1: \section{Theorem \ref{PNG}}
2:
3: \subsection{Multi-layer discrete PNG}
4:
5: Before we give the proof of Theorem \ref{PNG} we must present some preliminary
6: results.
7:
8: How does one get a hand on the process $h$ described in the introduction?
9: In \cite{Jo2} it is shown
10: that $h$ can be embedded as the top curve in a multi-layer process
11: given by a family of non-intersecting paths $\{ h_i, 0 \leq i < N \}$,
12: $h=h_0$. It turns out, see \cite{Jo2}, that
13: this multi-layer process is an example of a discrete determinantal
14: process.
15: \begin{theorem}[Johansson]
16: Let $u,v \in \Z$ be such that $|u|,|v| < N$ and let $q = \alpha^2$. Set
17: \begin{equation*}
18: G(z,w) = (1 - \alpha)^{2(v-u)}
19: \frac{(1-\alpha/z)^{N+u} (1- \alpha w)^{N-v}}
20: {(1 - \alpha z)^{N-u} (1- \alpha / w)^{N+v}}
21: \end{equation*}
22: and
23: \begin{equation*}
24: \widetilde{K}_N (2u,x;2v,y) =
25: \frac{1}{(2 \pi i)^2}
26: \int_{\gamma_{r_2}} \frac{\udz}{z}
27: \int_{\gamma_{r_1}} \frac{\udw}{w}
28: \frac{z}{z-w} G(z,w)
29: \end{equation*}
30: where $\gamma_r$ is the circle with radius $r$ centered around the origin,
31: $\alpha < r_1 < r_2 < 1/ \alpha$ and $x,y \in \Z$. Furthermore, define
32: \begin{equation*}
33: \phi_{2u,2v}(x,y) =
34: \frac{1}{2 \pi}
35: \int_{-\pi}^{\pi} e^{i(y-x) \theta} G \lp( e^{i \theta}, e^{i \theta}\rp)
36: d \theta
37: \end{equation*}
38: for $u<v$ and $\phi_{2u,2v}(x,y) = 0$ for $u \geq v$. Set
39: \begin{equation*}
40: K_N (2u,x;2v,y) = \widetilde{K}_N(2u,x;2v,y) - \phi_{2u,2v}(x,y).
41: \end{equation*}
42: Then,
43: \begin{multline*}
44: \Prob \big[ (2u,x_j^{2u}) \in \lp\{
45: (2t,h_i(2t,2N-1)); 0 \leq i < N , \, |t|<N \rp\},
46: \\ \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad
47: |u| <N,1 \leq j \leq k_u \big] \\
48: = \fdet
49: \lp( K_N(2u,x_i^{2u};2v,x_j^{2v}) \rp)_{|u|,|v|<N,1\leq i \leq k_u,1\leq j \leq k_v}
50: \end{multline*}
51: for any $x_j^{2u} \in \Z$ and any $k_u \in \{ 0,\ldots , N \}$.
52: \end{theorem}
53: The asymptotic information about the kernel $K_N$ needed to prove Theorem
54: \ref{PNG} is contained in two lemmas. The first
55: can be extracted from chapter four in \cite{Jo2} and the proof of the second
56: is provided at the end of this section. Please note that we make a slight
57: redefinition of the function $\phi$ from the last section.
58: However, for the purposes of this text $\phi$ acts as one and the same.
59: \begin{lemma}\label{Klemmat}
60: Let $\tau, \tau'$ be any real numbers such that
61: \begin{eqnarray*}
62: & & u = \frac{1 + \alpha}{1 - \alpha} d^{-1} N^{2/3} \tau \in \Z_+ \\
63: & & v = \frac{1 + \alpha}{1 - \alpha} d^{-1} N^{2/3} \tau' \in \Z_+.
64: \end{eqnarray*}
65: Let $x,y \in \Z_+$ and define $x',y'$ by
66: \begin{eqnarray*}
67: & & x = 2 \alpha (1 - \alpha)^{-1} N + (x' - \tau^2) d N^{1/3} \\
68: & & y = 2 \alpha (1 - \alpha)^{-1} N + (y' - \tau'^2) d N^{1/3} .
69: \end{eqnarray*}
70: For any $L \in \R$ there exist positive constants, $c$ and $C$, such that
71: \begin{equation*}
72: | \widetilde{K}_N(2u,x;2v,y) | \leq C N^{-1/3} e^{- c(x'+y')}
73: \end{equation*}
74: if $x',y' \geq L$.
75:
76: If $|x'|,|y'| \leq \log{N}$, then the exists $c > 0$ such that
77: \begin{equation*}
78: d N^{1/3} \widetilde{K}_N(2u,x;2v,y) =
79: e^{\frac{\tau^3 - \tau'^3}{3} + y' \tau' - x' \tau}
80: \widetilde{A}(\tau,x';\tau',y') + \Ordo(N^{-c}).
81: \end{equation*}
82: \end{lemma}
83: \begin{lemma}\label{philemmat}
84: Let $x,y \in \Z_+$ and define $x',y'$ by
85: \begin{eqnarray*}
86: & & x = 2 \alpha (1 - \alpha)^{-1} N + x' d N^{1/3} \\
87: & & y = 2 \alpha (1 - \alpha)^{-1} N + y' d N^{1/3} .
88: \end{eqnarray*}
89: Take $s > 0$, let $u \sim N^{2/3}$ and define $v$ by
90: \begin{equation*}
91: v = u + \frac{1 + \alpha}{1 - \alpha} d^{-1} s N^{\gamma}
92: \end{equation*}
93: where $0 < \gamma < \frac{2}{3}$.
94: There exists a constant $C > 0$ such that
95: \begin{equation*}
96: \phi_{2u,2v} (x,y) = \frac{1}{d N^{1/3}} \phi(x',y') + \phi_{E}(x',y')
97: \end{equation*}
98: where
99: \begin{equation*}
100: \phi(x',y') = \frac{1}{\sqrt{4 \pi s N^{\gamma - 2/3}}}
101: e^{-\frac{(x'-y')^2}{4 s N^{\gamma - 2/3}}}
102: \end{equation*}
103: and
104: \begin{equation*}
105: |\phi_{E}(x',y')| \leq \Bigg\{
106: \begin{array}{l}
107: C N^{-\frac{3 \gamma}{2}} \\
108: \frac{C}{N^{1/3} | x'-y'| N^{\gamma}}
109: \end{array}
110: \end{equation*}
111: for all $x,y$.
112: \end{lemma}
113:
114: \subsection{Proof of Theorem \ref{PNG}}
115:
116: This proof is really a discrete analog of the proof of Theorem \ref{theorem1}.
117: Unfortunately things are more involved in this case where
118: $N^{\gamma - 2/3}$ plays the role of $\epsilon$.
119:
120: Please recall that $J_1 = \mu N + \psi d N^{1/3}$ where
121: $\mu = 2 \alpha (1 - \alpha)^{-1}$ and $q = \alpha^2$.
122: Set $J_i = J_{i-1} + y_i d N^{\gamma / 2} \in \Z$, $i = 2, \ldots ,m$, and
123: \begin{equation*}
124: \widetilde{I}_i = \lp\{ z \in \Z | z > J_i \rp\}.
125: \end{equation*}
126: Here the $y_i$'s are arbitrary numbers such that $J_i \in \Z$. For later
127: convenience we also define $\psi_i$, $i = 1, \ldots ,m$, by
128: $J_i = \mu N + \psi_i d N^{1/3}$.
129:
130: We will prove that
131: \begin{multline*}
132: \Prob \lp[ \# J_2 = \ldots = \# J_m =1, \# \widetilde{I}_2 = \ldots =
133: \# \widetilde{I}_m = 0 | \# J_1 =1, \# \widetilde{I}_1 = 0 \rp] \\ =
134: \phi_{2K_1,2K_2}(J_1,J_2) \cdots \phi_{2K_{m-1},2K_m}(J_{m-1},J_m) \\ +
135: \Ordo \lp( ( N^{- \gamma /2} )^{m-1} N^{-c} \rp).
136: \end{multline*}
137: This implies Theorem \ref{PNG}:
138: \begin{multline*}
139: \phi_{2K_1,2K_2}(J_1,J_2) \cdots \phi_{2K_{m-1},2K_m}(J_{m-1},J_m) \\ =
140: \frac{1}{\sqrt{4 \pi s_2}} e^{-\frac{y_2^2}{4s_2}} \cdots
141: \frac{1}{\sqrt{4 \pi s_m}} e^{-\frac{y_m^2}{4s_m}}
142: \frac{1}{(d N^{\gamma / 2})^{m-1}} \lp( 1 + \Ordo \lp( N^{-c} \rp) \rp)
143: \end{multline*}
144: by Lemma \ref{philemmat}. The sum of this function over the sets $A_i$
145: is a Riemann sum that is well approximated by the integral in
146: Theorem \ref{PNG}.
147:
148: Define the finite integer intervals $I_i$, $1 \leq i \leq m$, by
149: \begin{equation*}
150: I_i = \lp\{ z \in \Z ; J_i < z < \lfloor \mu N \rfloor + N \rp\}.
151: \end{equation*}
152: The probability of finding a particle in $\widetilde{I}_i$ but outside of
153: $I_i$ is very small:
154: \begin{align*}
155: \Prob & \lp[ \# ( \widetilde{I}_i \setminus I_i) \geq 1 \rp] \leq
156: \sum_{x \in \widetilde{I}_i \setminus I_i} \Prob [\# x = 1] =
157: \sum_{x \in \widetilde{I}_i \setminus I_i} K(x,x) \\ &
158: = \sum_{k=0}^{\infty}
159: K \Bigg(\lfloor \mu N \rfloor + \lp( \frac{1}{d}N^{2/3} +
160: \frac{k}{dN^{1/3}} \rp)dN^{1/3}, \\ &
161: \qquad \qquad \qquad \qquad \qquad \qquad \qquad
162: \lfloor \mu N \rfloor + \lp( \frac{1}{d}N^{2/3} +
163: \frac{k}{dN^{1/3}} \rp)dN^{1/3} \Bigg)
164: \\ & \qquad \qquad \qquad \qquad
165: \leq C \e ^{-\frac{1}{d} N^{2/3}} \sum_{k=0}^{\infty} e^{- \frac{k}{d N^{1/3}}} =
166: \Ordo \lp( e^{-c N^{2/3}} \rp).
167: \end{align*}
168: This means that we can work with $I_i$ instead of $\widetilde{I}_i$. We now
169: proceed much like we did in the proof of Theorem \ref{theorem1}.
170: If we set
171: \begin{equation*}
172: A = \lp\{ \# J_1 = 1, \ldots, \# J_m = 1 \rp\},
173: \end{equation*}
174: then
175: \begin{multline*}
176: \Prob \lp[ A , \# I_1 = \ldots = \# I_m = 0 \rp] +
177: \Prob [ A , \# I_1 = 0, (\# I_2 = \ldots = \# I_m = 0)^c] \\ =
178: \Prob [ A , \# I_1 = 0]
179: \end{multline*}
180: where
181: \begin{multline*}
182: \Prob [ A , \# I_1 = 0, (\# I_2 = \ldots = \# I_m = 0)^c] \\ =
183: \Prob \lp[ A , \# I_1 = 0, \cup_{i=2}^m \{ \# I_i \neq 0 \} \rp] \\
184: \leq \sum_{i=2}^{m} \Prob [ A , \# I_1 = 0, \# I_i \neq 0 ]
185: \leq \sum_{i=2}^{m} \Prob [ A , \# I_i \neq \# I_1 ],
186: \end{multline*}
187: and
188: \begin{align*}
189: \Prob [ A , \# I_i \neq \# I_1 ] & =
190: \E [\chi_{\{ \# J_1 = 1 \}} \cdots \chi_{\{ \# J_m = 1 \}} \cdot
191: \chi_{\{ \# I_1 \neq \# I_i \}} ] \\
192: & = \E [ \# J_1 \cdots \# J_m \cdot \chi_{\{ \# I_1 \neq \# I_i \}} ] \\
193: & \leq \E [ \# J_1 \cdots \# J_m (\# I_1 - \# I_i )^2 ].
194: \end{align*}
195: The second equality holds since the probability of finding two particles
196: at the same place is zero.
197:
198: We need to prove three things:
199: \begin{eqnarray*}
200: & & 1. \quad \Prob [A, \# I_1 = 0] \\
201: & & \qquad \quad
202: = \phi_{2K_1,2K_2}(J_1,J_2) \cdots \phi_{2K_{m-1},2K_m}(J_{m-1},J_m) \\
203: & & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad
204: \times \Prob[ \# J_1 =1, \# I_1 = 0] \\
205: & & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad
206: + \Ordo \lp( N^{-1/3-c} (N^{- \gamma / 2})^{m-1}\rp) \\
207: & & 2. \quad \E [ \# J_1 \cdots \# J_m \, (\# I_1 - \# I_i )^2 ]
208: = \Ordo \lp( N^{-1/3-c} (N^{- \gamma / 2})^{m-1}\rp) \\
209: & & 3. \quad \Prob[ \# J_1 =1, \# I_1 = 0] \geq C N^{-1/3}
210: \end{eqnarray*}
211: Before giving the proofs we need some preliminaries.
212:
213: When summing a function $f(x)$ over, say, $I_1$ we can write
214: \begin{equation*}
215: \sum_{x \in I_1}f(x) = \sum_{l=1}^{T_1}
216: f\lp(\mu N + \lp( \psi_1 + \frac{l}{d N^{1/3}} \rp) d N^{1/3} \rp).
217: \end{equation*}
218: where $T_1 \sim N$. The next lemma will be frequently used later on.
219: \begin{lemma}\label{phi_sum}
220: There exists constants $C_1,C_2 > 0$ such that
221: \begin{equation*}
222: \sum_{k=1}^{\infty} \phi \lp( k / N^{1/3}, x \rp) N^{-1/3} \leq C_1
223: \end{equation*}
224: and
225: \begin{equation*}
226: \sum_{k=1}^{N^2} \phi_E \lp( k / N^{1/3}, x \rp) \leq C_2 N^{-\gamma / 2}
227: \end{equation*}
228: for any $x \in \R$.
229: \end{lemma}
230: \begin{proof}
231: \begin{multline*}
232: \sum_{k=1}^{\infty} \phi \lp( k / N^{1/3}, x \rp) N^{-1/3} =
233: \sum_{k=1}^{\infty} \phi \lp( \frac{k - x N^{1/3}}{N^{1/3}}, 0 \rp) N^{-1/3} \\
234: \leq \sum_{k= - \infty}^{\infty}
235: \phi \lp( \frac{k - x N^{1/3}}{N^{1/3}}, 0 \rp) N^{-1/3}
236: = \lp[ f := x N^{1/3} - \lfloor x N^{1/3} \rfloor \rp] \\ =
237: \sum_{k= - \infty}^{\infty}
238: \phi \lp( \frac{k - f}{N^{1/3}}, 0 \rp) N^{-1/3} \\
239: \leq \sum_{k= - \infty}^{0} \phi \lp( \frac{k}{N^{1/3}}, 0 \rp) N^{-1/3} +
240: \phi \lp( \frac{1 - f}{N^{1/3}}, 0 \rp) N^{-1/3} +
241: \sum_{k=2}^{\infty}
242: \phi \lp( \frac{k - 1}{N^{1/3}}, 0 \rp) N^{-1/3} \\ \leq
243: 2 \sum_{k=1}^{\infty}
244: \phi \lp( \frac{k}{N^{1/3}}, 0 \rp) N^{-1/3} + 2 \leq C
245: \end{multline*}
246: \begin{align*}
247: \sum_{k=1}^{N^2} & \phi_E \lp( k/N^{1/3},x \rp) \leq
248: C N^{-\gamma}\sum_{k=1}^{x N^{1/3} - N^{\gamma}} \frac{1}{x N^{1/3} - k} \\
249: & \qquad \qquad \qquad \qquad
250: + C \sum_{x N^{1/3} - N^{\gamma}}^{x N^{1/3} + N^{\gamma}} N^{-3 \gamma/2}
251: + C N^{-\gamma}\sum_{x N^{1/3} + N^{\gamma}}^{N^2} \frac{1}{k - x N^{1/3}} \\
252: & \qquad \qquad
253: \leq C N^{-\gamma} \log{N} + C N^{-\gamma / 2} + C N^{-\gamma} \log{N} \leq
254: C N^{- \gamma / 2}
255: \end{align*}
256: \end{proof}
257: We now turn to the proof of 1. As in the proof of Theorem \ref{theorem1} we get
258: \begin{equation} \label{probeq}
259: \Prob [A, \# I_1 = 0] = \sum_{k=0}^{\infty} \frac{(-1)^k}{k!}
260: \E \lp[ \# J_1 \cdots \# J_m \, \# I_1^{[k]} \rp] .
261: \end{equation}
262: For $0 \leq r \leq m-1$ set
263: \begin{multline*}
264: D_r(k) \\ = \phi_{2K_1,2K_2}(J_1,J_2) \phi_{2K_2,2K_3}(J_2,J_3) \cdots
265: \phi_{2K_r,2K_{r+1}}(J_r,J_{r+1})
266: \sum_{x_i \in I_1, 1 \leq i \leq k} \\ \times \lp|
267: \begin{array}{ccccc}
268: K(J_{r+1},J_1) & K(J_{r+1},J_{r+2}) & \ldots &
269: K(J_{r+1},J_{m}) & K(J_{r+1},x_j) \\
270: \vdots & \vdots & \, & \vdots & \vdots \\
271: K(J_{m},J_1) & K(J_{m},J_{r+2}) & \ldots &
272: K(J_{m},J_{m}) & K(J_{m},x_j) \\
273: K(x_i,J_1) & K(x_i,J_{r+2}) & \ldots &
274: K(x_i,J_{m}) & K(x_i,x_j)
275: \end{array} \rp| .
276: \end{multline*}
277: The indicies $i,j$ run from $1$ to $k$ and if $r=0$ the (empty) product of
278: $\phi$-functions is to be interpreted as $1$. Let $\widetilde{D}_r(k)$ be
279: like $D_r(k)$ but having $\widetilde{K}(J_{r+1},J_{r+2})$ in position
280: (1,2) in the matrix. We want to show that
281: \begin{equation*}
282: |D_0(k) - D_r(k)| \leq
283: N^{-1/3-c} \lp( N^{- \gamma/2} \rp)^{m-1} (Ck)^{\frac{k+m}{2}}
284: \end{equation*}
285: which, by the induction argument in the proof of Theorem \ref{theorem1},
286: follows if we can prove that
287: \begin{equation} \label{pr_2_Dtilde}
288: |\widetilde{D}_r(k)| \leq
289: N^{-1/3-c} \lp( N^{- \gamma/2} \rp)^{m-1} (Ck)^{\frac{k+m}{2}} .
290: \end{equation}
291: To show this we shall use Hadamard's inequality and therefore need to
292: estimate sums of column elements squared
293: (confer with the proof of Theorem 1). Lemmas \ref{Klemmat}, \ref{philemmat}
294: and \ref{phi_sum} will be frequently used below.
295:
296: Column 1:
297: \begin{equation*}
298: \sum_{i=r+1}^m K^2(J_i,J_1) + \sum_{i=1}^k K^2(x_i,J_1) \leq
299: C N^{-2/3}(m+k)
300: \end{equation*}
301: \indent Column 2:
302: \begin{equation*}
303: \widetilde{K}^2(J_{r+1},J_{r+2}) + \sum_{i=r+2}^m K^2(J_i,J_{r+2}) \leq
304: C N^{-2/3} m
305: \end{equation*}
306: and
307: \begin{multline*}
308: \sum_{i=1}^k K^2(x_i,J_{r+2}) \\ \leq
309: C N^{-2/3} \sum_{i=1}^k
310: \Big[ 1 + \phi \lp( l_i/dN^{1/3},\psi_1 - \psi_{r+2} \rp) \\ +
311: N^{1/3} \phi_E \lp(l_i/dN^{1/3},\psi_1 - \psi_{r+2} \rp) \Big]^2.
312: \end{multline*}
313:
314: Columns $3, \ldots , m-r$ ($r+3 \leq j \leq m$), if they exist:
315: \begin{equation*}
316: \sum_{i=r+1}^m K^2(J_i,J_j) + \sum_{i=1}^k K^2(x_i, J_j)
317: \leq C N^{-\gamma} (k+m)
318: \end{equation*}
319:
320: Last $k$ columns ($1 \leq j \leq k$):
321: \begin{equation*}
322: \sum_{i=r+1}^m K^2(J_i,x_j) + \sum_{i=1}^k K^2(x_i, x_j) \leq
323: C (k+m) N^{-2/3} e^{- c l_j N^{-1/3}}
324: \end{equation*}
325: Using Hadamard's inequality we get after some manipulations that
326: \begin{multline*}
327: |\widetilde{D}_r(k)| \\ \leq \sum_{l_1, \ldots ,l_k=1}^{T_1}
328: N^{-2/3} \lp(N^{-\gamma / 2}\rp)^{m-2}(Ck)^{\frac{k+m}{2}} \lp( N^{-1/3}\rp)^k
329: \prod_{i=1}^k e^{- c l_i N^{-1/3}} \\ \times
330: \sum_{i=1}^k \Big[ 1 + \phi \lp(l_i/dN^{1/3},\psi_1 - \psi_{r+2} \rp) \\ +
331: N^{1/3} \lp|\phi_E \lp(l_i/dN^{1/3},\psi_1 - \psi_{r+2} \rp) \rp| \Big] .
332: \end{multline*}
333: It follows from Lemma \ref{phi_sum} that
334: \begin{equation*}
335: \sum_{l_i=1}^{T_1} e^{-c l_i N^{-1/3}}
336: \phi \lp(l_i/dN^{1/3},\psi_1 - \psi_{r+2} \rp) N^{-1/3} \leq C
337: \end{equation*}
338: and also that
339: \begin{equation*}
340: \sum_{l_i=1}^{T_1} e^{-c l_i N^{-1/3}}
341: \lp| \phi_E \lp(l_i/dN^{1/3},\psi_1 - \psi_{r+2} \rp) \rp| \leq C N^{- \gamma / 2}
342: \end{equation*}
343: From this we get (\ref{pr_2_Dtilde}).
344:
345: To get 1 we also need to show that
346: \begin{multline}\label{det_sum}
347: \sum_{x_i \in I_1, 1 \leq i \leq k}
348: \fdet \lp[
349: \begin{array}{cc}
350: K(J_m,J_1) & K(J_m,x_j) \\
351: K(x_i,J_1) & K(x_i,x_j)
352: \end{array} \rp]_{1 \leq i,j \leq k} \\
353: = \sum_{x_i \in I_1, 1 \leq i \leq k}
354: \fdet \lp[
355: \begin{array}{cc}
356: K(J_1,J_1) & K(J_1,x_j) \\
357: K(x_i,J_1) & K(x_i,x_j)
358: \end{array} \rp]_{1 \leq i,j \leq k}
359: \\ + N^{-1/3 - c} \Ordo \lp( (Ck)^{\frac{k+m}{2}} \rp).
360: \end{multline}
361: Write
362: \begin{equation*}
363: x_i = \mu N + \lp( \psi_1 + \frac{l_i}{d N^{1/3}} \rp)d N^{1/3}
364: \end{equation*}
365: and consider first the case $1 \leq l_i \leq N^{1/3} \log{N}$. From
366: Lemmas \ref{K_approx} and \ref{Klemmat} it is straight forward to deduce that
367: if $z=x_i$ or $z=J_1$ then
368: \begin{equation*}
369: K(J_m,z) = K(J_1,z) + \Ordo (N^{-1/3-c} ).
370: \end{equation*}
371: We now expand the determinant in the sum to the left in (\ref{det_sum}).
372: \begin{multline*}
373: \fdet \lp[
374: \begin{array}{cc}
375: K(J_m,J_1) & K(J_m,x_j) \\
376: K(x_i,J_1) & K(x_i,x_j)
377: \end{array} \rp]_{1 \leq i,j \leq k} \\
378: = \fdet \lp[
379: \begin{array}{cc}
380: K(J_1,J_1) & K(J_1,x_j) \\
381: K(x_i,J_1) & K(x_i,x_j)
382: \end{array} \rp]_{1 \leq i,j \leq k} \\
383: + \Ordo (N^{-1/3-c} ) \sum_{p=1}^k
384: \fdet \lp[ K(x_i,J_1) \quad K(x_i,x_j) \rp]_{1 \leq i,j \leq k, j \neq p} \\
385: + \Ordo ( N^{-1/3-c} )\, \fdet \lp[ K(x_i,x_j) \rp]_{1 \leq i,j \leq k}
386: \end{multline*}
387: We now use Hadamard's inequality to get
388: \begin{multline*}
389: \sum_{l_i=1}^{N^{1/3} \log{N}}
390: |\fdet \lp[ K(x_i,J_1) \quad K(x_i,x_j) \rp]_{1 \leq i,j \leq k, j \neq p}| \\
391: \leq \sum_{l_i=1}^{N^{1/3} \log{N}}
392: (C k N^{-2/3} )^{k/2}
393: e^{-N^{-1/3}(l_1 + \ldots + l_{p-1} + l_{p+1} + \ldots + l_k)} \\
394: \leq \lp( C k \rp)^{k/2} \log{N}
395: \end{multline*}
396: and
397: \begin{multline*}
398: \sum_{l_i=1}^{N^{1/3} \log{N}}
399: |\fdet \lp[ K(x_i,x_j\rp]_{1 \leq i,j \leq k}| \\
400: \leq \sum_{l_i=1}^{N^{1/3} \log{N}}
401: \lp(C k N^{-2/3} \rp)^{k/2}
402: e^{-N^{-1/3}(l_1 + \ldots + l_k)} \leq \lp( C k \rp)^{k/2}.
403: \end{multline*}
404: This takes care of the summation over $1 \leq l_i \leq N^{1/3} \log{N}$,
405: $1 \leq i \leq m$. By using Hadamard's inequality once more one readily
406: shows that the contribution coming from the remaining terms in the
407: sums in (\ref{det_sum}) is small enough to make (\ref{det_sum}) hold.
408:
409: We now prove 2. Note that
410: \begin{equation*}
411: (\# I_1 - \#I_i)^2 = I_i^{[2]} + I_1^{[2]} + I_i + I_1 - 2 I_i I_1.
412: \end{equation*}
413: By arguing as in the proof of 1 above we obtain
414: \begin{equation*}
415: \E \lp[ \# J_1 \cdots \# J_m \# I_u^{[k]} \rp] =
416: \end{equation*}
417: \begin{multline*}
418: = \phi_{2K_1,2K_2}(J_1,J_2) \cdots \phi_{2K_{m-1},2K_m}(J_{m-1},J_m) \\
419: \qquad \qquad \qquad \qquad
420: \times \sum_{x_1, x_k \in I_u}
421: \fdet \lp[
422: \begin{array}{cc}
423: K(J_m,J_1) & K(J_m,x_s) \\
424: K(x_r,J_1) & K(x_r,x_s)
425: \end{array} \rp]_{1 \leq r,s \leq k}
426: \\ + \Ordo \lp( N^{-1/3-c - \frac{\gamma(m-1)}{2}} \rp)
427: \end{multline*}
428: where $u=1,i$ and $k=1,2$. One also gets
429: \begin{multline*}
430: \E [ \# J_1 \cdots \# J_m \# I_1 \# I_i] \\
431: = \phi_{2K_1,2K_2}(J_1,J_2) \cdots \phi_{2K_{m-1},2K_m}(J_{m-1},J_m) \\
432: \times \sum_{x \in I_1, y \in I_i}
433: \fdet \lp[
434: \begin{array}{ccc}
435: K(J_m,J_1) & K(J_m,x) & K(J_m,y) \\
436: K(x,J_1) & K(x,x) & K(x,y) \\
437: K(y,J_1) & K(y,x) & K(y,y)
438: \end{array} \rp]
439: \\ + \Ordo \lp( N^{-1/3-c - \frac{\gamma(m-1)}{2}} \rp).
440: \end{multline*}
441: We omit the details. Using Lemma \ref{Klemmat} and Lemma
442: \ref{K_approx} one readily gets
443: \begin{multline*}
444: \sum_{x_1, x_k \in I_i}
445: \fdet \lp[
446: \begin{array}{cc}
447: K(J_m,J_1) & K(J_m,x_s) \\
448: K(x_r,J_1) & K(x_r,x_s)
449: \end{array} \rp]_{1 \leq r,s \leq k} \\
450: = \E \lp[ \# J_1 \# I_1^{[k]} \rp] + \Ordo (N^{-1/3-c} )
451: \end{multline*}
452: for $k=1,2$ and
453: \begin{multline*}
454: \sum_{x \in I_1,y \in I_i} \fdet \lp[
455: \begin{array}{ccc}
456: K(J_m,J_1) & K(J_m,x) & K(J_m,y) \\
457: K(x,J_1) & K(x,x) & \widetilde{K}(x,y) \\
458: K(y,J_1) & K(y,x) & K(y,y)
459: \end{array} \rp] \\
460: = \E \lp[ \# J_1 \# I_1^{[2]} \rp] + \Ordo ( N^{-1/3-c} ).
461: \end{multline*}
462: We now see that 2 follows if
463: \begin{multline} \label{phi_sum2}
464: \sum_{x \in I_1,y \in I_i} \phi_{2K_1,2K_i}(x,y)
465: \fdet \lp[
466: \begin{array}{cc}
467: K(J_m,J_1) & K(J_m,x) \\
468: K(y,J_1) & K(y,x)
469: \end{array} \rp] \\
470: = \E [ \# J_1 \# I_1] + \Ordo (N^{-1/3-c} ).
471: \end{multline}
472: We shall prove this by showing that both sides are well approximated by
473: integrals. On the integral containing the function $\phi$ we can then apply
474: Lemma \ref{approx_delta}.
475:
476: By using Lemma \ref{phi_sum} we get rid of the error term associated
477: with $\phi_E$:
478: \begin{multline*}
479: \sum_{l_1,l_2=1}^N
480: \phi_E \lp(\psi_1 + \frac{l_1}{d N^{1/3}}, \psi_i + \frac{l_2}{d N^{1/3}} \rp)
481: e^{-\frac{l_1+l_2}{N^{1/3}}} N^{-2/3} \\
482: \leq C \sum_{l_2=1}^N e^{-\frac{l_2}{N^{1/3}}} N^{-2/3-\gamma / 2}
483: \leq C N^{-1/3 - \gamma / 2}
484: \end{multline*}
485: The following calculation, again using Lemma \ref{phi_sum},
486: shows that the main contribution to the
487: sums in (\ref{phi_sum2}) comes from summing over
488: $1 \leq l_1,l_2 \leq N^{1/3} \log{N}$.
489: \begin{multline*}
490: \sum_{l_1=N^{1/3}\log{N}}^N \sum_{l_2=1}^N
491: \phi \lp( l_2/dN^{1/3}, \psi_1 - \psi_i + l_2/dN^{1/3} \rp)
492: e^{-\frac{l_1+l_2}{N^{1/3}}} N^{-2/3} \\
493: \leq \sum_{l_1=N^{1/3}\log{N}}^N C e^{-l_1/N^{1/3}} N^{-1/3} \leq CN^{-1}
494: \end{multline*}
495: We shall use Euler's summation formula for two variables:
496: \begin{lemma}
497: Let $f(x,y)$ be a function of two variables such that its partial
498: derivatives up to second order are continuous in the rectangle
499: \begin{equation*}
500: \{ (x,y) |a \leq x \leq b, c \leq y \leq d \}
501: \end{equation*}
502: where $a,b,c,d$ are integers. Then
503: \begin{align*}
504: \sum_{a \leq m \leq b} \sum_{c \leq n \leq d} f(m,n) = &
505: \int_a^b \int_c^d f(x,y) \, \udx \udy \\
506: & + \int_a^b \int_c^d f_x(x,y)(x - \lfloor x \rfloor) \, \udx \udy \\
507: & + \int_a^b \int_c^d f_y(x,y)(y - \lfloor y \rfloor) \, \udx \udy \\
508: & + \int_a^b \int_c^d f_{xy}(x,y)(x - \lfloor x \rfloor)
509: (y - \lfloor y \rfloor)\, \udx \udy
510: \end{align*}
511: \end{lemma}
512: The case that we are interested in is when
513: \begin{equation*}
514: f(x,y) =
515: \phi \lp( \psi_1 + \frac{x}{d N^{1/3}},\psi_i + \frac{y}{d N^{1/3}} \rp)
516: g \lp(x/d N^{1/3},y/d N^{1/3} \rp) N^{-1}
517: \end{equation*}
518: where
519: \begin{equation*}
520: |g_x(x,y)|,|g_y(x,y)|, |g_x(x,y)| \leq C e^{-c(x+y)} .
521: \end{equation*}
522: We need to show that the integrals involving the absolute values of
523: $f_x(x,y)$, $f_y(x,y)$ and $f_{xy}(x,y)$ are negligible.
524: We only present the details for $|f_x(x,y)|$ here, the other terms
525: are treated similarly.
526: \begin{multline*}
527: \int_1^{N^{1/3} \log{N}} \int_1^{N^{1/3} \log{N}} |f_x(x,y)| \, \udx \udy \\
528: \leq (d N^{1/3})^2 \int_{\psi_1}^{\infty} \int_{\psi_i}^{\infty}
529: \lp|f_x \lp((x-\psi_1) d N^{1/3},(y-\psi_i) d N^{1/3} \rp) \rp| \, \udx \udy \\
530: \leq C N^{-2/3} \int_{\psi_1}^{\infty} \int_{\psi_i}^{\infty}
531: \lp( |\phi_x(x,y)| + \phi(x,y) \rp) e^{-c(x+y)} \, \udx \udy
532: \end{multline*}
533: By Lemma (\ref{approx_delta})
534: \begin{equation*}
535: \int_{\psi_1}^{\infty} \int_{\psi_i}^{\infty}
536: \phi(x,y) e^{-c(x+y)} \, \udx \udy \leq C.
537: \end{equation*}
538: The remaining term demands some analysis.
539: \begin{align*}
540: \int_{\psi_1}^{\infty} & \int_{\psi_i}^{\infty}
541: |\phi_x(x,y)| e^{-c(x+y)} \, \udx \udy \\ & =
542: \int_{\psi_1}^{\infty} \int_{\psi_i}^{\infty} \frac{|x-y|}{2 N^{\gamma - 2/3}}
543: \frac{1}{\sqrt{4 \pi N^{\gamma - 2/3}}}
544: e^{-\frac{(x-y)^2}{4 N^{\gamma - 2/3}}} e^{-c(x+y)} \, \udx \udy \\
545: & = \int_{\psi_1}^{\infty} \udx \lp( \int_{\psi_i}^x + \int_x^{\infty} \rp)
546: \frac{|x-y|}{2 N^{\gamma - 2/3}}
547: \frac{1}{\sqrt{4 \pi N^{\gamma - 2/3}}}
548: e^{-\frac{(x-y)^2}{4 N^{\gamma - 2/3}}} e^{-c(x+y)} \, \udy
549: \end{align*}
550: \begin{multline*}
551: \int_{\psi_1}^{\infty} \udx \int_{\psi_i}^x
552: \frac{x-y}{2 N^{\gamma - 2/3}}
553: \frac{1}{\sqrt{4 \pi N^{\gamma - 2/3}}}
554: e^{-\frac{(x-y)^2}{4 N^{\gamma - 2/3}}} e^{-c(x+y)} \, \udy \\
555: = \int_{\psi_1}^{\infty} \udx \lp(
556: \lp[ \phi(x,y) e^{-c(x+y)} \rp]_{\psi_i}^{x} +
557: c \int_{\psi_i}^x \phi(x,y) e^{-c(x+y)} \, \udy \rp) \\
558: \leq C N^{1/3 - \gamma / 2} + C \leq C N^{1/3 - \gamma / 2}
559: \end{multline*}
560: We can do the same calculation for the remaining integral. The
561: $|f_x(x,y)|$ integral is hence $\Ordo \lp(N^{-1/3 - \gamma / 2}\rp)$ and
562: the same goes for the $|f_y(x,y)|$ and $|f_{xy}(x,y)|$ integrals.
563:
564: Set
565: \begin{equation*}
566: A^{\tau_1}(x,y) = A \lp(\tau_1,x + \tau_1^2;\tau_1, y + \tau_1^2 \rp).
567: \end{equation*}
568: Applying the above calculations to the left hand side of (\ref{phi_sum2})
569: and using Lemmas \ref{princ_lemma}, \ref{approx_delta}, \ref{K_approx} and
570: \ref{Klemmat} we obtain
571: \begin{multline*}
572: \sum_{x \in I_1,y \in I_i} \phi_{2K_1,2K_i}(x,y)
573: \fdet \lp[
574: \begin{array}{cc}
575: K(J_m,J_1) & K(J_m,x) \\
576: K(y,J_1) & K(y,x)
577: \end{array} \rp] \\
578: = \sum_{l_1,l_2 = 1}^{N^{1/3 \log{N}}}
579: \frac{1}{dN^{1/3}} \phi \lp( \psi_1 + l_1/ dN^{1/3},\psi_i + l_2/ dN^{1/3}\rp)
580: \lp( \frac{1}{dN^{1/3}} \rp)^2 \\
581: \times \lp|
582: \begin{array}{cc}
583: A^{\tau_1}(\psi_1, \psi_1) &
584: e^{\tau_1 \frac{l_1}{dN^{1/3}}}
585: A^{\tau_1}(\psi_1,\psi_1 + l_1/dN^{1/3}) \\
586: e^{-\tau_1 \frac{l_2}{dN^{1/3}}}
587: A^{\tau_1}(\psi_1 + l_2/dN^{1/3},\psi_1) &
588: A^{\tau_1}(\psi_i + l_2/dN^{1/3},\psi_1 + l_1/dN^{1/3})
589: \end{array} \rp| \\
590: \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad
591: \qquad \qquad \qquad \qquad
592: + \Ordo(N^{-1/3-c})
593: \\= \frac{1}{d N^{1/3}} \int_{\psi_1}^{\infty}
594: \fdet \lp[
595: \begin{array}{cc}
596: A^{\tau_1}(\psi_1,\psi_1) & A^{\tau_1}(\psi_1,x) \\
597: A^{\tau_1}(x,\psi_1) & A^{\tau_1}(x,x)
598: \end{array} \rp] \udx
599: + \Ordo(N^{-1/3-c}).
600: \end{multline*}
601: We get the same expression for the right hand side of (\ref{phi_sum2})
602: when applying Euler's summation formula. This concludes the proof of 2.
603:
604: Let $F_2(t)$ be the Tracy-Widom distribution function corresponding to
605: the largest eigenvalue of the Gaussian Unitary Ensemble (GUE), \cite{TW}.
606: That 3 is true follows from the fact that $F_2'(t)>0$ $\forall t$, see
607: \cite{TW}, together with the next lemma.
608: \begin{lemma} \label{TWconv}
609: Let $J_1$ and $I_1$ be as above. It holds that
610: \begin{equation*}
611: \Prob [ \# J_1 = 0, \# I_1 = 0] =
612: \frac{1}{d N^{1/3}} F_2'(\psi_1 + \tau_1^2) + \Ordo(N^{-2/3}).
613: \end{equation*}
614: \end{lemma}
615: \noindent \textbf{Proof:} This will, again, be an exercise in using
616: Hadamard's inequality. We have the following representation for $F_2'$
617: (see the third equality in (\ref{twprimeq})):
618: \begin{equation}\label{TWdensity}
619: F_2'(t) = \sum_{k=0}^{\infty} \frac{(-1)^k}{k!}
620: \int_{(t,\infty)^k} \fdet(A(x_i,x_j))_{0 \leq i,j \leq k} \, d^kx
621: \end{equation}
622: where $x_0 = t$. In three steps we will now show that
623: \begin{equation*}
624: d N^{1/3}\sum_{k=0}^{\infty} \frac{(-1)^k}{k!}
625: \sum_{x_i \in I_1,1 \leq i \leq k} \fdet(K(x_i,x_j))_{0 \leq i,j \leq k}
626: \end{equation*}
627: where $x_0 = J_1$, is well approximated by the right hand side in
628: (\ref{TWdensity}). By (\ref{probeq}) this will prove the lemma. In steps
629: one and two we will use Lemma \ref{Klemmat} to insert the kernel $A$
630: instead of $K$. In in the last step we show that we can change from summation
631: to integration.
632:
633: First we show that we can sum over
634: $x_i = \mu N + \lp( \psi_1 + l_i/dN^{1/3} \rp)d N^{1/3}$ where
635: $1 \leq l_i \leq N^{1/3} \log{N}$, $1 \leq i \leq k$,
636: instead of over $I_1$. By Hadamard's inequality and Lemma \ref{Klemmat}
637: \begin{multline*}
638: \fdet \lp( K(x_i,x_j) \rp)_{0 \leq i,j \leq k} \leq
639: \lp( \prod_{j=0}^k \sum_{i=0}^k K^2(x_i,x_j) \rp)^{1/2} \\
640: \leq \lp( C(k+1) N^{-2/3}\prod_{j=1}^k C(k+1) N^{-2/3}
641: e^{-c l_j / N^{1/3}} \rp)^{1/2} \\
642: \leq N^{-1/3} (C(k+1))^{\frac{k+1}{2}} \prod_{j=0}^k
643: e^{-c l_j / N^{1/3}} N^{-1/3}.
644: \end{multline*}
645: We have that
646: \begin{multline*}
647: \sum_{\substack{l_i = 1 \\ 1 \leq i \leq k}}^{\infty}
648: \prod_{j=1}^k e^{-c l_j / N^{1/3}} N^{-1/3} -
649: \sum_{\substack{l_i = 1 \\ 1 \leq i \leq k}}^{N^{1/3} \log{N}}
650: \prod_{j=1}^k e^{-c l_j / N^{1/3}} N^{-1/3} \\
651: \leq k \sum_{l_1=1}^{N^{1/3} \log{N}}
652: \sum_{\substack{l_i = 1 \\ 2 \leq i \leq k}}^{\infty}
653: \prod_{j=1}^k e^{-c l_j / N^{1/3}} N^{-1/3}
654: \leq k \, C^{k} N^{-1}.
655: \end{multline*}
656: Since
657: \begin{equation*}
658: \sum_{k=1}^{\infty}\frac{1}{k!} N^{-1/3} (C(k+1))^{\frac{k+1}{2}}k N^{-1}
659: \leq C N^{-4/3}
660: \end{equation*}
661: we see that we can indeed restrict the summation.
662:
663: In the second step we replace $K$ by $A$. As before we shall use the
664: notation $A^{\tau}(x,y) = A(x+\tau^2,y+\tau^2)$. For
665: $1 \leq l_i \leq N^{1/3} \log{N}$ it holds by Lemma \ref{Klemmat} that
666: \begin{multline*}
667: \fdet(K(x_i,x_j))_{0 \leq i,j \leq k} \\ =
668: \frac{1}{(dN^{1/3})^{k+1}}
669: \fdet \lp(A^{\tau_1}(l_i/d N^{1/3},l_j/d N^{1/3}) +
670: \Ordo(N^{-c})\rp)_{0 \leq i,j \leq k}
671: \end{multline*}
672: where we let $l_0 = \psi_1 d N^{1/3}$. If we expand the determinant in the right
673: hand side we get $(k+1)^2$ error terms of type
674: \begin{equation*}
675: \frac{N^{-c}}{(dN^{1/3})^{k+1}}
676: \fdet \lp(A^{\tau_1}(l_i/d N^{1/3},l_j/d N^{1/3}) +
677: \Ordo(N^{-c})\rp)_{\substack{0 \leq i,j \leq k \\ i \neq i_0, j \neq j_0}}.
678: \end{equation*}
679: An application of Hadamard's inequality together with Lemma \ref{Klemmat}
680: shows that the total error we get when changing from $K$ to $A^{\tau_1}$
681: is of order $N^{-1/3-c}$. We omit the details.
682:
683: Finally we want to go from summation to integration. To do this we shall
684: use that
685: \begin{multline} \label{eulmacl1}
686: \sum_{l_i=1}^{N^{1/3} \log{N}}
687: A^{\tau_1}(l_i /dN^{1/3},x)A^{\tau_1}(y,l_i/dN^{1/3}) \\
688: = dN^{1/3} \int_0^{\infty} A^{\tau_1}(z,x) A^{\tau_1}(y,z) \udz
689: + \Ordo \lp(e^{-x-y} \rp)
690: \end{multline}
691: and
692: \begin{multline}\label{eulmacl2}
693: \sum_{l_i=1}^{N^{1/3} \log{N}} A^{\tau_1}(l_i/dN^{1/3},l_i/dN^{1/3}) \\ =
694: dN^{1/3} \int_0^{\infty}A^{\tau_1}(z,z)\udz + \Ordo(1).
695: \end{multline}
696: This follows from Euler-Maclaurins summation formula and Lemma \ref{K_approx}.
697: We will show that
698: \begin{multline}\label{deteq}
699: \sum_{\substack{l_i=1 \\ 1 \leq i \leq k}}^{N^{1/3} \log{N}}
700: \frac{1}{(d N^{1/3})^{k+1}}
701: \fdet \lp( A^{\tau_1}(l_i/d N^{1/3},l_j/d N^{1/3}) \rp)_{0 \leq i,j \leq k} \\
702: = \frac{1}{d N^{1/3}} \int_{(0,\infty)^{k}}
703: \fdet \lp( A^{\tau_1}(y_i,y_j) \rp)_{0 \leq i,j \leq k} \, d^ky
704: \\ + \Ordo \lp( (Ck)^{\frac{k+5}{2}} N^{-2/3}\rp)
705: \end{multline}
706: where $l_0 = d N^{1/3} \psi_1$ and $y_0 = \psi_1$.
707: This will prove the lemma since
708: \begin{equation*}
709: \sum_{k=1}^{\infty} \frac{1}{k!} (Ck)^{\frac{k+5}{2}} < \infty.
710: \end{equation*}
711: For $r = 0, \ldots ,k$ we set
712: \begin{equation*}
713: D_r = \frac{1}{(d N^{1/3})^{k-r+1}}
714: \fdet \lp( A^{\tau_1}(z_i,z_j) \rp)_{0 \leq i,j \leq k}
715: \end{equation*}
716: where
717: \begin{equation*}
718: z_i = \Bigg\{
719: \begin{array}{ll}
720: \psi_1 & \quad i=0 \\
721: y_i & \quad 1 \leq i \leq r \\
722: l_i/dN^{1/3} & \quad r+1 \leq i \leq k
723: \end{array}.
724: \end{equation*}
725: Please note that $D_0$ is what we sum over in (\ref{deteq}) and that $D_k$ is
726: what we integrate over. $D_r$ should roughly be what we
727: get after having changed summation over $l_1, \ldots ,l_r$ to integration over
728: $y_1, \ldots ,y_r$. We can expand $D_r$ in such a way that we get
729: $k^2$ terms of type
730: \begin{multline*}
731: \pm \frac{1}{(d N^{1/3})^{k-r+1}}
732: A^{\tau_1}(z_{i_0},l_{r+1}/d N^{1/3}) A^{\tau_1}(l_{r+1}/d N^{1/3},z_{j_0}) \\
733: \times \fdet \lp( A^{\tau_1}(z_i,z_j)\rp)_{\substack{0 \leq i,j \leq k \\
734: i \neq r+1,i_0 \\ j \neq r+1,j_0}}
735: \end{multline*}
736: and one term
737: \begin{equation*}
738: \frac{1}{(d N^{1/3})^{k-r+1}}
739: A^{\tau_1}(l_{r+1}/d N^{1/3},l_{r+1}/d N^{1/3})
740: \fdet \lp( A^{\tau_1}(z_i,z_j)\rp)_{\substack{0 \leq i,j \leq k \\
741: i,j \neq r+1}}.
742: \end{equation*}
743: We now apply (\ref{eulmacl1}) and (\ref{eulmacl2}) and therefore need to
744: deal with the corresponding errors.
745: \begin{multline*}
746: C (N^{1/3})^{k-r+1} e^{-z_{i_0} - z_{j_0}}
747: \fdet \lp( A^{\tau_1}(z_i,z_j) \rp)_{\substack{0 \leq i,j \leq k \\
748: i \neq i_0, r+1 \\ j \neq j_0, r+1}} \\
749: \leq C (N^{1/3})^{k-r+1} e^{-z_{i_0} - z_{j_0}}
750: \lp( \prod_{\substack{j=0 \\ j \neq j_0, r+1}}^k
751: C(k-1) e^{-c z_j} \rp)^{1/2} \\
752: \leq (N^{1/3})^{k-r+1} (C(k-1))^{\frac{k-1}{2}}
753: \prod_{\substack{j=1 \\ j \neq r+1}}^k e^{-c z_j}
754: \end{multline*}
755: Since
756: \begin{equation*}
757: \int_{(0,\infty)^r} d^r x
758: \sum_{\substack{l_i=1 \\ r+2 \leq i \leq k}}^{N^{1/3} \log{N}}
759: \prod_{\substack{j=1 \\ j \neq r+1}}^k e^{-c z_j}
760: \leq C^k (N^{1/3})^{k-(r+1)}
761: \end{equation*}
762: we find that the error from the $k^2$ terms of the first type is estimated by
763: \begin{equation*}
764: k^2 (C(k-1))^{\frac{k-1}{2}} N^{-2/3}.
765: \end{equation*}
766: The error coming from the remaining term can be treated in the same way.
767: Changing
768: from summation over $l_i$ to integration over $y_i$, $1 \leq i \leq k$,
769: hence results in an error estimated by
770: \begin{equation*}
771: k \, k^2 (C(k-1))^{\frac{k-1}{2}} N^{-2/3} = (Ck)^{\frac{k+5}{2}} N^{-2/3}
772: \end{equation*}
773: as needed.
774:
775: \noindent \textbf{Proof of Lemma \ref{philemmat}:}
776: By definition
777: \begin{equation*}
778: \phi_{2u,2v} (x,y) = \frac{(1-\alpha)^{2(v-u)}}{2 \pi}
779: \int_{- \pi}^{\pi} e^{i(y-x) \theta + (u-v)
780: \log{(1 + \alpha^2 - 2 \alpha \cos{\theta})}} d\theta.
781: \end{equation*}
782: Define
783: \begin{equation*}
784: g(\theta) = \log{(1 + \alpha^2 - 2 \alpha \cos{\theta})}
785: \end{equation*}
786: in $[-\pi,\pi]$.
787: This function is analytic in a neighbourhood of zero and a Maclaurin
788: expansion gives
789: \begin{equation*}
790: g(\theta) = \log{(1-\alpha)^2} + \frac{\alpha}{(1-\alpha)^2 } \theta^2
791: + c_2 \theta^4 + \mathcal{O}(\theta^6)
792: \end{equation*}
793: where $c_4 < 0$. It is easy to see that for any $\delta > 0$ there exists
794: $\epsilon > 0$ such that
795: \begin{equation*}
796: g(\theta) \geq \log{(1-\alpha)^2} + \epsilon
797: \end{equation*}
798: if $|\theta| \geq \delta$. Hence
799: \begin{equation*}
800: \lp| \int_{|\theta| > \delta} \frac{(1-\alpha)^{2(v-u)}}{2 \pi}
801: e^{i(y-x)\theta + (u-v)g(\theta)} d\theta \rp|
802: \leq \frac{1}{2 \pi}
803: \int_{\delta}^{\pi} e^{(u-v) \epsilon} d \theta
804: \sim e^{-\epsilon N^{\gamma}}.
805: \end{equation*}
806: We expect that the main contribution to $\phi_{2u,2v}$ will be
807: \begin{align*}
808: \frac{1}{2 \pi} \int_{-\delta}^{\delta}
809: & e^{i(y-x)\theta + (u-v) \frac{\alpha}{(1-\alpha)^2} \theta^2}
810: \, d \theta \\
811: & = \frac{1}{2 \pi} \int_{-\delta}^{\delta}
812: e^{i(y'-x') d N^{1/3} \theta - s d^2 N^{\gamma} \theta^2}
813: \, d \theta = \lp[ t = \sqrt{s}d N^{1/3}\theta \rp] \\
814: & = \frac{1}{2 \pi \sqrt{s} d N^{1/3}}
815: \int_{-\delta \sqrt{s} d N^{1/3}}^{\delta \sqrt{s} d N^{1/3}}
816: e^{i\frac{y'-x'}{\sqrt{s}} t - N^{\gamma- \frac{2}{3}} t^2}
817: \, d t \\
818: & = \frac{1}{2 \pi \sqrt{s} d N^{1/3}} \int_{-\infty}^{\infty}
819: e^{i\frac{y'-x'}{\sqrt{s}} t - N^{\gamma- \frac{2}{3}} t^2}
820: \, d t + \mathcal{O}\lp( e^{-N^{\gamma}} \rp) \\
821: & = \frac{1}{d N^{1/3}} \frac{1}{\sqrt{4 \pi s N^{\gamma - 2/3}}}
822: e^{-\frac{(x'-y')^2}{4 s N^{\gamma - 2/3}}} +
823: \mathcal{O}\lp( e^{-N^{\gamma}} \rp) .
824: \end{align*}
825: Below we will analyze the error. For simplicity we take $s=1$.
826:
827: Define $h(\theta)$ by
828: \begin{equation*}
829: g(\theta) = \log{(1-\alpha)^2} + \frac{\alpha}{(1-\alpha)^2}
830: \lp( \theta^2 + h(\theta) \rp).
831: \end{equation*}
832: This means that
833: \begin{equation*}
834: h(\theta) = \sum_{k=4}^{\infty} h_k \theta^k
835: \end{equation*}
836: where $h_4 < 0$. Note that $h$ is even since $g$ is and also that,
837: for $\delta$ small enough, $h(\theta) < 0$ if $|\theta| \leq \delta$.
838: The error becomes
839: \begin{equation*}
840: \textrm{Err} = \lp| \int_{-\delta}^{\delta}
841: e^{i(y'-x')d N^{1/3} \theta} F(\theta) d \theta \rp|
842: \end{equation*}
843: where
844: \begin{equation*}
845: F(\theta) = e^{- d^2 N^{\gamma} \theta^2} -
846: e^{-d^2 N^{\gamma} \theta^2 - d^2 N^{\gamma}h(\theta)} .
847: \end{equation*}
848: Next we integrate by parts.
849: \begin{multline*}
850: \textrm{Err} \leq
851: \lp| \lp[ \frac{1}{i(y'-x')d N^{1/3}}
852: e^{i(y'-x')d N^{1/3} \theta} F(\theta) \rp]_{-\delta}^{\delta} \rp| \\
853: \qquad \qquad + \frac{1}{|y'-x'|d N^{1/3}} \lp|
854: \int_{-\delta}^{\delta}
855: e^{i(y'-x')d N^{1/3} \theta} F'(\theta) d \theta \rp| \\
856: \leq \frac{3}{|y'-x'|d N^{1/3}} e^{-d^2 N^{\gamma} \delta^2} +
857: \frac{1}{|y'-x'|d N^{1/3}} \int_{-\delta}^{\delta}
858: \lp| F'(\theta) \rp| d \theta
859: \end{multline*}
860: The last integral will be easy to compute if we can find out where
861: $F'(\theta)$ changes sign.
862: \begin{equation*}
863: F'(\theta) = 2 d^2 N^{\gamma} \theta \, e^{-d^2 N^{\gamma}(\theta^2 + h(\theta))}
864: \lp( 1 + \frac{h'(\theta)}{2 \theta} - e^{d^2 N^{\gamma}h(\theta)} \rp)
865: \end{equation*}
866: A point in $[-\delta,\delta] \setminus \{0\}$ where $F'$ changes sign
867: will satisfy
868: \begin{equation*}
869: \frac{1}{d^2 N^{\gamma}} =
870: \frac{h(\theta)}{\log{\lp[ 1 + \frac{h'(\theta)}{2 \theta} \rp]}} =
871: \frac{\theta^2}{2} + \Ordo(\theta^4).
872: \end{equation*}
873: This shows that if $N$ is large then $F'$ has two zeros $\pm \theta_0$
874: in $[-\delta,\delta] \setminus \{0\}$. Moreover, $\theta_0$ is of order
875: $N^{-\gamma / 2}$.
876: Given this information we check which sign $F'$ has in different
877: intervals and get
878: \begin{align*}
879: \int_{-\delta}^{\delta} |F'(\theta)| d \theta & =
880: 2 \int_{0}^{\delta} |F'(\theta)| d \theta \\ & =
881: -\int_{0}^{\theta_0} F'(\theta) d \theta +
882: \int_{\theta_0}^{\delta} F'(\theta) d \theta \\ & =
883: F(0) - F(\theta_0) + F(\delta) - F(\theta_0) \\ & =
884: \Ordo (N^{-\gamma}).
885: \end{align*}
886: This almost finishes the proof of the second inequality in the lemma.
887: We should not forget the exponentially small error terms that appeared
888: above. They do not have the factor $|x'-y'|^{-1}$ in front of them.
889: However, a couple partial integrations can be used to take care of this
890: obstacle.
891:
892: The first inequality in the lemma follows from the following calculation.
893: \begin{multline*}
894: \int_{0}^{\delta} |F(\theta)| d \theta =
895: [ \theta = t N^{- \gamma / 2}] \\ =
896: N^{-\gamma / 2} \int_{0}^{N^{\gamma / 2} \delta}
897: e^{-d^2 t^2 - d^2 h(t N^{- \gamma / 2})}
898: \lp( 1 - e^{d^2 N^{\gamma} h(t N^{- \gamma / 2})} \rp) d t \\ \leq
899: N^{-\gamma / 2} \int_{0}^{N^{\gamma / 2} \delta} e^{- c_1 t^2}
900: \lp( 1 - e^{-c_2 N^{-\gamma} t^4} \rp) d t \\ \leq
901: N^{-\gamma / 2} \int_{1}^{N^{\gamma / 2} \delta} t e^{- c_1 t^2}
902: \lp( 1 - e^{-c_2 N^{-\gamma} t^4} \rp) d t + \Ordo(N^{-3\gamma / 2})
903: \end{multline*}
904:
905: We now use partial integration.
906: \begin{multline*}
907: \int_{1}^{N^{\gamma / 2} \delta} t e^{- c_1 t^2}
908: \lp( 1 - e^{-c_2 N^{-\gamma} t^4} \rp) d t \\ =
909: \lp[ -\frac{1}{2 c_1} e^{- c_1 t^2}
910: \lp( 1 - e^{-c_2 N^{-\gamma} t^4} \rp) \rp]_1^{N^{\gamma / 2} \delta} \\ +
911: \frac{2 c_2 N^{-\gamma}}{c_1} \int_{1}^{N^{\gamma / 2} \delta}
912: e^{- c_1 t^2} t^3 e^{-c_2 N^{-\gamma} t^4} d t \\
913: = \Ordo \lp( N^{-3 \gamma / 2} \rp)
914: \end{multline*}
915: This concludes the calculations in this section as well as in this paper.
916:
917:
918:
919:
920:
921: