27
SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT Local large deviations and the strong renewal theorem Francesco Caravenna Universit` a degli Studi di Milano-Bicocca Moscow, SPA 2017 Francesco Caravenna The strong renewal theorem 26 July 2017 1 / 27

Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Local large deviationsand the strong renewal theorem

Francesco Caravenna

Universita degli Studi di Milano-Bicocca

Moscow, SPA 2017

Francesco Caravenna The strong renewal theorem 26 July 2017 1 / 27

Page 2: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Joint work with Ron Doney

(Picture taken from MFO website)

Francesco Caravenna The strong renewal theorem 26 July 2017 2 / 27

Page 3: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Overview

Necessary and sufficient conditions

for the Strong Renewal Theorem (SRT)

Problem dates back to Garsia and Lamperti (1962)

First focus on renewal processes (random walks with positive increments)

then on general random walks (much harder!)

Key tool: local large deviation estimate of independent interest

For simplicity we stick to the lattice case, but everything applies tonon-lattice distributions

Francesco Caravenna The strong renewal theorem 26 July 2017 3 / 27

Page 4: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Outline

1. SRT for Renewal Processes

2. SRT for Random Walks

3. Local Large Deviations

4. Proof of the SRT

Francesco Caravenna The strong renewal theorem 26 July 2017 4 / 27

Page 5: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Outline

1. SRT for Renewal Processes

2. SRT for Random Walks

3. Local Large Deviations

4. Proof of the SRT

Francesco Caravenna The strong renewal theorem 26 July 2017 5 / 27

Page 6: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

The classical renewal theorem

Random walk Sk := X1 + X2 + . . .Xk with positive increments

(Xi ) i.i.d. Xi ∈ N = {1, 2, . . .} aperiodic

0 S1 S2 S3 ... 0 S1 S2 S3 ... n

Renewal (Green) function u(n) := P(S visits n) =∑k≥0

P(Sk = n)

Renewal Theorem (Erdos, Feller, Pollard 1949)

limn→∞

u(n) =1

E[X ]= 0 if E[X ] =∞

Francesco Caravenna The strong renewal theorem 26 July 2017 6 / 27

Page 7: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

The case of infinite mean

When E[X ] =∞, at which rate u(n)→ 0 ?

Define the truncated mean

m(n) := E[X ∧ n] = E[min{X , n}]

Guess (SRT) u(n) ∼n→∞

c

m(n)for some c ∈ (0,∞)

Assumptions are needed! But there is hope

Lemma (Erickson 1973)

n

m(n)≤

n∑i=1

u(i) ≤ 2n

m(n)

Francesco Caravenna The strong renewal theorem 26 July 2017 7 / 27

Page 8: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Asymptotically stable renewal processes

Tail Assumption

P(X > n) ∼n→∞

L(n)

nαα ∈ (0, 1) L(·) slowly varying

Explicit computation m(n) := (1− α) E[X ∧ n] ∼n→∞

L(n) n1−α

Theorem (Garsia, Lamperti 1962) (Erickson 1970)

lim infn→∞

u(n)

1/m(n)= cα :=

sinπα

π

I If α > 12 lim inf can be replaced by lim (the SRT holds)

I If α ≤ 12 there are counterexamples!

Francesco Caravenna The strong renewal theorem 26 July 2017 8 / 27

Page 9: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Sufficient conditions for the SRT

Local Assumption

P(X = n) ≤ CL(n)

n1+α

Theorem (Doney 1997) (Williamson 1968) (Vatutin, Topchii 2013)

Under Local Assumption, the SRT holds also for α ≤ 12

u(n) = P(S visits n) ∼n→∞

cαm(n)

=cα

L(n) n1−α

Sufficient conditions weaker than the Local Assumption (integral criteria)were given more recently in (Chi 2013, 2015)

Francesco Caravenna The strong renewal theorem 26 July 2017 9 / 27

Page 10: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

What can go wrong?

Without Local Assumption, we only know that

P(X = n) = o(P(X ≥ n)

)= o(1)

L(n)

where o(1) can be as slow as we wish (along subsequences)

P(S visits n) ≥ P(X1 = n) can be � 1

L(n) n1−αif α ≤ 1

2

The SRT fails because a single step (more generally, few steps)

can give “atypical” contribution to P(S visits n)

We must find optimal conditions on X to rule this out

Francesco Caravenna The strong renewal theorem 26 July 2017 10 / 27

Page 11: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Necessary and sufficient conditions

Define a weighted sum of P(X = n), P(X = n − 1), . . . , P(X = n − δn)

I+(δ; n) :=δn∑r=1

wr P(X = n − r) wr :=r2α−1

L(r)2

Definition

We say that I+(δ; n) is asymptotically negligible (a.n.) iff

I+(δ; n) � 1

m(n)as n→∞ , δ → 0

More precisely: limδ→0

lim supn→∞

I+(δ; n)

1/m(n)= 0

Francesco Caravenna The strong renewal theorem 26 July 2017 11 / 27

Page 12: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

SRT for renewal processes

Recall the Tail Assumption

P(X > n) ∼n→∞

L(n)

nαα ∈ (0, 1) L(·) slowly varying

Theorem (SRT for renewal processes) (Caravenna, Doney 2016)

I If α > 12 the SRT holds with no further assumption

I If α ≤ 12 the SRT holds if and only if I+(δ; n) is a.n.

The weighted sum I+(δ; n) is explicit, but slightly involved

We give a practical sufficient condition in terms of probability of intervals

Francesco Caravenna The strong renewal theorem 26 July 2017 12 / 27

Page 13: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

A nearly optimal condition

Corollary

For α ≤ 12 , a sufficient condition for the SRT is

P(n < X ≤ n + r) ≤ CL(n)

( rn

)γ∀r = 1, . . . , n (?)

for some C <∞ and γ > 1− 2α

P(X ≤ n + r |X > n) ≤ C( rn

)γ(?)

Condition (?) is actually nearly optimal

Corollary

A necessary condition for the SRT is that (?) holds for every γ < 1− 2α

Francesco Caravenna The strong renewal theorem 26 July 2017 13 / 27

Page 14: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Outline

1. SRT for Renewal Processes

2. SRT for Random Walks

3. Local Large Deviations

4. Proof of the SRT

Francesco Caravenna The strong renewal theorem 26 July 2017 14 / 27

Page 15: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

General random walks

Random walk Sk := X1 + X2 + . . .Xk on Z

(Xi ) i.i.d. Xi ∈ Z aperiodic

Tail Assumption

P(X > n) ∼n→∞

pL(n)

nαP(X < −n) ∼

n→∞qL(n)

p > 0, q ≥ 0 α ∈ (0, 1) L(·) slowly varying

We use the shorthandm(n) := L(n) n1−α

which is ' E[|X | ∧ n]

Francesco Caravenna The strong renewal theorem 26 July 2017 15 / 27

Page 16: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

A first renewal theorem

Renewal (Green) function u(n) :=∑k≥0

P(Sk = n)

Theorem (Williamson 1968) (Erickson 1971)

lim infn→∞

u(n)

1/m(n)= cα,p,q

I If α > 12 lim inf can be replaced by lim (the SRT holds)

I If α ≤ 12 there are counterexamples

The Local Assumption is again sufficient condition for the SRT

P(X = n) ≤ CL(n)

|n|1+α

Francesco Caravenna The strong renewal theorem 26 July 2017 16 / 27

Page 17: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

A first condition

Extend the previously introduced weighted sum (note the |r |)

I1(δ; n) :=δn∑|r |=1

wr P(X = n + r) wr :=|r |2α−1

L(|r |)2

I1(δ; n) depends only on X+, but SRT may also depend on X−

Proposition

For random walks, a sufficient (not necessary) condition for the SRT is

I1(δ; n) and I1(δ;−n) are both a.n.

In particular, a sufficient condition is that (?) holds for both n and −n

In general, assuming that I1(δ;−n) is a.n. is too strong

Francesco Caravenna The strong renewal theorem 26 July 2017 17 / 27

Page 18: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

SRT for random walks

For k ≥ 2 define a k-tuple weighted sum, depending on δ, η ∈ (0, 1)

Ik(δ, η; n) :=∑|r1|≤δn

|rj |≤η|rj−1| for 2≤j≤r

P(X = n + r1) P(X = −r1 + r2) · · ·

· P(X = −rk−1 + rk)|rk |(k+1)α−1

L(|rk |)k+1

Theorem (SRT for random walks) (Caravenna, Doney 2016)

Consider a random walk satisfying the Tail Assumption with p, q > 0

I If α > 12 the SRT holds with no further assumption

I If 13 < α < 1

2 the SRT holds if and only if I1(δ; n) is a.n.

I Given k ≥ 2, if 1k+2 < α < 1

k+1 the SRT holds if and only if

Ik(δ, η; n) is a.n., for every fixed η ∈ (0, 1)

Francesco Caravenna The strong renewal theorem 26 July 2017 18 / 27

Page 19: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Further considerations

I Necessary and sufficient conditions available also for α = 1k , k ∈ N,

in terms of slightly modified quantities Ik(δ, η; n)

I Given α ∈ (0, 1) let k ∈ N be such that 1k+2 < α < 1

k+1

SRT holds ⇐⇒ Ik is a.n. ⇐⇒ Ik′ is a.n. ∀k ′ ∈ N

I There are examples where I1(δ; x) is a.n. but I2(δ, η; x) is not a.n.

I When q = 0 the previous conditions are sufficient, but we don’texpect them to be necessary

Francesco Caravenna The strong renewal theorem 26 July 2017 19 / 27

Page 20: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Outline

1. SRT for Renewal Processes

2. SRT for Random Walks

3. Local Large Deviations

4. Proof of the SRT

Francesco Caravenna The strong renewal theorem 26 July 2017 20 / 27

Page 21: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Asymptotically stable random walks

Consider a random walk Sk = X1 + . . .+ Xk on Z

P(X > n) ∼n→∞

pL(n)

nαP(X < −n) ∼

n→∞qL(n)

We allow both α ∈ (0, 1) and α ∈ (1, 2) (with E[X ] = 0)

We know thatSkak

d−−−−−→k→∞

Y α-stable law(ak ∼ L(k) k

)Gnedenko LLT

P(Sk = n) =1

ak

(ϕ( n

ak

)+ o(1)

)=

o(1)

akfor n� ak

We improve on this bound, giving a quantitative estimate of o(1)

Francesco Caravenna The strong renewal theorem 26 July 2017 21 / 27

Page 22: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Local large deviations

Theorem (Local Large Deviations) (Caravenna, Doney 2016)

P(Sk = n) ≤ C

ak

k L(n)

For every γ ∈ (0, 1]

P(Sk = n, max{X1, . . . ,Xk} ≤ γn

)≤ Cγ

ak

(k L(n)

)d 1γ e

(�)

I Improves on Gnedenko’s bound with no extra assumptions

I Key tool for the SRT

I (�) is a local version of the Fuk-Nagaev inequality

I Extended to the Cauchy case α = 1 in (Berger 2017)

Francesco Caravenna The strong renewal theorem 26 July 2017 22 / 27

Page 23: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Outline

1. SRT for Renewal Processes

2. SRT for Random Walks

3. Local Large Deviations

4. Proof of the SRT

Francesco Caravenna The strong renewal theorem 26 July 2017 23 / 27

Page 24: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Probabilistic proof: strategy

Goal: SRT

u(n) ∼n→∞

cαm(n)

=cα

L(n) n1−α

u(n) =∑k≥0

P(Sk = n) =∑

k≤δ nα

L(n)︸ ︷︷ ︸(1)

+∑

δ nα

L(n)<k≤ 1δ

L(n)︸ ︷︷ ︸(2)

+∑

k> 1δ

L(n)︸ ︷︷ ︸(3)

I (3) is always negligible by Gnedenko’s LLT P(Sk = n) ≤ Cak

I (2) is a Riemann sum approximation: by P(Sk = n) ∼ 1akϕ( n

ak)

(. . . ) it gives full contribution to SRT as n→∞, δ → 0

SRT holds ⇐⇒ (1) � 1m(n) is asymptotically negligible

Francesco Caravenna The strong renewal theorem 26 July 2017 24 / 27

Page 25: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Necessity: SRT ⇒ I+(δ; n) is a.n.

P(Sk = n) ≥ P(Sk = n, ∃!Xi > Cak)

= (. . .)

≥ c∑

1≤r≤δn

P(X = n − r)k

ak1{ rα

L(r)<k≤2 rα

L(r)}

Then (1) =∑

k≤δ nα

L(n)

P(Sk = n)

≥ c∑

1≤r≤δn

P(X = n − r)∑

L(r)<k≤2 rα

L(r)

k

ak

≥ c∑

1≤r≤δn

P(X = n − r)r2α−1

L(r)2= I+(δ; n)

Francesco Caravenna The strong renewal theorem 26 July 2017 25 / 27

Page 26: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Sufficiency: I+(δ; n) is a.n. ⇒ SRT

Fix suitable γ ∈ (0, 1). Using local large deviation

P(Sk = n, max{X1, . . . ,Xk} ≤ γn) is negligible

P(Sk = n, max{X1, . . . ,Xk} > γn) ≤ k∑

r≤(1−γ)n

P(X = n−r) P(Sk−1 = r)

Then (1) ≤∑

r≤(1−γ)n

P(X = n − r)∑

k≤δ nα

L(n)

k P(Sk−1 = r)

I We can replace n by r in the inner sum, using that I+(δ; n) is a.n.

I Then the inner sum becomes (1) with an extra factor k

I This helps! (Backward) induction argument allows to conclude

Francesco Caravenna The strong renewal theorem 26 July 2017 26 / 27

Page 27: Local large deviations and the strong renewal theoremstaff.matapp.unimib.it/~fcaraven/download/slides/moscow-SPA-201… · SRT for Renewal ProcessesSRT for Random WalksLocal Large

SRT for Renewal Processes SRT for Random Walks Local Large Deviations Proof of the SRT

Thanks

Francesco Caravenna The strong renewal theorem 26 July 2017 27 / 27