math0304411/bst.tex
1: %auto-ignore
2: \chapter{Random permutation model: shape functional}
3: \label{cha:binary-search-trees}
4: As our first application of the theoretical results of
5: Chapters~\ref{cha:transfer-theorems} and~\ref{cha:sing-analys-hadam},
6: we demonstrate how to get a complete asymptotic expansion for the mean
7: of the shape functional (see Example~\ref{example:shape-functional})
8: for \( m \)-ary search trees and we determine the asymptotics of its
9: variance for binary search trees.  Our results generalize those
10: obtained in~\cite{MR97f:68021}.
11: 
12: 
13: \section{Complete asymptotic expansion of the mean}
14: \label{sec:bst_mean-shape-funct}
15: As noted in Example~\ref{example:shape-functional}, the shape
16: functional corresponds to the toll function \( t_n := \ln{
17:   \binom{n}{m-1} } \).  In order to obtain a complete asymptotic
18: expansion of its mean using the ETT (Theorem~\ref{theorem:ETT}), our
19: first task is to obtain a complete asymptotic expansion for its
20: generating function
21: \begin{equation*}
22:   T(z) := \sum_{n=m-1}^\infty t_n z^n.
23: \end{equation*}
24: 
25: 
26: \subsection*{The generating function of the toll sequence}
27: \label{sec:gener-funct-toll}
28: 
29: By definition
30: \begin{align*}
31:   T(z) &= \sum_{n=m-1}^\infty t_nz^n = \sum_{n=m-1}^\infty
32:   \left[\ln{\binom{n}{m-1}}\right] z^n\\ &= \sum_{n=m-1}^\infty \left[
33:   \sum_{l=0}^{m-2} \ln{(n-l)}\right] z^n -
34:   \left[\ln{(m-1)!}\right]\sum_{n=m-1}^\infty z^n = P(z) - Q(z),
35: \end{align*}
36: where
37: \begin{equation*}
38:   P(z) := \sum_{n=m-1}^\infty \left[ \sum_{l=0}^{m-2} \ln{(n-l)}
39:   \right] z^n
40: \end{equation*}
41: and
42: \begin{equation*}
43:   Q(z) := \left[\ln{(m-1)!}\right]
44:   \sum_{n=m-1}^\infty z^n = \left[\ln{(m-1)!}\right]\left[(1-z)^{-1} -
45:   \sum_{n=0}^{m-2} z^n\right].
46: \end{equation*}
47: But
48: \begin{align}
49:   P(z) = \sum_{l=0}^{m-2} \sum_{n=m-1}^\infty
50:   \left[\ln{(n-l)}\right]z^n
51:   &= \sum_{l=0}^{m-2} z^l \sum_{n=l+1}^\infty \left[ \ln{(n-l)}
52:   \right]z^{n-l} - 
53:   \sum_{l=0}^{m-2} \sum_{n=l+1}^{m-2} \left[ \ln{(n-l)} \right] z^{n} \notag\\
54:   \label{eq:5.1.4}
55:   &= \frac{1-z^{m-1}}{1-z} \Li_{0,1}(z) -
56:   \sum_{n=1}^{m-2}z^n \sum_{l=0}^{n-1} \ln{(n-l)},
57: \end{align}
58: where \(\Li\) denotes the generalized polylogarithm defined
59: at~\eqref{eq:4.1.50}.
60: The term subtracted in~\eqref{eq:5.1.4} equals
61: \begin{multline}
62:   \label{eq:5.1.1}
63:   \sum_{n=1}^{m-2} z^n \ln{n!}
64:   = \sum_{n=1}^{m-2} \sum_{k=0}^n (-1)^k \binom{n}{k}
65:   (1-z)^k \ln{n!}\\
66:   = \sum_{k=0}^{m-2} (-1)^k (1-z)^k \sum_{n=\max{\{k,1\}}}^{m-2}
67:   \binom{n}{k} \ln{n!} = \sum_{k=0}^{m-2} (-1)^k (1-z)^k \sum_{n=k}^{m-2}
68:   \binom{n}{k} \ln{n!},
69: \end{multline}
70: while the first term in~\eqref{eq:5.1.4} is the product of
71: \((1-z)^{-1} \Li_{0,1}(z)\) and
72: \begin{equation*}
73:   1-z^{m-1} 
74:   = 1 - \sum_{k=0}^{m-1}(-1)^k
75:   \binom{m-1}{k} (1-z)^k
76:   = \sum_{k=1}^{m-1} (-1)^{k+1} \binom{m-1}{k} (1-z)^k,
77: \end{equation*}
78: i.e., it is \(\Li_{0,1}(z)\) times
79: \begin{equation}
80:   \label{eq:5.1.2} \frac{1-z^{m-1}}{1-z} = \sum_{k=0}^{m-2} (-1)^k
81:   \binom{m-1}{k+1} (1-z)^k.
82: \end{equation}
83: Using Theorem~\ref{theorem:Li-sing} we know that
84: \begin{equation}
85:   \label{eq:5.1.3} \Li_{0,1}(z) \sim t^{-1}\ln{t^{-1}} + \Gamma'(1)t^{-1}
86:   - \sum_{j \geq 0} \frac{(-1)^j}{j!}\zeta'(-j)t^j
87: \end{equation}
88: as \(z \to 1\) in a suitable indented crown, with \(t :=
89: -\ln{z}\).  We note that \(\Gamma'(1) = -\gamma\), where \(\gamma
90: \doteq 0.5772\) is Euler's constant.
91: 
92: In order to carry out the integration required in the ETT, we need to
93: express \(T(z)\) in terms of powers of \((1-z)\) and
94: \(\ln{((1-z)^{-1})}\).  This is our first goal.  [For notational
95: convenience we will define $L(z) := \ln{((1-z)^{-1})}$.]  To this end,
96: expression \eqref{eq:5.1.1} can be written as
97: \begin{equation}
98:   \label{eq:5.1.5} \sum_{k=0}^{m-2} S_1(k) (1-z)^k
99:   \quad\text{ where }\quad
100:   S_1(k) := (-1)^k \sum_{n=k}^{m-2} \binom{n}{k}
101:   \ln{n!}.
102: \end{equation}
103: Expression \eqref{eq:5.1.2} can be written as
104: \begin{equation}
105:   \label{eq:5.1.2.3} \sum_{k=0}^{m-2} S_2(k) (1-z)^k,
106:   \quad\text{ where }\quad
107:   S_2(k) := (-1)^k \binom{m-1}{k+1}.
108: \end{equation}
109: For~\eqref{eq:5.1.3}, we will need the following two facts: First,
110: \begin{equation}
111:   \label{eq:5.1.6.5}
112: t^{j} = \sum_{r=j}^\infty T_1(r,j)(1-z)^r, \quad j \geq -1
113: \quad\text{ where }\quad
114: T_1(r,i) := 
115: \begin{cases}
116:   [z^r]L^i(z) & i \neq 0\\ \delta_{r,0} & i = 0.
117: \end{cases}
118: \end{equation}
119: Next,
120: \begin{equation}
121:   \label{eq:5.1.6.75}
122: \ln{t^{-1}} = L(z) + \sum_{r=1}^\infty T_2(r)(1-z)^r,
123: \quad\text{ where }\quad
124: T_2(r) := \sum_{i=1}^r \frac{(-1)^i}{i}
125: T_3(r+i,i),
126: \end{equation}
127: with
128: \begin{equation}
129:   \label{eq:5.1.6.875}
130: T_3(r,i) := [w^r]\left[L(w) - w \right]^i.
131: \end{equation}
132: The final ingredient in our expression for \(T(z)\) is
133: \begin{align}
134:   \label{eq:5.1.6} \sum_{n=0}^{m-2} z^n &= \sum_{n=0}^{m-2}
135:   (1-(1-z))^n
136:   = \sum_{n=0}^{m-2} \sum_{k=0}^n \binom{n}{k}
137:   (-1)^k (1-z)^k\notag\\
138:   &= \sum_{k=0}^{m-2} (1-z)^k (-1)^k\sum_{n=k}^{m-2} \binom{n}{k}
139:   = \sum_{k=0}^{m-2} S_3(k)(1-z)^k,
140: \end{align}
141: where
142: \begin{equation}
143:   \label{eq:5.1.3.5} S_3(k) := (-1)^k \sum_{n=k}^{m-2} \binom{n}{k}.
144: \end{equation}
145: 
146: We now express \(\Li_{0,1}(z)\) in terms of \((1-z)\):
147: \begin{align*}
148:   \Li_{0,1}(z) &\sim \left(\sum_{r=-1}^\infty
149:     T_1(r,-1)(1-z)^r\right)\left(L(z) + \sum_{r=1}^\infty
150:     T_2(r)(1-z)^r\right)\notag\\
151:   & \quad + \Gamma'(1) \sum_{r=-1}^\infty T_1(r,-1)(1-z)^r
152:    - \sum_{j \geq 0} \frac{(-1)^j}{j!} \zeta'(-j)
153:   \sum_{r=j}^\infty T_1(r,j)(1-z)^r\notag\\
154:   &= \sum_{r=-1}^\infty
155:   T_1(r,-1)(1-z)^r L(z) + \sum_{r=-1}^\infty
156:   T_1(r,-1)(1-z)^r \sum_{j=1}^\infty T_2(j) (1-z)^j\notag\\ & \quad
157:   + \Gamma'(1) \sum_{r=-1}^\infty T_1(r,-1)(1-z)^r
158:   + \sum_{r=0}^\infty (1-z)^r \sum_{j=0}^r \frac{(-1)^j}{j!} \zeta'(-j)
159:   T_1(r,j)\\
160:   &= \sum_{r=-1}^\infty
161:   T_1(r,-1)(1-z)^r L(z) + \sum_{r=-1}^\infty
162:   \sum_{j=1}^\infty T_1(r,-1) T_2(j)(1-z)^{r+j}\\
163:   & \quad +  \Gamma'(1) \sum_{r=-1}^\infty T_1(r,-1)(1-z)^r +
164:   \sum_{r=0}^\infty S_4(r)(1-z)^r
165: \end{align*}
166: where
167: \begin{equation}
168:   \label{eq:5.1.6.3}
169:   S_4(r) :=  \sum_{j=0}^r \frac{(-1)^j}{j!} \zeta'(-j) T_1(r,j).
170: \end{equation}
171: Hence
172: \begin{align*}
173:   \Li_{0,1}(z) &\sim \sum_{r=-1}^\infty
174:   T_1(r,-1)(1-z)^r L(z) +
175:   \sum_{r=-1}^\infty\sum_{j=r+1}^\infty T_1(r,-1)T_2(j-r)
176:   (1-z)^j\notag\\
177:   & \quad + \Gamma'(1) \sum_{r=-1}^\infty
178:   T_1(r,-1)(1-z)^r + \sum_{r=0}^\infty S_4(r)(1-z)^r\notag\\
179:   &= \sum_{r=-1}^\infty T_1(r,-1)(1-z)^r L(z) +
180:   \sum_{j=0}^\infty (1-z)^j \sum_{r=-1}^{j-1}
181:   T_1(r,-1)T_2(j-r) \\
182:   & \quad+  \Gamma'(1)\sum_{r=-1}^\infty
183:   T_1(r,-1)(1-z)^r + \sum_{r=0}^\infty S_4(r)(1-z)^r\notag\\ &=
184:   \sum_{r=-1}^\infty T_1(r,-1)(1-z)^r L(z) +
185:   \sum_{r=-1}^\infty S_5(r)(1-z)^r,
186: \end{align*}
187: where
188: \begin{equation}
189:   \label{eq:5.1.6.4}
190:   S_5(r) := 
191:   \begin{cases}
192:     \Gamma'(1)T_1(-1,1) & r=-1\\ \Gamma'(1)T_1(r,-1) + S_4(r) +
193:     \displaystyle\sum_{j=-1}^{r-1} T_1(j,-1)T_2(r-j) & r \geq 0.
194:   \end{cases}
195: \end{equation}
196: 
197: Now we multiply~\eqref{eq:5.1.2} and~\eqref{eq:5.1.3},
198: using~\eqref{eq:5.1.2.3} and~\eqref{eq:5.1.6.3}, to get
199: \begin{align*}
200:   & \left(\sum_{k=0}^{m-2} S_2(k)(1-z)^k\right)
201:   \left(\sum_{r=-1}^\infty
202:     T_1(r,-1)(1-z)^r L(z) + \sum_{r=-1}^\infty
203:     S_5(r)(1-z)^r\right)\notag\\
204:   &= \sum_{k=0}^{m-2} \sum_{r=-1}^\infty
205:   S_2(k)T_1(r,-1)(1-z)^{k+r} L(z) +
206:   \sum_{k=0}^{m-2} \sum_{r=-1}^\infty S_2(k) S_5(r)
207:   (1-z)^{k+r}\notag\\
208:   &= \sum_{k=0}^{m-2} \sum_{r=k-1}^\infty
209:   S_2(k)T_1(r-k,-1)(1-z)^r L(z) +
210:   \sum_{k=0}^{m-2}\sum_{r=k-1}^\infty S_2(k)S_5(r-k)(1-z)^r\notag\\
211:   &= \sum_{r=-1}^\infty (1-z)^r L(z) \sum_{k=0}^{\min{\{r+1,m-2\}}}
212:   S_2(k)T_1(r-k,-1)\\
213:   & \quad + \sum_{r=-1}^\infty
214:   (1-z)^r \sum_{k=0}^{\min{\{r+1,m-2\}}}S_2(k)S_5(r-k)\notag\\
215:   \label{eq:5.1.7.5}
216:   &= \sum_{r=-1}^\infty S_6(r)(1-z)^r L(z) + \sum_{r=-1}^\infty
217:   S_7(r)(1-z)^r
218: \end{align*}
219: where
220: \begin{equation}
221:   \label{eq:5.1.7.6}
222:   S_6(r) := \sum_{k=0}^{\min{\{r+1,m-2\}}} S_2(k)T_1(r-k,-1)
223:   \quad\text{ and }\quad
224:   S_7(r) := \sum_{k=0}^{\min{\{r+1,m-2\}}} S_2(k)S_5(r-k).
225: \end{equation}
226: We are now ready to plug~\eqref{eq:5.1.7.6}, \eqref{eq:5.1.5},
227: and~\eqref{eq:5.1.6} into the expression for \(T(z)\). This gives us
228: \begin{align}
229:   \label{eq:5.1.8.2}
230:   T(z) &\sim \sum_{r=-1}^\infty
231:   S_6(r)(1-z)^r L(z) + \sum_{r=-1}^\infty
232:   S_7(r)(1-z)^r \notag\\
233:   \quad &- \sum_{k=0}^{m-2} S_1(k)(1-z)^k -
234:   [\ln{(m-1)!}]\left[(1-z)^{-1} - \sum_{k=0}^{m-2}
235:   S_3(k)(1-z)^k\right]\notag\\
236:   &= \sum_{r=-1}^\infty
237:   S_6(r)(1-z)^r L(z) + \sum_{r=-1}^\infty S_8(r)(1-z)^r,
238: \end{align}
239: where
240: \begin{equation}
241:   \label{eq:5.1.8.3} S_8(r) :=
242:   \begin{cases}
243:     S_7(-1) - \ln{(m-1)!} & r=-1\\ S_7(r) - S_1(r) +
244:     [\ln{(m-1)!}]S_3(r) & 0 \leq r \leq m-2\\ S_7(r) & r \geq m-1.
245:   \end{cases}
246: \end{equation}
247: 
248: 
249: \subsection*{Application of the ETT}
250: \label{sec:application-ett}
251: We now apply the ETT (Theorem~\ref{theorem:ETT}).  For our toll
252: function,
253: \begin{equation*}
254:   \widehat{T}(z) := T(z) - \sum_{j=0}^{m-1} t_n z^j = T(z).
255: \end{equation*}
256: Thus, the generating function of the mean of the shape functional,
257: call it \(G_1\), is given by
258: \begin{equation}
259:   \label{eq:5.1.8.4}
260:   G_1(z) = T(z) + m! \sum_{j=1}^{m-1}
261:   \frac{(1-z)^{-\lambda_j}}{\psi'(\lambda_j)} \int_{\zeta=0}^z
262:   T(\zeta)(1-\zeta)^{\lambda_j-1} \, d\zeta.
263: \end{equation}
264: We rewrite~\eqref{eq:5.1.8.2} as
265: \begin{equation}
266:   \label{eq:5.1.9.4} T(z) \sim \sum_{r=-2}^\infty S_9(r)
267:   (1-z)^{r+1}L(z) + \sum_{r=-2}^\infty
268:   S_{10}(r)(1-z)^{r+1},
269: \end{equation}
270: where
271: \begin{equation}
272:   \label{eq:5.1.15.5}
273:   S_9(r) := S_6(r+1) \quad \text{and} \quad S_{10}(r) := S_8(r+1).
274: \end{equation}
275: For a fixed \(\lambda \in \mathbb{C}\), the terms
276: \((1-z)^{\lambda+r}L(z)\) and \((1-z)^{\lambda+r}\) are
277: both integrable (resp.,\ neither integrable) at \(z=1\) iff \(r \geq
278: -\lceil\Re(\lambda)\rceil\) (resp.,\ iff \(r <
279: -\lceil\Re(\lambda)\rceil\)).
280:     
281: We therefore write
282: \begin{subequations}
283:   \begin{align}
284:     &\int_{\zeta=0}^z T(\zeta)(1-\zeta)^{\lambda-1}d\zeta\notag\\
285:     \label{eq:5.1.10.1a}
286:     &= \sum_{-2 \leq r < -\lceil\Re(\lambda)\rceil} \int_{\zeta=0}^z
287:     [S_9(r)L(\zeta) +
288:     S_{10}(r)](1-\zeta)^{\lambda+r} \,d\zeta\\
289:     \label{eq:5.1.10.1b}
290:     & \quad + \int_{\zeta=0}^z \biggl[T(\zeta)(1-\zeta)^{\lambda-1}
291:      - \sum_{-2 \leq r <
292:       -\lceil\Re(\lambda)\rceil} [S_9(r) L(\zeta) +
293:     S_{10}(r)](1-\zeta)^{\lambda+r}\biggr] \,d\zeta.
294:   \end{align}
295: \end{subequations}
296: The integrand in~\eqref{eq:5.1.10.1b} is integrable at \(\zeta=1\) and is
297: asymptotically equivalent to
298: \[ \sum_{r \geq -\lceil\Re(\lambda)\rceil} [S_9(r) L(\zeta) +
299: S_{10}(r)](1-\zeta)^{\lambda+r},\] as \(\zeta \to 1\) in the sector
300: $-\pi + \epsilon < \arg{(1-\zeta)} < \pi - \epsilon$ for any $\epsilon
301: > 0$.
302: 
303: For~\eqref{eq:5.1.10.1a}, we have the following formulas, valid for any
304: \(\alpha\) with \(\Re(\alpha) > -1\):
305: \begin{align*}
306:   \int_{0}^z (1-\zeta)^\alpha \,d\zeta &=
307:   -\frac{1}{\alpha+1}[(1-z)^{\alpha+1} - 1],\\
308:   \intertext{and}
309:   \int_0^z (1-\zeta)^\alpha L(\zeta) \,d\zeta &=
310:   -\frac{1}{\alpha+1}(1-z)^{\alpha+1} L(z) -
311:   \frac{1}{(\alpha+1)^2}[(1-z)^{\alpha+1} - 1].\\
312: \end{align*}
313: In the special case that \(m\) is odd and \(\lambda=\lambda_{m-1}=-m\)
314: and \(r=m-1\), we also need the formulas
315: \begin{equation*}
316:   \int_{\zeta=0}^z (1-\zeta)^{-1} \,d\zeta = L(z)
317: \end{equation*}
318: and
319: \begin{equation*}
320: \int_{\zeta=0}^z (1-\zeta)^{-1} L(\zeta) \,d\zeta =
321:   \frac{1}{2} L^2(z).
322: \end{equation*}
323: We now conclude that~\eqref{eq:5.1.10.1a} equals, except in the special
324: case described above,
325: \begin{multline}
326:   \label{eq:5.1.11.4} \sum_{-2 \leq r < -\lceil\Re(\lambda)\rceil} \Biggl[
327:   -\frac{S_9(r)}{\lambda+r+1}(1-z)^{\lambda+r+1} L(z)
328:   - \frac{S_9(r)}{(\lambda+r+1)^2}(1-z)^{\lambda+r+1}\\
329:   + \frac{S_9(r)}{(\lambda+r+1)^2}
330:   + \frac{S_{10}(r)}{\lambda+r+1} - \frac{S_{10}(r)}{\lambda+r+1}
331:   (1-z)^{\lambda+r+1} \Biggr].
332: \end{multline}
333: In the special case (\(m\) is odd, \(\lambda=-m\), \(r=m-1\)),
334: \eqref{eq:5.1.10.1a} equals
335: \begin{multline}
336:   \label{eq:5.1.11.5}
337:   \sum_{r=-2}^{m-2} \Biggl[ -
338:   \frac{S_9(r)}{\lambda+r+1} (1-z)^{\lambda+r+1} L(z) -
339:   \frac{S_9(r)}{(\lambda+r+1)^2}(1-z)^{\lambda+r+1}
340:   + \frac{S_9(r)}{(\lambda+r+1)^2}\\
341:   + \frac{S_{10}(r)}{\lambda+r+1} -
342:   \frac{S_{10}(r)}{\lambda+r+1}(1-z)^{\lambda+r+1} \Biggr]
343:   + \frac{1}{2} S_9(m-1) L^2(z) +
344:   S_{10}(m-1) L(z).
345: \end{multline}
346:   
347: In the case \(m\) is even,
348: \begin{align}
349:   \label{eq:5.1.12.3} &
350:   \sum_{j=1}^{m-1}
351:   \frac{(1-z)^{-\lambda_j}}{\psi'(\lambda_j)} \times \eqref{eq:5.1.10.1a}
352:   \Biggl\rvert_{\lambda=\lambda_j} \notag\\ &=
353:   \sum_{j=1}^{m-1}\frac{1}{\psi'(\lambda_j)} \sum_{-2 \leq r <
354:   -\lceil\Re(\lambda_j)\rceil} \Biggl[ -
355:   \frac{S_9(r)}{\lambda_j+r+1}(1-z)^{r+1} L(z)\notag\\
356:   & \quad - \left(\frac{S_{10}(r)}{\lambda_j+r+1} +
357:   \frac{S_9(r)}{(\lambda_j+r+1)^2}\right)(1-z)^{r+1} 
358:   + \frac{S_9(r)}{(\lambda_j+r+1)^2} +
359:   \frac{S_{10}(r)}{\lambda_j+r+1}\Biggr]\notag\\ &= \sum_{r=-1}^{m-1}
360:   [S_{11}(r)(1-z)^r L(z) + S_{12}(r)(1-z)^r] + S_{13},
361: \end{align}
362: where
363: \begin{equation*}
364:   S_{11}(r) := -\sum_{j : \Re(\lambda_j) \leq -r}
365:   \frac{S_9(r-1)}{\psi'(\lambda_j)(\lambda_j+r)},
366: \end{equation*}
367: \begin{equation*}
368: S_{12}(r) := -
369:   \sum_{j: \Re(\lambda_j) \leq -r} \frac{1}{\psi'(\lambda_j)} \left(
370:     \frac{S_9(r-1)}{(\lambda_j+r)^2} + \frac{S_{10}(r-1)}{\lambda_j +
371:       r}\right),
372: \end{equation*}
373: and
374: \begin{equation}
375:   \label{eq:5.1.23.99}
376: S_{13} := \sum_{r=-1}^{m-1} \sum_{\Re(\lambda_j) \leq -r}
377:   \frac{1}{\psi'(\lambda_j)} \left[\frac{S_9(r-1)}{(\lambda_j+r)^2} +
378:     \frac{S_{10}(r-1)}{\lambda_j+r+1}\right].
379: \end{equation}
380: 
381: If \(m\) is odd, we need to add
382: \[ \frac{(1-z)^m}{\psi'(-m)} \left(\frac{1}{2}
383:   S_9(m-1) L^2(z) + S_{10}(m-1) L(z) \right)\] to the right-hand-side
384: of~\eqref{eq:5.1.12.3} to get the left-hand-size
385: of~\eqref{eq:5.1.12.3}.
386: 
387: For~\eqref{eq:5.1.10.1b}, we define
388: \begin{equation}
389:   \label{eq:5.1.19.5}
390: J_\lambda(\zeta) := T(\zeta)(1-\zeta)^{\lambda-1} - \sum_{-2 \leq r
391:   < -\lceil\Re(\lambda)\rceil}
392: [S_9(r) L(z) + S_{10}(r)](1-\zeta)^{\lambda+r}
393: \end{equation}
394: so that~\eqref{eq:5.1.10.1b} equals $\bar{K}(\lambda) -
395: \displaystyle\int_{\zeta=z}^1 J_\lambda(\zeta) \,d\zeta$, where
396: \begin{equation}
397:   \label{eq:5.1.70}
398:   \bar{K}(\lambda) := \int_{\zeta=0}^1 J_\lambda(\zeta) \,d\zeta  \in
399:   \mathbb{C}.
400: \end{equation}
401: Now we make use of
402: \begin{equation}
403:   \label{eq:5.1.13.5} \int_{\zeta=z}^1 J_\lambda(\zeta) \,d\zeta \sim
404:   \sum_{r \geq -\lceil\Re(\lambda)\rceil} \int_{\zeta=z}^1
405:   \Bigr[S_9(r) L(\zeta) +
406:   S_{10}(r)\Bigr](1-\zeta)^{\lambda+r} \,d\zeta.
407: \end{equation}
408: Again, exact integration is possible: for all \(\alpha \in
409: \mathbb{C}\) with \(\Re(\alpha) > -1\),
410: \begin{equation*}
411:   \int_{\zeta=z}^1 (1-\zeta)^\alpha \,d\zeta =
412:   \frac{1}{\alpha+1}(1-z)^{\alpha+1},
413: \end{equation*}
414: \begin{equation*}
415:   \int_{\zeta=z}^1
416:   (1-\zeta)^\alpha L(\zeta) \,d\zeta =
417:   \frac{1}{\alpha+1}(1-z)^{\alpha+1} L(z) +
418:   \frac{1}{(\alpha+1)^2}(1-z)^{\alpha+1}.
419: \end{equation*}
420: Therefore
421: \begin{multline*}
422:   \int_{\zeta=z}^1 J_\lambda(\zeta) \,d\zeta \sim \sum_{r \geq
423:     -\lceil\Re(\lambda)\rceil}\Bigl[\frac{S_9(r)}{\lambda+r+1}
424:   (1-z)^{\lambda+r+1} L(z) \\
425:   + \frac{S_9(r)}{(\lambda+r+1)^2}
426:   (1-z)^{\lambda+r+1}
427:   + \frac{S_{10}(r)}{\lambda+r+1} (1-z)^{\lambda+r+1} \Bigr].
428: \end{multline*}
429: 
430: We can thus conclude
431: \begin{align}
432:   \label{eq:5.1.14.3} &
433:   \sum_{j=1}^{m-1}
434:   \frac{(1-z)^{-\lambda_j}}{\psi'(\lambda_j)} \times
435:   \eqref{eq:5.1.10.1b}\Bigl\rvert_{\lambda=\lambda_j}\notag\\ & \sim
436:   \sum_{j=1}^{m-1} \frac{\bar{K}(\lambda_j)}{\psi'(\lambda_j)}
437:   (1-z)^{-\lambda_j}
438:    - \sum_{j=1}^{m-1}
439:   \frac{1}{\psi'(\lambda_j)} \sum_{r \geq -\lceil\Re(\lambda)\rceil}
440:   \Bigl[ \frac{S_9(r)}{\lambda_j+r+1} (1-z)^{r+1} L(z)\notag\\
441:   &\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
442:   + \Bigl( \frac{S_9(r)}{(\lambda_j+r+1)^2} +
443:   \frac{S_{10}(r)}{\lambda_j+r+1}\Bigr)(1-z)^{r+1} \Bigr]\notag\\ &=
444:   \sum_{j=1}^{m-1} \frac{\bar{K}(\lambda_j)}{\psi'(\lambda_j)}
445:   (1-z)^{-\lambda_j} + \sum_{r
446:     \geq -1} [S_{14}(r)(1-z)^r L(z) + S_{15}(r)(1-z)^r]
447: \end{align}
448: where
449: \begin{align*}
450:   S_{14}(r) &:= - \sum_{j: \Re(\lambda_j) > -r}
451:   \frac{S_9(r-1)}{(\lambda_j+r)\psi'(\lambda_j)}\\ S_{15}(r) &:= -
452:   \sum_{j:\Re(\lambda_j) > -r}
453:   \frac{1}{\psi'(\lambda_j)}\Bigl[\frac{S_9(r-1)}{(\lambda_j+r)^2} +
454:   \frac{S_{10}(r-1)}{\lambda_j+r} \Bigr].
455: \end{align*}
456: 
457: 
458: \subsection*{Case $m$ even}
459: \label{sec:case-m-even}
460: 
461: Using~\eqref{eq:5.1.12.3}, \eqref{eq:5.1.14.3},
462: and~\eqref{eq:5.1.8.2} in~\eqref{eq:5.1.8.4}, we get
463: \begin{equation}
464:   \label{eq:5.1.15.1} G_1(z) = \sum_{j=1}^{m-1}
465:   \frac{K(\lambda_j)}{\psi'(\lambda_j)} (1-z)^{-\lambda_j} + M +
466:   \sum_{r \geq -1}[ K_r(1-z)^r L(z) + L_r(1-z)^r],
467: \end{equation}
468: where %\marginal{Need to run tests to verify the value of \(M\).}
469: \begin{equation}
470:   \label{eq:5.1.15.2} K(\lambda) := m! \bar{K}(\lambda)
471:   \quad\text{ and }\quad
472:   M := m!S_{13},
473: \end{equation}
474: \begin{align}
475:   \label{eq:5.1.15.4}
476:   K_r &:= m![S_{11}(r) + S_{14}(r)] + S_6(r)
477:   = -m!S_9(r-1)\sum_{j=1}^{m-1} \frac{1}{(\lambda_j+r)\psi'(\lambda_j)} +
478:   S_6(r),\notag\\
479:   &= \frac{m!S_9(r-1)}{\rising{(-r)}{m-1} -m!} + S_6(r),
480: \end{align}
481: and
482: \begin{align}
483:   \label{eq:5.1.26}
484:   L_r &:= m![S_{12}(r) + S_{15}(r)] + S_8(r)\notag\\
485:   &= -m!\Biggl[S_9(r-1)\sum_{j=1}^{m-1}
486:   \frac{1}{(\lambda_j+r)^2\psi'(\lambda_j)} 
487:   + S_{10}(r-1) \sum_{j=1}^{m-1}
488:   \frac{1}{(\lambda_j+r)\psi'(\lambda_j)}\Biggr]
489:   + S_8(r)\notag\\
490:   &= -m!\Biggl[S_9(r-1)\sum_{j=1}^{m-1}
491:   \frac{1}{(\lambda_j+r)^2\psi'(\lambda_j)} -
492:   \frac{S_{10}(r-1)}{\rising{(-r)}{m-1} - m!}\Biggr] + S_8(r).
493: \end{align}
494: These constants can be obtained in a closed form using
495: Identities~\ref{identity:B.1} and~\ref{identity:B.2}.
496: %% \begin{equation}
497: %%   \label{eq:5.1.23.5}
498: %%   -\sum_{j=1}^{m-1}\frac{1}{(\lambda_j+r)^2\psi'(\lambda_j)} =
499: %%   \begin{cases}
500: %%     - \frac{(m-1)!H_{m-1}}{[(m-1)!]^2(m-1)^2} =
501: %%     \frac{-H_{m-1}}{(m-1)!(m-1)^2} & r = -1\\
502: %%     \frac{(-1)^{r+1}r!(m-2-r)!}{[\langle-r\rangle_{m-1}-m!]^2} & 0
503: %%     \leq r \leq m-2\\ \frac{(-1)^{m+1}(r)_{m-1}(H_r -
504: %%     H_{r-(m-1)})}{[(-1)^{m-1}(r)_{m-1}
505: %%       - m!]^2} & r \geq m-1.
506: %%   \end{cases}
507: %% \end{equation}
508: 
509: 
510: \subsection*{Case $m$ odd}
511: \label{case-m-odd}
512: 
513: In this case we have the same result as~\eqref{eq:5.1.15.1} except
514: that the \(r=m\) term must be replaced by
515: \begin{multline*}
516:   m! \biggl\{ \frac{(1-z)^m}{\psi'(-m)} \left[\frac{1}{2}
517:   S_9(m-1) L^2(z) + S_{10}(m-1) L(z)\right]\\
518:   + S_{14}(m)(1-z)^m L(z) + S_{15}(m)(1-z)^m\biggr\}
519:   + S_6(m)(1-z)^m L(z) + S_8(m)(1-z)^m\\
520:   = N_m(1-z)^m L^2(z) + K_m(1-z)^m L(z) +
521:   L_m(1-z)^m,
522: \end{multline*}
523: where
524: \begin{equation*}
525:   N_m  := \frac{m!S_9(m-1)}{2\psi'(-m)} = -
526:   \frac{S_9(m-1)}{2(H_m-1)},\quad K_m  := m! \Bigl[
527:   \frac{S_{10}(m)}{\psi'(-m)} + S_{14}(m) \Bigr] + S_6(m),
528: \end{equation*}
529: and
530: \begin{equation*}
531: L_m  :=  m!S_{15}(m) + S_8(m).
532: \end{equation*}
533: (We note that $K_m$ and $L_m$ defined here for $m$ odd are different
534: [compare~\eqref{eq:5.1.15.4} and~\eqref{eq:5.1.26}] from those defined
535: for $m$ even.)
536: The constants can be simplified using Identities~\ref{identity:B.3}
537: and~\ref{identity:B.4}.  Indeed,
538: \begin{align*}
539:   S_{14}(m) &= -\sum_{j:\Re(\lambda_j)>-m}
540:   \frac{S_9(m-1)}{(\lambda_j+m)\psi'(\lambda_j)} = -S_9(m-1)
541:   \sum_{j\neq m-1} \frac{1}{(\lambda_j+m)\psi'(\lambda_j)}\\
542:   & =  -\frac{S_9(m-1)}{2m!} \Bigl[1 - \frac{H_m^{(2)} -
543:     1}{(H_m-1)^2}\Bigr],
544: \end{align*}
545: and
546: \begin{align*}
547:   S_{15}(m) &= -S_9(m-1) \sum_{j \neq m-1}
548:   \frac{1}{(\lambda_j+m)^2\psi'(\lambda_j)}
549:   -  S_{10}(m-1)\sum_{j \neq m-1} \frac{1}{(\lambda_j+m)
550:     \psi'(\lambda_j)}\\
551:   &= -\frac{S_9(m-1)}{m!}\left[ \frac{H_m-1}{12} -
552:     \frac{1}{3}\frac{H_m^{(3)}-1}{(H_m-1)^2} + \frac{1}{4}
553:     \frac{(H_m^{(2)} - 1)^2}{(H_m-1)^3}\right]\\
554:   &\qquad
555:    - \frac{S_{10}(m-1)}{2 \cdot m!}\left[1 -
556:     \frac{H_m^{(2)} - 1}{(H_m-1)^2}\right].
557: \end{align*}
558: 
559: 
560: 
561: \subsection*{Some observations}
562: The following facts can be established for the constants.
563: \begin{remark}
564:   \label{remark:5.1.1}
565:   \(K_r = 0\) for \(0 \leq r \leq m-2\).  
566: \end{remark}
567: \begin{proof}
568:   \begin{equation*}
569:     K_r = \frac{m!S_9(r-1)}{\rising{(-r)}{m-1} - m!} + S_6(r)
570:     = \frac{m! S_6(r)}{\rising{(-r)}{m-1} - m!} + S_6(r) =
571:     S_6(r) \frac{\rising{(-r)}{m-1}}{\rising{(-r)}{m-1} -
572:       m!},
573:   \end{equation*} which contains the factor 0 if \(0 \leq r \leq m-2.\)
574: \end{proof}
575: \begin{remark}
576:   \label{remark:5.1.2}
577:   \(K_{-1} = -1\).
578: \end{remark}
579: \begin{proof}
580:   \begin{align*}
581:     K_{-1}                          & = S_6(-1)\frac{(m-1)!}{(m-1)!-m!}
582:     = \frac{S_6(-1)}{1-m} = \frac{(m-1)T_1(-1,-1)}{1-m}\\
583:                                     & = -T_1(-1,-1) = -[z^{-1}]L(z) = -1.
584:   \end{align*}
585: \end{proof}
586: \begin{remark}
587:   \label{remark:5.1.3}
588:   \(K_{m-1} = 0\) for all odd values of \(m\).
589: \end{remark}
590: \begin{proof}
591:   When \(m\) is odd,
592:   \begin{equation*}
593:     K_{m-1} = \frac{S_6(m-1) (-1)^{m-1} (m-1)!}{(-1)^{m-1} (m-1)! -
594:       m!} = -\frac{S_6(m-1)}{m-1}
595:   \end{equation*}
596:   Now \(S_6(m-1)\) is the coefficient of
597:   \((1-z)^{m-1}L(z) \) in the product \(\eqref{eq:5.1.2} \times
598:   \eqref{eq:5.1.3}\), i.e.,
599:   \begin{align*}
600:     S_6(m-1)                         & = [(1-z)^{m-1}L(z)]\left(
601:       \eqref{eq:5.1.2} \times \eqref{eq:5.1.3}\right) \\
602:     &= [(1-z)^{m-1} L(z)]
603:     \left(
604:       \frac{1-z^{m-1}}{1-z} (-\ln z)^{-1} L(z)
605:     \right)\\
606:        & = [(1-z)^{m-1}] \left( \frac{1-z^{m-1}}{1-z}
607:        (-\ln{z})^{-1}\right)       \\ 
608:      & = [z^{m-1}] \left( \frac{1 - (1-z)^{m-1}}{z L(z)}
609:          \right)
610:      = [z^m] \frac{1 - (1-z)^{m-1}}{L(z)} = 0.
611:   \end{align*}
612:   To see the last equality, observe that
613:   \[ \frac{1 - (1-z)^{m-1}}{L(z)} = \int_{x=0}^{m-1} (1-z)^x \,dx \]
614:   and
615:   \begin{equation*}
616:     [z^m] \int_{x=0}^{m-1} (1-z)^x \,dx = [z^m]\int_{x=0}^{m-1}
617:     \sum_{n=0}^\infty \frac{\rising{(-x)}{n}}{n!} z^n \,dx =
618:     \frac{1}{m!}\int_{x=0}^{m-1} \rising{(-x)}{m} \,dx.
619:   \end{equation*}
620:   But the last integral vanishes because its
621:   integrand is antisymmetric about the midpoint
622:   of the interval of integration:
623:   \[ \rising{-(m-1-x)}{m} = (-1)^m(m-1-x)_m =
624:   (-1)^m\rising{-x}{m} = -\rising{(-x)}{m},\]
625:   the last equality following from the assumption that \(m\) is odd.
626: \end{proof}
627: \begin{remark}
628:   \label{remark:5.1.5}
629:   \[
630:   L_{-1} = \frac{\ln{(m-1)!}}{m-1} - \frac{m}{m-1}H_{m-1} +
631:   \gamma.
632:   \]
633: \end{remark}
634: \begin{proof}
635:   \begin{align*}
636:     L_{-1} &= m!\Bigl[ S_9(-2)\Bigl(-
637:     \frac{H_{m-1}}{(m-1)!(m-1)^2}\Bigr) + S_{10}(-2) \Bigl(
638:     \frac{1}{(m-1)! - m!}\Bigr)\Bigr] + S_8(-1)\\
639:     &=  m!\Bigl[-S_6(-1) \frac{H_{m-1}}{(m-1)!(m-1)^2} - S_8(-1) \Bigl(
640:     \frac{1}{(m-1)!(m-1)}\Bigr)\Bigr] + S_8(-1).
641:   \end{align*}
642:   Now \(S_6(-1) = m-1\) as noted in the proof of Remark~\ref{remark:5.1.2},
643:   and \(S_8(-1) = -(m-1)\gamma - \ln{(m-1)!}\). Plugging
644:   these in the display above, we get the desired result.
645: \end{proof}
646: 
647: \subsection*{The complete asymptotic expansion}
648: \label{sec:compl-asympt-expans}
649: 
650: The computations in this section have thus established that the
651: generating function of the mean of the shape functional for \( m
652: \)-ary search trees under the random permutation model, denoted by \(
653: G_1(z)\), has the complete asymptotic expansion
654: \begin{multline}
655:   \label{eq:5.1.100}
656:   G_1(z) \sim \sum_{j=1}^{m-1} \frac{K(\lambda_j)}{\psi'(\lambda_j)}
657:   (1-z)^{-\lambda_j} + M - (1-z)^{-1} L(z) +
658:   L_{-1}(1-z)^{-1}\\
659:   + \sum_{r=0}^{m-2} L_r(1-z)^r
660:   + \sum_{r \geq m-1} [ K_r(1-z)^r L(z) + L_r (1-z)^{r} ]
661: \end{multline}
662: when \( m \) is even and
663: \begin{multline*}
664:   G_1(z) \sim \sum_{j=1}^{m-1} \frac{K(\lambda_j)}{\psi'(\lambda_j)}
665:   (1-z)^{-\lambda_j} + M - (1-z)^{-1} L(z) +
666:   L_{-1}(1-z)^{-1}
667:   + \sum_{r=0}^{m-1} L_r(1-z)^r \\ + N_m(1-z)^{m} L^2(z)  +
668:   K_m(1-z)^m L(z) 
669:   + L_{m} (1-z)^{m} + \sum_{r \geq m+1} [ K_r(1-z)^r L(z)
670:   + L_r (1-z)^{r} ]
671: \end{multline*}
672: when \( m \) is odd.
673: 
674: Using singularity analysis (Theorem~\ref{theorem:transfer}), the
675: complete asymptotic expansion can be translated on a term-by-term
676: basis to a complete asymptotic expansion of the mean itself, denoted
677: by \( \mu_n \).  Thus
678: \begin{equation}
679:   \label{eq:5.1.60}
680:   \mu_n \sim \sum_{j=1}^{m-1} \frac{K(\lambda_j)}{\psi'(\lambda_j)}
681:   \frac{\rising{\lambda_j}{n}}{n!} - H_n + L_{-1} + \sum_{r \geq m-1}
682:   K_r \Delta^r\left((\tfrac{1}{n})\right)
683: \end{equation}
684: when \( m \) is even and
685: \begin{equation*}
686:   \mu_n \sim \sum_{j=1}^{m-1} \frac{K(\lambda_j)}{\psi'(\lambda_j)}
687:   \frac{\rising{\lambda_j}{n}}{n!} - H_n + L_{-1} + N_m \Delta^m
688:   \left( \left( \tfrac{2H_{n-1}}{n} \right) \right) + K_m
689:   \Delta^m\left((\tfrac{1}{n})\right) + \sum_{r \geq m+1}
690:   K_r \Delta^r\left((\tfrac{1}{n})\right)
691: \end{equation*}
692: when \( m \) is odd, where \( \Delta^r(s) \) denotes the \( r \)th-order
693: difference of a sequence $s = (s_n)$.  Note that $\Delta^r((1/n)) =
694: \Theta(n^{-(r+1)})$ and $\Delta^m((2H_{n - 1}/n)) =
695: \Theta(n^{-(m+1)}\log n)$.
696: 
697: 
698: \subsection*{The binary search tree}
699: \label{sec:binary-search-tree}
700: 
701: In preparation for the computation of the variance of the shape
702: functional for the binary search trees we state a special case of the
703: complete asymptotic expansion derived above.  When \( m=2 \), the toll
704: function $t_n$ equals $\ln{n}$ and the
705: indicial polynomial $\psi(\lambda)$ equals $\lambda - 2$, with root
706: $\lambda_1 = 2$.  Thus~\eqref{eq:5.1.60},
707: Remark~\ref{remark:5.1.5}, and the well-known expansion for the
708: harmonic numbers~\cite[1.2.7-(3)]{knuth97} immediately yield
709: \begin{equation}
710:   \label{eq:5.1.61}
711:   \mu_n = K(2) \frac{\rising{2}{n}}{n!} - H_n - 2 + \gamma + O(n^{-2})
712:   = K(2)(n+1) - \ln{n} - 2 - \frac{1}{2n} + O(n^{-2}),
713: \end{equation}
714: where, by~\eqref{eq:5.1.15.2},~\eqref{eq:5.1.70},
715: and~\eqref{eq:5.1.19.5},
716: \begin{equation*}
717:   K(2) = 2 \bar{K}(2) = 2\int_{\zeta=0}^1 T(\zeta) (1-\zeta)\,d\zeta =
718:   2 \int_{\zeta=0}^1 \sum_{k=1}^\infty (\ln{k}) \zeta^k
719:   (1-\zeta)\,d\zeta = 2 \sum_{k=1}^\infty \frac{\ln{k}}{(k+1)(k+2)}.
720: \end{equation*}
721: 
722: 
723: \section{The variance for binary search trees}
724: \label{sec:bst-vari-shape-funct}
725: 
726: Let \( X_n \) denote the mean of the shape functional for binary
727: search trees under the random permutation model.  In the sequel it
728: will be convenient to work with the approximately centered random
729: variable \( \widetilde{X}_n \)
730: defined as
731: \begin{equation*}
732:   \widetilde{X}_n := X_n - C_1(n+1) \quad \text{for \( n \geq 0 \)},
733: \end{equation*}
734: where $C_1 := K(2)$, and to define
735: \begin{equation*}
736:   \tilde{\mu}_n(k) := \E{\widetilde{X}_n^k} \quad \text{for \( k
737:   \geq 1 \)},
738: \end{equation*}
739: with \( \tilde{\mu}_n := 0 \).
740: Then, following the arguments leading to~\eqref{eq:2.3}
741: and~\eqref{eq:2.4}, we have
742: \[
743: \tilde{\mu}_n(2) = \frac{2}{n} \sum_{j=0}^{n-1} \tilde{\mu}_j(2) +
744: r_n(2),
745: \]
746: where
747: \begin{align*}
748:  r_n(2)
749:   &:=  \sum_{\substack{k_1,k_2,k_3 \geq 0\\k_1+k_2+k_3=2\\k_1 < 2,\ k_2 <
750:       2}} \binom{2}{k_1,k_2,k_3}(\ln{n})^{k_3} \frac{1}{n}
751:   \sum_{\substack{j_1+j_2=n-1\\ j_1,j_2 \geq 0}}
752:     \tilde{\mu}_{j_1}(k_1)\tilde{\mu}_{j_2}(k_2)\\
753: &= (\ln{n})^2 + 4(\ln{n})\frac{1}{n}\sum_{j=0}^{n-1}
754: \tilde{\mu}_j(1)
755: + \frac{2}{n}\sum_{j=0}^{n-1}\tilde{\mu}_j(1)\tilde{\mu}_{n-1-j}(1)\\
756: &= (\ln{n})^2 + 4(\ln{n})\frac{1}{n} \sum_{j=0}^{n-1}(\mu_j-C_1(j+1))\\
757:  &\qquad+ \frac{2}{n}
758:   \sum_{j=0}^{n-1}(\mu_j-C_1(j+1))(\mu_{n-1-j}-C_1(n-j)),
759: \end{align*}
760: and $r_0(2) = \E{\widetilde{X}_0^2} = C_1^2.$  Note that knowing \(
761: \tilde{\mu}_n(1) \) and \( \tilde{\mu}_n(2) \) is enough to compute
762: the variance of the random variable \( X_n \), call it \( \sigma_n^2
763: \), since
764: \begin{equation}
765:   \label{eq:5.2.70}
766:   \sigma_n^2 = \Var{\widetilde{X}_n} = \tilde{\mu}_n(2) - \tilde{\mu}_n(1)^2.
767: \end{equation}
768: 
769: Our proposal is to use the ETT~(Theorem~\ref{theorem:ETT}) again.  To
770: that end, let $B(z)$ be the generating function of $r_n(2)$.  Then
771: using the expression for \( r_n(2) \) above we get
772: \begin{multline}
773:   \label{eq:5.2.V3.1}
774:   B(z) = C_1^2 + \Li_{0,2}(z)+ 4\Li_{0,1}(z)\odot \Bigl[ \int_0^z
775:   (1-\zeta)^{-1}\bigl(G_1(\zeta) - C_1(1-\zeta)^{-2}\bigr)\,d\zeta\Bigr]\\
776:   + 2\int_0^z [G_1(\zeta) - C_1(1-\zeta)^{-2}]^2\,d\zeta.
777: \end{multline}
778: Our goal is to get each term in this sum to $O(|1-z|^{-\epsilon})$ for
779: arbitrarily small $\epsilon > 0$.
780: 
781: Using Theorem~\ref{theorem:Li-sing} we get
782: \[
783: \Li_{\alpha,0}(z) \sim \Gamma(1-\alpha)t^{\alpha-1} + \sum_{j \geq 0}
784: \frac{(-1)^j}{j!}\zeta(\alpha-j)t^j,
785: \]
786: \[
787: \Li_{\alpha,1}(z) \sim \Gamma'(1-\alpha)t^{\alpha-1} -
788: \Gamma(1-\alpha)t^{\alpha-1}\ln{t} - \sum_{j \geq 0}
789: \frac{(-1)^j}{j!} \zeta'(\alpha-j)t^j,
790: \]
791: and
792: \begin{multline*}
793: \Li_{\alpha,2}(z) \sim \Gamma''(1-\alpha)t^{\alpha-1} -
794: 2\Gamma'(1-\alpha)t^{\alpha-1}\ln{t} \\
795: +
796: \Gamma(1-\alpha)t^{\alpha-1}(\ln{t})^2 + \sum_{j\geq 0}
797: \frac{(-1)^j}{j!} \zeta''(\alpha-j)t^j,
798: \end{multline*}
799: where
800: \begin{equation*}
801:   t = -\ln{z} = (1-z) + \frac{(1-z)^2}{2} + O(|1-z|^3).
802: \end{equation*}
803: Hence
804: \[
805: t^{-1} = (1-z)^{-1} - \frac12 + O(|1-z|)
806: \]
807: and [recalling $L(z) := \ln((1-z)^{-1})$]
808: \[
809: \ln{t} = -L(z) + \frac12 (1-z) + O(|1-z|^2).
810: \]
811: This gives
812: \begin{equation*}
813:   (\ln{t})^2  = L^2(z) - (1-z)L(z) + O(|(1-z)^2L(z)|),
814: \end{equation*}
815: so that
816: \[
817: t^{-1}(\ln{t})^2 = (1-z)^{-1}L^2(z) - \frac{1}{2}L^2(z) - L(z) +
818:   O(|(1-z)L(z)|)
819: \]
820: and
821: \[
822: t^{-1}\ln{t} = -(1-z)^{-1}L(z) + \frac{1}{2}L(z) + O(1).
823: \]
824: Combining these expansions we get
825: \begin{equation}
826:   \label{eq:5.2.V5.4}
827: \Li_{0,1}(z) = (1-z)^{-1}L(z) + \Gamma'(1)(1-z)^{-1} - \frac{1}{2}L(z)
828: + O(1)
829: \end{equation}
830: and
831: \begin{align}
832:   \label{eq:5.2.V5.7}
833:   \Li_{0,2}(z)
834:   &= (1-z)^{-1}L^2(z) - \frac{1}{2}L^2(z) - L(z)\notag\\
835:   & \quad - 2\Gamma'(1)\bigl[-(1-z)^{-1}L(z)
836:   + \frac12L(z)\bigr]
837:   + \Gamma''(1)(1-z)^{-1} + O(1)\notag\\
838:   &= (1-z)^{-1}L^2(z) + 2\Gamma'(1)(1-z)^{-1}L(z)\notag\\
839:   & \quad +  \Gamma''(1)(1-z)^{-1}
840:   - \frac{1}{2}L^2(z) - (\Gamma'(1)+1)L(z) + O(1).
841: \end{align}
842: 
843: We also need an asymptotic expansion for the generating function \(
844: G_1(z) - C_1(1-z)^{-2}\) of $\tilde{\mu}_n(1)$.  This is available to
845: us using~\eqref{eq:5.1.100}.  Indeed, using $M=0$
846: from~\eqref{eq:5.1.15.2}, $L_0 = 1/4$ from~\eqref{eq:5.1.26},
847: $L_{-1}=-(2-\gamma)$ from Remark~\ref{remark:5.1.5}, and $K_1 = -1/36$
848: from~\eqref{eq:5.1.15.4} [and the
849: definitions~\eqref{eq:5.1.15.5},~\eqref{eq:5.1.7.6},~\eqref{eq:5.1.2.3},~\eqref{eq:5.1.6.5},~\eqref{eq:5.1.23.99}],
850: we get
851: \begin{equation*}
852: G_1(z) - C_1(1-z)^{-2} = -(1-z)^{-1}L(z) +
853: L_{-1}(1-z)^{-1}
854: + \frac{1}{4} - \frac{1}{36}(1-z)L(z) +O(|1-z|).
855: \end{equation*}
856: Thus, invoking Theorem~\ref{theorem:2B}, we get
857: \begin{align*}
858:   &\int_0^z (1-\zeta)^{-1}[G_1(\zeta) - C_1(1-\zeta)^{-2}]\,d\zeta\notag\\
859:   &= \int_0^z [-(1-\zeta)^{-2}L(\zeta) + L_{-1}(1-\zeta)^{-2}
860:   + \frac{1}{4}(1-\zeta)^{-1}
861:   - \frac{1}{36}L(\zeta) + O(1)]\,d\zeta \notag\\
862:   &= -[(1-z)^{-1}L(z) - (1-z)^{-1}] + L_{-1}(1-z)^{-1}
863:   + \frac{1}{4}L(z) + O(1) \notag\\
864: &= -(1-z)^{-1}L(z) + (L_{-1}+1)(1-z)^{-1} + \frac14L(z) + O(1),
865: \end{align*}
866: and
867: \begin{multline}
868:   \label{eq:5.2.2}
869:   [G_1(z) - C_1(1-z)^{-2}]^2
870:   = (1-z)^{-2}L^2(z) - 2L_{-1}(1-z)^{-2}L(z) + L_{-1}^2(1-z)^{-2}\\
871:   \qquad -\frac12(1-z)^{-1}L(z) + \frac{1}{2}L_{-1}(1-z)^{-1} + O(|L^2(z)|).
872: \end{multline}
873: We now integrate~\eqref{eq:5.2.2} term by term.  Integration by parts
874: establishes
875: \begin{multline*}
876: \int_0^z (1-\zeta)^rL^2(\zeta)\,d\zeta =\\
877: \begin{cases}
878:   -\displaystyle\frac{1}{r+1}(1-z)^{r+1}L^2(z) -
879:    \displaystyle\frac{2}{(r+1)^2} (1-z)^{r+1}L(z) - 
880:    \displaystyle\frac{2}{(r+1)^3}[(1-z)^{r+1} - 1] & (r \ne -1)\\
881:    \frac{1}{3}L^3(z) & (r = -1)
882: \end{cases}
883: \end{multline*}
884: and
885: \[
886: \int_0^z (1-\zeta)^rL(\zeta)\,d\zeta =
887: \begin{cases}
888:   -\displaystyle\frac{1}{r+1}(1-z)^{r+1}L(z) -
889:    \displaystyle\frac{1}{(r+1)^2}(1-z)^{r+1} + 
890:    \displaystyle\frac{1}{(r+1)^2} & (r \ne -1)\\
891:    \displaystyle\frac{1}{2} L^2(z) & (r = -1).
892: \end{cases}
893: \]
894: By Theorem~\ref{theorem:2B} the integral
895: of the remainder in~\eqref{eq:5.2.2} is $O(1) = O(|1-z|^{-\epsilon})$. Hence
896: \begin{align}
897:   &\int_0^z [G_1(\zeta) - C_1(1-\zeta)^{-2}]^2\,d\zeta \notag\\
898:   &= (1-z)^{-1}L^2(z) - 2(1-z)^{-1}L(z) + 2(1-z)^{-1}\notag\\
899:    & \quad - 2L_{-1}[(1-z)^{-1}L(z) - (1-z)^{-1}]
900:   + L_{-1}^2(1-z)^{-1} + O(|1-z|^{-\epsilon})\notag\\
901:   &= (1-z)^{-1}L^2(z) - 2(L_{-1}+1)(1-z)^{-1}L(z)
902:   \label{eq:5.2.V9.1}
903:   + (2 + 2L_{-1} + L_{-1}^2)(1-z)^{-1} + O(|1-z|^{-\epsilon}).
904: \end{align}
905: 
906: We now have an asymptotic expansion for each term in the
907: sum~\eqref{eq:5.2.V3.1} but the one involving the Hadamard product.
908: We make use of the following facts.
909: \begin{enumerate}
910: \item $(1-z)^{-1}$ is the identity element for Hadamard products,
911:   i.e., \( (1-z)^{-1} \odot f(z) = f(z) \).
912: \item $L(z) = O(|1-z|^{-\epsilon})$ for any $\epsilon > 0$, so that
913:   $(1-z)^{-1}L(z) = O(|1-z|^{-1-\epsilon})$. Hence, by
914:   Theorem~\ref{thm:hadamard}, $[(1-z)^{-1}L(z)] \odot O(1) =
915:   O(|1-z|^{-\epsilon})$.
916: \item Similarly,
917:   \begin{equation*}
918:     [(1-z)^{-1}L(z)] \odot L(z) = O(|1-z|^{-\epsilon}),\qquad
919:     L(z) \odot L(z) = O(1),
920:   \end{equation*}
921:   \begin{equation*}
922:     L(z) \odot O(1) = O(1), \qquad
923:     O(1) \odot O(1) = O(1).
924:   \end{equation*}
925: \item The following lemma will also be used.
926:   \begin{lemma}
927:     \label{lem:V1}
928:     \(
929:     [(1-z)^{-1}L(z)] \odot [(1-z)^{-1}L(z)] = (1-z)^{-1}L^2(z) +
930:     (1-z)^{-1}\Li_{2,0}(z).
931:     \)
932:   \end{lemma}
933: \begin{proof}
934:   Let $H_n$ denote the $n$th harmonic number.  Then
935:   \begin{align*}
936:     &[(1-z)^{-1}L(z)] \odot [(1-z)^{-1}L(z)]
937:     = \sum_{n \geq 1} H_n^2 z^n
938:     = \sum_{n \geq 1} \left(\sum_{k=1}^n \frac{1}k \right)^2z^n\\
939:     &= \sum_{n \geq 1} \left(\sum_{k=1}^n \frac1{k^2}\right)z^n + 2
940:     \sum_{n \geq 1} \Biggl[ \sum_{k=2}^n \frac1k \sum_{j=1}^{k-1}
941:     \frac1j\Biggr] z^n
942:     = \sum_{k \geq 1} \frac{1}{k^2} \Biggl(\sum_{n \geq k} z^n\Biggr) +
943:     2\sum_{n \geq 1} \Biggl[ \sum_{k=2}^n \frac{H_{k-1}}{k}\Biggr] z^n\\
944:     &= \sum_{k \geq 1} \frac{z^k}{k^2} (1-z)^{-1} + 2\sum_{n \geq 1}
945:     \Biggl[\sum_{k=2}^n \frac{H_{k-1}}{k}\Biggr] z^n.
946:   \end{align*}
947:   The first sum is $(1-z)^{-1}\Li_{2,0}(z)$.  For the second sum,
948:   observe that
949:   \begin{align*}
950:     \sum_{n \geq 1}\Biggl[ \sum_{k=2}^n \frac{H_{k-1}}{k}\Biggr] z^n 
951:     &= \sum_{k \geq 2} \frac{H_{k-1}}k \sum_{n \geq k}z^n = (1-z)^{-1}
952:     \sum_{k \geq 1} \frac{H_{k}}{k+1} z^{k+1}\\
953:     &= (1-z)^{-1}\int_0^z(1-\zeta)^{-1}L(\zeta)\,d\zeta =
954:     \frac{1}{2}(1-z)^{-1}L^2(z), 
955:   \end{align*}
956:   and the claim follows.
957: \end{proof}
958: \end{enumerate}
959: 
960: In light of these facts we can now conclude that
961: \begin{align}
962:   \label{eq:5.2.3}
963:   &\Li_{0,1}(z) \odot \int_0^z(1-\zeta)^{-1}[G_1(\zeta) -
964:   C_1(1-\zeta)^{-2}]\,d\zeta\\
965:   &= [(1-z)^{-1}L(z) + \Gamma'(1)(1-z)^{-1} - \frac12L(z) + O(1)]\notag\\
966:   &\qquad\odot [-(1-z)^{-1}L(z) + (L_{-1}+1)(1-z)^{-1} +\frac14 L(z) +
967:   O(1)]\notag\\ 
968:   &=-\bigl[(1-z)^{-1}L^2(z) + (1-z)^{-1}\Li_{2,0}(z)\bigr] +
969:   (L_{-1}+1)(1-z)^{-1}L(z) \notag\\
970:   &\qquad - \Gamma'(1)(1-z)^{-1}L(z)
971:   + \Gamma'(1)(L_{-1}+1)(1-z)^{-1} + O(|1-z|^{-\epsilon}). \notag
972: \end{align}
973: To get an asymptotic expansion for \( \Li_{2,0} \) we
974: use~\eqref{eq:4.1.posintalpha}:
975: \begin{equation*}
976:   \Li_{2,0}(z) = t(\ln{t} - H_1) + \frac{\pi^2}{6} + O(|1-z|^2) =
977:   \frac{\pi^2}6 - (1-z)L(z) + O(|1-z|).
978: \end{equation*}
979: Hence we see that the Hadamard product~\eqref{eq:5.2.3} has the asymptotic
980: expansion
981: \begin{align}
982:   &-(1-z)^{-1}L^2(z) - (1-z)^{-1}\Bigl(\frac{\pi^2}{6} - (1-z)L(z) +
983:   O(|1-z|)\Bigr)\notag\\
984:   &\quad + (L_{-1}  + 1 - \Gamma'(1))(1-z)^{-1}L(z) +
985:   \Gamma'(1)(L_{-1} + 1)(1-z)^{-1} + O(|1-z|^{-\epsilon})\notag\\
986:   &= -(1-z)^{-1}L^2(z) + (L_{-1} + 1-\Gamma'(1))(1-z)^{-1}L(z)\notag\\
987:   &\quad + \Bigl[\Gamma'(1)(L_{-1} + 1)-\frac{\pi^2}6\Bigr](1-z)^{-1} +
988:   O(|1-z|^{-\epsilon}).
989:   \label{V10.5}
990: \end{align}
991: 
992: Using~\eqref{eq:5.2.V5.7},~\eqref{eq:5.2.V9.1}, and~\eqref{V10.5}
993: in~\eqref{eq:5.2.V3.1}, we get
994: \begin{align}
995:   &B(z)\notag\\ &= C_1^2 + (1-z)^{-1}L^2(z) + 2\Gamma'(1)(1-z)^{-1}L(z) +
996:   \Gamma''(1)(1-z)^{-1}\notag\\
997:   &+ 4\Bigl[-(1-z)^{-1}L^2(z) +
998:   (L_{-1}+1-\Gamma'(1))(1-z)^{-1}L(z)
999:   +\bigl[\Gamma'(1)(L_{-1}+1)-\frac{\pi^2}{6}\bigr](1-z)^{-1}
1000:   \notag\Bigr]\\ 
1001:   &+ 2\Bigl[(1-z)^{-1}L^2(z) - 2(L_{-1}+1)(1-z)^{-1}L(z)
1002:    + (2+2L_{-1}+ L_{-1}^2)(1-z)^{-1}\Bigr] +
1003:   O(|1-z|^{-\epsilon})\notag\\ 
1004:   \label{eq:5.2.V12.3}
1005:   &= -(1-z)^{-1}L^2(z) + V_1(1-z)^{-1}L(z) + V_2(1-z)^{-1} +
1006:   O(|1-z|^{-\epsilon}), 
1007: \end{align}
1008: where
1009: \begin{equation}
1010:   \label{eq:5.2.defV1}
1011:   V_1 := -2\Gamma'(1) = 2\gamma
1012: \end{equation}
1013: and
1014: \begin{equation}
1015:   \label{eq:5.2.defV2}
1016:   V_2 := \Gamma''(1) + 4\Bigl( \Gamma'(1)(L_{-1}+1) - \frac{\pi^2}6\Bigr) +
1017:   2(2+2L_{-1}  + L_{-1}^2) = 4 - \gamma^2 - \frac{\pi^2}2.
1018: \end{equation}
1019: 
1020: Let $\widehat{B}(z) := B(z) - r_0(2) = B(z) - C_1^2$.  By direct
1021: application of the ETT, $A(z)$, defined to be the generating function
1022: of $\bigl(\tilde{\mu}_n(2)\bigr)$, is given by
1023: \[
1024: A(z) = c_1(1-z)^{-2} + \widehat{B}(z) + 2(1-z)^{-2}\int_0^z
1025: \widehat{B}(\zeta)(1-\zeta)\,d\zeta, 
1026: \]
1027: where $c_1 = r_0(2) = C_1^2$.  The integral here can be broken up as
1028: \[
1029: \int_0^z \widehat{B}(\zeta)(1-\zeta)\,d\zeta = V - \int_z^1
1030: \widehat{B}(\zeta)(1-\zeta)\,d\zeta,
1031: \]
1032: where the first term on the right is the constant
1033: \begin{equation*}
1034: V := \int_0^1 \widehat{B}(\zeta)(1-\zeta)\,d\zeta = \int_0^1 \left(
1035:   \sum_{k=1}^\infty r_k(2)\zeta^k\right) (1-\zeta)\,d\zeta
1036: = \sum_{k=1}^\infty \frac{r_k(2)}{(k+1)(k+2)}.
1037: \end{equation*}
1038: For the second term we use the following consequences of
1039: integration by parts:
1040: \[
1041: \int_z^1 L^2(\zeta)\,d\zeta = (1-z)L^2(z) + 2(1-z)L(z) + 2(1-z),
1042: \]
1043: \[
1044: \int_z^1 L(\zeta)\,d\zeta = (1-z)L(z) + (1-z)
1045: \]
1046: and the crude remainder estimate
1047: %% \[
1048: %% \int_z^1 (1-\zeta)L^2(\zeta)\,d\zeta = \frac12(1-z)^2L^2(z) +
1049: %% \frac12(1-z)^2L(z) - \frac14(1-z)^2.
1050: %% \]
1051: \[
1052: \int_z^1 O(|1-\zeta|^{1-\epsilon})\,d\zeta = O(|1-z|^{2-\epsilon}).
1053: \]
1054: Thus
1055: \begin{align*}
1056:   &(1-z)^{-2}\int_z^1 \widehat{B}(\zeta)(1-\zeta)\,d\zeta\\
1057:   &= (1-z)^{-2}\bigl[ -(1-z)L^2(z) - 2(1-z)L(z) - 2(1-z) + V_1(1-z)L(z)\\
1058:  & \qquad\qquad + V_1(1-z) + V_2(1-z) + O(|1-z|^{2-\epsilon})\bigr]\\
1059:   &= -(1-z)^{-1}L^2(z) + (V_1-2)(1-z)^{-1}L(z)  +
1060:   (V_1+V_2-2)(1-z)^{-1}  +  O(|1-z|^{-\epsilon}), 
1061: \end{align*}
1062: and
1063: \begin{align}
1064:   A(z) &= C_1^2(1-z)^{-2} - (1-z)^{-1}L^2(z) + V_1(1-z)^{-1}L(z) + V_2(1-z)^{-1} +
1065:   O(|1-z|^{-\epsilon}) \notag\\
1066:   &\quad + 2\Bigl[ V(1-z)^{-2} + (1-z)^{-1}L^2(z) -
1067:   (V_1-2)(1-z)^{-1}L(z)\notag \\ 
1068:   &\qquad - (V_1+V_2-2)(1-z)^{-1} + O(|1-z|^{-\epsilon})\notag\Bigr]\\
1069:   &= (C_1^2+2V)(1-z)^{-2} + (1-z)^{-1}L^2(z) + (V_1 -
1070:   2(V_1-2))(1-z)^{-1}L(z) \notag\\
1071:   & \qquad + (V_2 - 2(V_1+V_2-2))(1-z)^{-1} + O(|1-z|^{-\epsilon})\notag\\
1072:   &= (C_1^2+2V)(1-z)^{-2} + (1-z)^{-1}L^2(z) + (4-V_1)(1-z)^{-1}L(z)\notag\\
1073:   \label{eq:5.2.V16.3}
1074:   &\qquad + (4-2V_1-V_2)(1-z)^{-1} + O(|1-z|^{-\epsilon}).
1075: \end{align}
1076: 
1077: We can now transfer this asymptotic expansion for the generating
1078: function to one for its Taylor coefficients, invoking
1079: Theorem~\ref{theorem:transfer} for the remainder estimate.  We know that
1080: \[
1081: [z^n](1-z)^{-1}L(z) = H_n = \ln{n} + \gamma + O(n^{-1})
1082: \]
1083: and, from the proof of Lemma~\ref{lem:V1},
1084: \begin{align*}
1085: [z^n]\bigl[(1-z)^{-1} L^2(z)\bigr] &= H_n^2 - \sum_{k=1}^n \frac{1}{k^2} =
1086: \Bigl(\ln{n}+\gamma+O(n^{-1})\Bigr)^2 - \sum_{k=1}^\infty
1087: \frac{1}{k^2} + O(n^{-1})\\
1088: &= (\ln{n})^2 + 2\gamma\ln{n} + \biggl(\gamma^2 - \frac{\pi^2}6\biggr) +
1089: O\biggl(\frac{\ln{n}}{n}\biggr).
1090: \end{align*}
1091: So, by singularity analysis (Theorem~\ref{theorem:transfer}),
1092: \begin{multline}
1093:   \label{eq:5.2.V17.3}
1094:   \tilde{\mu}_n(2) = (C_1^2+2V)(n+1) + (\ln{n})^2 + 2\gamma\ln{n} +
1095:   \biggl(\gamma^2 - \frac{\pi^2}6 \biggr) +  
1096:   (4-V_1)(\ln{n} + \gamma)\\  + (4-2V_1 - V_2) + O(n^{-1+\epsilon}).
1097: \end{multline}
1098: 
1099: In order to obtain the variance using~\eqref{eq:5.2.70} we also need
1100: [recall~\eqref{eq:5.1.61}]
1101: \begin{equation*}
1102:   \tilde{\mu}_n(1)^2 = (\ln{n})^2 + 4\ln{n} + 4 + O(n^{-1+\epsilon}).
1103: \end{equation*}
1104: Hence, using~\eqref{eq:5.2.defV1} and~\eqref{eq:5.2.defV2} we obtain
1105: an expansion for the variance:
1106: \begin{align}
1107:   \sigma_n^2 &= (C_1^2+2V)(n+1) + (2\gamma-V_1)\ln{n}
1108:   + \biggl[\Bigl(\gamma^2 -
1109:   \frac{\pi^2}6\Bigr) + (4-V_1)\gamma  - 2V_1 - V_2 \biggr] +
1110:   O(n^{-1+\epsilon})\notag\\
1111:   \label{eq:5.2.V19.1}
1112:   &= (C_1^2+2V)(n+1) 
1113:   - \Bigl(4 - \frac{\pi^2}3\Bigr) +  O(n^{-1+\epsilon}).
1114: \end{align}
1115: Numerical calculations using \texttt{Mathematica} suggest that the
1116: next term in the expansion is $-\tfrac1{2n}$.  One could possibly
1117: obtain this by carrying out the program of this section, keeping track
1118: of more terms in the asymptotic expansions.
1119: 
1120: 
1121: %%% Local Variables: 
1122: %%% mode: latex
1123: %%% TeX-master: "main"
1124: %%% End: 
1125: