒ[๐‘˜ โ‰ค ๐‘‡ < โˆž] โ‰ค 1โˆ’๐‘’ ๐ผ where ๐ผ = sup(๐‘ก โˆ’ log ๐ธ[๐‘’ ๐‘ก๐‘‹ ]) > 0
๐œ†
Remark: In ๐บ (๐‘›, ๐‘›) , ๐œ† > 1, we will see that โˆƒ๐‘˜ > 0 ๐‘ . ๐‘ก. there are no components of size
1
between ๐‘˜ log ๐‘› and ๐‘˜ โˆ™ ๐‘›, with high probability as ๐‘› โ†’ โˆž.
Lemma 1.11: (Cramer/Chernoff bound):
Let {๐‘‹๐‘– }โˆž
๐‘–=1 be iid.
1. For any ๐‘Ž โˆˆ (๐ธ[๐‘‹1 ], โˆž), ๐‘ƒ[โˆ‘๐‘›๐‘–=1 ๐‘‹๐‘– โ‰ฅ ๐‘›๐‘Ž] โ‰ค ๐‘’ โˆ’๐‘›๐ผ(๐‘Ž)
2. For any ๐‘Ž โˆˆ (โˆ’โˆž, ๐ธ[๐‘‹1 ]), ๐‘ƒ[โˆ‘๐‘›๐‘–=1 ๐‘‹๐‘– โ‰ค ๐‘›๐‘Ž] โ‰ค ๐‘’ โˆ’๐‘›๐ผ(๐‘Ž)
Where ๐ผ(๐‘Ž) = sup(๐‘ก๐‘Ž โˆ’ log ๐ธ[๐‘’ ๐‘ก๐‘‹1 ]) in (1)
tโ‰ฅ0
and ๐ผ(๐‘Ž) = sup(๐‘ก๐‘Ž โˆ’ log ๐ธ[๐‘’ ๐‘ก๐‘‹1 ]) in (2)
tโ‰ค0
Proof of Lemma 1.11:
๐‘›
๐‘›
๐‘›
๐‘ƒ [โˆ‘ ๐‘‹๐‘– โ‰ฅ ๐‘›๐‘Ž] = ๐‘ƒ [exp (โˆ’๐‘ก๐‘›๐‘Ž + ๐‘ก โˆ‘ ๐‘‹๐‘– ) โ‰ฅ 1] = ๐ธ[1๐ด ] โ‰ค ๐ธ 1๐ด โˆ™ exp (โˆ’๐‘ก๐‘›๐‘Ž + ๐‘ก โˆ‘ ๐‘‹๐‘– )
โŸ
๐‘–=1
๐‘–=1
๐‘–=1
โˆ€๐‘กโ‰ฅ0
[
]
โ‰ฅ1
๐‘›
๐‘›
โ‰ค ๐ธ exp (โˆ’๐‘ก๐‘›๐‘Ž + ๐‘ก โˆ‘ ๐‘‹๐‘– ) = ๐‘’
โŸ
๐‘–=1
[
]
โ‰ฅ0
๐‘ก๐‘‹๐‘– ])
= exp(โˆ’๐‘›(๐‘ก๐‘Ž โˆ’ log ๐ธ[๐‘’
)
Now optimize over ๐‘ก โ‰ฅ 0 โ‡’(1).
โˆ’๐‘ก๐‘›๐‘Ž
๐ธ [โˆ ๐‘’
๐‘›
๐‘ก๐‘‹๐‘–
]=๐‘’
โˆ’๐‘ก๐‘›๐‘Ž
๐‘–=1
โˆ ๐ธ[๐‘’ ๐‘ก๐‘‹๐‘– ]
๐‘–=1
(2) is left as an exercise!
Proof of theorem 1.10:
โˆž
โˆž
โˆž
๐‘ƒ[๐‘˜ โ‰ค ๐‘‡ < โˆž] = โˆ‘ ๐‘ƒ [ ๐‘‡
โŸ= ๐‘› ] โ‰ค โˆ‘ ๐‘ƒ[๐‘‹1 + โ‹ฏ + ๐‘‹๐‘› = ๐‘› โˆ’ 1] = โˆ‘ ๐‘ƒ[๐‘‹1 + โ‹ฏ + ๐‘‹๐‘› โ‰ค ๐‘›]
๐‘›=๐‘˜
โŠ†{๐‘†๐‘› =0}
๐‘›=๐‘˜
๐‘›=๐‘˜
๐‘†๐‘› = (๐‘‹1 + โ‹ฏ + ๐‘‹๐‘› ) โˆ’ (๐‘› โˆ’ 1)
๐‘™๐‘’๐‘š๐‘š๐‘Ž 1.11,(2),๐‘Ž=1
๐‘ƒ[๐‘˜ โ‰ค ๐‘‡ < โˆž]
โ‰ค
โˆž
โˆ‘ ๐‘’ โˆ’๐‘›๐ผ(1)
๐‘›=๐‘˜
Define โ€“ ๐‘“(๐‘ก) = ๐‘ก โˆ’ log ๐ธ[๐‘’ ๐‘ก๐‘‹ ] , ๐‘ก โ‰ค 0
๐‘“(0) = 0 โˆ’ log 1 = 0
1
๐‘“ โ€ฒ (๐‘ก) | = 1 โˆ’
โˆ™ ๐ธ[๐‘‹๐‘’ 0๐‘‹ ] = 1 โˆ’ ๐ธ[๐‘‹] < 0
0๐‘‹ ]
๐ธ[๐‘’
๐‘ก=0
Since the derivative is negative, there is some negative value where ๐‘“(๐‘ก) approaches from
above.
So โˆƒ๐‘ก < 0: ๐‘“(๐‘ก) > 0 โ‡’ ๐ผ(1) > 0
Duality Principle
Assume ๐œ‚ โˆˆ (0,1)
(๐œ‚ = ๐‘ƒ[๐‘‡ < โˆž])
Condition the branching process to die out. (only consider the probability ๐‘ƒ[๐ด|๐‘‡ < โˆž] =
๐‘ƒ[๐ดโˆฉ{๐‘‡<โˆž}]
๐‘ƒ[๐‘‡<โˆž]
(A event)
Question: How does this change the branching process?
Definition 1.12: (Conjugate Distribution)
Let ๐‘๐‘ฅ = ๐‘ƒ[๐‘‹ = ๐‘ฅ], ๐‘ฅ โˆˆ โ„คโ‰ฅ0 and assume ๐‘0 > 0 (โ‡” ๐œ‚ > 0).
Then the conjugate distribution (๐‘๐‘ฅโ€ฒ )๐‘ฅโˆˆโ„คโ‰ฅ0 associated to ๐‘ is defined as ๐‘๐‘ฅโ€ฒ = ๐œ‚ ๐‘ฅโˆ’1 ๐‘๐‘ฅ , ๐‘ฅ โˆˆ โ„คโ‰ฅ0
Lemma 1.13:
โˆ‘ ๐‘๐‘ฅโ€ฒ = 1
๐‘ฅโ‰ฅ0
Proof:
โˆ‘ ๐‘๐‘ฅโ€ฒ = โˆ‘ ๐œ‚ ๐‘ฅโˆ’1 ๐‘๐‘ฅ =
๐‘ฅโ‰ฅ0
๐‘ฅโ‰ฅ0
1
1
1
โˆ‘ ๐œ‚ ๐‘ฅ ๐‘๐‘ฅ = ๐บ๐‘‹ (๐œ‚) = โˆ™ ๐œ‚ = 1
๐œ‚
๐œ‚
๐œ‚
๐‘ฅโ‰ฅ0
Denote the branching process with offspring distribution ๐‘โ€ฒ (๐‘โ€ฒ branching process) by
โ€ฒ
{๐‘‹๐‘›,๐‘–
}๐‘›โ‰ฅ1,๐‘–โ‰ฅ1
Lemma 1.14 (๐œ‚ > 0) ๐‘ƒ[๐‘‡ โ€ฒ < โˆž] = 1
Proof: Let ๐‘  โˆˆ [0,1], assume that ๐บ๐‘‹ โ€ฒ (๐‘ ) = ๐‘ . Want to show โ‡’ ๐‘  = 1, use theorem 1.1.
โˆž
โˆž
1
๐บ๐‘‹ โ€ฒ (๐‘ ) = โˆ‘ ๐‘  ๐‘ƒ[๐‘‹ = ๐‘–] = โˆ‘ ๐‘  ๐‘– (๐œ‚ ๐‘–โˆ’1 ๐‘๐‘– ) = ๐บ๐‘‹ (๐‘ ๐œ‚) = ๐‘ 
๐œ‚
๐‘–
โ€ฒ
๐‘–=0
๐‘–=0
So
๐‘ ๐œ‚โ‰ฅ๐œ‚
๐บ๐‘‹ (๐‘ ๐œ‚) = ๐‘ ๐œ‚ โ‡’
โˆŽ
------ End of lesson 3
๐‘ ๐œ‚ โ‰ฅ ๐œ‚ โ‡’ ๐‘  = 1
Duality Principle
Assume that ๐œ‚๐‘ƒ[๐‘‡ < โˆž] โˆˆ (0,1)
Given ๐‘๐‘ฅ = ๐‘ƒ[๐‘‹ = ๐‘ฅ], ๐‘ฅ โ‰ฅ 0, ๐‘0 > 0
Define conjugate distribution ๐‘โ€ฒ by ๐‘๐‘ฅโ€ฒ = ๐œ‚ ๐‘ฅโˆ’1 ๐‘๐‘ฅ , ๐‘ฅ โˆˆ โ„คโ‰ฅ0
โ€ฒ
{๐‘‹๐‘›,๐‘–
}๐‘›โ‰ฅ1,๐‘–โ‰ฅ1 ๐‘โ€ฒ - A branching process
(๐‘‡ โ€ฒ , ๐‘๐‘›โ€ฒ , ๐‘‹1โ€ฒ , ๐‘‹2โ€ฒ โ€ฆ ) then ๐‘ƒ[๐‘‡ โ€ฒ < โˆž] = 1
Theorem 1.15: (duality princinple)
Let ๐‘ = (๐‘๐‘ฅ )๐‘ฅโˆˆโ„คโ‰ฅ0 be a distribution on โ„คโ‰ฅ0 with ๐‘0 > 0 and let ๐‘โ€ฒ be its conjugate distribution.
Let (๐‘‹๐‘˜ )๐‘˜โ‰ฅ1 be a branching process.
Let (๐‘‹๐‘˜โ€ฒ )๐‘˜โ‰ฅ1 be a branching process
Then ๐‘ƒ[(๐‘‹1 , โ€ฆ , ๐‘‹ ๐‘‡ โ€ฒ ) = (๐‘ฅ1 , โ€ฆ , ๐‘ฅ๐‘ก )|๐‘‡ < โˆž] = ๐‘ƒ[(๐‘‹1โ€ฒ , โ€ฆ , ๐‘‹๐‘‡โ€ฒ ) = (๐‘ฅ1 , โ€ฆ , ๐‘ฅ๐‘ก )|๐‘‡ < โˆž]
For ๐‘ฅ1 , โ€ฆ , ๐‘ฅ๐‘› โˆˆ โ„คโ‰ฅ0 .
In words, ๐‘-branching process conditioned to die out is distributed as a ๐‘โ€ฒ branching process.
Proof: Denote ๐ป = (๐‘‹1 , โ€ฆ , ๐‘‹๐‘‡ ), ๐ป โ€ฒ = (๐‘‹1โ€ฒ , โ€ฆ , ๐‘‹๐‘‡โ€ฒ โ€ฒ ), โ„Ž = (๐‘ฅ1 , โ€ฆ , ๐‘ฅ๐‘ก )
Recall:
๐‘†๐‘– = (๐‘‹1 + โ‹ฏ + ๐‘‹๐‘– ) โˆ’ (๐‘– โˆ’ 1)
๐‘‡ = inf{๐‘– โ‰ฅ 1|๐‘†๐‘– = 0}
๐‘ ๐‘– = (๐‘ฅ1 , โ€ฆ , ๐‘ฅ๐‘– ) โˆ’ (๐‘– โˆ’ 1)
Assume that ๐‘†๐‘– > 0 for 1 โ‰ค ๐‘– < ๐‘ก and ๐‘†๐‘ก = 0.
Otherwise (1) otherwise the equation is true trivially (0 = 0)
๐‘ก
๐‘ก
๐‘–=1
๐‘–=1
๐‘ƒ[๐ป = โ„Ž, ๐‘‡ < โˆž] ๐ต๐‘ฆ ๐‘™๐‘’๐‘š๐‘š๐‘Ž 1.7 1
1
๐‘ƒ[๐ป = โ„Ž|๐‘‡ < โˆž] =
=
โˆ ๐‘๐‘ฅ๐‘– = โˆ ๐‘๐‘ฅโ€ฒ ๐‘– ๐œ‚1โˆ’๐‘ฅ๐‘–
๐‘ƒ[๐‘‡ < โˆž]
๐œ‚
๐œ‚
๐‘ก
=
1 ๐‘กโˆ’โˆ‘๐‘ก ๐‘ฅ
โˆ™ ๐œ‚ ๐‘–=1 ๐‘– โˆ ๐‘๐‘ฅโ€ฒ ๐‘–
๐œ‚
โŸ
๐‘–=1
๐‘ƒ[๐ป โ€ฒ =โ„Ž]
โˆ‘๐‘ก๐‘–=1 ๐‘ฅ๐‘–
But ๐‘ก โˆ’
= ๐‘ก โˆ’ (๐‘ ๐‘ก + (๐‘ก โˆ’ 1)) = ๐‘ก โˆ’ ๐‘ก + 1 = 1
โ€ฒ
So it equals: ๐‘ƒ[๐ป = โ„Ž]
Which is what we needed to prove.
Exercise 1.16:
Let ๐‘‹ be ๐‘ƒ๐‘œ๐‘–(๐œ†) - distributed, ๐œ† > 0.
Recall from Lemma 1.11
๐ผ(๐‘Ž) = sup(๐‘ก๐‘Ž โˆ’ log ๐ธ[๐‘’ ๐‘ก๐‘ฅ ]) , ๐‘Ž > 0
(๐‘ก โ‰ฅ 0 if ๐‘Ž > ๐œ†)
(๐‘ก โ‰ค 0 if ๐‘Ž < ๐œ†)
Show that in both cases (๐‘Ž > ๐œ†, ๐‘Ž < ๐œ†)
๐‘Ž
๐ผ(๐‘Ž) = sup (๐‘ก๐‘Ž โˆ’ log ๐ธ[๐‘’ ๐‘ก๐‘ฅ ]) = ๐œ† โˆ’ ๐‘Ž + ๐‘Ž log ( )
๐œ†
tโˆˆRโ„
Exercise 1.17 (coupling)
Let (๐‘๐‘ฅ )๐‘ฅโˆˆโ„ค , (๐‘ž๐‘ฅ )๐‘ฅโˆˆโ„ค be different probability distributions on โ„ค.
We want a random variable (๐‘‹, ๐‘Œ), โ„ค2 valued, such that
๐‘ƒ[๐‘‹ = ๐‘ฅ] = ๐‘๐‘ฅ , ๐‘ฅ โˆˆ โ„คโ‰ฅ0
๐‘ƒ[๐‘Œ = ๐‘ฅ] = ๐‘ž๐‘ฅ , ๐‘ฅ โˆˆ โ„คโ‰ฅ0
๐‘ƒ[๐‘‹ โ‰  ๐‘Œ] as small as possible.
Prove that:
1
2
(๐‘๐‘ฅ โˆ’min{๐‘๐‘ฅ ,๐‘ž๐‘ฅ })(๐‘ž๐‘ฅ โˆ’min{๐‘๐‘ฅ ,๐‘ž๐‘ฅ })
๐‘‘๐‘‡๐‘‰ (๐‘,๐‘ž)
(i)
โˆ‘๐‘ฅโˆˆโ„ค(๐‘๐‘ฅ โˆ’ min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ }) = โˆ‘๐‘ฅโˆˆโ„ค(๐‘ž๐‘ฅ โˆ’ min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ }) = โˆ‘๐‘ฅโˆˆโ„ค|๐‘๐‘ฅ โˆ’ ๐‘ž๐‘ฅ | = ๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž)
(ii)
๐‘๐‘ฅ,๐‘ฆ = min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ } if ๐‘ฅ = ๐‘ฆ, and
(iii)
otherwise. Is a
distribution on โ„ค2
With ๐‘ƒ[(๐‘‹, ๐‘Œ) = (๐‘ฅ, ๐‘ฆ)] = ๐‘๐‘ฅ,๐‘ฆ , (๐‘ฅ, ๐‘ฆ) โˆˆ โ„ค2
get ๐‘ƒ[๐‘‹ โ‰  ๐‘Œ] = ๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž) and this is the best.
if ๐‘ƒ[(๐‘‹, ๐‘Œ) = (๐‘ฅ, ๐‘ฆ)] = ๐‘โ€ฒ๐‘ฅ,๐‘ฆ , ๐‘ƒ[๐‘‹ = ๐‘ฅ] = ๐‘๐‘ฅ , ๐‘ƒ[๐‘Œ = ๐‘ฅ] = ๐‘ž๐‘ฅ โ‡’ ๐‘ƒ[๐‘‹ โ‰  ๐‘Œ] โ‰ฅ
๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž)
Possible for any random variable (๐‘‹ โ€ฒ , ๐‘Œโ€ฒ) with values in โ„ค2 , with ๐‘ƒ[๐‘‹ โ€ฒ = ๐‘ฅ] =
๐‘๐‘ฅ , ๐‘ƒ[๐‘Œ โ€ฒ = ๐‘ฅ] = ๐‘ž๐‘ฅ โ‡’ ๐‘ƒ[๐‘‹ โ€ฒ โ‰  ๐‘Œ โ€ฒ ] > ๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž)
Poisson branching processes
Notation: Use superscript โ€œ*โ€ for Poisson branching processes.
(๐‘‹๐‘›โˆ— , ๐‘‹๐‘–โˆ— , ๐‘‡ โˆ— , โ€ฆ )
Recall:
๐œ†๐‘˜ โˆ’๐œ†
๐‘’ ,
๐‘˜!
-
๐‘ƒ[๐‘‹ โˆ— = ๐‘˜] =
-
๐ธ[๐‘‹ โˆ— ] = ๐œ†
๐บ๐‘‹ โˆ— (๐‘ ) = ๐‘’ ๐œ†(๐‘ โˆ’1) , ๐‘  โ‰ฅ 0
๐‘˜ โˆˆ โ„คโ‰ฅ0
Definition 2.1: Call 0 < ๐œ‡ < 1 < ๐œ† a conjugate pair if ๐œ‡๐‘’ โˆ’๐œ‡ = ๐œ†๐‘’ โˆ’๐œ†
Note:
๐‘“: ๐‘ฅ โ†’ ๐‘ฅ๐‘’ โˆ’๐‘ฅ is strictly increasing on [0,1]
It is strictly decreasing on [1, โˆž]
๐‘“(0) = 0, ๐‘“(โˆž) = 0
TODO: Draw functionโ€ฆ
So for any ๐œ† > 1, there is a unique conjugate ๐œ‡ โˆˆ [0,1) and vice-versa.
Theorem 2.2: (Poisson duality principle):
Let ๐œ‡ < 1 < ๐œ† be a conjugate pair
The ๐‘ƒ๐‘œ๐‘–(๐œ†) - a branching process conditioned on extinction is distributed as ๐‘ƒ๐‘œ๐‘–(๐œ‡) - a
branching process.
Proof: By theorem 1.15, The conditioned process is a branching process with offspring
distribution ๐‘๐‘ฅโ€ฒ = ๐œ‚๐œ†๐‘ฅโˆ’1
๐œ†๐‘ฅ โˆ’๐œ†
๐‘’ ,
๐‘ฅ!
๐‘ฅ > 0 where ๐œ‚๐œ† = extinction probability for ๐‘ƒ๐‘œ๐‘–(๐œ†)- branching
process. ๐œ‚๐œ† โˆˆ (0,1)
i.e.
(๐œ†๐œ‚๐œ† )๐‘ฅ โˆ’๐œ† 1
๐‘๐‘ฅโ€ฒ =
๐‘’ ( )
๐‘ฅ!
๐œ‚๐œ†
By theorem 1.1:
๐œ‚๐œ† = ๐บ๐‘‹ โˆ— (๐œ‚๐œ† ) = ๐‘’ ๐œ†(๐œ‚๐œ† โˆ’1)
So ๐‘๐‘ฅโ€ฒ =
(๐œ†๐œ‚๐œ† )๐‘ฅ
๐‘ฅ!
๐‘’ โˆ’๐œ† โˆ™
1
๐‘’
๐œ†(๐œ‚๐œ† โˆ’1)
=
(๐œ†๐œ‚๐œ† )๐‘ฅ
๐‘ฅ!
๐‘’ โˆ’๐œ†๐œ‚๐œ† , ๐‘ฅ โ‰ฅ 0
Hence, ๐‘โ€ฒ = ๐‘ƒ๐‘œ๐‘–(๐œ†๐œ‚๐œ† )
?
๐œ†๐œ‚๐œ† ๐‘’ โˆ’๐œ†๐œ‚๐œ† = ๐œ‡๐‘’ โˆ’๐œ‡
Have to prove ๐œ†๐œ‚๐œ† = ๐œ‡
๐œ†๐œ‚๐œ† ๐‘’ โˆ’๐œ†๐œ‚๐œ† = ๐œ†๐‘’ ๐œ†(๐œ‚๐œ† โˆ’1) ๐‘’ โˆ’๐œ†๐œ‚๐œ† = ๐œ†๐‘’ โˆ’๐œ†
So, ๐œ†๐œ‚๐œ† โ‰  ๐œ† (since ๐œ† > ๐‘ž (theorem 1.1) โ‡’ ๐ธ[๐‘‹ โˆ— ] > 1 โ‡’ ๐œ‚๐œ† > 1)
And (๐œ†๐œ‚๐œ† )๐‘’ โˆ’(๐œ†๐œ‚๐œ† ) = ๐œ†๐‘’ โˆ’๐œ†
Since ๐‘ฅ๐‘’ โˆ’๐‘ฅ โ†‘ on [0,1], โ†“ on [1, โˆž]
There is at most 1 ๐‘ฅ diferent from ๐œ† such that ๐‘ฅ๐‘’ โˆ’๐‘ฅ = ๐œ†๐‘’ โˆ’๐œ† = ๐œ‡
Hence, ๐œ†๐œ‚๐œ† = ๐œ‡ โˆŽ
Theorem (Cayleyโ€™s formula):
The number of labeled trees on ๐‘› vertices is equal to ๐‘›๐‘›โˆ’2
TODO: Draw a tree
Theorem 3.15: In reference
---- End of lesson 4
โˆž
โˆ—
๐‘‡ = โˆ‘ ๐‘๐‘›โˆ— = inf{๐‘– โ‰ฅ 0: (๐‘‹1 + โ‹ฏ + ๐‘‹๐‘– ) โˆ’ (๐‘– โˆ’ 1) = 0}
๐‘›=1
Theorem 2.3 (Total Progeny of Poisson Branching Process):
For ๐‘ƒ๐‘œ๐‘–(๐œ†) - branching process. ๐œ† > 0
(๐œ†๐‘› )๐‘›โˆ’1 โˆ’๐œ†๐‘›
๐‘ƒ[๐‘‡ โˆ— = ๐‘›] =
๐‘’
,
๐‘›!
๐‘›โ‰ฅ1
Proof: By Induction on ๐‘›.
โˆ—
Check for ๐‘› = 1: ๐‘ƒ[๐‘‡ โˆ— = 1] = ๐‘ƒ[๐‘‹1,1
= 0] = ๐‘’ โˆ’๐œ† which fully agrees with the formula.
๐‘› โ‰ฅ 2, assume the result holds for all ๐‘›โ€ฒ < ๐‘›.
๐‘›โˆ’1
โˆ—
๐‘ƒ[๐‘‡ = ๐‘›] = โˆ‘ ๐‘ƒ[๐‘‡ โˆ— = ๐‘›, ๐‘‹1,1
= ๐‘–]
โˆ—
๐‘–=1
If we observe the children that โ€œsprungโ€ 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.
โˆ—
๐‘‡ โˆ— is distributed as 1 + ๐‘‡ (1) + โ‹ฏ + ๐‘‡ (๐‘‹1,1 ) , where {๐‘‡ (2) }๐‘–โ‰ฅ1 are iid with the same distribution as
๐‘‡โˆ—
๐‘›โˆ’1
โˆ—
๐‘ƒ[๐‘‡ = ๐‘›] = โˆ‘ ๐‘ƒ[๐‘‡ (1) + โ‹ฏ + ๐‘‡ (๐‘–) = ๐‘› โˆ’ 1, ๐‘‹1,1
= ๐‘–]
โˆ—
๐‘–=1
But these random variables are independent (by our assumption)
๐‘›โˆ’1
=โˆ‘
โˆ—
๐‘ƒ[๐‘‡ (1) = ๐‘›1 , โ€ฆ , ๐‘‡ (๐‘–) = ๐‘›๐‘– , ๐‘‹1,1
= ๐‘–] =
โˆ‘
๐‘–=1 (๐‘›1 ,โ€ฆ,๐‘›๐‘– )
๐‘›๐‘˜ โ‰ฅ1,
โˆ‘๐‘˜ ๐‘›๐‘˜ =๐‘›โˆ’1
๐‘›โˆ’1
โˆ‘
โˆ‘
โˆ—
๐‘ƒ[๐‘‡ (1) = ๐‘›1 ] โˆ™ โ€ฆ โˆ™ ๐‘ƒ[๐‘‡ (๐‘–) = ๐‘›๐‘– ] โˆ™ ๐‘ƒ[๐‘‹1,1
= ๐‘–]
๐‘–=1 (๐‘›1 ,โ€ฆ,๐‘›๐‘– )
๐‘›๐‘˜ โ‰ฅ1,
โˆ‘๐‘˜ ๐‘›๐‘˜ =๐‘›โˆ’1
Now we can use our induction hypothesis!
Each of the ๐‘›๐‘– โ€™s is smaller than ๐‘›.
๐‘›โˆ’1
โˆ‘
๐‘–
โˆ‘
๐‘›โˆ’1โˆ’๐‘– โˆ’๐œ†(๐‘›โˆ’1)
๐œ†
๐‘’
โˆ™ โˆ(
๐‘–=1 (๐‘›1 ,โ€ฆ,๐‘›๐‘– )
๐‘›๐‘˜ โ‰ฅ1,
โˆ‘๐‘˜ ๐‘›๐‘˜ =๐‘›โˆ’1
So, so far we have shown that:
๐‘˜=1
(๐‘›๐‘˜ )๐‘›๐‘˜โˆ’1 ๐œ†๐‘– โˆ’๐œ†
)โˆ™ ๐‘’
๐‘›๐‘˜ !
๐‘–!
๐‘›โˆ’1
โˆ—
๐‘›โˆ’1 โˆ’๐œ†๐‘›
๐‘ƒ[๐‘‡ = ๐‘›] = ๐œ†
๐‘’
1
โˆ‘
๐‘–!
๐‘–=1
โŸ
๐‘–
โˆ‘
โˆ(
(๐‘›1 ,โ€ฆ,๐‘›๐‘– ) ๐‘˜=1
๐‘›๐‘˜ โ‰ฅ1
โˆ‘๐‘˜ ๐‘›๐‘˜ =๐‘›โˆ’1
(๐‘›๐‘˜ )๐‘›๐‘˜ โˆ’1
) = ๐œ†๐‘›โˆ’1 ๐‘’ โˆ’๐œ†๐‘› โˆ™ ๐‘Ž๐‘›
๐‘›๐‘˜ !
=:๐‘Ž๐‘›
Lemma 2.4: Let ๐ฟ๐‘› be the number of labeled trees on ๐‘› vertices.
Then, ๐ฟ๐‘› = ๐‘›๐‘›โˆ’2 = (๐‘› โˆ’ 1)! ๐‘Ž๐‘›
๐‘›๐‘›โˆ’2
With the lemma 2.4, we get ๐‘ƒ[๐‘‡ โˆ— = ๐‘›] = ๐œ†๐‘›โˆ’1 ๐‘’ โˆ’๐œ†๐‘› โˆ™ (๐‘›โˆ’1)! =
(๐œ†๐‘›)๐‘›โˆ’1
๐‘›!
๐‘’ โˆ’๐œ†๐‘› โˆŽ (theorem 2.3).
How do we prove lemma 2.4? We use lemma 2.5.
Lemma 2.5 (Cayleyโ€™s Formula):
๐ฟ๐‘› = ๐‘›๐‘›โˆ’2
Example:
๐‘›=5
TODO: Draw graph
Proof that lemma 2.5 โ‡’ Lemma 2.4:
For ๐‘– โ‰ฅ 1, ๐‘›1 , โ€ฆ , ๐‘›๐‘– , ๐‘›๐‘˜ โ‰ฅ 1,
โˆ‘๐‘˜ ๐‘›๐‘˜ = ๐‘› โˆ’ 1, define ๐‘ก(๐‘›1 ,โ€ฆ,๐‘›๐‘–) =number of labeled trees of ๐‘› vertices such that ๐‘ฃ1 has ๐‘–
neighbors, (tree\{๐‘ฃ1 }) โ€“ has components with ๐‘›1 , โ€ฆ , ๐‘›๐‘– vertices.
To choose such a tree:
1. Split {๐‘ฃ2 , โ€ฆ , ๐‘ฃ๐‘› } into sets of size ๐‘›1 , โ€ฆ , ๐‘›๐‘–
2. Choose ๐‘– labeled trees of size ๐‘›1 , โ€ฆ , ๐‘›๐‘–
3. Choose a vertex in each of the ๐‘– trees, and connect it to ๐‘ฃ1
๐‘–
๐‘ก(๐‘›1 ,โ€ฆ,๐‘›๐‘–)
(๐‘›๐‘˜ )๐‘›๐‘˜โˆ’1
๐‘›โˆ’1
๐‘› โˆ’2
๐‘› โˆ’2
=(
) โˆ™ (๐‘›1 1 โˆ™ โ€ฆ โˆ™ ๐‘›๐‘– ๐‘– ) โˆ™ (๐‘›1 โˆ™ โ€ฆ โˆ™ ๐‘›๐‘– ) = (๐‘› โˆ’ 1)! โˆ (
)
๐‘›1 , โ€ฆ , ๐‘›๐‘–
๐‘›๐‘˜ !
๐‘˜=1
Hence,
๐‘›โˆ’1
๐ฟ๐‘› = โˆ‘
๐‘›โˆ’1
โˆ‘
๐‘›๐‘˜ โ‰ฅ1
๐‘–=1
โˆ‘๐‘˜ ๐‘›๐‘˜ =๐‘›โˆ’1
(๐‘ˆ๐‘›๐‘œ๐‘Ÿ๐‘‘๐‘’๐‘Ÿ๐‘’๐‘‘!)
Now weโ€™re done!
โ‡’ ๐ฟ๐‘› = (๐‘› โˆ’ 1)! โˆ™ ๐‘Ž๐‘›
๐‘ก(๐‘›1 ,โ€ฆ,๐‘›๐‘–) = โˆ‘
๐‘–=1
1
๐‘–!
โˆ‘
๐‘›๐‘˜ โ‰ฅ1
โˆ‘๐‘˜ ๐‘›๐‘˜ =๐‘›โˆ’1
๐‘‚๐‘Ÿ๐‘‘๐‘’๐‘Ÿ๐‘’๐‘‘!
๐‘ก(๐‘›1 ,โ€ฆ,๐‘›๐‘–)
Lemma 2.5 (๐ฟ๐‘› = ๐‘›๐‘›โˆ’2 , ๐‘› โ‰ฅ 1):
Definition: A rooted tree is a tree with one distinguished vertex called โ€œthe rootโ€ and the edges
are all oriented, and oriented away from the root.
Example:
TODO: Draw an oriented tree
Itโ€™s not possible that a vertex will have 2 incoming edges. So a vertex has 1 incoming degree.
For labeled vertices ๐‘ฃ1 , โ€ฆ , ๐‘ฃ๐‘› , ๐ธ = {โŒฉ๐‘ฃ๐‘– , ๐‘ฃ๐‘— โŒช, ๐‘– โ‰  ๐‘—}
๐‘ ๐‘› = |{(๐‘’1 , โ€ฆ , ๐‘’๐‘› ) โˆˆ ๐ธ ๐‘›โˆ’1 |({๐‘ฃ1 , โ€ฆ , ๐‘ฃ๐‘› }, {๐‘’1 , โ€ฆ , ๐‘’๐‘›โˆ’1 }) โˆ’ ๐‘–๐‘  ๐‘Ž ๐‘Ÿ๐‘œ๐‘œ๐‘ก๐‘’๐‘‘ ๐‘ก๐‘Ÿ๐‘’๐‘’}|
Example:
TODO: Draw vertices example
To choose such a sequence as (was supposed to be drawn) above:
1. Choose a labeled tree on {๐‘ฃ1 , โ€ฆ , ๐‘ฃ๐‘› }
2. Choose a root
3. Choose order in which to add the ๐‘› โˆ’ 1 edges.
Hence, ๐‘ ๐‘› = ๐ฟ๐‘› โˆ™ ๐‘› โˆ™ (๐‘› โˆ’ 1)! = ๐ฟ๐‘› โˆ™ ๐‘›!
Alternatively, ๐‘ ๐‘› = ๐‘ ๐‘›โ€ฒ , where ๐‘†๐‘›โ€ฒ is the number of rooted trees after you remove the last ๐‘˜
edges.
๐‘ ๐‘›โ€ฒ
= |{(๐‘’1 , โ€ฆ , ๐‘’๐‘› ) โˆˆ ๐ธ ๐‘›โˆ’1 |({๐‘ฃ1 , โ€ฆ , ๐‘ฃ๐‘› }, {๐‘’1 , โ€ฆ , ๐‘’๐‘›โˆ’1โˆ’๐‘˜ }) โˆ’ ๐‘–๐‘  ๐‘Ž ๐‘Ÿ๐‘œ๐‘œ๐‘ก๐‘’๐‘‘ ๐‘“๐‘œ๐‘Ÿ๐‘’๐‘ ๐‘ก ๐‘ค๐‘–๐‘กโ„Ž ๐‘˜ + 1 ๐‘๐‘œ๐‘š๐‘๐‘œ๐‘›๐‘’๐‘›๐‘ก๐‘  โˆ€๐‘˜ = 0, โ€ฆ , ๐‘› โˆ’ 1}|
๐‘ ๐‘› โ‰ค ๐‘ ๐‘›โ€ฒ obviously (every sequence in a set of ๐‘ ๐‘› can generate a sequence in ๐‘ ๐‘›โ€ฒ )
But also ๐‘ ๐‘›โ€ฒ โ‰ค ๐‘ ๐‘› , since for ๐‘˜ = 1, every forest is a rooted tree.
Suppose edges ๐‘’1 , โ€ฆ , ๐‘’๐‘›โˆ’1โˆ’๐‘˜ have been added.
Number of choices for ๐‘’๐‘›โˆ’๐‘˜ = (๐‘ข, ๐‘ฃ)
โŸ
๐‘›
โˆ™
โŸ
๐‘˜
๐‘ข ๐‘Ž๐‘Ÿ๐‘๐‘–๐‘ก๐‘Ÿ๐‘Ž๐‘Ÿ๐‘ฆ ๐‘ฃ ๐‘Ž๐‘›๐‘ฆ ๐‘Ÿ๐‘œ๐‘œ๐‘ก ๐‘œ๐‘กโ„Ž๐‘’๐‘Ÿ ๐‘กโ„Ž๐‘Ž๐‘› ๐‘กโ„Ž๐‘’ ๐‘œ๐‘›๐‘’ ๐‘œ๐‘“ ๐ถ(๐‘ข)
Hence
๐‘›โˆ’1
๐‘†๐‘› =
๐‘ ๐‘›โ€ฒ
= โˆ(๐‘›๐‘˜) = ๐‘›๐‘›โˆ’1 โˆ™ (๐‘› โˆ’ 1)! = ๐‘›๐‘›โˆ’2 ๐‘›! = ๐ฟ๐‘› (๐‘›!)
๐‘˜=1
--- end of lesson 5
Theorem 2.3: For ๐‘ƒ๐‘œ๐‘–(๐œ†) = ๐ต. ๐‘ƒ.
๐‘ƒ[๐‘‡ โˆ— = ๐‘›] =
(๐œ† > 0),
(๐œ†๐‘› )๐‘›โˆ’1 โˆ’๐œ†๐‘›
๐‘’
,
๐‘›!
๐‘›โ‰ฅ1
Proof of Lemma 2.4:
(Correction):
๐‘– โ‰ฅ 1,
๐‘›1 , โ€ฆ , ๐‘›๐‘– ,
๐‘›๐‘˜ โ‰ฅ 1,
โˆ‘ ๐‘›๐‘˜ = ๐‘› โˆ’ 1
๐‘˜
TODO: Draw correction drawing
๐‘ก(๐‘›1 , โ€ฆ , ๐‘›๐‘– ) = number of labeled trees on ๐‘› vertices {๐‘ฃ1 , โ€ฆ , ๐‘ฃ๐‘› }, where ๐‘ฃ1 belongs to labeled
edges 1, โ€ฆ , ๐‘–.
And ๐‘ฃ1 has ๐‘›๐‘˜ descendants attached to edge ๐‘˜, ๐‘˜ โˆˆ {1, โ€ฆ , ๐‘–}
Reminder of what we did last time:
๐‘›โˆ’1
(
)
โŸ๐‘›1 , โ€ฆ , ๐‘›๐‘–
๐‘ก(๐‘›1 , โ€ฆ , ๐‘›๐‘– ) =
๐‘œ๐‘“ ๐‘ค๐‘Ž๐‘ฆ๐‘  ๐‘ก๐‘œ
๐‘๐‘Ž๐‘Ÿ๐‘ก๐‘–๐‘ก๐‘–๐‘œ๐‘› {๐‘ฃ2 ,โ€ฆ,๐‘ฃ๐‘› }
๐‘–๐‘›๐‘ก๐‘œ ๐‘”๐‘Ÿ๐‘œ๐‘ข๐‘๐‘  ๐‘œ๐‘“
๐‘ ๐‘–๐‘ง๐‘’ ๐‘›1 ,โ€ฆ,๐‘›๐‘–
๐‘›1 โˆ’2
๐‘› โˆ’2
=
๐‘›
× โ€ฆ × ๐‘›๐‘– ๐‘–
โŸ1
๐‘œ๐‘“ ๐‘โ„Ž๐‘œ๐‘–๐‘๐‘’๐‘  ๐‘œ๐‘“
๐‘™๐‘Ž๐‘๐‘’๐‘™๐‘’๐‘‘ ๐‘ ๐‘ข๐‘๐‘ก๐‘Ÿ๐‘’๐‘’๐‘ 
×
(๐‘›1 × โ€ฆ × ๐‘›๐‘– )
โŸ
๐‘œ๐‘“ ๐‘โ„Ž๐‘œ๐‘–๐‘๐‘’๐‘  ๐‘œ๐‘“
๐‘›๐‘’๐‘–๐‘”โ„Ž๐‘๐‘œ๐‘Ÿ๐‘  ๐‘œ๐‘“ ๐‘ฃ1
๐‘Ž
Remark: (๐‘ , โ€ฆ , ๐‘ ) - is the number of ways to partition ๐‘Ž objects into ๐‘˜ labeled bins of sizes
1
๐‘˜
๐‘1 , โ€ฆ , ๐‘๐‘˜
And the number of labeled trees on ๐‘› vertices equals
๐‘›โˆ’1
โˆ‘
๐‘–=1
1
๐‘–!
โˆ‘
๐‘ก(๐‘›1 ,โ€ฆ,๐‘›๐‘–) โˆŽ
(๐‘›1 ,โ€ฆ,๐‘›๐‘– )
โŸ๐‘›๐‘˜โ‰ฅ1,โˆ‘๐‘˜ ๐‘›โˆ’1
# ๐‘œ๐‘“ ๐‘ก๐‘Ÿ๐‘’๐‘’๐‘  ๐‘ค๐‘–๐‘กโ„Ž deg(๐‘ฃ1 )=๐‘–
Corollary 2.6 (differentiability of the extinction probability):
Let ๐œ‚(๐œ†) = ๐‘ƒ๐œ† [๐‘‡ โˆ— < โˆž], ๐œ† > 1
And let 0 < ๐œ‡ < 1 be the conjugate of ๐œ† (cf. definition 2.1)
So ๐œ‡๐‘’ โˆ’๐œ‡ = ๐œ†๐‘’ โˆ’๐œ†
Then
โˆ’๐œ‚โ€ฒ (๐œ†) =
๐œ‚(๐œ†)(๐œ† โˆ’ ๐œ‡)
<โˆž
๐œ†(1 โˆ’ ๐œ‡)
Proof: Assume ๐œ† > 1 + ๐œ–, ๐œ– > 0
By theorem 2.3:
โˆž
๐œ‚(๐œ†) = โˆ‘
๐‘›=1
(๐œ†๐‘›)๐‘›โˆ’1 โˆ’๐œ†๐‘›
๐‘’
๐‘›!
By using the mean-value theorem, and the dominant convergence theorem, we can check that
we can differentiate term-by-term.
So:
โˆž
๐œ‚
โ€ฒ (๐œ†)
=โˆ‘
๐‘›=1
(๐œ†๐‘›)๐‘›โˆ’1 โˆ’๐œ†๐‘› (๐‘› โˆ’ 1)๐œ†๐‘›โˆ’2 ๐‘›๐‘›โˆ’1 โˆ’๐œ†๐‘›
๐‘›๐‘’
โˆ’
๐‘’
=
๐‘›!
๐‘›!
๐‘๐‘ฆ ๐‘กโ„Ž๐‘’๐‘œ๐‘Ÿ๐‘’๐‘š
1.15 (๐‘‘๐‘ข๐‘Ž๐‘™๐‘–๐‘ก๐‘ฆ)
1
1
1
1
๐ธ๐œ† [๐‘‡ 1{๐‘‡ โˆ— <โˆž} ] (1 โˆ’ ) + ๐œ‚(๐œ†) = ๐ธ๐œ† [๐‘‡ โˆ— |๐‘‡ โˆ— < โˆž] โˆ™ ๐œ‚(๐œ†) (1 โˆ’ ) + ๐œ‚(๐œ†)
=
๐œ†
๐œ†
๐œ†
๐œ†
1
1
๐œ‚(๐œ†) ๐‘๐‘ฆ ๐ธ๐‘ฅ.1.3(?) ๐œ‡(๐œ†) (1 โˆ’ ๐œ†) ๐œ‚(๐œ†) ๐œ‚(๐œ†)(๐œ† โˆ’ 1) + ๐œ‚(๐œ†)(1 โˆ’ ๐œ‡)
โˆ—
๐ธ๐œ‡ [๐‘‡ ] โˆ™ ๐œ‚(๐œ†) (1 โˆ’ ) +
=
+
=
๐œ†
๐œ†
1โˆ’๐œ‡
๐œ†
๐œ†(1 โˆ’ ๐œ‡)
๐œ‚(๐œ†)(๐œ† โˆ’ ๐œ‡)
=
โˆŽ
๐œ†(1 โˆ’ ๐œ‡)
โˆ—
Recall exercise 1.17:
๐‘, ๐‘ž be different distributions on โ„ค, then there exists a distribution ๐‘ on โ„ค2 such that if
(๐‘‘)
(๐‘‘)
(๐‘‘)
(๐‘‹, ๐‘Œ) = ๐‘, then ๐‘‹ = ๐‘, ๐‘Œ = ๐‘ž and ๐‘ƒ[๐‘‹ โ‰  ๐‘Œ] = ๐‘‘ ๐‘‡๐‘‰ (๐‘, ๐‘ž) = 1 โˆ’ โˆ‘๐‘ฅโˆˆโ„ค min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ }
Exercise 2.7: Let (๐‘‹๐‘– )๐‘›๐‘–=1 be iid ๐‘ฅ๐‘– ~๐ต๐‘’(๐‘), i.e. ๐‘ƒ[๐‘‹๐‘– = 1] = ๐‘, ๐‘ƒ[๐‘‹๐‘– = 0] = 1 โˆ’ ๐‘, ๐‘ โˆˆ [0,1],
then show that โˆ‘๐‘›๐‘–=1 ๐‘ฅ๐‘– ~๐ต๐‘–๐‘›(๐‘›, ๐‘)
Exercise 2.8: Let (๐‘‹๐‘– )๐‘›๐‘–=1 be independent, ๐‘ฅ๐‘– ~๐‘ƒ๐‘œ๐‘– (๐œ†๐‘– ), ๐œ†๐‘– > 0
Show that โˆ‘๐‘›๐‘–=1 ๐‘‹๐‘– ~๐‘ƒ๐‘œ๐‘–(โˆ‘๐‘›๐‘–=1 ๐œ†๐‘– )
Theorem 2.9: (coupling of Binomial and Poisson random variables)
For ๐œ† > 0, there exists a โ„ค2 values random variable (๐‘‹, ๐‘Œ) such that:
๐œ†
-
๐‘‹~๐ต๐‘–๐‘› (๐‘›, ๐‘›)
-
๐‘Œ~๐‘ƒ๐‘œ๐‘–(๐œ†)
-
๐‘ƒ[๐‘‹ โ‰  ๐‘Œ] โ‰ค
๐œ†2
๐‘›
Proof: By Ex.1.17, there are iid random variables {(๐ผ๐‘˜ , ๐ฝ๐‘˜ )}๐‘›๐‘˜=1 such that:
๐œ†
-
๐ผ๐‘˜ ~๐ต๐‘’ (๐‘›)
-
๐ฝ๐‘˜ ~๐‘ƒ๐‘œ๐‘– (๐‘›)
-
๐‘ƒ[๐ผ๐‘˜ = ๐ฝ๐‘˜ ] = โˆ‘๐‘ฅโˆˆโ„คโ‰ฅ0 min{๐‘ƒ[๐ผ1 = ๐‘ฅ], ๐‘ƒ[๐ฝ1 = ๐‘ฅ]} =
๐œ†
๐œ†
๐œ† ๐œ† ๐œ† 1โˆ’๐‘ฅโ‰ค๐‘’ โˆ’๐‘ฅ
๐œ† ๐œ† ๐œ†
min {1 โˆ’ , ๐‘’ โˆ’๐‘› } + min { , ๐‘’ โˆ’๐‘› } = 1 โˆ’ + ๐‘’ โˆ’๐‘›
๐‘›
๐‘› ๐‘›
๐‘› ๐‘›
Define:
๐œ†
๐‘›
(๐‘‹, ๐‘Œ) = โˆ‘(๐ผ๐‘˜ , ๐ฝ๐‘˜ )
๐‘˜=1
๐œ†
By exercise 2.7, ๐‘‹~๐ต๐‘–๐‘› (๐‘›, ๐‘›)
By exercise 2.8: ๐‘Œ~๐‘ƒ๐‘œ๐‘–(๐œ†)
Finally:
๐‘›
๐œ† ๐œ† ๐œ†
๐œ†
๐‘ƒ[๐‘‹ โ‰  ๐‘Œ] โ‰ค โˆ‘ ๐‘ƒ[๐ผ๐‘˜ โ‰  ๐ฝ๐‘˜ ] = ๐‘› โˆ™ ( โˆ’ ๐‘’ โˆ’๐‘› ) โ‰ค ๐œ† โˆ™
๐‘› ๐‘›
๐‘›
๐‘˜=1
๐œ†
๐‘›
Exercise 2.10: Let (๐‘‹๐‘› )๐‘›โ‰ฅ1 , ๐‘‹๐‘› ~๐ต๐‘–๐‘› (๐‘›, ) , ๐‘Œ~๐‘ƒ๐‘œ๐‘–(๐œ†), ๐œ† โ‰ฅ 0.
(๐‘‘)
๐‘›โ†’โˆž
Prove that ๐‘‹๐‘› โ†’ ๐‘Œ, meaning: For any ๐ด โŠ† โ„ค, ๐‘ƒ[๐‘‹๐‘› โˆˆ ๐ด] โ†’
๐‘ƒ[๐‘Œ โˆˆ ๐ด]
Theorem 2.11 (Poisson and Binomial branching processes):
๐œ†
๐‘›
Let ๐‘‡ and ๐‘‡ โˆ— be the total progenies of a ๐ต๐‘–๐‘› (๐‘›, ) - branching process ๐‘‡ and of a ๐‘ƒ๐‘œ๐‘–(๐œ†) branching process ๐‘‡ โˆ—, ๐œ† > 0. Then for ๐‘˜ โ‰ฅ 1 ๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜] = ๐‘ƒ[๐‘‡ โˆ— โ‰ฅ ๐‘˜] + ๐‘’(๐‘˜, ๐‘›)
Where |๐‘’(๐‘˜, ๐‘›)| โ‰ค
2๐œ†2 ๐‘˜โˆ’1
โˆ‘๐‘ =1 ๐‘ƒ[๐‘‡ โˆ—
๐‘›
> ๐‘ ] โ‰ค
2๐œ†2 ๐‘˜
๐‘›
Proof: By theorem 2.9, we can define the branching processes:
(๐‘‹1 , ๐‘‹2 , โ€ฆ ) (Bin)
(๐‘‹1โˆ— , ๐‘‹2โˆ— , โ€ฆ ) (Poi)
(random walk perspective)
Such that ๐‘ƒ[๐‘‹๐‘– โ‰  ๐‘‹๐‘–โˆ— ] โ‰ค
๐œ†2
,
๐‘›
๐‘–โ‰ฅ1
๐‘˜ โ‰ฅ 1:
{๐‘‡ โ‰ฅ ๐พ} depends only on {๐‘‹1 , โ€ฆ , ๐‘‹๐‘˜โˆ’1 } - A key observation!
Now:
๐‘˜โˆ’1
๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜, ๐‘‡ < ๐‘˜] โ‰ค โˆ‘ ๐‘ƒ[๐‘‹๐‘– = ๐‘‹๐‘–โˆ— ๐‘“๐‘œ๐‘Ÿ 1 โ‰ค ๐‘– โ‰ค ๐‘  โˆ’ 1, ๐‘‹๐‘  โ‰  ๐‘‹๐‘ โˆ— ๐‘‡ โ‰ฅ ๐‘˜ โ‰ฅ ๐‘ ]
โˆ—
๐‘ =1
We can deduce from ๐‘‡ โ‰ฅ ๐‘˜ โ‰ฅ ๐‘  that ๐‘‡ โˆ— โ‰ฅ ๐‘ !
๐‘˜โˆ’1
โ‰ค โˆ‘ ๐‘ƒ[๐‘‹๐‘  โ‰  ๐‘‹๐‘ โˆ— , ๐‘‡ โˆ— โ‰ฅ ๐‘ ]
๐‘ =1
But these two events are independent
๐‘˜โˆ’1
= โˆ‘ ๐‘ƒ[๐‘‹๐‘  โ‰ 
๐‘ =1
Similarly:
๐‘˜โˆ’1
๐‘‹๐‘ โˆ— ]๐‘ƒ[๐‘‡ โˆ—
๐œ†2
โ‰ฅ ๐‘ ] โ‰ค โˆ‘ ๐‘ƒ[๐‘‡ โˆ— โ‰ฅ ๐‘ ]
๐‘›
๐‘ =1
๐‘˜โˆ’1
๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜, ๐‘‡ < ๐‘˜] โ‰ค โˆ‘ ๐‘ƒ[๐‘‹๐‘– = ๐‘‹๐‘–โˆ— , ๐‘“๐‘œ๐‘Ÿ ๐‘– โ‰ค ๐‘  โˆ’ 1, ๐‘‹๐‘– โ‰  ๐‘‹๐‘–โˆ— , ๐‘‡ โˆ— โ‰ฅ ๐‘˜ โ‰ฅ ๐‘ ] โ‰ค
โˆ—
๐‘ =1
๐‘˜โˆ’1
โˆ‘โŸ
๐‘ƒ[๐‘‹๐‘  โ‰  ๐‘‹๐‘ โˆ— ] ๐‘ƒ[๐‘‡ โˆ— โ‰ฅ ๐‘ ]
๐‘ =1
โ‰ค
๐œ†2
๐‘›
We are interested in: |๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜] โˆ’ ๐‘ƒ[๐‘‡ โˆ— โ‰ฅ ๐‘˜]| โ‰ค
|๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜, ๐‘‡ โˆ— โ‰ฅ ๐‘˜] + ๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜, ๐‘‡ โˆ— < ๐‘˜] โˆ’ ๐‘ƒ[๐‘‡ โˆ— โ‰ฅ ๐‘˜, ๐‘‡ โ‰ฅ ๐‘˜] โˆ’ ๐‘ƒ[๐‘‡ โˆ— โ‰ฅ ๐‘˜, ๐‘‡ < ๐‘˜]| โ‰ค
๐‘˜โˆ’1
2๐œ†2
โˆ‘ ๐‘ƒ[๐‘‡ โˆ— โ‰ฅ ๐‘ ] โˆŽ
๐‘›
๐‘ =1
--- end of lesson 6
Exercise 1.17 ๐‘, ๐‘ž different distributions on โ„ค
(๐‘๐‘ฅ โˆ’ min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ })(๐‘ž๐‘ฆ โˆ’ min{๐‘๐‘ฆ , ๐‘ž๐‘ฆ })
๐‘œ๐‘“ ๐‘ฅ = ๐‘ฆ
๐‘๐‘ฅ,๐‘ฆ = {
๐‘‘ ๐‘‡๐‘‰ (๐‘, ๐‘ž)
min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ } ๐‘–๐‘“ ๐‘ฅ = ๐‘ฆ
(๐‘ฅ, ๐‘ฆ) โˆˆ โ„ค2
1
1
๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž) = โˆ‘|๐‘๐‘ฅ โˆ’ ๐‘ž๐‘ฅ | = ( โˆ‘ (๐‘๐‘ฅ โˆ’ ๐‘ž๐‘ฅ ) + โˆ‘ (๐‘ž๐‘ฅ โˆ’ ๐‘๐‘ฅ )) =
2
2
๐‘ฅโˆˆโ„ค
๐‘ฅ:๐‘๐‘ฅ โ‰ฅ๐‘ž๐‘ฅ
๐‘ฅ:๐‘ž๐‘ฅ <๐‘๐‘ฅ
1
(โˆ‘ ๐‘๐‘ฅ โˆ’ โˆ‘ ๐‘๐‘ฅ โˆ’ โˆ‘ ๐‘ž๐‘ฅ + โˆ‘ ๐‘ž๐‘ฅ โˆ’ โˆ‘ ๐‘๐‘ฅ ) =
2
๐‘ฅโˆˆโ„ค
๐‘ฅ:๐‘๐‘ฅ <๐‘ž๐‘ฅ
๐‘ฅ:๐‘๐‘ฅ โ‰ฅ๐‘ž๐‘ฅ
๐‘ฅ:๐‘๐‘ฅ <๐‘ž๐‘ฅ
๐‘ฅ:๐‘๐‘ฅ <๐‘๐‘ฅ
1
(1 โˆ’ โˆ‘ ๐‘๐‘ฅ โˆ’ โˆ‘ ๐‘ž๐‘ฅ + โˆ‘ ๐‘ž๐‘ฅ โˆ’ โˆ‘ ๐‘ž๐‘ฅ โˆ’ โˆ‘ ๐‘๐‘ฅ ) =
2
๐‘ฅ:๐‘๐‘ฅ <๐‘๐‘ฅ
๐‘ฅ:๐‘๐‘ฅ โ‰ฅ๐‘ž๐‘ฅ
๐‘ฅโˆˆโ„ค
๐‘ฅ:๐‘ž๐‘ฅ โ‰ค๐‘๐‘ฅ
๐‘ฅ:๐‘๐‘ฅ <๐‘ž๐‘ฅ
1
(2 โˆ’ 2 โˆ‘ min{๐‘๐‘ฅ โˆ’ ๐‘ž๐‘ฅ }) = 1 โˆ’ โˆ‘ min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ }
2
๐‘ฅโˆˆโ„ค
๐‘ฅโˆˆโ„ค
๐‘ is a distribution on โ„ค2 :
โˆ‘ โˆ‘ ๐‘๐‘ฅ,๐‘ฆ = โˆ‘ ๐‘๐‘ง,๐‘ง + โˆ‘ โˆ‘ ๐‘๐‘ฅ,๐‘ฆ = โˆ‘ min{๐‘๐‘ง , ๐‘๐‘ง } + โˆ‘(๐‘๐‘ฅ โˆ’ min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ }) = 1
๐‘ฅ
๐‘ฆ
๐‘งโˆˆโ„ค
๐‘ƒ[๐‘‹ = ๐‘ฅ] = โˆ‘ ๐‘๐‘ฅ,๐‘ฆ
๐‘ฆโˆˆโ„ค
๐‘ฅโˆˆโ„ค ๐‘ฆโ‰ ๐‘ฅ
๐‘งโˆˆโ„ค
๐‘ฅโˆˆโ„ค
(๐‘‹, ๐‘Œ)~๐‘
(๐‘๐‘ฅ โˆ’ min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ })๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž)
= min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ } +
= ๐‘๐‘ฅ
๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž)
Similarly, ๐‘ƒ[๐‘Œ = ๐‘ฅ] = ๐‘ž๐‘ฅ
๐‘ƒ[๐‘‹ โ‰  ๐‘Œ] = 1 โˆ’ ๐‘ƒ[๐‘‹ = ๐‘Œ] = 1 โˆ’ โˆ‘ ๐‘๐‘ฅ,๐‘ฅ = 1 โˆ’ โˆ‘ min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ } = ๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž)
๐‘ฅโˆˆโ„ค
๐‘ฅโˆˆโ„ค
Cannot do better!
Assume (๐‘‹ โ€ฒ , ๐‘Œ โ€ฒ ): ๐‘‹ โ€ฒ ~๐‘, ๐‘Œ โ€ฒ ~๐‘ž
๐‘ƒ[๐‘‹ โ€ฒ = ๐‘Œ โ€ฒ ] = โˆ‘ ๐‘ƒ[๐‘‹ โ€ฒ = ๐‘Œ โ€ฒ = ๐‘ฅ]
๐‘ฅโˆˆโ„ค
But ๐‘ƒ[๐‘‹ โ€ฒ = ๐‘Œ โ€ฒ = ๐‘ฅ] โ‰ค ๐‘ƒ[๐‘‹ โ€ฒ = ๐‘ฅ] = ๐‘๐‘ฅ
And also ๐‘ƒ[๐‘‹ โ€ฒ = ๐‘Œ โ€ฒ = ๐‘ฅ] โ‰ค ๐‘ƒ[๐‘Œ โ€ฒ = ๐‘ฅ] = ๐‘ž๐‘ฅ
And both are smaller or equal to min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ } โ‡’
๐‘ƒ[๐‘‹ โ€ฒ โ‰  ๐‘Œ โ€ฒ ] = 1 โˆ’ ๐‘ƒ[๐‘‹ โ€ฒ = ๐‘Œ โ€ฒ ] โ‰ฅ 1 โˆ’ โˆ‘ min{๐‘๐‘ฅ , ๐‘ž๐‘ฅ } = ๐‘‘๐‘‡๐‘‰ (๐‘, ๐‘ž)
๐‘ฅโˆˆโ„ค
3. The Eedos-Renyi random graph
Definition 3.1: E.R. random graphs, ๐บ(๐‘›, ๐‘) for ๐‘› โ‰ฅ 1, ๐‘ โˆˆ [0,1].
Let (๐ผ{๐‘–,๐‘—} )1โ‰ค๐‘–<๐‘—โ‰ค๐‘› be iid random variables with distribution ๐ต๐‘’(๐‘).
Then ๐บ(๐‘›, ๐‘) = ([๐‘›], ๐ธ)
Where [๐‘›] = {1, โ€ฆ , ๐‘›}, ๐ธ = {{๐‘–, ๐‘—} โ‰ค [๐‘›]: ๐‘– โ‰  ๐‘—, ๐ผ{๐‘–,๐‘—} = 1}
Exercise 3.2:
Let ๐‘‹ โ‰” number of triangles in ๐บ(๐‘›, ๐‘)
๐‘›
๐ธ[๐‘‹] = ( ) ๐‘3
3
1
Let ๐‘ = ๐‘›๐›ผ , ๐›ผ > 0
Show that:
-
๐‘›โ†’โˆž
For ๐›ผ > 1, ๐‘ƒ[๐‘‹ > 0] โ†’
๐‘›โ†’โˆž
For ๐›ผ < 1, ๐‘ƒ[๐‘‹ > 0] โ†’
๐›ผ = 1?
What about ๐พ๐‘™ , ๐‘™ โ‰ฅ 4?
(hint, prove that
๐‘ฃ๐‘Ž๐‘Ÿ(๐‘‹) ๐‘›โ†’โˆž
โ†’ 0
๐ธ[๐‘‹]2
0 (can be proven using a chebishev inequality)
1 (should use a second moment method! โ€“ the variant of ๐‘‹)
and use chebishev).
Notation:
- For ๐‘ข, ๐‘ฃ โˆˆ [๐‘›], "๐‘ข โ†” ๐‘ฃโ€ means โ€œThere is a nearest-neighbor path from ๐‘ข to ๐‘ฃ in
๐บ(๐‘›, ๐‘)โ€
- For ๐‘ฃ โˆˆ [๐‘›], ๐ถ(๐‘ฃ) = {๐‘ค โˆˆ [๐‘›]: ๐‘ฃ โ†” ๐‘ค}
Consider the following exploration algorithm for ๐ถ(1) set ๐ด0 = {1}, ๐‘1 = [๐‘›]\{1}
-
For ๐‘ก > 0, if ๐ด๐‘กโˆ’1 = โˆ…, then ๐ด๐‘ก = โˆ…, ๐‘๐‘ก = โˆ… (terminate)
If ๐ด๐‘กโˆ’1 โ‰  โˆ…, let ๐‘ค๐‘กโˆ’1 be the smallest vertex in ๐ด๐‘กโˆ’1 and define:
๐‘ฃ๐‘ก = {๐‘ค โˆˆ ๐‘๐‘กโˆ’1 |{๐‘ค๐‘กโˆ’1 , ๐‘ค} โˆˆ ๐ธ}
๐ด๐‘ก โ‰” (๐ด๐‘กโˆ’1 \{๐‘ค๐‘กโˆ’1 }) โˆช ๐‘‰๐‘ก
๐‘๐‘ก โ‰” ๐‘๐‘กโˆ’1 \๐‘‰๐‘ก
Example:
Draw graphical example.
Note that:
- |๐ด0 | = 1, |๐ด๐‘ก | = |๐ด๐‘กโˆ’1 | + |๐‘‰๐‘ก | โˆ’ 1 for ๐‘ก โ‰ฅ 1
- |๐‘‡๐‘ก | = ๐‘› โˆ’ |๐ด๐‘ก | โˆ’ ๐‘ก, ๐‘ก โ‰ฅ 0
- |๐ถ(1)| = inf{๐‘ก โ‰ฅ 1 | |๐ด๐‘ก | = 0}
Lemma 3.3: For ๐‘ฅ1 , โ€ฆ , ๐‘ฅ๐‘˜ โˆˆ โ„คโ‰ฅ0 , let:
- ๐‘ 0 = 1, ๐‘ ๐‘ก = ๐‘ ๐‘กโˆ’1 + ๐‘ฅ๐‘ก โˆ’ 1 for ๐‘ก โ‰ฅ 1
- ๐‘›๐‘ก = ๐‘› โˆ’ ๐‘ ๐‘ก โˆ’ ๐‘ก
If ๐‘ 1 , ๐‘ 2 , โ€ฆ , ๐‘ ๐‘˜ โ‰ฅ 1, ๐‘›0 , ๐‘›1 , โ€ฆ , ๐‘›๐‘˜โˆ’1 โ‰ฅ 0
Then
๐‘˜
๐‘ƒ[|๐‘‰1 | = ๐‘ฅ1 , โ€ฆ , |๐‘‰๐‘˜ | = ๐‘ฅ๐‘˜ ] = โˆ(๐‘(๐‘›๐‘กโˆ’1 , ๐‘, ๐‘ฅ๐‘ก ))
๐‘›
Where ๐‘(๐‘›, ๐‘, ๐‘˜) = ( ) ๐‘๐‘˜ (1 โˆ’ ๐‘)๐‘›โˆ’๐‘˜
๐‘˜
๐‘ก=1
Proof: By picture.
TODO: Draw picture proof.
๐‘ƒ[|๐‘‰1 | = ๐‘ฅ1 , โ€ฆ , |๐‘‰๐‘˜ | = ๐‘ฅ๐‘˜ ] =
โˆ‘ โˆ‘ ๐‘ƒ [|๐‘‰1 | = ๐‘ฅ1 , โ€ฆ , |๐‘‰๐‘˜โˆ’1 | = ๐‘ฅ๐‘˜โˆ’1 , ๐ด๐‘˜โˆ’1 = ๐ด, ๐‘๐‘˜โˆ’1 = ๐‘, โˆ‘ ๐ผ{๐‘Ž,๐‘ค} = ๐‘ฅ๐‘˜ ]
๐ดโŠ†[๐‘›] ๐‘โŠ†[๐‘›]
๐ดโ‰ โˆ…
๐‘คโˆˆ๐‘
๐‘Ž is the smallest vertex in ๐ด.
{|๐‘‰1 } = ๐‘ฅ1 , โ€ฆ , ๐‘๐‘กโˆ’1 = ๐‘} โˆˆ ๐œŽ(๐ผ๐‘’ : ๐‘’ ๐‘›๐‘œ๐‘ก ๐‘Ž๐‘› ๐‘’๐‘‘๐‘”๐‘’ ๐‘“๐‘Ÿ๐‘œ๐‘š ๐‘Ž ๐‘ก๐‘œ ๐‘)
{ โˆ‘ ๐ผ{๐‘Ž,๐‘ค} = ๐‘ฅ๐‘˜ } โˆˆ ๐œŽ(๐ผ๐‘’ : ๐‘’ ๐‘’๐‘‘๐‘”๐‘’ ๐‘“๐‘Ÿ๐‘œ๐‘š {๐‘Ž} ๐‘ก๐‘œ ๐‘)
๐‘คโˆˆ๐‘
So these two events must be independent!
So:
๐‘ƒ[|๐‘‰1 | = ๐‘ฅ1 , โ€ฆ , |๐‘‰๐‘˜ | = ๐‘ฅ๐‘˜ ] = โˆ‘ โˆ‘ ๐‘ƒ[|๐‘‰1 | = ๐‘ฅ1 , โ€ฆ , ๐‘๐‘˜โˆ’1 = ๐‘] โˆ™ ๐‘(๐‘›๐‘˜โˆ’1 , ๐‘, ๐‘ฅ๐‘˜ )
If the probability of ๐‘ƒ[|๐‘‰1 | = ๐‘ฅ1 , โ€ฆ , ๐‘๐‘˜โˆ’1 = ๐‘] is positive then necessarily |๐‘| = |๐‘๐‘˜โˆ’1 | = ๐‘›๐‘˜โˆ’1
โ‡’ ๐‘ƒ[|๐‘‰1 | = ๐‘ฅ1 , โ€ฆ , |๐‘‰๐‘˜ | = ๐‘ฅ๐‘˜ ] = ๐‘ƒ[๐‘‰1 = ๐‘ฅ! , โ€ฆ , |๐‘‰๐‘˜โˆ’1 | = ๐‘ฅ๐‘˜โˆ’1 ] × ๐‘(๐‘›๐‘˜โˆ’1 , ๐‘, ๐‘ฅ๐‘˜ )
So now we can use induction to complete the proof.
--- end of lesson
Definition: 3.4:
๐‘Œ
๐‘กโˆ’1
Let (๐ผ๐‘ก,๐‘– )๐‘ก,๐‘–โ‰ฅ1 be iid ๐ต๐‘’(๐‘). Define ๐‘Œ0 = ๐‘› โˆ’ 1, for ๐‘ก โ‰ฅ 1 : โˆ‘๐‘–=1
๐ผ๐‘ก,๐‘– , ๐‘Œ๐‘ก = ๐‘Œ๐‘ก โˆ’ ๐‘‹๐‘ก , ๐‘†๐‘ก =
(๐‘‹1 + โ‹ฏ + ๐‘‹๐‘ก ) โˆ’ (๐‘ก โˆ’ 1)
๐‘‹๐‘ก ~|๐‘‰๐‘ก |,
๐‘Œ๐‘ก ~|๐‘๐‘ก |,
๐‘†๐‘ก ~|๐ด๐‘ก |
Lemma 3.5 For ๐‘ฅ๐‘– , ๐‘ ๐‘– , ๐‘›๐‘– as in lemma 3.3,
๐‘˜
โˆ ๐‘(๐‘›๐‘กโˆ’1 , ๐‘, ๐‘ฅ๐‘ก ) ,
๐‘ƒ[๐‘‹1 = ๐‘ฅ1 , โ€ฆ , ๐‘‹๐‘˜ = ๐‘ฅ๐‘˜ ] = {
๐‘–๐‘“ ๐‘›1 , โ€ฆ , ๐‘›๐‘กโˆ’1 โ‰ฅ 0
๐‘ก=0
0,
๐‘œ๐‘กโ„Ž๐‘’๐‘Ÿ๐‘ค๐‘–๐‘ ๐‘’
Proof: {๐‘‹1 = ๐‘ฅ1 , โ€ฆ , ๐‘‹๐‘˜ = ๐‘ฅ๐‘˜ } โŠ† {๐‘Œ1 = ๐‘›1 , โ€ฆ , ๐‘Œ๐‘˜โˆ’1 = ๐‘›๐‘˜โˆ’1 }
So if ๐‘›๐‘– < 0, then ๐‘ƒ[โ€ฆ ] = 0. Otherwise:
๐‘ƒ[๐‘‹1 = ๐‘ฅ1 , โ€ฆ , ๐‘‹๐‘˜โˆ’1 = ๐‘ฅ๐‘˜โˆ’1 , ๐‘‹๐‘˜ = ๐‘ฅ๐‘˜ ] = ๐‘ƒ[๐‘‹1 = ๐‘ฅ1 , โ€ฆ , ๐‘‹๐‘˜โˆ’1 = ๐‘ฅ๐‘˜โˆ’1 ] โˆ™ ๐‘(๐‘›๐‘˜โˆ’1 , ๐‘, ๐‘ฅ๐‘˜ )
๐‘๐‘ฆ ๐‘‘๐‘’๐‘“
๐‘›๐‘˜โˆ’1
Because ๐‘‹๐‘˜ = โˆ‘๐‘–=1
๐ผ๐‘˜,๐‘–
We can continue by induction. And that finishes the proof. โˆŽ
Corollary 3.6: For ๐‘˜ โ‰ฅ 1, and any ๐‘“: โ„ค๐‘˜โ‰ฅ0 โ†’ โ„
๐ธ[๐‘“(|๐‘‰1 |, โ€ฆ , |๐‘‰๐‘˜ |)1{|๐ด1 |>0,โ€ฆ,|๐ด๐‘˜โˆ’1 |>0} ] = ๐ธ[๐‘“(๐‘‹1 , โ€ฆ , ๐‘‹๐‘˜ )1{๐‘†1 >0,โ€ฆ,๐‘†๐‘˜โˆ’1 >0} ]
Proof: Combine Lemma 3.3 and Lemma 3.5. โˆŽ
Theorem 3.7: Let ๐‘‡ โˆ’ and ๐‘‡ + be the total progenies of a ๐ต๐‘–๐‘›(๐‘› โˆ’ ๐‘˜, ๐‘) and ๐ต๐‘–๐‘›(๐‘›, ๐‘) branching
processes.
Where 1 โ‰ค ๐‘˜ โ‰ค ๐‘› and ๐‘ โˆˆ [0,1] then ๐‘ƒ[๐‘‡ โˆ’ โ‰ฅ ๐‘˜] โ‰ค ๐‘ƒ[|๐ถ(1)| โ‰ฅ ๐‘˜] โ‰ค ๐‘ƒ[๐‘‡ + โ‰ฅ ๐‘˜].
+/โˆ’
Proof: For (๐ผ๐‘ก,๐‘– ) as in Definition 3.4, define ๐‘‹๐‘ก
+/โˆ’
Then (๐‘‹๐‘ก ) are iid~๐ต๐‘–๐‘›(๐‘›/๐‘› โˆ’ ๐‘˜, ๐‘).
๐‘กโ‰ฅ1
+/โˆ’
+/โˆ’
+/โˆ’
๐‘†๐‘ก : = (๐‘‹1 + โ‹ฏ + ๐‘‹๐‘ก ) โˆ’ (๐‘ก โˆ’ 1),
(๐‘‘)
+/โˆ’
Then ๐‘‡ +/โˆ’ = inf {๐‘ก โ‰ฅ 1|๐‘†๐‘ก = 0}
๐‘›/๐‘›โˆ’๐‘˜
= โˆ‘๐‘–=1
๐‘ก โ‰ฅ 1,
๐ผ๐‘ก,๐‘– , ๐‘ก โ‰ฅ 1.
+/โˆ’
๐‘†0
=1
๐‘ƒ[|๐ถ(1)| โ‰ฅ ๐‘˜] = ๐‘ƒ[|๐ด1 | > 0, โ€ฆ , |๐ด๐‘˜โˆ’1 | > 0] = ๐‘ƒ[๐‘†1 > 0, โ€ฆ , ๐‘†๐‘˜โˆ’1 > 0] โ‰ค
+
๐‘ƒ[๐‘†1+ > 0, โ€ฆ , ๐‘†๐‘˜โˆ’1
> 0] = ๐‘ƒ[๐‘‡ + โ‰ฅ ๐‘˜]
(by cor 3.6 if ๐‘“ โ‰ก 1) ๐‘†๐‘ก โ‰ค ๐‘†๐‘ก+ , ๐‘ก โ‰ค ๐‘˜ โˆ’ 1
On the other hand, ๐‘ƒ[|๐ถ(1)| โ‰ค ๐‘˜] = ๐‘ƒ[โ‹ƒ๐‘˜๐‘ก=1{|๐‘‰๐‘ก | = 0, |๐ด๐‘กโˆ’1 | = 1, |๐ด๐‘กโˆ’2 | > 0, โ€ฆ , |๐ด1 | > 0}]
(exploration stops at ๐‘ก)
๐‘˜
= โˆ‘ ๐‘ƒ[๐‘‹๐‘ก = 0, ๐‘†๐‘กโˆ’1 = 1, ๐‘†๐‘กโˆ’2 > 0, โ€ฆ , ๐‘†1 > 0]
๐‘ก=1
(by Cor 3.6)
= ๐‘ƒ[๐‘‡ โ‰ค ๐‘˜] where ๐‘‡ = inf{๐‘ก โ‰ฅ 1|๐‘†๐‘ก = 0}
Claim: If {๐‘‡ โ‰ค ๐‘˜} โŠ† {๐‘‡ โˆ’ โ‰ค ๐‘˜}
Indeed, if ๐‘‡ โ‰ค ๐‘˜ ๐‘Œ๐‘‡ = ๐‘› โˆ’ 1 โˆ’ ๐‘‹1 โˆ’ โ‹ฏ โˆ’ ๐‘‹๐‘‡ = ๐‘› โˆ’ ๐‘‡ โˆ’ ๐‘†๐‘‡ โ‰ฅ ๐‘› โˆ’ ๐‘˜
So ๐‘Œ1 โ‰ฅ ๐‘Œ2 โ‰ฅ โ‹ฏ โ‰ฅ ๐‘Œ๐‘‡ โ‰ฅ ๐‘› โˆ’ ๐‘˜
Hence, ๐‘‹๐‘กโˆ’ โ‰ค ๐‘‹๐‘ก , ๐‘ก โ‰ค ๐‘‡
๐‘†๐‘กโˆ’ โ‰ค ๐‘†๐‘ก ,
๐‘ก โ‰ค ๐‘‡ โ‡’ ๐‘‡โˆ’ โ‰ค ๐‘˜
Which proves this claim. โˆŽ
Exercise 3.8:
Consider random variables ๐‘, ๐‘€ such that ๐‘~๐ต๐‘–๐‘›(๐‘›, ๐‘) and conditionally on ๐‘,
๐‘€~๐ต๐‘–๐‘›(๐‘, ๐‘ž), ๐‘› โ‰ฅ 1, ๐‘, ๐‘ž โˆˆ [0,1]
๐‘˜
I.e. ๐‘ƒ[๐‘€ = ๐‘™|๐‘ = ๐‘˜] = ( ) ๐‘ž ๐‘™ (1 โˆ’ ๐‘ž)๐‘˜โˆ’๐‘™
๐‘™
(0 โ‰ค ๐‘™ โ‰ค ๐‘˜ โ‰ค ๐‘›)
Prove that ๐‘€~๐ต๐‘–๐‘›(๐‘›, ๐‘๐‘ž)
It would be probable to assume: |๐‘๐‘ก |~๐ต๐‘–๐‘›(๐‘› โˆ’ 1, (1 โˆ’ ๐‘)๐‘ก )
But this is not correctโ€ฆ
Proposition 3.9:
For 0 โ‰ค ๐‘ก โ‰ค ๐‘›, ๐‘Œ๐‘ก ~๐ต๐‘–๐‘›(๐‘› โˆ’ 1, (1 โˆ’ ๐‘)๐‘ก )
Proof: By induction.
๐‘Œ0 = ๐‘› โˆ’ 1. This is true.
Assume that ๐‘Œ๐‘ก ~๐ต๐‘–๐‘›(๐‘› โˆ’ 1, (1 โˆ’ ๐‘)๐‘ก )
Need to check that conditionally on ๐‘Œ๐‘ก , ๐‘Œ๐‘ก+1 ~๐ต๐‘–๐‘›(๐‘Œ๐‘ก , 1 โˆ’ ๐‘)
(By Ex 3.6) 0 โ‰ค ๐‘™ โ‰ค ๐‘˜ โ‰ค ๐‘› โˆ’ 1
๐‘ƒ[๐‘Œ๐‘ก+1 = ๐‘™. ๐‘Œ๐‘ก = ๐‘˜],
๐‘Œ๐‘ก+1 = ๐‘Œ๐‘ก โˆ’ ๐‘‹๐‘ก+1
= ๐‘ƒ[ ๐‘‹
โŸ
๐‘ก+1
= ๐‘˜ โˆ’ ๐‘™, ๐‘Œ๐‘ก = ๐‘˜]
๐‘–๐‘›๐‘‘๐‘’๐‘๐‘’๐‘›๐‘‘๐‘’๐‘›๐‘๐‘’
=
๐‘(๐‘˜, ๐‘, ๐‘˜ โˆ’ ๐‘™)๐‘ƒ[๐‘Œ๐‘ก = ๐‘˜] =
=โˆ‘๐‘˜
๐‘–=1 ๐ผ๐‘ก+1,๐‘–
๐‘(๐‘˜, (1 โˆ’ ๐‘), ๐‘™)๐‘ƒ[๐‘Œ๐‘ก = ๐‘˜] โˆŽ
๐‘˜
(
) ๐‘๐‘˜โˆ’๐‘™ (1 โˆ’ ๐‘)๐‘™
๐‘˜โˆ’๐‘™
๐ผ๐œ† = ๐œ† โˆ’ 1 โˆ’ log ๐œ† ,
(see Ex. 1.16, ๐ผ(๐œ†))
๐œ†>0
๐œ†
Lemma 3.8: For ๐‘ = ๐‘› , ๐œ† โˆˆ (0,1)
๐‘˜ โ‰ฅ 1,
๐‘ƒ[|๐ถ(1)| > ๐‘˜] โ‰ค ๐‘’ โˆ’๐ผ๐œ† ๐‘˜
Proof: Let ๐‘Ÿ > 0, ๐‘Ÿ = โˆ’ log ๐œ† > 0
๐‘ƒ[|๐ถ(1)| > ๐‘˜] = ๐‘ƒ[|๐ด1 | > 0, โ€ฆ , |๐ด๐‘˜ | > 0]
๐ถ๐‘œ๐‘Ÿ 3.6
= โ‰ค ๐‘ƒ[๐‘†1 > 0, โ€ฆ , ๐‘†๐‘˜ > 0] โ‰ค ๐‘ƒ[๐‘†๐‘˜ > 0] =
๐‘ƒ ๐‘†
โŸ๐‘˜ + ๐‘˜ โˆ’ 1 โ‰ฅ ๐‘˜
๐‘’๐‘ž๐‘ข๐‘–๐‘ฃ๐‘Ž๐‘™๐‘’๐‘›๐‘ก ๐‘ก๐‘œ:
[
]
๐‘†๐‘˜ + ๐‘˜ โˆ’ 1 = ๐‘‹1 + โ‹ฏ + ๐‘‹๐‘˜ = ๐‘› โˆ’ 1 โˆ’ ๐‘Œ๐‘˜ ~๐ต๐‘–๐‘›(๐‘› โˆ’ 1,1 โˆ’ (1 โˆ’ ๐‘)๐‘˜ )
๐‘’ ๐‘Ÿ(๐‘†๐‘˜ +๐‘˜โˆ’1) โ‰ฅ๐‘’ ๐‘Ÿ๐‘˜
So ๐‘ƒ[|๐ถ(1)| > ๐‘˜] โ‰ค ๐‘ƒ[๐‘’ ๐‘Ÿ(๐‘†๐‘˜ +๐‘˜โˆ’1) โ‰ฅ ๐‘’ ๐‘Ÿ๐‘˜ ]
๐‘๐‘ฆ ๐‘€๐‘Ž๐‘Ÿ๐‘˜๐‘œ๐‘ฃ
โ‰ค
๐‘’ โˆ’๐‘Ÿ๐‘˜ ๐ธ[๐‘’ ๐‘Ÿ(๐‘†๐‘˜ +๐‘˜โˆ’1) ]
๐‘›โˆ’1
๐ธ[๐‘’ ๐‘Ÿ(๐‘†๐‘˜ +๐‘˜โˆ’1) ] = ((1 โˆ’ ๐‘ž) + ๐‘ž๐‘’ ๐‘Ÿ )
(๐‘‘)
๐‘†๐‘˜ + ๐‘˜ โˆ’ 1 = ๐œ‰1 + โ‹ฏ + ๐œ‰๐‘›โˆ’1 , ๐œ‰๐‘– iid ~๐ต๐‘’(๐‘ž)
๐‘›โˆ’1
โ‡’ ๐ธ[๐‘’
๐‘Ÿ(๐‘†๐‘˜ +๐‘˜โˆ’1)
] = ๐ธ [๐‘’
๐‘Ÿ(โˆ‘๐‘›โˆ’1
๐‘–=1 ๐œ‰๐‘– )
๐‘›โˆ’1
] = โˆ(๐ธ[๐‘’ ๐‘Ÿ๐œ‰๐‘– ]) = (๐ธ[๐‘’ ๐‘Ÿ๐œ‰1 ])
๐‘›โˆ’1
= ((1 โˆ’ ๐‘ž) + ๐‘ž๐‘’ ๐‘Ÿ )
๐‘–=1
So
๐‘ƒ[|๐ถ(1)| > ๐‘˜] โ‰ค ๐‘’
โˆ’๐‘Ÿ๐‘˜
๐‘Ÿ ๐‘›โˆ’1
((1 โˆ’ ๐‘ž) + ๐‘ž๐‘’ )
(1โˆ’๐‘ฅ)โ‰ค๐‘’ โˆ’๐‘ฅ
โ‰ค
๐‘Ÿ
exp(โˆ’๐‘Ÿ๐‘˜ + ๐‘ž(๐‘’ โˆ’ 1)(๐‘› โˆ’ 1)) โ‰ค exp(โˆ’๐‘Ÿ๐‘˜ + ๐‘ž(๐‘’ ๐‘Ÿ โˆ’ 1)๐‘›)
๐‘ž = 1 โˆ’ (1 โˆ’ ๐‘)๐‘˜ โ‰ค ๐‘๐‘˜ (โˆ—)
(comment: Seems like ๐‘ก turned into ๐‘˜)
โ‡’ ๐‘ƒ[|๐ถ(1)| > ๐‘˜] โ‰ค exp(โˆ’๐‘Ÿ๐‘˜ + ๐‘ž(๐‘’ ๐‘Ÿ โˆ’ 1)๐‘›) โ‰ค exp(โˆ’๐‘Ÿ๐‘˜ + ๐‘˜๐‘(๐‘’ ๐‘Ÿ โˆ’ 1)๐‘›) =
๐‘Ÿ
๐‘=
๐œ†
๐‘›
exp (โˆ’๐‘˜(๐‘Ÿ โˆ’ ๐‘๐‘›(๐‘’ โˆ’ 1))) = exp (โˆ’๐‘˜(๐‘Ÿ โˆ’ ๐œ†(๐‘’ ๐‘Ÿ โˆ’ 1)))
If we try to minimize the expression, we get that ๐‘Ÿ = โˆ’ log ๐œ†
So letโ€™s just set it in:
๐‘’ โˆ’๐‘˜๐ผ๐œ†
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œ†
1
Fix ๐œ† โˆˆ (0,1), ๐‘ = ๐‘› , ๐‘Ž > ๐ผ
๐œ†
Then there is a ๐›ฟ = ๐›ฟ(๐‘Ž, ๐œ†) such that ๐‘ƒ[|๐ถ๐‘š๐‘Ž๐‘ฅ | > ๐‘Ž log ๐‘›] < ๐‘›โˆ’๐›ฟ
Proof:
๐‘ƒ[|๐ถ๐‘š๐‘Ž๐‘ฅ | > ๐‘Ž log ๐‘›] = ๐‘ƒ [ โ‹ƒ {|๐ถ(๐‘ฃ)| > ๐‘Ž log ๐‘›}] โ‰ค โˆ‘ ๐‘ƒ[|๐ถ(๐‘ฃ)| > ๐‘Ž log ๐‘›] =
๐‘ฃโˆˆ[๐‘›]
๐‘๐‘ฆ ๐ฟ๐‘’๐‘š๐‘š๐‘Ž 3.8
๐‘›๐‘ƒ[|๐ถ(1)| > ๐‘Ž log ๐‘›]
โ‰ค
๐‘ฃโˆˆ[๐‘›]
๐‘› โˆ™ ๐‘›โˆ’๐ผ๐œ†๐‘Ž ,
๐ผ๐œ† ๐‘Ž > 1 = โˆŽ
Next, lower bound:
Lemma 3.10:
Let ๐‘โ‰ฅ๐‘˜ = โˆ‘๐‘›๐‘ฃ=1 1{|๐ถ(๐‘ฃ)|โ‰ฅ๐‘˜}
๐œ’โ‰ฅ๐‘˜ = ๐ธ[|๐ถ(1)|1{|๐ถ(1)|โ‰ฅ๐‘˜} ]
Then ๐‘ฃ๐‘Ž๐‘Ÿ(๐‘โ‰ฅ๐‘˜ ) โ‰ค ๐‘›๐œ’โ‰ฅ๐‘˜
2
2
Proof: ๐‘ฃ๐‘Ž๐‘Ÿ(๐‘โ‰ฅ๐‘˜ ) = ๐ธ [(โˆ‘๐‘›๐‘ฃ=1 1{|๐ถ(๐‘ฃ)|โ‰ฅ๐‘˜} ) ] โˆ’ ๐ธ[โˆ‘๐‘›๐‘ฃ=1 1{|๐ถ(๐‘ฃ)|โ‰ฅ๐‘˜} ] =
๐‘›
๐‘›
โˆ‘ โˆ‘ ๐‘ƒ[|๐ถ(๐‘–)| โ‰ฅ ๐‘˜, |๐ถ(๐‘—)| โ‰ฅ ๐‘˜] โˆ’ ๐‘ƒ[|๐ถ(๐‘–)| โ‰ฅ ๐‘˜]๐‘ƒ[|๐ถ(๐‘—)| โ‰ฅ ๐‘˜]
๐‘–=1 ๐‘—=1
Split according to ๐‘– โ†” ๐‘— and ๐‘– โ†” ๐‘—
๐‘– connected to ๐‘— and ๐‘– not connected to ๐‘—
๐‘›
๐‘›
๐‘ƒ[|๐ถ(๐‘–)| โ‰ฅ ๐‘˜, |๐ถ(๐‘—)| โ‰ฅ ๐‘˜, ๐‘– โ†” ๐‘—] = โˆ‘ โˆ‘ ๐‘ƒ[|๐ถ(๐‘–)| = ๐‘™, ๐‘– โ†” ๐‘—, |๐ถ(๐‘—)| = ๐‘š] =
๐‘™=๐‘˜ ๐‘š=๐‘˜
๐‘›
โˆ‘ ๐‘ƒ[|๐ถ(๐‘–) = ๐‘™|, ๐‘– โ†” ๐‘—] โˆ™ ๐‘ƒ [|๐ถ(๐‘—)| = ๐‘š] = โˆ‘ ๐‘ƒ[|๐ถ(๐‘–)| = ๐‘™, ๐‘– โ†” ๐‘—] โˆ™ ๐‘ƒ [|๐ถ(๐‘—)| โ‰ฅ ๐‘˜] โ‰ค
(๐‘›โˆ’๐‘™)
๐‘™,๐‘š
๐‘›
(๐‘›โˆ’๐‘™)
๐‘™=๐‘˜
๐‘›
โˆ‘ ๐‘ƒ[|๐ถ(๐‘–)| = ๐‘™, ๐‘– โ†” ๐‘—] โˆ™ ๐‘ƒ [|๐ถ(๐‘—)| โ‰ฅ ๐‘˜] = โˆ‘ ๐‘ƒ[|๐ถ(๐‘–)| = ๐‘™] โˆ™ ๐‘ƒ[|๐ถ(๐‘—)| โ‰ฅ ๐‘˜]
(๐‘›โˆ’๐‘™)
๐‘™=๐‘˜
๐‘™=๐‘˜
(we can strike these out since it only increases the probability of the event)
So letโ€™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‘›
๐‘ฃ๐‘Ž๐‘Ÿ(๐‘โ‰ฅ๐‘˜ ) โ‰ค โˆ‘ โˆ‘ ๐‘ƒ[|๐ถ(๐‘–)| โ‰ฅ ๐‘˜, ๐ถ(๐‘—) โ‰ฅ ๐‘˜, ๐‘– โ†” ๐‘—]
๐‘–=1 ๐‘—=1
๐‘–โ†”๐‘—โ‡’
๐‘‡โ„Ž๐‘’๐‘ ๐‘’ ๐‘ ๐‘๐‘’๐‘๐‘–๐‘“๐‘ฆ ๐‘›
๐‘กโ„Ž๐‘’ ๐‘ ๐‘Ž๐‘š๐‘’ ๐‘’๐‘ฃ๐‘’๐‘›๐‘ก
=
๐‘›
โˆ‘ โˆ‘ ๐‘ƒ[|๐ถ(๐‘–)| โ‰ฅ ๐‘˜] =
๐‘–=1 ๐‘—=1
๐‘›
๐‘›
โˆ‘ ๐ธ 1{|๐ถ(๐‘–)|โ‰ฅ๐‘˜} โˆ™ โˆ‘ 1{๐‘—โˆˆ๐ถ(๐‘–)} = ๐‘›๐œ’โ‰ฅ๐‘˜
๐‘–=1
๐‘—=1
โŸ
[
]
=|๐ถ(๐‘–)|
Theorem 3.11 (lower bound on largest subcritical component):
๐œ†
Fix ๐œ† โˆˆ (0,1), ๐‘œ = ๐‘›
1
Then for every ๐‘Ž < ๐ผ there exists ๐›ฟ = ๐›ฟ(๐‘Ž, ๐œ†), and ๐‘ = ๐‘(๐‘Ž, ๐œ†) such that
๐œ†
๐‘ƒ[|๐ถ๐‘š๐‘Ž๐‘ฅ | โ‰ค ๐‘Ž log ๐‘›] โ‰ค ๐‘๐‘›โˆ’๐›ฟ
Proof: Let ๐‘˜ = โŒŠ๐‘Ž log ๐‘›โŒ‹, ๐‘ denotes a constant depending only on ๐œ†, ๐‘Ž changing from place to
place.
Claim 1: For any 0 < ๐›ผ < 1 โˆ’ ๐ผ๐œ† ๐‘Ž
๐ธ[๐‘โ‰ฅ๐‘˜ ] โ‰ฅ ๐‘›๐›ผ , for ๐‘› โ‰ฅ ๐‘
Claim 2: ๐‘ฃ๐‘Ž๐‘Ÿ(๐‘โ‰ฅ๐‘˜ ) โ‰ค ๐‘๐‘˜๐‘›1โˆ’๐‘Ž๐ผ๐œ†
From these claims, it follows that we can prove this theorem.
Why?
๐‘ƒ[|๐ถ๐‘š๐‘Ž๐‘ฅ | โ‰ค ๐‘Ž log ๐‘›] = ๐‘ƒ[๐‘โ‰ฅ๐‘˜ = 0] โ‰ค ๐‘ƒ[|๐‘๐‘˜โ‰ฅ0 โˆ’ ๐ธ[๐‘โ‰ฅ๐‘˜ ]| โ‰ฅ ๐ธ[๐‘โ‰ฅ๐‘˜ ]] =
๐‘๐‘ฆ ๐‘กโ„Ž๐‘’
๐‘ƒ[|๐‘๐‘˜โ‰ฅ0 โˆ’ ๐ธ[๐‘โ‰ฅ๐‘˜
]|2
โ‰ฅ ๐ธ[๐‘โ‰ฅ๐‘˜
]2 ]
๐‘ฃ๐‘Ž๐‘Ÿ(๐‘โ‰ฅ๐‘˜ ) ๐‘๐‘™๐‘Ž๐‘–๐‘š๐‘  ๐‘๐‘˜๐‘›1โˆ’๐‘Ž๐ผ๐œ†
โ‰ค
โ‰ค
๐ธ[๐‘โ‰ฅ๐‘˜ ]2
๐‘›2๐›ผ
2
(For instance, if ๐›ผ = 3 (1 โˆ’ ๐‘Ž๐ผ๐œ† ) then itโ€™s โ‰ค
๐‘ log ๐‘›
๐‘›
(1โˆ’๐‘Ž๐ผ๐œ† )
3
)
Letโ€™s prove claim 1:
๐œ†
Let ๐‘‡, ๐‘‡ โˆ— be the total progenies of ๐ต๐‘–๐‘› (๐‘› โˆ’ ๐‘˜, ๐‘›) branching process, and ๐‘ƒ๐‘œ๐‘–(๐œ†๐‘› ) branching
process wher ๐œ†๐‘› =
๐œ†(๐‘›โˆ’๐‘˜)
๐‘›
By theorem 3.7, ๐‘ƒ[|๐ถ(1)| โ‰ฅ ๐‘˜] โ‰ฅ ๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜]
๐‘๐‘ฆ
๐‘กโ„Ž๐‘’๐‘œ๐‘Ÿ๐‘’๐‘š 2.11
โ‰ฅ
๐œ†
1
๐œ†(๐‘› โˆ’ ๐‘˜)
=
(
)
๐‘› ๐‘›โˆ’๐‘˜โŸ ๐‘›
๐œ†๐‘›
โˆ—
โˆ—
๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜] โ‰ฅ ๐‘ƒ[๐‘‡ = ๐‘˜]
๐‘๐‘ฆ
๐‘กโ„Ž๐‘’๐‘œ๐‘Ÿ๐‘’๐‘š 2.3
=
We can use sterlingโ€™s formula:
(๐œ†๐‘› ๐‘˜)๐‘˜โˆ’1 โˆ’๐œ† ๐‘˜
๐‘’ ๐‘›
๐‘˜!
๐‘ƒ[๐‘‡ โˆ— โ‰ฅ ๐‘˜] โˆ’
2๐œ†2๐‘› ๐‘˜
๐‘›
๐‘˜ ๐‘˜
๐‘˜! = ( ) โˆ™ โˆš2๐œ‹๐‘˜(1 + ๐‘œ(1))
๐‘’
And then:
โ‰ฅ
๐œ†๐‘˜๐‘›
๐‘’
โˆ’๐œ†๐‘› ๐‘˜ ๐‘˜
๐‘’ (1 + ๐‘œ(1)) โ‰ฅ ๐‘’
โˆ’๐ผ๐œ†๐‘› ๐‘˜(1+๐œ–)
๐‘“๐‘œ๐‘Ÿ ๐‘ ๐‘ข๐‘“๐‘“๐‘–๐‘๐‘–๐‘’๐‘›๐‘ก๐‘™๐‘ฆ
๐‘™๐‘Ž๐‘Ÿ๐‘”๐‘’ ๐‘›
๐‘˜โˆš2๐œ‹๐‘˜๐œ†๐‘›
(๐ผ๐œ† = ๐œ† โˆ’ 1 โˆ’ log ๐œ†)
=
๐‘’ โˆ’๐ผ๐œ† ๐‘˜(1+๐œ–)
For ๐œ– โ‰ฅ 0: 1 โˆ’ ๐‘Ž๐ผ๐œ† (1 + ๐œ–) > ๐›ผ
for ๐‘› โ‰ฅ ๐‘(๐‘Ž, ๐œ†, ๐œ–)
>๐›ผ
โˆ—
๐ธ[๐‘โ‰ฅ๐‘˜ ] = ๐‘›๐‘ƒ[|๐ถ(1)| โ‰ฅ ๐‘˜] โ‰ฅ ๐‘›๐‘ƒ[๐‘‡ โ‰ฅ ๐‘˜] โˆ’
๐‘๐‘˜
๐‘›
โ‰ฅ ๐‘› (๐‘’
โˆ’๐ผ๐œ† ๐‘˜(1+๐œ–)
โˆ’
๐‘๐‘˜
)
๐‘›
=๐‘›
โž
1โˆ’๐ผ๐œ† ๐‘Ž(1+๐œ–)
โˆ’ ๐‘ log ๐‘› โ‰ฅ
๐‘›๐›ผ for ๐‘› โ‰ฅ ๐‘.
(For ๐œ– chosen sufficiently small)
Now letโ€™s prove claim 2:
Proof of claim 2:
|๐ถ(1)|
We can write |๐ถ(1)| = โˆ‘๐‘–=1 1{|1|>๐‘–} = โˆ‘๐‘›๐‘–=1 1{|1|>๐‘–}
So:
๐‘›
๐‘›
๐œ’โ‰ฅ๐‘˜ = โˆ‘ ๐‘ƒ[|๐ถ(1)| โ‰ฅ 1, |๐ถ(1)| โ‰ฅ ๐‘˜] = ๐‘˜๐‘ƒ[|๐ถ(1)| โ‰ฅ ๐‘˜] + โˆ‘ ๐‘ƒ[|๐ถ(1)| โ‰ฅ ๐‘–]
๐‘–=1
๐‘–=๐‘˜+1
โˆž
๐‘˜๐‘’ โˆ’๐ผ๐œ† (๐‘˜โˆ’1) + โˆ‘ ๐‘’ โˆ’๐ผ๐œ† (๐‘–โˆ’1) โ‰ค ๐‘๐‘˜๐‘›โˆ’๐‘Ž๐ผ๐œ† + ๐‘๐‘’
โŸ โˆ’๐ผ๐œ† ๐‘˜ โ‰ค ๐‘๐‘˜๐‘›โˆ’๐‘Ž๐ผ๐œ†
โ‰ค๐‘›โˆ’๐‘Ž๐ผ๐œ†
๐‘–=๐‘˜+1
So by lemma 3.10 get claim 2. โˆŽ
Exercise 3.12 (second moment method):
Let ๐‘‹ be a random variable such that 0 โ‰ค ๐ธ[๐‘‹] < โˆž and 0 < ๐ธ[๐‘‹ 2 ] < โˆž
Let 0 โ‰ค ๐œ† < 1.
๐ธ[๐‘ฅ]2
Prove that: ๐‘ƒ[๐‘‹ > ๐œ†๐ธ[๐‘‹]] โ‰ฅ (1 โˆ’ ๐œ†)2 ๐ธ[๐‘ฅ 2 ]
In our proof, we used it with ๐œ† = 0.
Hint: Prove first that (1 โˆ’ ๐œ†)๐ธ[๐‘‹] โ‰ค ๐ธ[๐‘‹ โˆ™ 1{๐‘‹>๐œ†๐ธ[๐‘‹]} ]
๐‘๐‘ฆ
๐‘™๐‘’๐‘š๐‘š๐‘Ž 3.8
โ‰ค