1: %\documentstyle[preprint,revtex]{aps}
2: %\documentstyle[preprint,eqsecnum,psfig]{aps}
3: \documentstyle[multicol,aps,psfig]{revtex}
4: \setlength{\topmargin}{-0.25in}
5: \begin{document}
6:
7: \draft
8:
9: \preprint{\vbox{\hbox{DRAFT of U. of Iowa preprint 2000-2503}}}
10:
11: \title{Small Numerators Canceling Small Denominators: Is Dyson's
12: Hierarchical Model Solvable? }
13:
14: \author{Y. Meurice \\
15: {\it Department of Physics and Astronomy, The University of Iowa,
16: Iowa City, Iowa 52242, USA}}
17:
18: \maketitle
19:
20: \begin{abstract}
21:
22: We present an analytical method to solve
23: Dyson's hierarchical model, involving
24: the scaling variables
25: near the high-temperature fixed point. The procedure
26: seems plagued by the presence of small denominators
27: as in perturbative expansions near integrable systems in Hamiltonian mechanics.
28: However, in all cases considered,
29: a zero denominator always comes with a zero numerator.
30: We conjecture that these cancellations occur in general,
31: suggesting that the model has remarkable features reminiscent
32: of the integrable systems.
33: \end{abstract}
34: \pacs{PACS: 05.50.+q, 11.10.Hi, 64.60.Ak, 75.40.Cx}
35: %\narrowtext
36: \begin{multicols}{2}\global\columnwidth20.5pc
37: \multicolsep=8pt plus 4pt minus 3pt
38:
39: In many physical problems involving nonlinear flows, a common
40: strategy consists in constructing a system of coordinates where
41: the flow becomes linear. The action-angle variables in Hamiltonian mechanics
42: provide well-know examples of such a procedure. Whenever the angle
43: variables can be constructed, they evolve linearly with time and
44: expressing the original variables in terms of the new ones
45: solves completely the original problem.
46: %Can every problem of Hamiltonian mechanics be solved this way?
47: %The answer to this question is definitely no.
48: The problems for which
49: well-defined angle-action variables can be constructed
50: (e.g. Kepler's problem or the free rigid body) are
51: very distinguished and
52: called integrable systems.
53: %Well-known examples are Kepler's problem,
54: %the free double pendulum or
55: %the free rigid body.
56: A large number of numerical experiment has lead us to
57: believe that in a generic way,
58: small perturbations destroy integrability.
59: This point of view was was first inferred
60: by H. Poincar\'e who pointed out the existence of small denominators
61: in the canonical transformation designed to eliminate the angle
62: dependence of a perturbed Hamiltonian.
63: %Later, the small denominators were used
64: %in numerical experiment as a guide to understand the destruction
65: %of inv
66: %More recently,
67: %the ``destructive'' effect of small perturbations in selected
68: %regions of phase space has been illustrated by a large number
69: %of numerical experiments. A well-known example is the chaotic behavior
70: %of the double pendulum when subjected to
71: %a gravitational interactions.
72: %This is telling us that integrable systems are quite ``isolated'' and
73: %distinguished in the context of Hamiltonian mechanics.
74:
75: In this letter, we discuss the question of small denominators for
76: renormalization group (RG) flows.
77: The variables which play the role of angle variables are the
78: scaling variables introduced by Wegner \cite{wegner72}. Near a fixed point,
79: the RG flows can be linearized. The problem of expressing the physical
80: quantities in terms of
81: variables which transform as in the linear approximation
82: when the non-linear
83: terms are taken into account, is analogous to removing the
84: angle dependence of a perturbed Hamiltonian. If the task
85: can be carried through, one obtains analytical expressions
86: for the RG flows.
87: However, as we will show, small denominators appear.
88: Does this mean that, as in Hamiltonian mechanics,
89: in generic situations the construction
90: will fail?
91:
92: Surprisingly, we found in a numerical calculation performed with
93: Dyson's
94: hierarchical model \cite{dyson69,baker72},
95: that zero denominators were systematically
96: canceled by zero numerators. These remarkable cancellations
97: suggest that either the model considered
98: is as distinguished
99: as the integrable systems of classical mechanics or
100: that there exists a general mechanism that allows us to
101: circumvent the small denominator problem for RG flows.
102: The example of the two-dimensional Ising model shows the
103: importance of having a non-trivial model which can
104: solved in closed form. The results presented below indicate
105: that Dyson's hierarchical can be solved analytically.
106:
107: During the last decades, the RG method has been successfully applied
108: to many important problems in field theory and statistical mechanics:
109: the critical behavior of ferromagnets and superconductors,
110: the confinement of quarks or the generation of mass for the W and Z
111: bosons.
112: However, its practical implementation
113: is still a formidable enterprise.
114: Typically, expansions are often available near fixed points,
115: but not to orders large enough to allow one to extrapolate between
116: fixed points. Unfortunately, the calculation of physical quantities
117: (e. g. the magnetic susceptibility) beyond an
118: order of magnitude estimation, requires a calculation of the
119: flows in crossover regions.
120: One has then to rely to Monte Carlo simulations to achieve this goal.
121: One also needs to select a small set of interactions which closes
122: reasonably well under RG transformation near both fixed points.
123: Interesting examples of such lattice Monte Carlo calculations are given in
124: Refs. \cite{gonzalez87} for scalar theory
125: and \cite{taro00} for gauge theories.
126:
127: In order to get analytical results, further approximations are
128: needed. One possibility consists in using hierarchical
129: approximations such as the one derived by Wilson in Ref. \cite{wilson71b}
130: and resulted into the ``approximate recursion formula''.
131: In this approximation, only the local interactions get renormalized
132: and the flow can be calculated from a simple integral formula.
133: Retracing Wilson's construction from the beginning, one can restore
134: the other renormalizations perturbatively. In order to perform this task,
135: one would like to have a closed form solution in the hierarchical
136: approximation. In order to achieve this goal, we have proposed \cite{scaling}
137: to use the Fourier representation of the integral formula and to
138: construct the scaling variables in this basis. In this process, we
139: identified the existence of small denominators possibly
140: ruining the whole approach. This
141: problem can be avoided in special circumstances, for instance
142: for flows starting exactly along the unstable direction
143: of a non-trivial fixed point. With this restriction, we found \cite{scaling}
144: expansions with overlapping domains of convergence in the crossover region.
145:
146: In the following, we discuss the problem of small denominators
147: in the construction of the scaling variables near the
148: high-temperature (HT) fixed point of Dyson's hierarchical model.
149: The treatment of the this model is
150: mathematically similar to the one of the approximate recursion formula.
151: However there exists a large literature on
152: Dyson's model \cite{baker77,collet78,bleher75}, the non-trivial
153: fixed point is known very precisely \cite{koch95} and the HT
154: expansion studied to a very large order \cite{osc1}.
155:
156: For a description of this model as a spin model, we refer to
157: Ref. \cite{finite,hyper}, while the details of
158: the derivation of the RG flows as expressed
159: below can be found in \cite{scaling}. To make a long story short, all
160: the information regarding the local interactions
161: after $n$ RG transformations is encoded in a function
162: $R_n(k)=1+a_{n,1}k^2+a_{n,2}k^4+\dots$. The logarithm of this function
163: generates the connected zero-momentum Green's functions at finite volume.
164: The recursion formula reads
165: \begin{equation}
166: R_{n+1}(k) = C_{n+1} \exp \left[ -{1\over 2}
167: {{\partial ^2} \over
168: {\partial k ^2}} \right]\left[R_{n} \left({\sqrt{c}k\over 2} \right) \right]^2 \ .
169: \label{eq:rec}\end{equation}
170: We fix the normalization constant $C_{n}$ so that $R_{n}(0) = 1$.
171: We use the parametrization $c=2^{1-2/D}$ which implies that a free
172: massless field scales in the same way as in a
173: usual $D$-dimensional theory.
174: In this formulation, the temperature dependence has been absorbed in the
175: initial $R_0(k)$.
176: For an Ising measure, $R_{0}(k) = \cos(\sqrt{\beta}k)$, while in general,
177: we have to numerically integrate the Fourier transform of
178: the local measure to determine the coefficients of
179: $R_{0}(k)$ expanded in terms of $k$. In general, $a_{n,l}$ is of order
180: $\beta^l$ in the HT expansion.
181:
182: In the HT phase, polynomial truncations of order $l_{max}$
183: in $k^2$ provide rapidly converging approximations \cite{finite}.
184: The RG flows can be expressed in terms of the quadratic map
185: \begin{equation}
186: a_{n+1, l} = \frac{u_{n,l}}{u_{n,0}} \ ,
187: \label{eq:aofu}
188: \end{equation}
189: with
190: \begin{equation}
191: u_{n,\sigma} = \Gamma_{\sigma}^{ \mu \nu} a_{n,\mu} a_{n,\nu} \ ,
192: \end{equation}
193: and
194: \begin{equation}
195: \Gamma_{\sigma}^{ \mu \nu}
196: = (c/4)^{\mu+\nu}\
197: \frac{(-1/2)^{\mu + \nu - \sigma}(2(\mu+\nu))!}{
198: (\mu+\nu-\sigma)!(2 \sigma)!} \ ,
199: \label{eq:struct}
200: \end{equation}
201: for $\mu+\nu \geq\sigma$ and zero otherwise.
202: We use ``relativistic'' notations.
203: Repeated indices mean summation.
204: The greek indices $\mu$ and $\nu$
205: go from $0$ to $l_{max}$, while latin indices $i$, $j$ go from 1 to
206: $l_{max}$.
207:
208: The diagonalization of the linear RG transformation near the HT fixed
209: point is quite simple because it is of the upper triangular form.
210: From Eq. (\ref{eq:struct}),
211: one finds the spectrum
212: \begin{equation}
213: \lambda_{(r)}=2(c/4)^{r} \ .
214: \label{eq:hteigenv}
215: \end{equation}
216: in agreement with Ref. \cite{collet78}.
217: Using the matrix of right eigenvectors, ${\cal M}_{l}^{i} \psi^r_i = \lambda_{(r)} \psi^r_l$,
218: we introduce new coordinates such that
219: $a_{n,l} = \psi^r_l h_{n,r}$. This diagonalizes the linear RG transformation.
220: Note that the form of the eigenvectors guarantees that
221: $h_{n,l}$ is also of order $\beta^l$.
222: More details are given Ref. \cite{scaling}, where all the quantities
223: related to the HT fixed point are dressed with a ``tilde'' omitted here .
224: In summary, the RG flows in the new coordinates can be written as
225: \begin{equation}
226: h_{n+1,l} = \frac{\lambda_{(l)} h_{n,l}
227: + \Delta_{l}^{ p q} h_{n,p} h_{n,q} }
228: {1 + \Lambda^{p} h_{n,p}
229: + \Delta_{0}^{ p q} h_{n,p} h_{n,q}}\ ,
230: \label{eq:hrules}
231: \end{equation}
232: with coefficients calculable from Eq. (\ref{eq:struct}).
233:
234: We now express the $h_{n,l}$ in terms of
235: the scaling variables $y_{n,1},\ldots,
236: y_{n,l_{max}}$ which
237: transform under a RG transformation as $y_{n+1,i} = \lambda_{i} y_{n,i}$.
238: If we can construct functions $h_l$ and $y_l$ such that
239: $h_{n,l}=h_l({\mathbf{y}}_{n})$ and $y_{n,l}=y_l({\mathbf{h}}_{n})$,
240: then we get a complete analytical expression of $h_{n,l}$ (which contains
241: all the thermodynamical quantities)
242: in terms
243: of $h_{0,l}$ (which depends on the initial energy density):
244: \begin{equation}
245: h_{n,l}=h_l(\lambda_1^ny_1({\mathbf h}_0),\lambda_2^ny_2({\mathbf h}_0),
246: \dots)\ .
247: \label{eq:solve}
248: \end{equation}
249: The feasibility of this approach is demonstrated for a
250: one-dimensional example in
251: Ref. \cite{dual}.
252:
253: We now discuss the construction of the $h_l$. We use the expansion
254: \begin{equation}
255: h_{l} = y_l+\sum_{i_{1},i_{2},\ldots} s_{l,i_{1} i_{2} \ldots} y_{1}^{i_{1}}
256: y_{2}^{i_{2}} \ldots \ ,
257: \end{equation}
258: where the sums over the $i$'s run from $0$ to infinity in each variable
259: with at least two non-zero indices.
260: In the following, we use the notation ${\mathbf{i}}$ for $(i_1,i_2,\dots)$.
261: Plugging the expansion into Eq. (\ref{eq:hrules}),
262: and requiring that the one step advance
263: is obtained by rescaling the scaling variables by their associated
264: eigenvalue, we obtain
265: \begin{equation}
266: s_{l, \mathbf{i}} = \frac{N_{l,\mathbf{i}}}
267: {\left(\prod_{m} \lambda_{(m)}^{i_{m}} - \lambda_{r}\right)} \ .
268: \label{eq:denom}
269: \end{equation}
270: with
271: \begin{eqnarray}
272: \label{eq:num}
273: N_{l, \mathbf{i}} =& \sum_{\mathbf{j}+\mathbf{k} = \mathbf{i}}
274: ( \Delta_{l}^{ p q} s_{p,\mathbf{j}}
275: s_{q,\mathbf{k}} - s_{l,\mathbf{j}} \prod_{m} \lambda_{(m)}^{j_{m}}
276: \Lambda^{p} s_{p, \mathbf{k}} )\\ \nonumber
277: &-\sum_{\mathbf{j}
278: +\mathbf{k}+\mathbf{r}=\mathbf{i}}
279: s_{l, \mathbf{j}} \prod_{m} \lambda_{(m)}^{j_{m}} \Delta_{0}^{ p q}
280: s_{p, \mathbf{k}} s_{q, \mathbf{r}} \ .
281: \end{eqnarray}
282: For a given set of indices $\mathbf{i}$, we
283: introduce the notation
284: \begin{equation}
285: {\cal{I}}_q ({\mathbf i})=\sum_m i_m m^q \ .
286: \label{eq:indices}
287: \end{equation}
288: One sees that ${\cal{I}}_0$ is the degree of the associated
289: monomial and ${\cal{I}}_1$
290: its order in the HT expansion (since $y_l$ is also of order $\beta ^l$).
291: Given that all the indices are positive and that at least one index
292: is not zero,
293: one can see that if ${\mathbf{j}}+{\mathbf{k}}={\mathbf{i}}$ then
294: ${\cal I}_q({\mathbf{j}})<{\cal I}_q({\mathbf{i}})$ and
295: ${\cal I}_q({\mathbf{k}})<{\cal I}_q({\mathbf{i}})$. Consequently,
296: Eq. (\ref{eq:num}) provides a solution order by order in ${\cal{I}}_0$ or
297: in ${\cal{I}}_1$ (since the r.h.s is always of lower order)
298: assuming that the small denominator problem can be avoided.
299:
300: Using the parametrization $c=2^{1-2/D}$,
301: the zero denominators in Eq. (\ref{eq:denom}) appear when
302: \begin{equation}
303: D-l(D+2)=D{\cal{I}}_0 -(D+2){\cal{I}}_1 \ .
304: \end{equation}
305: If $D$ and $D+2$ have no common factors, this equation has non-trivial
306: solutions for indices such that ${\cal{I}}_0=(2+D)q+1$, for $q$ a strictly
307: positive integer and $l$ such
308: that ${\cal{I}}_1=Dq+l$. If $D$ and $D+2$ have common factors, we can proceed
309: in the same way but with these common factors removed from both $D$ and $D+2$.
310:
311: We have investigated numerically the well-studied case $D=3$
312: with a polynomial truncation $l_{max}=25$.
313: For practical reasons, we have limited our study the case where
314: $h_l$ depends only on
315: the two leading variables $y_1$ and $y_2$. This restriction is self-consistent
316: since if we start for instance with $y_3=0$, the multiplicative
317: renormalization of the scaling variables implies that this condition
318: stays valid after $n$ iterations. Since $i_3=i_4=\dots=0$,
319: we will use notations such as $s_{l,i_1,i_2}$ or $h_l(y_1,y_2)$.
320: In addition, we have concentrated our attention to the large order
321: behavior in the leading variable $y_1$ and calculated the
322: first order corrections in the subleading variable $y_2$
323: If $i_2=0$, we have the following sets of small denominators:
324: $i_1=5q+1$ and $l=2q+1$. If $i_2=1$, we have $i_1=5q$ and $l=2q+2$.
325: We have calculated the numerators corresponding to these zero
326: denominators for $q=1$, 2, 3 and 4.
327:
328: We have checked our calculation of the $h_l$
329: with two different methods. First we have
330: considered a configuration
331: $h_{n,l}=h_l(y_1,y_2)$
332: for particular values of $y_1$ and $y_2$ and then calculated
333: the one step backward configuration
334: $h_{n-1,l}=h_l(\lambda_1^{-1}y_1,\lambda_2^{-1}y_2)$.
335: We then calculated the one step forward using these $h_{n-1,l}$
336: and the exact formula Eq. (\ref{eq:hrules}).
337: Comparison between the two $h_{n,l}$
338: for values of $y_1$ and $y_2$ varying between 0.1 and 0.001 and
339: for various truncations in the power of $y_1$ and $y_2$ considered,
340: showed errors scaling like the powers neglected with coefficients
341: compatible with the order of magnitude of the coefficients involved in
342: the expansion. Second, we used the expression $y_1(h_l)$ calculated in
343: Ref. \cite{scaling} up to order 11 in $\beta$
344: and checked $y_1(h_l(y_1,y_2))=y_1$ with errors
345: ranging between
346: $10^{-14}$ and $10^{-16}$ for the various coefficients of the
347: higher order terms.
348:
349: In Fig. 1 we show the absolute value of $N_{3,j,0}$. The calculations
350: have been performed with different arithmetic precisions.
351: We have used $Mathematica$ 4.0
352: with $\$MaxPrecision$ and $\$ MinPrecision$
353: both set to the same value $pr$.
354: We have considered the cases $pr$=20, 30 and 40.
355: The graphs shows $N_{3,6,0}$ is
356: more than twenty orders of magnitudes smaller than its peers when
357: $pr=20$ and then drops by ten orders of magnitude each time the precision
358: is increased by 10, while all the other values stay stable. This
359: is a strong evidence for $N_{3,6,0}=0$
360: \begin{figure}
361: \centerline{\psfig{figure=36.EPS,width=3in}}
362: \caption{$Log_{10}(|N_{3,j,0}|)$ versus $j$.}
363: \label{fig:36}
364: \end{figure}
365: In Figs. \ref{fig:all0} and \ref{fig:all2}, we show
366: $N_{l,i_1,0}$
367: and $N_{l,i_1,1}$ calculated with
368: $pr$=20.
369: One sees that $N_{2q+1,5q+1,0}$ and $N_{2q+2,5q,1}$ are more than
370: 20 orders of magnitude smaller than the naive interpolation.
371: We have checked in each of these cases that
372: the small value drops by ten order of magnitudes each time the
373: precision is increased just as in Fig. \ref{fig:36}.
374: \begin{figure}
375: \centerline{\psfig{figure=all0.EPS,width=3in}}
376: \caption{$Log_{10}(|N_{l,j,0}|)$ versus $j$ for $l=1,\dots 9$.}
377: \label{fig:all0}
378: \end{figure}
379: \begin{figure}
380: \centerline{\psfig{figure=all2.EPS,width=3in}}
381: \caption{$Log_{10}(|N_{l,j,1}|)$ versus $j$ for $l=1,\dots 10$.}
382: \label{fig:all2}
383: \end{figure}
384:
385: In summary, for
386: each zero denominator considered, we found a
387: zero numerator.
388: We thus conjecture that all numerators corresponding to zero denominators
389: are zero.
390: If this conjecture is correct, the $s_{l,{\bf i }}$ corresponding to
391: non-zero denominators are calculable from Eq. (\ref{eq:num})
392: and those corresponding to
393: zero denominators are
394: undetermined. The calculations
395: performed above have been done with these coefficients treated as undetermined.
396: The figures have been drawn with these coefficients set to zero.
397: This choice is not essential, we have considered other choices where
398: the undetermined coefficients have been set to values of
399: the same order as the coefficients above and below (in $i_1$) and reached
400: identical conclusions. This is a non-trivial statement. For instance,
401: the undetermined coefficient
402: $s_{3,6,0}$ appears explicitly in the numerator of the
403: equation for $s_{5,11,0}$, however it is multiplied by a very small
404: number which drops when $pr$ is increased.
405:
406: We have checked that the individual terms in the
407: zero numerators were
408: not zero.
409: The fact that we were able to obtain very precise cancellations
410: without any fine-tuning
411: suggests the existence of
412: closed form formulas or of a symmetry forbidding these terms. More
413: generally, the equally spaced spectrum (on a log scale), the number of terms
414: at a given level $m$ (the order in HT expansion)
415: equal to the number of partitions of $m$
416: \cite{scaling}
417: and the existence of ``nested'' ambiguities are somehow reminiscent
418: of string theory. A possible starting point to discover these
419: hypothetical symmetries would be to exploit the fact that since
420: for instance $y_3$ and $y_1^6$ transform the same way under
421: a RG transformation, there exist ambiguities in the
422: construction
423: of $h_l$ in terms of the scaling variables.
424:
425: We should say a few words about the small denominators
426: near other fixed points. The spectrum of the linearized RG transformation
427: near the non-trivial fixed point \cite{koch95}
428: can be calculated numerically \cite{gam3rapid}.
429: Investigating the small denominators
430: of the form $\lambda_l^k\simeq \lambda_m^r$ for the first twenty eigenvalues
431: and for powers not larger than 100. The best solution we found was
432: $\lambda_2^9\simeq\lambda_4$ with two parts in a thousand.
433: A more detailed study is necessary to decide if the rate of decay
434: of the coefficients is sufficient to take care of the smaller denominator
435: which will appear at larger orders.
436: On the other hand, the spectrum at the Gaussian fixed point \cite{collet78}
437: is $\lambda_j=2c^{-j}$. There are many zero denominators in integer
438: dimensions,
439: e.g., $\lambda_1=\lambda_2^2$ for $D=3$.
440: If the numerators are not zero, one
441: can ``repair'' \cite{wegner72} the situation by considering $n$-dependent
442: coefficients. This excludes a solution of the form of
443: Eq. (\ref{eq:solve}) but it
444: generates logarithmic corrections which are
445: necessary. These can even be observed in the large order
446: of the HT temperature expansion in Ref. \cite{ht4}.
447:
448: In conclusion, we have found remarkable cancellations of small
449: denominators by small numerators. If these cancellations occur
450: in general, it is possible to calculate analytically the
451: thermodynamical quantities of Dyson's hierarchical model in the HT phase
452: using Eq. (\ref{eq:num}).
453: Our results suggest that this model
454: has features analogous to the integrable systems in Hamiltonian
455: mechanics. As such, the model would stand out as a first
456: approximation to be used in situations where the conventional
457: perturbative expansions are not reliable.
458:
459: This research was supported in part by the Department of Energy
460: under Contract No. FG02-91ER40664.
461: Y. M. thanks the
462: Aspen Center for Physics for its hospitality in Summer 2000 while
463: this work was in progress and for a conversation there with L. Kadanoff.
464:
465: %\bibliography{da}
466: %\bibliographystyle{unsrt}
467: %\bibliographystyle{prsty}
468:
469: \begin{thebibliography}{10}
470: \bibitem{wegner72}
471: F. Wegner, Phys. Rev. B {\bf 3}, 4529 (1972).
472:
473: \bibitem{dyson69}
474: F. Dyson, Comm.\ Math.\ Phys.\ {\bf 12}, 91 (1969).
475:
476: \bibitem{baker72}
477: G. Baker, Phys.\ Rev.\ B {\bf 5}, 2622 (1972).
478:
479: \bibitem{gonzalez87}
480: A. Gonzalez-Arroyo and M. Okawa, Phys. Rev. D {\bf 35}, 672 (1987).
481:
482: \bibitem{taro00}
483: P. de~Forcrand~et al., Nucl. Phys. B {\bf 577}, 263 (2000).
484:
485: \bibitem{wilson71b}
486: K. Wilson, Phys. Rev. B. {\bf 4}, 3185 (1971).
487:
488: \bibitem{scaling}
489: Y. Meurice and S. Niermann, u. of Iowa Preprint, hep-lat/00077037.
490:
491: \bibitem{baker77}
492: G. Baker and G. Golner, Phys. Rev. B {\bf 16}, 2080 (1977).
493:
494: \bibitem{collet78}
495: P. Collet and J. Eckmann, {\em A Renormalization Group Analysis of the
496: Hierarchical Model in Statistical Mechanics} (Springer-Verlag, Berlin, 1978).
497:
498: \bibitem{bleher75}
499: P. Bleher and Y. Sinai, Comm. Math. Phys. {\bf 45}, 247 (1975).
500:
501: \bibitem{koch95}
502: H. Koch and P. Wittwer, Math. Phys. Electr. Jour. {\bf 1}, Paper 6 (1995).
503:
504: \bibitem{osc1}
505: Y. Meurice, G. Ordaz, and V.~G.~J. Rodgers, Phys.\ Rev.\ Lett. {\bf 75}, 4555
506: (1995).
507:
508: \bibitem{finite}
509: J. Godina, Y. Meurice, M. Oktay, and S. Niermann, Phys. Rev. D {\bf 57}, 6326
510: (1998).
511:
512: \bibitem{hyper}
513: J.~J. Godina, Y. Meurice, and M. Oktay, Phys. Rev. D {\bf 61}, 114509 (2000).
514:
515: \bibitem{dual}
516: Y. Meurice and S. Niermann, Phys. Rev. E {\bf 60}, 2612 (1999).
517:
518: \bibitem{gam3rapid}
519: J. Godina, Y. Meurice, and M. Oktay, Phys. Rev. D {\bf 57}, R6581 (1998).
520:
521: \bibitem{ht4}
522: J.~J. Godina, Y. Meurice, and S. Niermann, Nucl. Phys. B {\bf 519}, 737
523: (1998).
524:
525: \end{thebibliography}
526: \end{multicols}
527: \end{document}