In this section we explain how Corollaries2.6,2.7and2.9follow from Theorem2.5.
Proof of Corollary 2.6. Given ϑ>0, we define the set
Gϑ:={λˆ∈RdN:|λ`i−γi/N` |6Nϑ−2/3min{i, N+1−i}1/3for alliand`}. (5.1) As proved in [29] in the special case of the Gaussian ensembles and then generalized in [13, Theorem 2.4] to potentialsWk satisfying much weaker conditions than the ones assumed here, the following rigidity estimate holds: for all ϑ>0 there exist ¯c>0 and C <∞ such that, for allN>0,
PeβN,0 RN\Gϑ
6Ce −Nc¯. (5.2)
Also, due to the fact thatµ0k has a density which is strictly positive inside its support [a0k, b0k] except at the two boundary points where it goes to zero as a square root (see Lemma3.2), we deduce that
m N > 1
C
γk(i+m)/N
γi/Nk
minp
s−a0k,p
b0k−s ds,
from which it follows easily that
|γk(i+m)/N−γi/Nk |6 C N2/3min
m2/3, m
min{i, N+1−i}1/3
. (5.3)
Since
|λki+m−λki|6|λki−γi/Nk |+|λki+m−γ(i+m)/Nk |+|γ(i+m)/Nk −γki/N|, (5.4) using (5.2) and (5.3) and recalling that by assumptionmN, we deduce that
|N(λki
k+j−λki
k)|6C(Nϑ+m) for all ˆλ∈Gϑ,ik∈[N ε, N(1−ε)], j= 1, ..., m, (5.5) and
|N2/3(λkj−a0k)|6C(Nϑ+m2/3) for all ˆλ∈Gϑ,j= 1, ..., m. (5.6) Z
Now, given a bounded function χ:RdN!R, applying (2.8) to 1
2
1+ χ kχk∞
withk=0 andη=ϑ, we deduce that
χTNdPβN,0− χ dPβN,aV
6CNϑ−1kχk∞. (5.7)
Recall that the mapTN is given byX1N,whereXtN is the flow of the vector fieldYNt that has the very special form (4.13) (see Proposition4.13). In particular, since the functions y0k,t,y1k,t,ζk`,t(·, y) are uniformly Lipschitz, we see that
|( ˙XtN)ki−( ˙XtN)kj|6L|(XtN)ki−(XtN)kj| for alli, j= 1, ..., N andk= 1, ..., d.
Hence, sinceX1N=TN andX0N=Id, Gr¨onwall’s inequality yields
e−L(λki−λkj)6(TN)ki(ˆλ)−(TN)kj(ˆλ)6eL(λki−λkj) for allλki >λkj. (5.8) We now remark that the lawPeβN,aV is obtained as the image of the law ofλk=(λk1, ..., λkN), 16k6dunderPβN,aV under the map
R:b RdN!RdN, R(λb 1, ..., λk, ..., λd) := (R(λ1), ...,R(λk), ...,R(λd)), (5.9) whereR:RN!RN is defined as
[R(x1, ..., xN)]i:= min
#J=imax
j∈J xj for alli= 1, ..., N. (5.10) Hence, due to (5.8), it follows thatTN andRb commute, namely
RbTN=TNR.b (5.11)
We now consider a test functionχof the form
χ(ˆλ) =f((N(λkik+1−λkik), ..., N(λkik+m−λkik))dk=1). (5.12) Then
f((N(λki
k+1−λkik), ..., N(λki
k+m−λkik))dk=1)dPeβN,aV = χRbdPβN,aV, and it follows by (5.7) and (5.11) that
χ dPeβN,aV− χTNRbdPβN,0
6CNϑ−1kfk∞.
Z Z
Z
Z Z
Z
LetX0,t,X1,t,andX2,tbe as in Proposition4.13, and note the following fact: whenever
SinceX0,t` is a smooth diffeomorphism which sends the quantiles ofµ∗`,0onto the quantiles ofµ∗`,t, we deduce that and by the same argument as the one used in the proof of Proposition4.13to show (4.39) and (4.40) we get
(this follows by the same proof as the one of (5.8), compare also with [5, Equation (5.2)]).
In addition,
and hence, by the definition ofGϑ, Hence, we have proved that(1)
For the second statement we choose
χ(ˆλ) =f((N2/3(λk1−aaVk ), ..., N2/3(λkm−aaVk ))dk=1),
(1) This estimate, as well as the one at the edge that we shall prove below, should be compared with the one obtained in [5, Theorem 1.5]. While the estimates here are considerably stronger than the ones in [5, Theorem 1.5] (this follows from the fact that we have better bounds on our approximate transport maps), as a small “loss” we now haveNϑ−1instead of a term (logN)3/N. The reason for this small difference comes from the fact that we decided to apply (2.8) to deduce (5.7). It is worth noticing that the argument in§4combined with [5, Lemma 2.2] proves that also the stronger bound
holds. However, since in general (2.8) is much more powerful than the estimate above (as it allows to deal with functions that grow polynomially with respect to the dimension) and the improvement between (logN)3/NandNϑ−1 is minimal, we have decided not to state also this second estimate.
Z
and we note thatT0k(a0k)=aaVk . Then, due to (4.39) and (5.13), we get using the rigidity estimate (5.6), we may replace
N2/3(T0k(λk1)−T0k(a0k)) by (T0k)0(a0k)N2/3(λk1−a0k) which proves the second statement by choosingϑ6θ.
Proof of Corollary 2.7. We first note that the proof of Corollary 2.6 could be re-peated verbatim in the context of [5] to show that [5, Theorem 1.5] holds with the same estimates as we obtained here. Hence, by combining this result with Corollary 2.6, we have
whereγik/N satisfies µsc((−∞, γik/N))=ik/N. Note that the transport relations (2.10) and (2.11) imply thatT0kS0k(γik/N)=γik
k/N,a, whereγik
k/N,a satisfies µaVk ((−∞, γki
k/N,a)) =ik
N, and hence (again by (2.10) and (2.11))
(T0kS0k)0(γik/N) = %sc(γik/N)
%aVk (γik
k/N,a).
Finally, since |σk−ik/N|6C/N and σk∈(0,1), arguing as we did for proving (5.3), we deduce that|γik/N−γσk|6C/Ne , so up to another small error we may replace
%sc(γik/N)
%aVk (γik
k/N,a) by %sc(γσk)
%aVk (γσk,k).
This concludes the proof of of the first statement, while the second one is just a conse-quence of Corollary2.6(2) and [5, Theorem 1.5 (2)].
Proof of Corollary 2.8. As is clear by looking at the proof of Corollaries2.6and2.7, the fact of dealing at the same time with the eigenvalues of different matrices does not complicate the proof. For this reason, since the proof of Corollary 2.8 is already very involved, to make the argument more transparent we shall prove the result when the test function is of the form
−
Rk(E)+N−ζR0k(E)
Rk(E)−N−ζR0k(E)
X
ijdistinct
f N(λki
1−E), ..., Ne (λki
m−E)e dEe
for someE∈(−2,2), the proof in the general case being completely analogous and just notationally heavier.
To simplify the notation, we set gEe(ˆλ) := X
i16=...6=im
f(N(λki1−E), ..., N(λe kim−E)),e
Ak:=
−
Rk(E)+N−ζR0k(E)
Rk(E)−N−ζR0k(E)
g
EedEe
dPβN,aV.
It follows by (2.8) withη=θthat
|log(1+Ak)−log(1+A1,k)|6CNθ−1, (5.15) Z
Z Z
where
A1,k:=
−
Rk(E)+N−ζR0k(E)
Rk(E)−N−ζR0k(E)
gEe(TN)kdEe
dPβN,0
=
−
Rk(E)+N−ζR0k(E)
Rk(E)−N−ζR0k(E)
X
i16=...6=im
f(N((TN)ki1(ˆλ)−E), ..., N((Te N)kim(ˆλ)−E))e dEe
dPβN,0.
Define the quantilesγi/Nk ∈(Sk0(−2), Sk0(2)) as in Corollary2.6, and given ϑ>0 small (to be fixed later) we consider the setGϑ defined in (5.1).
Since the integrandg
Ee(TN)k is pointwise bounded bykfk∞Nm, it follows by (5.2) that
A1,k=A2,k+O(e−Nc), (5.16)
where
A2,k:=
Gϑ
−
Rk(E)+N−ζR0k(E)
Rk(E)−N−ζR0k(E)
gEe(TN)kdEe
dPβN,0. Observe that if ˆλ∈Gϑ then, by definition,
|λki−λkj|>|γki/N−γj/Nk |−N−2/3+ϑmin{i, N+1−i}−1/3−N−2/3+ϑmin{j, N+1−j}−1/3. Hence, sinceγ(i+1)/Nk −γi/Nk >c0N−2/3min{i, N+1−i}−1/3 for alli, we deduce that
|λki−λkj|>Nϑ−1 provided|i−j|>C0Nϑ, which, combined with (5.8) yields, for ˆλ∈Gϑ,
|(TN)ki(ˆλ)−(TN)kj(ˆλ)|>e−LNϑ−1 provided|i−j|>C0Nϑ. (5.17) We now notice that, sincef is compactly supported, the quantity
f(N((TN)ki
1(ˆλ)−E), ..., Ne ((TN)ki
m(ˆλ)−E))e can be non-zero only if
|(TN)kij(ˆλ)−E|e 6C1
N for allj= 1, ..., m.
Therefore, if ¯i∈{1, ..., N}is an index (depending on ˆλandE) such thate
|(TN)k¯i(ˆλ)−E|e 6C1
N, Z Z
Z Z
Z Z
then (5.17) yields
|(TN)ki(ˆλ)−E|e 6C1
N =⇒ |i−¯i|6C0Nϑ. This proves that, for any ˆλ∈Gϑ, there exists a set of indices
Jˆλ,Ee⊂ {(i1, ..., im)∈ {1, ..., N}m:i16=...6=im} such that #Jλ,ˆEe6CNmϑand
A2,k=
Gϑ
−
Rk(E)+N−ζR0k(E)
Rk(E)−N−ζR0k(E)
ˆ
gEe(TN)kdEe
dPβN,0, where
ˆ
gEe(ˆλ) := X
(i1,...,im)∈Jλ,fˆE
f(N(λki1−E), ..., N(λe kim−E))e
satisfies|ˆgTk
0(E)e |6Ckfk∞Nmϑ.
We now perform the change of variableEe7!T0k(E), which givese
Rk(E)+N−ζR0k(E)
Rk(E)−N−ζR0k(E)
ˆ g
Ee(TN)kdEe=
(T0k)−1[Rk(E)+N−ζR0k(E)]
(T0k)−1[Rk(E)−N−ζRk0(E)]
ˆ gTk
0(E)e (TN)k(T0k)0(E)e dE.e Recalling thatRk=T0kS0k and that these maps are all smooth diffeomorphisms ofR, we see that for
Ee∈[(T0k)−1[Rk(E)−N−ζR0k(E)],(T0k)−1[Rk(E)+N−ζR0k(E)]]
it holds
|(T0k)0(E)−(Te 0k)0S0k(E)|6CN−ζ, R0k(E) = [(T0k)0S0k(E)] (S0k)0(E), and
(T0k)−1[Rk(E)±N−ζR0k(E)] =S0k(E)±N−ζ(S0k)0(E)+O(N−2ζ).
Hence, since|ˆgTk
0(E)e |6CNmϑ,
−
(T0k)−1[Rk(E)+N−ζR0k(E)]
(T0k)−1[Rk(E)−N−ζR0k(E)]
ˆ gTk
0(E)e (TN)k(T0k)0(E)e dEe
=−
Sk0(E)−N−ζ(Sk0)0(E)
S0k(E)−N−ζ(S0k)0(E)
ˆ gTk
0(E)e (TN)kdEe+O(Nmϑ−ζ),
Z Z
Z Z
Z
Z
which proves that
Due to Theorem2.5, we can write ˆ arguing as we did for (5.13), we get
|∂EeX1,kˆλ(E)|e 6CNϑ, |(X1,1N )ki(ˆλ)−X1,kλˆ(E)|e 6CNϑ|λki−E|e for all ˆλ∈Gϑ. (5.20)
Z Z
Z
Z Z
Z
In addition, by the same reasoning,
maxi,k ∂1zk`,t(X0,tk (λki), y)dMXN`
0,t(y) =O(Nϑ) for all ˆλ∈Gϑ, and the argument used to prove (4.39) (see in particular (4.50)) yields
max
i,k |(X2,1N )ki|6CN2ϑ for all ˆλ∈Gϑ. Hence, since #Jλ,ˆEe6CNmϑwe immediately deduce that
O 1
N Gϑ
X
(i1,...,im)∈Jλ,fˆE
|(X2,1N )kij|dPβN,0
=O N(m+2)ϑ−1
. (5.21)
Now, to get rid of the term Xk
1,λˆ(E) insidee h
Ee we take advantage of (5.20) and the average with respect toE: more precisely, we consider the change of variablee
Ee7−!Φˆλ(E) := (Te 0k)−1
T0k(E)+e 1 NXk
1,ˆλ(E)e
so that
−
S0k(E)−N−ζ(S0k)0(E)
Sk0(E)−N−ζ(Sk0)0(E)
h
EedEe
=−
S0k(E)−N−ζ(S0k)0(E)
Sk0(E)−N−ζ(Sk0)0(E)
X
(i1,...,im)∈Jλ,fˆE
f(N(T0k(λki
1)−T0k(E))+[(Xe 1,1N )ki
1(ˆλ)−Xk
1,λˆ(E)e , ..., N(T0k(λkim)−T0k(E))+[(Xe 1,1N )kim(ˆλ)−X1,kλˆ(E)])∂e
EeΦλˆ(E)e dE.e Therefore, since∂
EeΦλˆ(E)=1+O Ne ϑ−1
(due to (5.20)),|h
Ee|6CNmϑ, and the interval [S0k(E)−N−ζ(S0k)0(E), S0k(E)−N−ζ(S0k)0(E)] has length of orderN−ζ, we deduce that
−
S0k(E)−N−ζ(S0k)0(E)
Sk0(E)−N−ζ(Sk0)0(E)
h
EedEe
=−
S0k(E)−N−ζ(S0k)0(E)
Sk0(E)−N−ζ(Sk0)0(E)
X
(i1,...,im)∈Jλ,fˆE
f(N(T0k(λki1)−T0k(E))+[(Xe 1,1N )ki1(ˆλ)−X1,kλˆ(E)],e ..., N(T0k(λkim)−T0k(E))+[(Xe 1,1N )kim(ˆλ)−X1,kˆλ(E)])e dEe+O(NζNmϑNϑ−1).
(5.22) We now observe that, since T0k:R!R is a diffeomorphism with (T0k)0>e−L>0 (see (5.14)), it follows by (5.20) that
|(X1,1N )ki1(ˆλ)−X1,kλˆ(E)|e 6CNϑ|T0k(λki)−T0k(E)|.e Z
Z
Z Z
Z Z
Therefore, sincef is compactly supported, we see that the expression f(N(T0k(λki1)−T0k(E))+[(Xe 1,1N )ki1(ˆλ)−X1,kλˆ(E)],e
..., N(T0k(λki
m)−T0k(E))+[(Xe 1,1N )ki
m(ˆλ)−Xk
1,λˆ(E)])e is non-zero only if
|T0k(λkij)−T0k(E)|e 6C1
N for allj= 1, ..., m.
In particular, using again that (T0k)0>e−L>0, this implies that|λkij−E|e 6C/N. Thus
|T0k(λkij)−T0k(E)−(Te 0k)0(E)[λkij−E]|e =O 1
N2
and
Nϑ|T0k(λkij)−T0k(E)|e =O Nϑ−1 , and we get
f(N(T0k(λki
1)−T0k(E)e
+[(X1,1N )ki
1(ˆλ)−Xk
1,λˆ(E)],e
..., N(T0k(λkim)−T0k(E))+[(Xe 1,1N )kim(ˆλ)−X1,kλˆ(E)])e
=f((T0k)0(E)N(λkij−E), ...,e (T0k)0(E)N(λkij−E))+O(k∇fe k∞Nϑ−1).
Combining this estimate with (5.22) and the fact that #Jλ,ˆEe6CNmϑwe conclude that
−
Sk0(E)−N−ζ(S0k)0(E)
Sk0(E)−N−ζ(Sk0)0(E)
hEedEe= ¯gE+O N(m+1)ϑ+ζ−1 , where
¯ gE(ˆλ)
:=−
S0k(E)−N−ζ(S0k)0(E)
Sk0(E)−N−ζ(Sk0)0(E)
X
(i1,...,im)∈Jλ,fˆE
f((T0k)0(E)N(λkij−E), ...,e (T0k)0(E)N(λkij−E))e dE.e
Also, by the argument above it follows that we can add back into the sum all the in-dices outsideJλ,ˆEe (since, up to infinitesimal errors, the function above vanishes on such indices), and therefore
−
Sk0(E)−N−ζ(Sk0)0(E)
S0k(E)−N−ζ(Sk0)0(E)
h
EedEe= ¯¯gE+O(N(m+1)ϑ+ζ−1), Z
Z
Z
with Combining this bound with (5.15), (5.16), (5.18), (5.19), and (5.21), we conclude that
|log(1+Ak)−log(1+ ¯A¯k)|6C(Nmϑ−ζ+N(m+2)ϑ−1+N(m+1)ϑ+ζ−1), (5.23) Combining this estimate with (5.23), we get
|log(1+Ak)−log(1+ ˆAk)|6C(Nmϑ−ζ+N(m+2)ϑ−1+N(m+1)ϑ+ζ−1).
Choosingϑsmall enough so that (m+2)ϑ<θ, this gives
|log(1+Ak)−log(1+ ˆAk)|6C(Nθ+ζ−1+Nθ−1/2+Nθ−ζ)6C(Nθ+ζ−1+Nθ−ζ), and since ˆAk is uniformly bounded inN (see for instance [65]) and the right-hand side is infinitesimal (recall thatθ<min{ζ,1−ζ}), we conclude that
|Ak−Aˆk|6C(Nθ+ζ−1+Nθ−ζ).
Recalling the definition ofAk and ˆAk, this proves that
which corresponds to our statement when f depends only on the eigenvalues of one matrix. As explained at the beginning of the proof, the very same argument presented above extends also to the general case.
Z
Z Z
Z Z
Z Z
Z
Proof of Corollary 2.9. We begin by noticing that the proof of Theorem 2.5 could be repeated verbatim in the context of [5] to show that [5, Theorem 1.4] holds with the same estimates as we obtained here.
To prove the gap estimates, it is enough to show that the approximate transport maps do not change gaps in the bulk uniformly (away from the edges). Due to Theo-rem2.5and [5, Theorem 1.4], we have the expansions
(TN)ki(ˆλ) =T0k(λki)+ 1
N(X1,1N )ki(ˆλ)+ 1
N2(X2,1N )ki(ˆλ), (SkN)i(λk) =S0k(λki)+ 1
N(Sk,1)i(λk)+ 1
N2(Sk,2)i(λk),
where (Sk,1)i and (Sk,2)i satisfy the same estimates as (X1N)ki and (X2N)ki. Hence, by the formulas above we deduce that
(TN)ki S1N(λ1), ..., SdN(λd)
=T0kS0k(λki)+ 1
N[(T0k)0S0k(λki)](Sk,1)i(λk) +1
N(X1,1N )ki
S01(λ11)+ 1
N(S1,1)1(λ1), ..., S0d(λdN)+ 1
N(SN,d)N(λd)
+Ei,
(5.24)
where the errorEi satisfies (due to the bounds in Theorem2.5and [5, Theorem 1.4]) s
X
i
kEik2L2(PGVE,βN )=O
(logN)2 N3/2
(5.25)
Also, by using again Theorem2.5 and [5, Theorem 1.4], with probability greater than 1−e−c(logN)2 and uniformly with respect toi∈{1, ..., N}, we have
|[(T0k)0S0k(λki+1)](Sk,1)i+1(λk)−[(T0k)0S0k(λki)](Sk,1)i(λk)|
6C(logN)N1/(σ−15)|λki+1−λki|, and
|(X1,1N )ki+1−(X1,1N )ki|
(S01)⊗N+ 1
NS1,1, ...,(S0d)⊗N+ 1 NSd,1
(ˆλ) 6C(logN)N1/(σ−15)
|S0k(λki+1)−S0k(λki)|+1
N|(Sk,1)i+1(λk)−(Sk,1)i(λk)|
6ClogN N1/(σ−15)|λki+1−λki|, while
T0kS0k(λki+1)−T0kS0k(λki) = (T0kS0k)0(λki)[λki+1−λki]+O(|λki+1−λki|2).
Recalling that, with probability greater than 1−e−N¯c, |λki+1−λki|6CNθ−1 when the {λki}Ni=1 are ordered andi∈[εN,(1−ε)N] (see (5.2) and (5.4)), we conclude that, with probability greater than 1−e−c(logN)2, and uniformly with respect to i∈[εN,(1−ε)N], we have
[(TN)ki+1−(TN)ki](S1N(λ1), ..., SdN(λd))
= (T0kS0k)0(λki)[λki+1−λki]+O
(logN)N1/(σ−15) N2−θ
. Combining this estimate with (5.25) and noticing that
N4/3
(logN)N2/(σ−15)
N2−θ +(logN)2 N3/2
!0 as N!∞,
providedθ<16 (recall that by assumptionσ>36; see Hypothesis2.1), the two statements follow from the fact thatTN(S1N, ... SNd ):RdN!RdN is an approximate transport map from (PGVE,βN )⊗d to PβN,aV and that the results are true underPGVE,βN due to [6, Theo-rem 1.3 and Corollary 1.5].