nlin0001026/lds.tex
1: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2: %%% Cohen and Kottos (December 1999, revised September 2000). 
3: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
4: 
5: % FIGURES:
6: % Fig.1: lds_contours.eps; lds_poincare.eps
7: % Fig.2: comega.eps
8: % Fig.3: profile_cl.eps
9: % Fig.4: profiles.eps 
10: % Fig.5: lds_anlys.eps
11: 
12: \documentstyle[aps,pre,multicol,epsf]{revtex}
13: \textheight9.6in
14: \begin{document}
15: \newcommand{\tbox}[1]{\mbox{\tiny #1}}
16: \newcommand{\half}{\mbox{\small $\frac{1}{2}$}}
17: \newcommand{\mbf}[1]{{\mathbf #1}}
18: 
19: \title{Parametric dependent Hamiltonians, wavefunctions, \\ 
20: random-matrix-theory, and quantal-classical correspondence}
21: 
22: \author
23: {
24: Doron Cohen$^1$ and Tsampikos Kottos$^2$\\
25: %
26: \footnotesize
27: %
28: $^1$
29: Department of Physics, Harvard University, Cambridge, MA 02138 \\
30: %
31: $^2$
32: Max-Planck-Institut f\"ur Str\"omungsforschung,
33: 37073 G\"ottingen, Germany
34: }
35: 
36: \date{December 1999, revised September 2000}
37: 
38: \maketitle
39: 
40: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
41: 
42: \begin{abstract}
43: 
44: We study a classically chaotic system which is described 
45: by a Hamiltonian ${\cal H}(Q,P;x)$ where $(Q,P)$ are 
46: the canonical coordinates of a particle in a 2D well, 
47: and $x$ is a parameter. By changing $x$ we can deform 
48: the `shape' of the well. The quantum-eigenstates of the 
49: system are $|n(x)\rangle$. We analyze numerically how the 
50: parametric kernel $P(n|m)= |\langle n(x)|m(x_0)\rangle|^2$ 
51: evolves as a function of $\delta x \equiv (x{-}x_0)$. 
52: This kernel, regarded as 
53: a function of $n-m$, characterizes the shape of the wavefunctions, 
54: and it also can be interpreted as the local density of states (LDOS). 
55: The kernel $P(n|m)$ has a well defined classical limit, 
56: and the study addresses the issue of 
57: quantum-classical correspondence (QCC).  
58: Both the perturbative and the non-perturbative regimes 
59: are explored. The limitations of the random-matrix-theory (RMT) 
60: approach are demonstrated. 
61: 
62: \end{abstract}
63: 
64: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
65: \begin{multicols}{2}
66: 
67: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
68: \section{Introduction}
69: 
70: Consider a system whose total Hamiltonian is ${\cal H}(Q,P;x)$, 
71: where $(Q,P)$ is a set of canonical coordinates, 
72: and $x$ is a constant parameter. This parameter 
73: may represent the effect of some externally controlled field. 
74: We assume that both ${\cal H}_0 = {\cal H}_0(Q,P;x_0)$ 
75: and ${\cal H} = {\cal H}(Q,P;x)$ 
76: generate classically chaotic dynamics of similar nature. 
77: Moreover, we assume that $\delta x \equiv (x{-}x_0)$ 
78: is {\em classically small}, 
79: meaning that it is possible to apply linear analysis 
80: in order to describe how the energy surfaces ${\cal H}(Q,P;x)=E$ 
81: are deformed as a result of changing the value of $x$. 
82: %
83: %
84: Quantum mechanically, we can use a basis where 
85: ${\cal H}_0 = \mbf{E}_0$ has a diagonal representation, 
86: while 
87: %
88: \begin{eqnarray} \label{e1} 
89: {\cal H} \ \  = \ \ \mbf{E}_0 \ + \ \delta x \ \mbf{B}
90: \end{eqnarray}
91: %
92: %
93: For reasonably small $\hbar$, it follows from general 
94: semiclassical considerations \cite{mario},  
95: that $\mbf{B}$ is a {\em banded matrix}. 
96: Generically, this matrix {\em looks random}, 
97: as if its off-diagonal elements were {\em independent} random numbers. 
98: 
99: 
100: It was the idea of Wigner \cite{wigner} forty years ago,  
101: to study a simplified model, where the Hamiltonian is 
102: given by Eq.(\ref{e1}), and where 
103: $\mbf{B}$ is a {\em random} banded matrix. This is 
104: known as Wigner's banded random matrix (WBRM) model.     
105: The applicability of such a model is a matter 
106: of {\em conjecture}. Obviously this conjecture 
107: should be tested$^\dag$. The most direct way to 
108: test it, which we are going to apply, is to 
109: take the matrix $\mbf{B}$ of a `physical' Hamiltonian, 
110: and then to randomize the signs of its off-diagonal 
111: elements. The outcome of such operation will be referred to  
112: as the {\em effective} WBRM model that is associated 
113: with the {\em physical} Hamiltonian. One issue of this paper 
114: is to make a comparison between the eigenstates of 
115: the physical Hamiltonian, and those of the associated  
116: effective WBRM model. 
117:  
118: 
119: The {\em standard} WBRM model (unlike the `effective' one) 
120: involves an additional simplification. Namely, one 
121: assumes that $\mbf{B}$ has a rectangular band profile. 
122: The theory of eigenstates for the standard WBRM model 
123: is well known \cite{wigner,casati,fyodo}. 
124: %
125: Increasing $x$, starting from $\delta x=0$
126: the eigenstates of Eq.(\ref{e1}) change their nature. 
127: %
128: The general questions to address are: 
129: %
130: \begin{enumerate}
131: \item 
132: What are the {\em parametric regimes} in the 
133: parametric evolution of the eigenstates;  
134: \item
135: How the structure of the eigenstates 
136: changes as we go via the subsequent regimes. 
137: \end{enumerate}
138: %
139: %
140: Recently some ideas have been introduced \cite{frc,vrn,wls} 
141: how to go beyond Wigner's theory in case of physical Hamiltonians. 
142: It has been suggested that there are at least three generic 
143: parametric scales   
144: $\delta x_c^{\tbox{qm}}\ll
145: \delta x_{\tbox{prt}}\ll
146: \delta x_{\tbox{SC}}$ 
147: that control the parametric evolution of the eigenstates. 
148: We shall define these parametric scales later. 
149: Accordingly one should distinguish between 
150: the standard perturbative regime ($\delta x \ll \delta x_c^{\tbox{qm}}$), 
151: the core-tail regime 
152: ($\delta x_c^{\tbox{qm}} \ll \delta x \ll \delta x_{\tbox{prt}}$), 
153: and the semiclassical regime  
154: ($\delta x \gg \delta x_{\tbox{SC}}$).  
155: 
156: 
157: 
158: The purpose of this paper is not just to numerically 
159: establish (for the first time) the existence of the 
160: parametric regimes suggested in \cite{frc,vrn,wls}, 
161: but mainly to address question {\bf (2)} above \cite{prm}. 
162: Namely, we would like to study how the structure of the 
163: eigenstates changes as we go via the subsequent regimes. 
164: In particular we would like to understand the significance
165: of RMT assumptions in the general theoretical considerations. 
166: The latter issue has been left unexplored in the 
167: `quantum chaos' literature. (Note however that literally 
168: the same question is addressed in numerous publication once 
169: spectral statistics of eigenvalues, 
170: rather than eigenstate structure, is concerned).   
171: We also suggest a new procedure for `region analysis' of the 
172: eigenstate structure. We are going to distinguish between 
173: first-order tail regions (FOTRs), higher-order far-tail 
174: regions, and non-perturbative (core) region. Our main 
175: conclusion is going to be that RMT is inadequate for 
176: the analysis of any features that go beyond 
177: first-order perturbation theory.  
178: 
179: 
180: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
181: \section{The model Hamiltonian}
182: 
183: We study the Hamiltonian 
184: %
185: \begin{eqnarray} \label{e2} 
186: {\cal H}(Q,P;x) = \half(P_1^2{+}P_2^2 + Q_1^2{+}Q_2^2)  
187: + x \cdot Q_1^2 Q_2^2
188: \end{eqnarray}
189: %
190: with $x=x_0+\delta x$ and $x_0=1$. 
191: This Hamiltonian describes the motion of 
192: a particle in a 2D well (see Fig.1). 
193: The units are chosen such that the mass is equal to one, 
194: the frequency for small oscillations is one, 
195: and for $\delta x=0$ the coefficient of the 
196: anharmonic term is also one. The energy $E$ 
197: is the only dimensionless parameter of the 
198: classical motion. Our numerical study is 
199: focused on an energy window around $E \sim 3$ 
200: where the motion is mainly chaotic. 
201: 
202: 
203: In the classical analysis there is only one 
204: parametric scale, which is $\delta x_c^{\tbox{cl}} \sim 1$. 
205: This scale determines the regime of 
206: (classical) linear analysis. 
207: For $\delta x \ll \delta x_c^{\tbox{cl}}$ 
208: the deformation of the energy surface  
209: ${\cal H}_0(Q,P;x)=E$ can be described 
210: as a linear process. Later we are going to 
211: give a precise mathematical formulation of 
212: this idea. From now on assume that we 
213: are in the classical linear regime.  
214: 
215: 
216: 
217: \ \\
218: %%%%%%%%%%%%%%%%%%%%%%%%%%%
219: \epsfysize=1.3in
220: \epsffile{lds_contours.eps} 
221: \ \ \ 
222: \epsfysize=1.3in
223: \epsffile{lds_poincare.eps}
224: \noindent \\
225: {\footnotesize {\bf FIG.1}:
226: {\em Left:} equipotential contours of the model 
227: Hamiltonian (\ref{e2}) with $x=x_0=1$.  
228: {\em Right:} A Poincare section of a long trajectory 
229: ($0<t<1300$) that we have picked in order to get the 
230: fluctuating quantity ${\cal F}(t)$. 
231: The initial conditions are \mbox{$(Q_1,Q_2,P_1,P_2)=(1,0,1,2)$} 
232: corresponding to $E=3$. The trajectory is quite ergodic.  
233: It avoids some small quasi-integrable islands 
234: (the main one is around $(0,0)$). } \\
235: %%%%%%%%%%%%%%%%%%%%%%%%%%%
236: 
237: 
238: 
239:   
240: Let us pick a very long ergodic trajectory 
241: $(Q(t),P(t))$ that covers densely the energy 
242: surface $E$. See Fig.1. 
243: Let us define the fluctuating quantity 
244: %
245: \begin{eqnarray} \label{e_2} 
246: {\cal F}(t) \ \equiv \ 
247: -(\partial {\cal H}/\partial x)
248: \ = \ -Q_1^2 Q_2^2
249: \end{eqnarray} 
250: %
251: For the later analysis it is important to know 
252: the distribution of the variable ${\cal F}$, 
253: and to characterize its temporal correlations. 
254: The average value is $F = \langle {\cal F} \rangle$. 
255: The angular brackets stand for microcanonical average 
256: over $(Q(0),P(0))$, which should be the same as 
257: time ($t$) average (due to the assumed ergodicity).   
258: The auto-correlation function of ${\cal F}(t)$ is
259: %
260: \begin{eqnarray} \label{e_3} 
261: C(\tau) \ \ = \ \ 
262: \langle \ ({\cal F}(t)-F) \ ({\cal F}(t{+}\tau)-F) \ \rangle
263: \end{eqnarray} 
264: %
265: Note that $C(\tau)$ is independent of $t$, and 
266: that average over $t$ should give the same result 
267: as a microcanonical average over $(Q(0),P(0))$.  
268: 
269: 
270: 
271: The variance of the fluctuations 
272: is $C(0) = \langle ({\cal F}-F)^2 \rangle$. 
273: The correlation time will be denoted by 
274: $\tau_{\tbox{cl}}$. Note that with our choice 
275: of units $\tau_{\tbox{cl}} \sim 1.0$  
276: within the energy range of interest. 
277: The power spectrum $\tilde{C}(\omega)$
278: of the fluctuating ${\cal F}(t)$, is obtained  
279: via a Fourier transform of $C(\tau)$. See Fig.2.
280: The average $F$ and the variance $C(0)$ determine 
281: just the first two moments of the ${\cal F}$ 
282: distribution. The probability density of ${\cal F}$
283: will be denoted by $P_{\tbox{F}}({\cal F})$. 
284: 
285: 
286: 
287: All the required information for the 
288: subsequent semiclassical analysis is contained 
289: in the functions $C(\tau)$ and $P_{\tbox{F}}({\cal F})$ 
290: as defined above. All we have to do in 
291: order to numerically determine them is 
292: to generate one very long ergodic trajectory (see Fig.1),  
293: to compute the respective ${\cal F}(t)$, 
294: and from it to extract the desired information 
295: (see Fig.2 and Fig.3).     
296: It is convenient to express $P_{\tbox{F}}({\cal F})$
297: in terms of a scaling function as follows
298: %
299: \begin{eqnarray} \label{e_4} 
300: P_{\tbox{F}}({\cal F}) \ = \ \frac{1}{\sqrt{C(0)}} 
301: \hat{P}_{\tbox{cl}}\left(-\frac{{\cal F}-F}{\sqrt{C(0)}}\right)
302: \end{eqnarray} 
303: % 
304: By this definition the scaled distribution  
305: $\hat{P}_{\tbox{cl}}(f)$ is characterized 
306: by a zero average ($\langle f \rangle = 0$), 
307: a unit variance ($\langle f^2 \rangle = 1$), 
308: and it is properly normalized. Note that 
309: $\hat{P}_{\tbox{cl}}(-f)$  rather than 
310: $\hat{P}_{\tbox{cl}}(f)$ correspond 
311: to $P_{\tbox{F}}({\cal F})$. This has been done 
312: for later convenience. 
313: 
314: 
315: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
316: \section{The quantized Hamiltonian}
317: 
318: 
319: Upon quantization we have a second 
320: dimensionless parameter $\hbar$.  
321: For obvious reasons we are considering
322: a de-symmetrized ($1/8$) well with Dirichlet 
323: boundary conditions on the lines
324: $Q_1{=}0$ and $Q_2{=}0$ and $Q_1{=}Q_2$. 
325: The matrix representation of ${\cal H}={\cal H}(Q,P;x)$ 
326: in the basis which is determined 
327: by ${\cal H}(Q,P;0)$ is very simple.
328: The eigenstates ($n=1,2,3\cdots$) of the chaotic 
329: Hamiltonian ${\cal H}_0={\cal H}(Q,P;1)$ 
330: has been found numerically. 
331: 
332: 
333: The phase space volume ($dQdP$ integral) 
334: which is enclosed by an energy surface 
335: ${\cal H}(Q,P;x)=E$  is given by a function 
336: $n=\Omega(E,x)$. It is convenient to measure 
337: phase space volume in units of $(2\pi\hbar)^d$, 
338: where $d=2$ is the dimensionality of our system. 
339: Upon quantization the phase space volume $n$ 
340: corresponds to the level index ($n=1,2,3\cdots$). 
341: This is known as Weyl law. It follows that 
342: $g(E)=\partial_E\Omega(E,x)$ corresponds  
343: to the density of states, and 
344: $\Delta = 1/g(E) \propto \hbar^d $ 
345: is the mean level spacing. 
346: 
347: 
348: In the following presentation we are 
349: going to assume the our interest is restricted 
350: to an energy window which is `classically small' 
351: but `quantum mechanically large'.  
352: In the numerical analysis of our model Hamiltonian 
353: the energy window was $2.8<E<3.1$, where the 
354: classical motion is predominantly chaotic.   
355: The mean level spacing for $E\sim 3$ 
356: is given approximately by the formula 
357: $\Delta\approx 4.3*\hbar^2$. 
358: Our numerical analysis has been carried  
359: out for $\hbar=0.03$ and for $\hbar=0.015$. 
360: Smaller values of $\hbar$ where beyond our 
361: numerical capabilities since the maximal 
362: matrix that we can handle is of size $5000 \times 5000$.  
363:  
364:  
365: 
366: The representation of $Q_1^2 Q_2^2$, 
367: in the basis which is determined by the chaotic 
368: Hamiltonian ${\cal H}_0$, gives 
369: the matrix $\mbf{B}$ of Eq.(\ref{e1}). 
370: The banded matrix $\mbf{B}$ and the band profile 
371: are illustrated in Fig.2.
372: The band profile is implied by the semiclassical 
373: relation \cite{mario}:  
374: %
375: \begin{eqnarray} \label{e_5}  
376: |\mbf{B}_{nm}|^2 \ \ \approx \ \ 
377: \frac{\Delta}{2\pi\hbar} 
378: \tilde{C}\left(\frac{E_n-E_m}{\hbar}\right)  
379: \end{eqnarray} 
380: %
381: As we see from Fig.2 the agreement  
382: with this formula is remarkable.
383: For the bandwidth Eq.(\ref{e_5}) implies that 
384: $\Delta_b=2\pi\hbar/\tau_{\tbox{cl}}$. 
385: It is common to define $b=\Delta_b/\Delta$. 
386: 
387: \ \\
388: %%%%%%%%%%%%%%%%%%%%%%%%%%%
389: \epsfysize=1.65in
390: \epsffile{comega.eps}
391: \noindent \\
392: {\footnotesize {\bf FIG.2}:
393: The band profile $(2\pi\hbar/\Delta)\cdot|\mbf{B}_{nm}|^2$ 
394: versus $\omega = (E_n{-}E_m)/\hbar$ is compared 
395: with the classical power spectrum $C(\omega)$. 
396: {\em Inset}: An image of a piece of the $\mbf{B}$ matrix. } \\
397: %%%%%%%%%%%%%%%%%%%%%%%%%%%
398: 
399: 
400: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
401: \section{Definition of the LDOS profile}
402: 
403: 
404: The quantum-eigenstates of the Hamiltonian 
405: ${\cal H}(Q,P;x)$ are $|n(x)\rangle$, 
406: and the ordered eigen-energies are $E_n(x)$.  
407: We are interested in the parametric kernel 
408: %
409: \begin{eqnarray} \label{e3} 
410: P(n|m) \ = \ |\langle n(x)|m(x_0)\rangle|^2 
411: \ = \ \mbox{trace}(\rho_n\rho_m)
412: \end{eqnarray}
413: %
414: In the equation above $\rho_m(Q,P)$ and $\rho_n(Q,P)$ 
415: are the Wigner functions that correspond to the 
416: eigenstates $|m(x_0)\rangle$ and $|n(x)\rangle$ 
417: respectively. The trace stands for $dQdP/(2\pi\hbar)^d$ 
418: integration. 
419: 
420: 
421: 
422: We can identify $P(n|m)$ as the local density of states (LDOS), 
423: by regarding it as a function of $n$, where $m$ is considered 
424: to be a fixed reference state. An average of $P((m+r)|m)$ over 
425: several $m$-states leads to the LDOS profile $P(r)$. 
426: %
427: Alternatively, fixing $n$, the vector $P(n|m)$ describes 
428: the shape of the $n$-th eigenstate in the ${\cal H}_0$ representation. 
429: By averaging $P(n|(n-r))$ over few eigenstates one obtains 
430: the average shape of the eigenstate (ASOE).  
431: The ASOE is just $P(-r)$. Thus the ASOE and the LDOS are 
432: given by the same function. One would have to be more careful 
433: with these definitions if ${\cal H}_0$ were  
434: integrable while ${\cal H}$ non-integrable.  
435:  
436:  
437: 
438: The kernel $P(n|m)$ gives the overlap between 
439: the $n$th eigenstate of ${\cal H}$ and the 
440: $m$th eigenstate of ${\cal H}_0$.  
441: For $\delta x=0$ we have simply $P(n|m)=\delta_{nm}$. 
442: For $\delta x > 0$ the kernel develops a structure, 
443: which is described by the LDOS profile $P(r)$. 
444: If $\delta x$ is very small then evidently 
445: $P(r)$ consists of Kronecker delta (at $r=0$) 
446: and tail regions ($|r|>0$). Later we are going to 
447: distinguish between first-order tail regions (FOTRs), 
448: and higher order far-tail regions. 
449: As $\delta x$ becomes larger a non-perturbative
450: core region appears around $r=0$.  
451: Namely, the profile exhibits a bunch of states 
452: (rather than one) that share most of the probability. 
453: If $\delta x$ becomes even larger, the distinction between 
454: core and tail regions become meaningless, and 
455: the LDOS profile becomes purely non-perturbative. 
456: We are going to explain that the non-perturbative 
457: profile reflects the underlying classical 
458: phase space structure.
459: 
460: 
461: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
462: \section{The classical approximation for the LDOS}
463: 
464: 
465: The classical approximation \cite{felix2,frc,vrn,wls} 
466: for $P(n|m)$ follows naturally from the definition Eq.(\ref{e3}).  
467: It is obtained if we approximate $\rho_n(Q,P)$ 
468: by a microcanonical distribution that 
469: is supported by the energy surface 
470: ${\cal H}(Q,P;x)=E_n(x)$. Namely, 
471: %
472: \begin{eqnarray} \nonumber 
473: \rho_n(Q,P) \ &=& \ 
474: \frac{1}{g(E)}\delta({\cal H}(Q,P;x)-E_n(x)) 
475: \\  \label{e_mc} 
476: \ &=& \ \delta(\Omega({\cal H}(Q,P;x)) - n)
477: \end{eqnarray}
478: %
479: and a similar expression (with $x=x_0$) for 
480: $\rho_m(Q,P)$. In the classical limit $n$ is 
481: the phase space volume by which we label energy surfaces.
482: Each energy surface $n$ is associated 
483: with a microcanonical state $\rho_n(Q,P)$. 
484: The classical LDOS profile will be denoted 
485: by $P_{\tbox{cl}}(r)$. The $\delta x$ regime where 
486: the classical approximation $P(r) \approx P_{\tbox{cl}}(r)$
487: applies will be discussed in a later section. 
488: 
489:  
490:     
491: By definition, for $\delta x \ll \delta x_c^{\tbox{cl}}$ 
492: the deformed energy surfaces departs linearly 
493: from the $\delta x=0$ surfaces. As already 
494: stated in the Introduction, being in this classical 
495: linear regime is a fixed assumption of this 
496: paper. Now we want to explain the consequences 
497: of this assumption. One may consider these consequences 
498: as giving an operational definition for the classical 
499: linear regime.     
500: %
501: The dispersion (square-root of the variance) of 
502: the classical profile in the classical linear regime is 
503: %
504: \begin{eqnarray} \label{e_6} 
505: \delta E_{\tbox{cl}} \ \ = \ \ \sqrt{C(0)} \times \delta x  
506: \end{eqnarray}
507: %
508: (This should be divided by $\Delta$ if we want the 
509: dispersion in proper $r$ units. See (\ref{e_7}) below).
510: For our model Hamiltonian, for energies $E \sim 3$, 
511: we have found that $\delta E_{\tbox{cl}}\approx 0.38*\delta x$.
512: Eq.(\ref{e_6}) can be regarded as a special consequence 
513: of the following scaling relation which we are going to derive below:
514: %
515: \begin{eqnarray} \label{e4} 
516: P_{\tbox{cl}}(r) \ = \ 
517: \frac{\Delta}{\sqrt{C(0)}\ \delta x} \cdot 
518: \hat{P}_{\tbox{cl}}\left(\frac{\Delta \cdot r}{\sqrt{C(0)}\ \delta x}\right) 
519: \end{eqnarray}
520: %
521: The scaling function has already been defined 
522: in Eq.(\ref{e_4}), and it is illustrated in Fig.3.
523: The classical profile $P_{\tbox{cl}}(r)$ is in general 
524: non-symmetric, but it follows from Eq.(\ref{e4}) 
525: that it must be characterized by $\langle r \rangle = 0$. 
526: [By definition the scaling function of Eq.(\ref{e_4}) 
527: gives zero average]. Another obvious feature 
528: is having sharp cutoffs, beyond which $P_{\tbox{cl}}(r)=0$. 
529: The existence of these outer `classically forbidden' regions 
530: follows from the observation that for large enough $r$ 
531: there is no longer classical overlap between the 
532: energy surfaces that correspond to $|m(x_0)\rangle$ 
533: and $|n(x)\rangle$ respectively.    
534:  
535: 
536: The rest of this section is dedicated to technical 
537: clarifications of Eq.(\ref{e4}), and it can be skipped 
538: in first reading. The derivation is done in two steps. 
539: The first step is to establish a relation 
540: between $P_{\tbox{cl}}(r)$ and its trivially related 
541: version $P_{\tbox{E}}(\epsilon)$. The second step is to  
542: demonstrate that $P_{\tbox{E}}(\epsilon)$ is related 
543: to $P_{\tbox{F}}({\cal F})$ of Eq.(\ref{e_4}).  
544: It is also possible to make a one-step derivation 
545: that relates $P_{\tbox{cl}}(r)$ to $P_{\tbox{F}}({\cal F})$, 
546: but we find the derivation below more physically 
547: appealing.
548: 
549: 
550: \ \\
551: %%%%%%%%%%%%%%%%%%%%%%%%%%%
552: \epsfysize=1.3in
553: \epsffile{profile_cl.eps} \\
554: \noindent
555: {\footnotesize {\bf FIG.3}:
556: The scaled classical profile $\hat{P}_{\tbox{cl}}()$. 
557: One unit on the horizontal axis correspond to 
558: energy difference $\delta E_{\tbox{cl}}\approx 0.38*\delta x$.
559: Note that $r = 0$ implies $(E_n(x){-}E_m(x_0))>0$. 
560: The caustic is located at $(E_n(x){-}E_m(x_0))=0$, 
561: while the anti-caustic is located at $(E_n(x){-}E_m(x_0)) = 1.65*x$.   
562: The ``forbidden regions'' are defined 
563: as those regions where $P_{\tbox{cl}}(r) = 0$.
564: They are located to the left of the caustic 
565: and to the right of the anti-caustic. } \\
566: %%%%%%%%%%%%%%%%%%%%%%%%%%%
567: 
568: 
569: By differentiating of $n=\Omega(E,x)$, 
570: keeping $n$ constant, we get the relation 
571: $\delta E = - F(x) \delta x$, where 
572: $F(x) = \partial_x\Omega(E,x)/g(E)$ 
573: is known as the (generalized) conservative force.
574: Using the latter expression it is a straightforward 
575: exercise to prove that 
576: $F(x) = \langle {\cal F} \rangle \equiv F$.  
577: Alternatively, we can eliminate $E$ from 
578: the relation $n=\Omega(E,x)$, and write the 
579: result as $E=E_n(x)$. Accordingly  
580: $F(x) = -(\partial E_n(x)/\partial x)$. 
581: Now we can write the following relation:
582: %
583: \begin{eqnarray} \nonumber
584: E_n(x) - E_m(x_0) \ = \ 
585: \left. \frac{\partial E}{\partial x} \right|_n \delta x
586: + \left. \frac{\partial E}{\partial n} \right|_x (n-m)
587: \end{eqnarray}
588: %
589: which can be re-written in the following form
590: %
591: \begin{eqnarray} \label{e_7}
592: \epsilon \ \  = \ \  
593: -F(x) \ \delta x \ \ + \ \ ({1}/{g(E)}) \ r
594: \end{eqnarray}
595: %   
596: Whenever we regard the kernel $P(n|m)$ as a function 
597: of $n-m$ we use the notation $P(r)$. 
598: But sometimes it is convenient to regard $P(n|m)$ 
599: as an energy distribution $P_{\tbox{E}}(\epsilon)$.  
600: Due to the change of variables (\ref{e_7}) we have 
601: the following relation:
602: %
603: \begin{eqnarray} \label{e_8}
604: P(r) \ \ = \ \ 
605: \frac{1}{g(E)} \ P_{\tbox{E}}\left( 
606: \frac{1}{g(E)} r - F(x)\delta x \right)
607: \end{eqnarray}
608: %
609: The energy distribution $P_{\tbox{E}}(\epsilon)$ 
610: can be formally defined as follows:
611: %
612: \begin{eqnarray} \label{e_9}
613: P_{\tbox{E}}(\epsilon) \ = \ 
614: \sum_n P(n|m) \ \delta(\epsilon-(E_n(x){-}E_m(x_0)))
615: \end{eqnarray}
616: % 
617: In the classical limit the summation over 
618: $n$ should be interpreted as a $dn$ integral.
619: For $P(n|m)$ in the above expression 
620: we can substitute the definition Eq.(\ref{e3}) 
621: with $\rho_n$ and $\rho_m$ approximated 
622: as in Eq.(\ref{e_mc}). A straightforward manipulation 
623: leads to the result:
624: %
625: \begin{eqnarray} \nonumber
626: P_{\tbox{E}}(\epsilon) \ &=& \ 
627: \langle \delta(\epsilon - 
628: ({\cal H}(Q,P;x)-{\cal H}(Q,P;x_0))) \rangle
629: \\  \nonumber
630: \ & = & \
631: \langle \delta(\epsilon + \delta x {\cal F}(t) ) \rangle
632: \  =  \ 
633: \frac{1}{\delta x} 
634: P_{\tbox{F}}\left(-\frac{1}{\delta x}\epsilon\right)
635: \end{eqnarray}
636: %  
637: Together with (\ref{e_4}) and (\ref{e_8}), 
638: we get Eq.(\ref{e4}) along with the implied 
639: special result (\ref{e_6}). 
640: 
641: 
642: 
643: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
644: \section{Numerical determination of LDOS profiles}
645:  
646:  
647: Given $\delta x$ we can determine numerically 
648: the LDOS profile $P(r)$. Representative profiles 
649: are displayed in Fig.4.  For the purpose 
650: of further discussion we introduce  the following definitions:
651: %
652: \begin{itemize}
653: \item
654: The classical LDOS profile $P_{\tbox{cl}}(r)$   
655: \item
656: The quantum mechanical LDOS profile $P(r)$
657: \item
658: The effective WBRM LDOS profile $P_{\tbox{RMT}}(r)$
659: \item
660: The first-order perturbative profile $P_{\tbox{prt}}(r)$ 
661: \end{itemize} 
662: %
663: We have already discussed the classical LDOS profile. 
664: Below we explain how we numerically determine the 
665: quantum mechanical LDOS profiles $P(r)$ and $P_{\tbox{RMT}}(r)$,  
666: and we also define the profile $P_{\tbox{prt}}(r)$.  
667: 
668: 
669: The numerical procedure for finding $P(r)$ 
670: is straightforward. For a given $\delta x$ we  
671: have to diagonalize the matrix~(\ref{e1}).  
672: The columns of the diagonalization matrix $\mbf{T}_{mn}$ are 
673: the eigenstates of the Hamiltonian, and 
674: by definition we have $P(n|m)= |\mbf{T}_{mn}|^2$. 
675: Then $P(r)$ is computed by averaging over 
676: roughly 300 reference states that are located 
677: within the classically-small energy window $2.8<E<3.1$.    
678: Fig.4  displays typical profiles. 
679: 
680: 
681: The effective WBRM Hamiltonian is obtained by randomizing 
682: the signs of the off-diagonal elements in the $\mbf{B}$ matrix. 
683: For the effective WBRM Hamiltonian exactly the same 
684: procedure (as for $P(r)$) is applied leading to $P_{\tbox{RMT}}(r)$. 
685: 
686: 
687: 
688: 
689: In order to analyze the structure of either $P(r)$ or 
690: $P_{\tbox{RMT}}(r)$ we have defined the first-order 
691: perturbative profile as follows:
692: %
693: \begin{eqnarray} \label{e5} 
694: P_{\tbox{prt}}(r) \ = \  
695: \frac{\delta x^2 \ |\mbf{B}_{nm}|^2}{\Gamma^2 + (E_n{-}E_m)^2}
696: \end{eqnarray}
697: %
698: It is implicit in this definition that 
699: $(E_n{-}E_m)$ and $|\mbf{B}_{nm}|^2$ 
700: should be regarded as a function of $r$.  
701: The $r=0$ value of the band-profile should be  
702: re-defined by an interpolation. 
703: The parameter $\Gamma \equiv b_0\Delta$ is determined 
704: (for a given $\delta x$) such that the $P_{\tbox{prt}}(r)$ 
705: has a unit normalization. Note that Wigner's Lorentzian 
706: would be obtained if the band profile were flat. 
707: 
708: 
709: 
710: 
711: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
712: \section{Region analysis for the quantal LDOS}
713: 
714: 
715: By comparing $P(r)$ to $P_{\tbox{prt}}(r)$ as in Fig.4, 
716: we can determine$^\ddag$ 
717: the range $b_1\mbox{[left]} < r < b_1\mbox{[right]}$ 
718: where $P_{\tbox{prt}}(r)$ is a reasonable approximation for $P(r)$. 
719: Loosely speaking (avoiding the distinction between 
720: the `left' and the `right' sides of the profile) we shall say 
721: that $P_{\tbox{prt}}(r)$ is a reasonable approximation for $|r| < b_1$. 
722: The core is defined as the region $|r|<b_0$. 
723: The FOTRs are $b_0<|r|<b_1$. The far-tail regions are $|r|>b_1$. 
724: 
725: 
726: \ \\
727: %%%%%%%%%%%%%%%%%%%%%%%%%%%
728: \epsfysize=5.8in
729: \epsffile{profiles.eps} \\
730: \noindent
731: {\footnotesize {\bf FIG.4}:  
732: The quantal profile $P(r)$ is compared with 
733: $P_{\tbox{prt}}(r)$ and with $P_{\tbox{RMT}}(r)$. 
734: We are using here the $\hbar=0.015$ output. 
735: The insets are normal plots while 
736: the main figures are semilog plots.
737: In the lower plot ($\delta x = 0.2123$) 
738: the classical LDOS profile $P_{\tbox{cl}}(r)$ 
739: is represented by heavy dashed line. } \\
740: %%%%%%%%%%%%%%%%%%%%%%%%%%%
741: 
742: \ \\
743: 
744: 
745: The results of this region analysis are summarized by Fig.5.
746: In the following sections we are going to present a detailed 
747: discussion of this analysis.   
748: For the convenience of the reader we summarize: 
749: %
750: \begin{itemize}
751: \item
752: $b_0 \ = \ $ border of the core region
753: \item
754: $b_1 \ = \ $ border of the first order tail region (FOTR)
755: \end{itemize}
756: %
757: Having $b_0 \ll 1$ implies a standard perturbative 
758: structure. Having $1 \ll b_0 \ll b_1$ implies that 
759: we have a well developed core-tail structure.  
760: Having $b_0 \sim b_1$ implies a purely 
761: non-perturbative structure. In the latter case 
762: the distinction between core and tail regions 
763: become meaningless.
764: 
765: 
766: 
767: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
768: \section{The standard perturbative regime}
769: 
770: The standard perturbative regime  
771: $\delta x \ll \delta x_c^{\tbox{qm}}$ 
772: is defined by the requirement $b_0(\delta x) \ll 1$.
773: This condition implies that $P(n|m) \sim \delta_{nm}$. 
774: For numerical purpose it is convenient 
775: to define $\delta x_c^{\tbox{qm}}$ as the 
776: value of $\delta x$ for which $P(r=0) \approx 0.5$.
777: The theoretical considerations of \cite{frc} 
778: imply that $\delta x_c^{\tbox{qm}} \propto \hbar^{(1{+}d)/2}$. 
779: The prefactor is a classical quantity whose 
780: precise value depends on the operational 
781: definition of $\delta x_c^{\tbox{qm}}$. 
782: With the operational definition given above 
783: we have extracted the result  
784: $\delta x_c^{\tbox{qm}} \approx 3.8*\hbar^{3/2}$.
785: 
786:   
787: In the standard perturbative regime we can write  
788: schematically 
789: %
790: \begin{eqnarray}
791: P(n|m) \ \ \approx \ \ \delta_{nm} + \mbox{Tail}
792: \end{eqnarray}
793: %
794: The `Tail' is composed of FOTRs and far-tail 
795: regions. The former are given by Eq.(\ref{e5}), 
796: while the latter are determined by higher orders 
797: of perturbation theory. Note that for the {\em standard} 
798: WBRM we have by construction $b_1 \equiv b$, and more generally  
799: $n$-th order perturbation theory becomes essential 
800: for $(n{-}1) \times b < |r| < n \times b$. 
801: In case of our physical Hamiltonian, as well as 
802: for the associated {\em effective} WBRM model, the 
803: boundary $b_1$ is $\delta x$ dependent. 
804: 
805: 
806: By comparing $P(r)$ with $P_{\tbox{RMT}}(r)$ we can see 
807: that RMT cannot be trusted for the analysis of the far-tails, 
808: because system-specific interference phenomena becomes 
809: important there. Namely, the RMT profile $P_{\tbox{RMT}}(r)$  
810: is almost indistinguishable from $P_{\tbox{prt}}(r)$. 
811: In contrast to that, the far-tails of $P(r)$ are dominated 
812: by either destructive interference (left tail), 
813: or by constructive interference (right tail).   
814: 
815: 
816: 
817: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
818: \section{The core-tail regime}
819: 
820: The core-tail regime  
821: $\delta x_c^{\tbox{qm}} \ll \delta x \ll \delta x_{\tbox{prt}}$ 
822: is defined by the requirement $1 \ll  b_0 \ll b_1$. 
823: The theoretical considerations of \cite{frc} 
824: imply that $\delta x_{\tbox{prt}} \propto \hbar$. 
825: The prefactor is a classical quantity whose 
826: precise value depends on the operational 
827: definition of $\delta x_{\tbox{prt}}$. 
828: In our numerical analysis we have defined $\delta x_{\tbox{prt}}$ 
829: as the $\delta x$ for which the contribution of the FOTRs 
830: to the variance becomes less than $80\%$. 
831: With this operational definition 
832: we have extracted (using the lower subplot of Fig.5) 
833: the result $\delta x_{\tbox{prt}} \approx 5.3*\hbar$.
834: 
835: 
836: In the core-tail regime we can write schematically
837: %
838: \begin{eqnarray}
839: P(n|m) \ \ \approx \ \ \mbox{Core} + \mbox{Tail}
840: \end{eqnarray}
841: %
842: Disregarding the far-tail regions, the large-scale
843: behavior of $P(r)$ can be approximated by that 
844: of $P_{\tbox{prt}}(r)$.  As in the standard 
845: perturbative regime one observes that the far-tails are 
846: dominated by either destructive interference (left tail), 
847: or by constructive interference (right tail).  
848: 
849: 
850: 
851: The core is a non-perturbative region. It means, 
852: that unlike the far-tail, it cannot be obtained 
853: from any finite-order perturbation theory. 
854: %
855: Once the core appears, the validity of first-order 
856: perturbation theory becomes a non-trivial matter.  
857: In \cite{frc} a non-rigorous argument 
858: is suggested in order to support the claim 
859: that, disregarding smoothing effect, 
860: the local mixing of neighboring levels does not 
861: affect the growth of the tail. 
862: An important ingredient in this argumentation is 
863: the (self consistent) assumption that  
864: most of the probability is well-contained in 
865: the core region. Indeed the analysis which is 
866: presented in Fig.5 is in agreement with this 
867: assumption.   
868: 
869: 
870: The observation that the local mixing of 
871: neighboring levels does not affect the growth 
872: of the tail, implies that the tail grows 
873: as $\delta x^2$ and not like say $\delta x$. 
874: (The latter type of dependence is implied 
875: by an over-simplified argumentation). 
876: Having indeed $\delta x^2$ behavior is implied by 
877: observing that $P(r) \approx P_{\tbox{prt}}(r)$ 
878: for the FOTRs. 
879: 
880: 
881: Finally, it should be emphasized that the local mixing of levels on 
882: the small scale $b_0$ is not reflected by Eq.(\ref{e5}). 
883: In particular, one should not expect Eq.(\ref{e5})  
884: to be literally valid within the core region ($|r|<b_0$).
885: 
886: 
887: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
888: \section{The non-perturbative regime }
889: 
890: In the non-perturbative regime  
891: ($\delta x \gg \delta x_{\tbox{prt}}$) 
892: one may say that the core spills over the FOTRs  
893: and therefore $P(n|m)$ becomes purely non-perturbative. 
894: As an example for non-perturbative profile let 
895: us consider the lower plot of Fig.4, 
896: corresponding to $\delta x = 0.2123$.
897: We see that there is poor resemblance 
898: between $P(r)$ and $P_{\tbox{prt}}(r)$. 
899: The LDOS profile $P(r)$ no longer contains a predominant FOTRs.  
900: This claim can be quantified using the analysis 
901: in Fig.5. The lower subfigure there 
902: displays the FOTR contribution to the dispersion. 
903: For $\delta x > \delta x_{\tbox{prt}}$ 
904: the dispersion is no longer determined 
905: by the FOTR contribution.  
906: 
907: 
908: The complete disappearance of FOTRs is guaranteed 
909: only for $\delta x \gg \delta x_{\tbox{prt}}$. 
910: Evidently, for $\delta x \gg \delta x_{\tbox{prt}}$ 
911: the FOTRs must disappear, because $P(r)$ goes on expanding, 
912: while $P_{\tbox{prt}}(r)$ saturates. 
913: This is not captured by our numerics since 
914: for $\hbar=0.015$ we cannot satisfy the strong inequality 
915: $\delta x \gg \delta x_{\tbox{prt}}$, and have 
916: a classically small $\delta x$ at the same time. 
917: 
918: 
919: 
920: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
921: \section{The semiclassical regime}
922: 
923: Looking back at the lower plot of Fig.4, 
924: we see that detailed QCC with the classical profile 
925: (represented by heavy dashed line) starts to develop.  
926: The right far tail contains a component where $P(r)$ 
927: and $P_{\tbox{cl}}(r)$ are indistinguishable. 
928: This detailed QCC obviously does not hold for the RMT profile. 
929: 
930: Being in the non-perturbative regime does not imply 
931: detailed QCC \cite{crs,frc,vrn}. Detailed QCC means 
932: that $P(r)$ can be approximated by $P_{\tbox{cl}}(r)$. 
933: Having $\delta x \gg \delta x_{\tbox{prt}}$ is a necessary 
934: rather than sufficient condition for detailed QCC.
935: 
936: A sufficient condition for detailed QCC is 
937: $\delta x \gg \delta x_{\tbox{SC}}$. 
938: The parametric scale $\delta x_{\tbox{SC}}$  
939: is defined in \cite{frc}, and for our system we can 
940: obtain the (theoretical) rough estimate 
941: $\delta x_{\tbox{SC}}\approx 4*\hbar^{2/3}$. 
942: 
943: In our numerical study we could not make 
944: $\hbar$ small enough such that  
945: $\delta x_{\tbox{SC}} \ll \delta x_c^{\tbox{cl}}$.  
946: Therefore, the lower profile in Fig.4 
947: is neither reasonably approximated by 
948: $P_{\tbox{prt}}(r)$, nor by $P_{\tbox{cl}}(r)$. 
949: However, we have verified (by comparing the $\hbar=0.03$
950: output to the $\hbar=0.015$ output) that 
951: detailed QCC between $P(r)$ and $P_{\tbox{cl}}(r)$
952: is easily improved by making $\hbar$ smaller. 
953: Comparing $P(r)$ to $P_{\tbox{cl}}(r)$ on the one 
954: hand, and $P_{\tbox{rmt}}(r)$ to $P_{\tbox{cl}}(r)$ 
955: on the other hand, leaves no doubt regarding the 
956: manifestation of underlying classical structures.  
957: 
958: 
959: Using a phase-space picture \cite{frc,vrn} it is evident 
960: that larger $\delta x$ leads to better QCC. 
961: The WBRM model does not have a classical limit, 
962: and one finds a quite different scenario \cite{casati}. 
963: For large enough $\delta x$  the eigenstates of 
964: Eq.(\ref{e1}) become Anderson localized. 
965: This localization shows up in the ASOE
966: {\em provided} the eigenstates are properly centered 
967: prior to averaging. In the (non-averaged) LDOS 
968: localization manifests itself as sparsity, 
969: and therefore the various moments of the LDOS profile 
970: are not affected. This latter remark should be kept 
971: in mind while reading the next section.
972: 
973: 
974: 
975: 
976: \ \\ \ \\
977: %%%%%%%%%%%%%%%%%%%%%%%%%%%
978: \epsfysize=2.6in
979: \epsffile{lds_anlys.eps} \\
980: \noindent
981: {\footnotesize {\bf FIG.5}
982: The results of region analysis. 
983: The common horizontal axis is $\delta x$. 
984: The {\em upper subfigure} presents the $r$ boundaries  
985: as a function of $\delta x$.  
986: The dotted lines $\pm b_0$ define the core region ($|r|<b_0$). 
987: The solid lines define the $r$ region in which $50\%$ 
988: of the probability is concentrate. 
989: The dashed lines are $b_1\mbox{[left]}$ and $b_1\mbox{[right]}$. 
990: The FOTRs are the regions where $b_0<|r|<b_1$. 
991: The light solid lines and the light dashed lines 
992: are for the effective WBRM model.
993: %
994: The {\em lower subfigure} displays the 
995: dependence of $\delta E_{\tbox{cl}}$ and $\delta E_{\tbox{qm}}$ 
996: and $\delta E_{\tbox{prt}}$ on $\delta x$. 
997: The quantal and the classical results 
998: are almost indistinguishable, whereas $\delta E_{\tbox{prt}}$ 
999: approaches saturation. 
1000: The contribution of the FOTRs to $\delta E_{\tbox{qm}}$ 
1001: is also displayed. } \\
1002: %%%%%%%%%%%%%%%%%%%%%%%%%%%
1003: 
1004: 
1005: 
1006: 
1007:      
1008: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1009: \section{Restricted QCC}
1010: 
1011: 
1012: It is important to distinguish between detailed QCC 
1013: and restricted QCC. Let us denote the dispersion of 
1014: the quantal LDOS profile by $\delta E_{\tbox{qm}}$, 
1015: The corresponding classical quantity is given 
1016: by Eq.(\ref{e_6}). The two types of QCC are defined 
1017: as follows: 
1018: %
1019: \begin{itemize}
1020: \item
1021: Detailed QCC means \ \ \ $P(r) \approx P_{\tbox{cl}}(r)$  
1022: \item
1023: Restricted QCC means \ \ $\delta E_{\tbox{qm}} \approx \delta E_{\tbox{cl}}$
1024: \end{itemize} 
1025: %
1026: Obviously restricted QCC is a trivial consequence 
1027: of detailed QCC, but the converse is not true. 
1028: It turns out that restricted QCC is much more 
1029: robust than detailed QCC.  In Fig.5 we see that 
1030: the dispersion $\delta E_{\tbox{qm}}$ of either 
1031: $P(r)$ or $P_{\tbox{RMT}}(r)$ is almost indistinguishable
1032: from $\delta E_{\tbox{cl}}$.  This is quite remarkable 
1033: becuase the corresponding LDOS profiles 
1034: (quantal versus classical) are very different!  
1035: 
1036: 
1037: 
1038: It is important to realize that restricted QCC 
1039: is implied by first order perturbation theory. 
1040: If we use Eq.(\ref{e5}) and take into accound 
1041: the FOTR dominace which is implied by 
1042: $\delta x \ll \delta x_{\tbox{prt}}$, then we get simply
1043: %
1044: \begin{eqnarray} \label{e_pr}
1045: \delta E_{\tbox{qm}}  =   
1046: \sum_n P(n|m) \ (E_n{-}E_m)^2  =
1047: \delta x^2 \sum_n'|\mbf{B}_{nm}|^2 
1048: \end{eqnarray}
1049: %
1050: where prime indicates omission of the $n=m$ term.   
1051: Using Eq.(\ref{e_5}) one realizes that this 
1052: result is in complete agreement with Eq.(\ref{e_6}).
1053: In contrast to that higher moments of the perturbative 
1054: profile are vanishingly small compared with the corresponding 
1055: classical result. The latter fact is just a reflection 
1056: of the absence of detailed QCC.  
1057: 
1058: 
1059: 
1060: One may wonder what happens with Eq.(\ref{e_pr})   
1061: if we try to do a better work, taking into account the core width, 
1062: as well as higher order far-tails contributions. 
1063: One may think that Eq.(\ref{e_pr}) is only 
1064: the lowest order approximation, which would imply 
1065: that restricted QCC should become worse as $\delta x$ grows. 
1066: However, the latter speculation turns out to be wrong.
1067: 
1068: 
1069: 
1070: We already saw that restricted QCC is implied 
1071: on the one hand (for small $\delta x$) 
1072: by first-order perturbation theory, and on the other 
1073: hand (for large $\delta x$) by detailed QCC. 
1074: Now we would like to argue that 
1075: restricted QCC holds in general. 
1076: It simply follows from the observation 
1077: that $\delta E_{\tbox{qm}}$ is determined 
1078: just by the band profile.
1079: %
1080: The prove is very simple \cite{casati}. 
1081: The variance of $P(n|m)$ is determined 
1082: by the first two moments of the Hamiltonian 
1083: in the unperturbed basis. Namely
1084: %
1085: \begin{eqnarray} \nonumber 
1086: \delta E_{\tbox{qm}}^2  =  
1087: \langle m | {\cal H}^2 |m \rangle - \langle m| {\cal H} |m \rangle^2 
1088: \\ \nonumber    
1089: = \delta x^2
1090: (\langle m | \mbf{B}^2 |m \rangle - \langle m| \mbf{B} |m \rangle^2)
1091: \end{eqnarray}
1092: % 
1093: Thus, we get the same result as in first order 
1094: perturbation theory without invoking any special 
1095: assumptions regarding the nature of the profile. 
1096: Having  $\delta E_{\tbox{qm}}$ that is determined 
1097: only by the band profile, is the reason for detailed QCC, 
1098: and is also the reason why restricted QCC is not sensitive 
1099: to the RMT assumption.   
1100: 
1101: \ \\
1102: 
1103: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1104: 
1105: We thank Felix Izrailev for suggesting to study the 
1106: model (\ref{e2}). We also thank ITAMP for their support. 
1107: 
1108: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1109: 
1110: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1111: \begin{thebibliography}{99}
1112: 
1113: \bibitem[\dag]{note1}
1114: To be more specific, one should be aware that there 
1115: is an hierarchy of challenges where the applicability 
1116: of the RMT approach should be tested. Namely: 
1117: The study of spectral statistics; 
1118: The study of eigenstates; 
1119: The study of quantum dynamics. 
1120: In a previous work \cite{wbr} we have argued that  
1121: the RMT approach does not generally apply to the 
1122: study of wavepacket dynamics, since it leads to 
1123: a contradiction with the QCC principle.     
1124: On the other hand, it is well known that spectral statistics 
1125: is much more robust. In most of the RMT literature 
1126: (including the later works by Wigner himself), 
1127: it is assumed that for the purpose of 
1128: `quantum chaos' studies one can consider 
1129: full (rather than banded) matrices, and   
1130: the first term in Eq.(\ref{e1}) is generally neglected. 
1131: In spite of these enormous simplifications, 
1132: it turns out that the so-called Gaussian invariant ensembles
1133: (GOE,GUE) provide a valid description of some major 
1134: spectral properties. 
1135: 
1136: 
1137: \bibitem[\ddag]{note2}
1138: The determination of $b_1$ has been done using the following 
1139: numerical procedure. We define relative error function 
1140: $\mbox{RE}(r) = (P-P_{\tbox{prt}})/(P+P_{\tbox{prt}})$ 
1141: and then cumulative error 
1142: function $\mbox{CRE}(r) = |\sum_0^r \mbox{RE}(r')|$.   
1143: Note that by this definition `positive' relative error can 
1144: be compensated by 'negative' relative error. 
1145: As we go away from $r=0$, the function $\mbox{CRE}(r)$ fluctuates, 
1146: and later shoots up. The regime $|r|<b_1$ 
1147: has been determined  by the condition 
1148: $\mbox{CRE}(r)<\mbox{Threshold}$. 
1149: The threshold has been determined using adaptively procedure.  
1150: 
1151: 
1152: \bibitem{mario}
1153: M. Feingold and A. Peres, Phys. Rev. A {\bf 34} 591, (1986).
1154: M. Feingold, D. Leitner, M. Wilkinson, Phys. Rev. Lett. {\bf 66}, 986 (1991). 
1155: M. Wilkinson, M. Feingold, D. Leitner, J. Phys. A {\bf 24}, 175 (1991). 
1156: M. Feingold, A. Gioletta, F. M. Izrailev, L. Molinari, Phys. Rev. Lett. 
1157: {\bf 70}, 2936 (1993).
1158: 
1159: \bibitem{wigner}
1160: E. Wigner, Ann. Math {\bf 62} 548 (1955); {\bf 65} 203 (1957).
1161: 
1162: \bibitem{casati}
1163: G. Casati, B.V. Chirikov, I. Guarneri, F.M. Izrailev, 
1164: Phys. Rev. E {\bf 48}, R1613 (1993);
1165: \ Phys. Lett. A {\bf 223}, 430 (1996).
1166: V.V. Flambaum, A.A. Gribakina, G.F. Gribakin and M.G. Kozlov, 
1167: Phys. Rev. A {\bf 50} 267 (1994).
1168: 
1169: \bibitem{fyodo}
1170: Y.V. Fyodorov, O.A. Chubykalo, F.M. Izrailev and G. Casati, 
1171: Phys. Rev. Lett. {\bf 76}, 1603 (1996).
1172: 
1173: \bibitem{frc}
1174: D. Cohen, Ann. Phys. {\bf 283}, 175-231 (2000).
1175:  
1176: \bibitem{vrn} D. Cohen, lecture notes, 
1177: {\em in} Proceedings of the International 
1178: School of Physics `Enrico Fermi' Course CXLIII 
1179: ``New Directions in Quantum Chaos'', 
1180: Edited by G. Casati, I. Guarneri and U. Smilansky, 
1181: IOS Press, Amsterdam, 2000. 
1182: 
1183: \bibitem{wls} 
1184: D. Cohen and E.J. Heller,  
1185: Phys. Rev. Lett. {\bf 84}, 2841 (2000). 
1186: 
1187: \bibitem{prm} 
1188: In a later work regarding "Parametric Evolution for a Deformed Cavity",    
1189: (D. Cohen, A. Barnett and E.J. Heller, nlin.CD/0008040), 
1190: further study of the various parametric regimes is introduced. 
1191: In our present paper the emphasis is on generic features of 
1192: parametric evolution, while in nlin.CD/0008040 the emphasize is 
1193: on non-generic features which have been discussed in \cite{wls}.     
1194:  
1195: \bibitem{felix2}
1196: F. Borgonovi, I. Guarneri and F.M. Izrailev, 
1197: Phys. Rev. E {\bf 57}, 5291 (1998). 
1198: L. Benet, F.M. Izrailev, T.H. Seligman and A. Suarez-Moreno, 
1199: chao-dyn/9912035. 
1200: 
1201: \bibitem{crs}
1202: D. Cohen, Phys. Rev. Lett. {\bf 82}, 4951 (1999) 
1203: 
1204: \bibitem{wbr} 
1205: D. Cohen, F.M. Izrailev and T. Kottos,  
1206: Phys. Rev. Lett. {\bf 84} 2052 (2000). 
1207: 
1208: \end{thebibliography}
1209: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1210: \end{multicols}
1211: \end{document}
1212: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1213: