Chapter IV: Weak and Strong Ties in Social Network
A.2 Long-term behavior of public belief
Then by Lemma 29, the sequence(ππ‘)is well approximated by π(π‘), the solution to the corresponding differential equation:
π‘ββlim ππ‘ π(π‘) =1.
Now, conditional on π = +1, all agents take action +1 from some point on with probability 1. Thus, with probability 1,
βπ‘+
1=βπ‘+π·+(βπ‘) for all sufficiently largeπ‘. Further, by Lemma 27,
π₯ββlim
π·+(π₯) πΊβ(βπ₯) =1. So by Lemma 28,
π‘limββ
βπ‘ ππ‘
=1 with probability 1. Thus, we have
π‘ββlim βπ‘
π(π‘) =π‘ββlim βπ‘ ππ‘
Β· ππ‘ π(π‘) =1 with probability 1.
Proofs of Lemmas 28 and 29
Proof of Lemma 28. We prove the claim in two steps. First, we show that for every π > 0 there are infinitely many timesπ‘such that
(1βπ)ππ‘ β€ ππ‘ β€ (1+π)ππ‘. (A.3) Second, we show that if (A.3) holds for some π‘ large enough, then it holds for all π‘0> π‘, proving the claim.
We start with step 1. Assume without loss of generality thatππ‘ β€ ππ‘ for infinitely many values ofπ‘. Fix π > 0. To show that (1βπ)ππ‘ β€ ππ‘ β€ (1+π)ππ‘ holds for infinitely many values ofπ‘, letπ₯
0> 1 be such that for allπ₯ > π₯
0it holds that π΄and π΅are monotone decreasing,
π΄(π₯), π΅(π₯) < π < 1 and
(1βπ/2)π΄(π₯) < π΅(π₯) < (1+π/2)π΄(π₯). (A.4)
Assume thatππ‘, ππ‘ > π₯
0; this will indeed be the case forπ‘large enough, since π΄and π΅are positive and continuous, and so bothππ‘ and ππ‘ are monotone increasing and tend to infinity. So
π΅(ππ‘) < (1+π/2)π΄(ππ‘) β€ (1+π/2)π΄(ππ‘),
where the first inequality follows from (A.4), and the second follows from the fact that π΄ is monotone decreasing and ππ‘ < ππ‘. Since π΅(π(π‘)) = ππ‘+
1 β π(π‘) and π΄(ππ‘) =ππ‘+
1βπ(π‘) we have shown that ππ‘+
1βππ‘ < (1+π/2) (ππ‘+
1βππ‘),
and so eventually ππ‘ β€ (1+π)ππ‘. Also, notice that the first time this obtains, we also have that the left inequality in (A.3) holds at the same moment:
ππ‘ > ππ‘β
1> ππ‘β
1=ππ‘ β (ππ‘ βππ‘β
1) > ππ‘βπ > ππ‘βπ ππ‘ =(1βπ)ππ‘.
This completes the first step. Now we go to step 2. Here we show that if (A.3) holds for large enoughπ‘ then it holds for allπ‘0> π‘.
Fixπ > 0, and letπ₯
0be defined as above. Suppose that(1βπ)ππ‘ < ππ‘ < (1+π)ππ‘, with ππ‘, ππ‘ > π₯
0. Assume without loss of generality that ππ‘ β₯ ππ‘. Then our assumptions and (A.4) imply
ππ‘+
1=ππ‘+π΅(ππ‘)
< (1+π)ππ‘+ (1+π)π΄(ππ‘). Becauseππ‘ β€ ππ‘and π΄is decreasing we have
ππ‘+
1< (1+π)ππ‘+ (1+π)π΄(ππ‘)
= (1+π)ππ‘+
1. For the other direction, note first that
ππ‘+
1> ππ‘ β₯ ππ‘,
by assumption. We can write ππ‘ = (1 β π)ππ‘ + π ππ‘, and since ππ‘ > π₯
0 > 1, π ππ‘ > (1βπ)π, and so
ππ‘+
1 > (1βπ)ππ‘+ (1βπ)π .
Now,π > π΄(ππ‘) sinceππ‘ > π₯
0, and so ππ‘+
1 > (1βπ)ππ‘+ (1βπ)π΄(ππ‘)
=(1βπ)ππ‘+
1. Thus
(1βπ)ππ‘+
1 < ππ‘+
1< (1+π)ππ‘+
1, (A.5)
as required.
Proof of Lemma 29. We restrict the domain of π to the interval(π‘
0,β)such that for π‘ > π‘
0 it already holds that π0(π‘) = π΄(π(π‘)). Since π΄is continuous, limπ‘ββ π(π‘) =
β, and so we can also assume that in the interval(π(π‘
0),β)it holds thatπ΄is convex and differentiable.
Since π is strictly increasing in (π‘
0,β), it has an inverse πβ1. Forπ₯ large enough defineπ΅(π₯) = π(πβ1(π₯) +1) βπ₯.
Now, let(ππ‘)be any sequence satisfying the recurrence relation ππ‘+
1=ππ‘+π΅(ππ‘). In order to apply Lemma 28, we will first show that
π₯limββ
π΅(π₯) π΄(π₯) =1.
Letπ‘ = πβ1(π₯). Such aπ‘ exists and is unique for all sufficiently largeπ₯, because π is monotone. Notice that by the definitions ofπ΅(π₯)and π0(π₯)
π΅(π₯)= π(πβ1(π₯) +1) βπ₯
= π(πβ1(π₯) +1) βπ₯β π0(πβ1(π₯)) + π0(πβ1(π₯))
= π(π‘+1) β π(π‘) β π0(π‘) +π΄(π(π‘)),
where in the last equality we substitute π‘ = πβ1(π₯). Because π0 is positive and decreasing (π is concave) then π(π‘ +1) β π(π‘) β₯ π0(π‘+1), and so
π΅(π₯) β₯ π0(π‘+1) β π0(π‘) +π΄(π(π‘)).
By the definition of π, π0(π‘) = π΄(π(π‘)), and so
π΅(π₯) β₯ π΄(π(π‘+1)) βπ΄(π(π‘)) +π΄(π(π‘))= π΄(π(π‘+1)).
Again, due to concavity of π we have π(π‘+1) β€ π(π‘) + π0(π‘)and asπ΄is decreasing and convex we get
π΅(π₯) β₯ π΄(π(π‘) + π0(π‘))
β₯ π΄0(π(π‘))π0(π‘) +π΄(π(π‘))
= π΄0(π(π‘))π΄(π(π‘)) +π΄(π(π‘)). We now substitute backπ₯ = π(π‘):
π΅(π₯) β₯ π΄0(π₯)π΄(π₯) +π΄(π₯)
= π΄(π₯) (π΄0(π₯) +1) so in particular, since π΄0(π₯) β0 asπ₯ β β,
lim inf
π₯ββ
π΅(π₯) π΄(π₯) β₯ 1. Now we are going to show that lim supπ₯ββ
π΅(π₯)
π΄(π₯) β€ 1 which will conclude the proof.
By the definitions of πβ1(π₯)andπ΅(π₯)
π΅(π₯) =π΅(π(π‘))= π(π‘+1) β π(π‘) =
β« π‘+1 π‘
π0(π)dπ . As π0is decreasing it follows that
π΅(π₯) β€
β« π‘+1 π‘
π0(π‘)dπ = π0(π‘) = π΄(π(π‘)) = π΄(π₯). Therefore,
lim sup
π₯ββ
π΅(π₯) π΄(π₯) β€ 1. Hence, from these two inequalities we get that
π₯ββlim π΅(π₯) π΄(π₯) =1.
Now notice that, by construction, π(π‘+1) = π(π‘) +π΅(π(π‘)). Thus, by Lemma 28,
πlimββ
π(π‘) ππ‘
=1.
Monotonicity of solutions to a differential equation
We now prove a general lemma regarding differential equations of the formπ0(π‘) = π΄(π(π‘)). It shows that the solutions to this equation are monotone in π΄. This is useful for calculating approximate analytic solutions whenever it is impossible to find analytic exact solutions, as is the case of Gaussian signals, in which we use this lemma.
Lemma 30. Let π΄, π΅: R>0βR>0be continuous, and letπ, π: R>0βR>0satisfy π0(π‘) = π΄(π(π‘))andπ0(π‘) =π΅(π(π‘))for all sufficiently largeπ‘.
Suppose that
lim inf
π₯ββ
π΄(π₯) π΅(π₯) > 1. Thenπ(π‘) > π(π‘) for all sufficiently largeπ‘.
Proof. Notice that π(π‘) and π(π‘) are eventually monotone increasing and tend to infinity asπ‘ tends to infinity. Thus for allπ₯ greater than someπ₯
0 > 0 large enough, πandπhave inverses that satisfy the following differential equations:
d dπ₯
πβ1(π₯) = 1 π΄(π₯) d
dπ₯
πβ1(π₯) = 1 π΅(π₯)Β·
Since lim infπ₯ π΄(π₯)/π΅(π₯) > 1, we can furthermore chooseπ₯
0so that for allπ₯ β₯ π₯
0, π΄(π₯) > (1+π)π΅(π₯) for someπ >0. Thus, forπ₯ > π₯
0
πβ1(π₯)=πβ1(π₯
0) +
β« π₯
π₯0
1 π΄(π₯) dπ₯ πβ1(π₯)=πβ1(π₯
0) +
β« π₯
π₯0
1 π΅(π₯)dπ₯ and so
πβ1(π₯) < πβ1(π₯
0) + 1 1+π
β« π₯ π₯0
1 π΅(π₯) dπ₯
=πβ1(π₯
0) + 1 1+π
(πβ1(π₯) βπβ1(π₯
0)) and thus
πβ1(π₯) βπβ1(π₯) <β π 1+π
πβ1(π₯) +
πβ1(π₯
0) β 1 1+π
πβ1(π₯
0)
.
Sinceπβ1(π₯)tends to infinity asπ₯tends to infinity, it follows that for all sufficiently largeπ₯, πβ1(π₯) < πβ1(π₯). Thus, for all sufficiently largeπ‘
π‘ =πβ1(π(π‘)) < πβ1(π(π‘)), and so, sinceπ(π‘) is monotone increasing,
π(π‘) < π(π‘).
Eventual monotonicity of public belief update
We end this section with a lemma that shows that under some technical conditions on the left tail of πΊβ, the function π’+(π₯) = π₯ + π·+(π₯) (i.e., the function that determines how the public log-likelihood ratio is updated when the action +1 is taken) is eventually monotone increasing.
Lemma 31. Suppose πΊβ has a convex and differentiable left tail. Then the map π’+(π₯) =π₯+π·+(π₯)is monotone increasing for all sufficiently largeπ₯.
Proof. Recall that
π·+(π₯)=log1βπΊ+(βπ₯) 1βπΊβ(βπ₯)Β·
Since πΊβ has a differentiable left tail, it has a derivative πβ(βπ₯) for all π₯ large enough. It then follows from (A.1) thatπΊ+also has a derivative in this domain, and
π’0+(π₯) =1+ π+(βπ₯)
1βπΊ+(βπ₯) β πβ(βπ₯) 1βπΊβ(βπ₯)
=1+ eβπ₯πβ(βπ₯)
1βπΊ+(βπ₯) β πβ(βπ₯) 1βπΊβ(βπ₯)Β· Since 1βπΊβ(βπ₯)and 1βπΊ+(βπ₯)tend to 1 asπ₯tends to infinity,
π₯ββlim
π’0+(π₯) =π₯ββlim 1+eβπ₯πβ(βπ₯) βπβ(βπ₯). SinceπΊβ is eventually convex,πβ(βπ₯) tends to zero, and therefore
π₯limββ
π’0+(π₯) =1.
In particular, π’0+(π₯) is positive for π₯ large enough, and hence π’+(π₯) is eventually monotone increasing.
A.3 Gaussian private signals