HAL Id: jpa-00246886
https://hal.archives-ouvertes.fr/jpa-00246886
Submitted on 1 Jan 1994
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Towards an experimental determination of the number of metastable states in spin-glasses?
J. Bouchaud, E. Vincent, J. Hammann
To cite this version:
J. Bouchaud, E. Vincent, J. Hammann. Towards an experimental determination of the number of metastable states in spin-glasses?. Journal de Physique I, EDP Sciences, 1994, 4 (1), pp.139-146.
�10.1051/jp1:1994127�. �jpa-00246886�
Classification Physics Abstracts
75.50L 05.40
Towards
anexper1nlental determination of the nunlber of nletastable states in spin-glasses?
J-P-
Bouchaud(~,~),
E.Vincent(~)
and J.Hammann(~)
(~) TCM Group, Cavendish Laboratory, Madingley Road, Cambridge CB3 OHE, U-K-
(~) Service de Physique de l'Etat Condensé, Orme des Merisiers, CEA-Saclay, 91191 Gif
s/Yvette
cedex, France(Received
25 August 1993, accepted 15October1993)
Abstract. We argue that trie systematic deviation from perfect aging
(which
means thattrie relaxation function only depend on the ratio
£,
where twis trie waiting
time)
observed in spin-glasses can be accounted for if trienumberlf
available metastable statesis finite. A
significant fraction of these metastable states will be visited after an "ergodic" time which we estimate from trie experiments to be ci 10~ seconds. This corresponds ta m 10~~-10~~ dilferent
metastable states per independent subsystem.
1. Introduction.
The
experimental investigations
of thedynamics
of thespin-glass phase
have revealed that the observed properties areevolving
with the timeelapsed
from thequench:
this is theaging
phenomenon
[1, 2]. A well knownexample
is the relaxation of the thermc-remanentmagneti-
sation
(TRM):
afterfield-cooling
thesample
from aboveTg,
thelonger
thewaiting
time tw before the field iscut-off,
the slower thesubsequent magnetisation decay,
as if thespin-glass
had become "stilfer" with time. Similar elfects are observed in the mechanical
properties
ofpolymer glasses
[3, 4].Aging experiments
also olfer aunique
way of"scanning"
thephase
space structure of
complex
systems [5], and a detailedunderstanding
of theseexperiments
willprobably
oifer akey
for thislong
debatedproblem
[6-9].Recently,
one of usproposed
[10] aphenomenological theory
ofaging
based on the assumption that energy barriers are distributedexponentially,
assuggested by
mean field solutions of thespin-glass problem.
Theprobability
to find the system in a
given
metastable state with lifetime T is found to beeztremely
broad in thespin-glass phase.
As usual with broaddistributions,
one isonly
sensitive to the fewlargest
events encountered
(see
e-g.[Il]).
In the present case, thedynamics
isentirely
controlledby
thelongest "trap"
encountered after time tw, which is of order tw itself.Only
metastable states with lifetime of the order of tw have a finiteprobability
to beobserved;
inparticular
140 JOURNAL DE PHYSIQUE I N°1
-~ 0 6
j
t~
(se~)
*5 0 14
ÎÎÎ
m 31o~
é
Tj ((Î
à 0 12 j
~ ~
lC
~ ~ o t.Ot~+t
~j
HoJà oo~
~ Ag Mn 2 6z
~
't'"' ' T=9K=0 87 TgoJ
E 0.06
10-5 io-4 io-3 io-2 io-i ioo 101 10?
t / tw
_
0 16
1
~§ 0 14
~
§
ù 12àÎ ~
$
0 lu1
7J0 08 Aq.Mn2.6z
1
1=9K=0 87 Tgfi
~ÎÎ-5
io-4 io-3 jo-2 io-1 10Ù loi '02
t / tw
Fig.
l.a)
TRM relaxation inAg:Mn
2.6% ai 9 K, as a function oflog(
), for tw= 300, 1000, iw3000, 10000 and 30000 seconds. The systematic shift towards faster relaxation
as tw increases is clearly
seen. Insert: sketch of the experimental TRM procedure.
b)
Same data as in figure la, but with the"interrupted aging" correction, equation
(3),
expanded to second order in~i~.
Theerror bar is an
ierg
estimate of the systematic uncertainty in the vertical axis.
the
microscopic
time scaletotally disappears
from theproblem [loi. Hence, quite naturally,
aildynamical phenomena
indeed takeplace
on the scale of tw.Simple
arguments,building
uponthese
ideas,
show that say in thermoremanentmagnetisation
relaxation experiments themagnetisation decays
asfi ),
where t is the time after the switch off of themagnetic field,
tw
tw is the
"waiting
time"during
which trie field is on, andf(it)
is a function which behaves as a"stretched
exponent1al"
for small it, and as a power law forlarge
it[10,
12]. Thisprediction
isin very
good
agreement with experiments overnearly
rive time decades for agiven
tw. However,curves for different tw fait to
collapse
onto the same curve whenplotted
versus trie variable(see Fig. la):
there is a systematicdeviation,
which suggests that trie relaxation of older twsystems is "faster" thon it should be
(although
of course, whenplotted
versits t, trie older trie system, trie slower trierelaxation).
Trie sonne effect is observed both in metallic andinsulating
spin-glasses
[2], but also on trie viscoelasticproperties
ofpolymeric glasses
[3]. In order to finda
unique
"Master" curve, trie authors of [2]proposed
to take into account thissystematic
shiftby introducing
a relaxation time distribution of trie form p ~,
where /t is around
ii
+tw)"
0.9. This allows one to rescale ail data
quite satisfactonly
s
a
unct~n
of a reduced time variable which is
equalto )
in trie t < tw limit.(A
similarprocedure
was in fact
proposed by
Struik[3]).
From a fundamentalwpoint
of view thisprocedure
bas anintriguing
consequence:on dimensional
grounds,
triescaling
variable shouldreally
be written as~~ ~
Trie value
T -"tw
of trie time scale T* con be extracted from trie
experiments
and isextremely large
[2]: m 10~°secondsl This is
surprising
sinceOrly
two time scales are apriori
available: triemicroscopic
time scale To OE 10~~~ seconds
(which,
asexplained above,
shoulddisappear
in triespin-glass phase)
and trieexperimental waiting
time tw.Trie aim of this letter is to
give
aphysical meaning
to T* and to propose a new way to estimate it within trie framework of triethecry developed
in[loi.
We argue that triesystematic
deviation from asimple scaling
con beexplained provided
one takes into accoitnt thejinite
nitmbertw
of
available metastable stores in realsarnples
made of"grains"
of finite size. Inshort,
trie idea is triefollowing:
a finite size system willeventually
find trie"deepest trop"
in itsphase
space,which
corresponds
to trieequilibrium
state[loi.
This will take along
but finite timeterg,
whentrie
waiting
time tw exceeds this"ergodic"
time t~rg,aging
is"interrupted"
because triephase
space bas been
faithfully probed. Beyond
this timescale,
conventionalstationary dynamics
resume. A comparison of this
eqitilibriitm dynamics
with trie numerous available results [fil,like e-g- a-c-
susceptibility experiments
at net toc lowfrequency,
is howeverbeyond
trie scope of trie presentsimple
model.It is reasonable to assume that trie
experimental sarnples
are collections ofindependent subsystems,
each of whichevolving
within its owncomplicated phase
space. There will thus be a distribution ofergodic
times some shorter and somelonger.
Triesystematic
deviationfrom a
perfect scaling
is trie consequence of triepart1al
"death" of trie system, 1-e- that twaging
isinterrupted
in a certain fraction ofsubsystems.
This is what was mimickedby
trieprevious
~ ~
scaling proposed
in [2]. Here we show how one mayquantify
thiseffect,
T* -"tw
which allows us to extract from trie
spin-glass
data atypical ergodic
time of trie system whichis in tum related to trie number of available metastable states. In
fact,
this"interrupted aging"
phenomenon
bas been observed in 2-Dspin-glasses
[13], inpolymer glasses
[14] andpossibly
in
glassy Charge Density
Wave systems [15]. Ail these systems however reveal importantqualitative
andquantitative
differences. Forexample,
trieergodic
time scale in CDW appears to be much shorter thon inspin-glasses.
2. Trie mortel.
In order to be
self-contained,
let us first recall trie results of reference[loi.
Anexponent1al
distribution of barrier
heights (suggested by
trie mean-field models ofspin-glasses [16,
7] inducesa
power-law
distribution for trie lifetime of metastable states:ifilT)
= ~~fi lT
» To)Ii)
where To is a
microscopic
time scale (~- 10~~~s). ~(T)
is a temperaturedependent
parameterdescribing
trie structure ofphase
space: trie characteristicfree-energy
barriers are of trie order142 JOURNAL DE PHYSIQUE I N°1
of ~
I?i,
andmight
thus be temperaturedependent,
x is between 0 and 1 in triespir~-glass
~(T)
phase,
and thus, fromequation il),
or~e finds that the averageexploration
time is infir~ite. Theprobability P(T)
to fir~d the system in a metastable state of lifetime T isproportional
to T~fi(T)which,
for ~ <l,
is r~on-normalisable [9, 10,1?i.
For a fir~ite observation time tw,however,
trie distribution is cut-off above T t tw henceleading
to a twdeper~dent
normalisation. Trie~~-i
resulting P(T)
is then found to beindependent of
ToiP(T)
cc ~(for
T <tw).
Theseonly ingredients
are suilicier~t to induceaging
effects [18]. For x > 1, onT~ trie otherbond,
trie averageexploration
time is finite and cor~ventionalstationary
relaxation occurs.There
might
however bephysical
mechanisms which limit triepossible
values of T. Forexarnple if,
as discussedabove,
trie total number of metastable states of a giver~subsystem
isjinite,
=
S,
trie"ergodic
time" is of trie order of trielongest
T drawn from distributionil):
terg(S)
ceTOSÎ.
Trie relaxation of triemagnetisatior~ mit, tw,terg(S))
after agiven waiting
time tw is found to be, for tw <
terg(S), fi
withf(it)
= exp
~ it~~~ for
it < 1, and
tw
il
~
f lit)
cc it~?II for it » 1[loi,
~fI is a numberquantifying
trietendency
of trie system to torse itsmagnetisation
rather thon continue to evolve withinmagnetized
states. ~fII isproportional
to~fi, with an x
dependent
coefficient(which
was, forsimplicity,
taken to be in[loi
seebelow).
For
longer
times tw »t~rg(S), mit,
tw,t~rg(S))
becomes a function of This functiont~rg(S)
is in fact very
nearly
trie samef [loi land
trie small difference does net affect triefollowing
conclusions:only
its monotonic character is ofimportance). Hence,
triemagnetisation
of agiven subsystem
can be written in a compact forrn as:Îllil,
ÎW>Îergis))
" iÎÀQIS)1
l~[~ ~~leml
12)with
t*[
= tw for t + tw <t~rg(S)
and t*I,j =terg(S)
for t + tw »t~rg(S).
Inwords, equation (2) simply
means that once equilibrium is achievedii
+ tw »terg(S)),
themagnetisation
ismuch smaller than its "short" time
extrapolation, fi ).
tw
Now, assuming
that the number of metastable states varies from onesubsystem
to ar~otheraccording
to a distributionPi
~(So
is thus a"typical"
number of metastable states persubsystem),
the total(observed)
omagnetisation Mit, tw)
is obtained as trie average over P ofequation (2).
With theseonly assumptions,
onefinally
finds:Mit, tw)
~j
t~
~
jjt
+ tw~j
j~~M t to
o w
erg
where
t(r~
+TOS)
is trie"ergodic
time" of trie system, andFij
is anir~creasir~g
functior~ whichdepends
on trie choice ofmo(S)
and P. Beforetaking
aspecific example,
it should be noted thatequation (3) qualitatively
agrees with trieexperimental
trend: older systems indeed relax '~faster" whenplotted
versustw
3.
Analysis
of trieexperimental
data.The
simplest assumption
is that mû ar~d P behaveregularly
when their arguments are small.In this case, one
finds,
for t + tw < t(~~F(it)
= Ait + Bit~ +,
where A is a constant which
Table I. Parameters ~, nfI, nfII, t(~~ and
"complezity"
flfor
theAg:Mn
2.6$isample
[2]. ~, nfI were e~tracted byjitting
the "Initial" partof
trie TRM relaxation(t
<0.3tw)
ta a stretchede~ponential,
while ~fII is obtainedfrom
thelong
lime part(t
> 2tw) of
the rela~ations. Note thonQ(T
= 10 K)
coitld net be estimated becaitseof
oint lackof precision
in the verticalpositionning of
the rela~ation citrves. This là aproblem
close ta Tg since the magnetisationfaits rapidly
tasmall valites.
Table I: AgbIn: 2.6% (T =10.4 K)
TII<) z 11 ~/jj t~r«(10~ s.) Q
8 0.66+0.05 0.083+0.010 0 090+0 003 4+2 12.1+0.9
9 0.65+0 06 0.0124+0.016 0.127+0.003 2+1 11 9+0 9
g-à 0.68+0.06 0 153+0.020 0 lôâ+0.005 0.8+0.3 12.2+0.9
10 0.îâ+0.05 0.221+0.030 0.26â+0.005
Table II. Same as in table
I,
bittfor
theCdcri.7Ino_354 sample
[2].Here,
the initial partof
the rela~ation was extended ta t < 0.stw. The valite
of
the parametersgiven
here arecompatible
with those obtained in
reference [loi,
where the additionnai assitmption ~fI +'fII was made.Table II. Cdcrj 7In0 354 lT =16.î Ii)
TII<) x ni iii t~~~j10 s Q
10 0 76+0 05 0.059+0.008 0 058+0.003 5+2 14.2+0.9
12 0 76+0 0â 0.100+0 01 0 100+0 003 3+1 14.0+0.9
14 0 î3+0 0â 0 200+0 02 0 182+0.010 0 6+0.2 13.0+0.9
con be reabsorbed in the definition of
t$~~ [19]. We show in
figure
16 the same data as infigure
la, but with the"interrupted aging"
correction, equation(3).
Theonly
free parameters arethe
ergodic
time of the systemt$~~(S)
and the value of B(the
value of ~ is fixedby fitting
the
"youngest"
curve see[19]).
We have howeverimposed
trie same value of B= -3
ii-e-
the same functional form forP)
for ail temperatures andsamples.
Triecollapse
of trie curves is quitegood
in ail cases(see Fig. lb). Hence,
and this is trie main result of this paper, ointinterritpted aging
scenario là able ta mimic a " "scaling
with /t < 1.tw
The
corresponding
values of the parameters~,~fi,'fII,t$~~,
and of trie"complexity"
fl ex
logio( ~f)
are
given
in tables I and II, for different temperatures and two differentsamples:
Ag:Mn
2.6$i(Agmn),
which is metallic andCdcri.7Ino.354 (CrIn)
which is aninsulating
spir~glass.
It isinteresting
to see that bothsamples
lead to similar conclusions:a)
Theergodic
time t$~~ is of trie order of106seconds,
andthus,
within our interpretation, trie number of metastable states per"subsystem"
is found to be ce 10~~ -10~~(see
Tabs.I, II).
This number is however not easy to discuss sir~ce we do not kr~owprecisely
what these"subsystems"
are.
They
maybe,
forexample, magr~etically
disconnectedregions (like grains)
trie size of whichbeing
determinedby
triesample preparation,
and thus is temperatureindependent.
It could also be that trie
phase
space of 3dspin-glasses
is broken ir~tomutually
inaccessibleregions ("true" ergodicity breaking).
In this case, S would denote the number of metastablestates within one such
ergodic
component. This number would then be an intrinsic propertyof 3d
spin-glasses.
We however think that numerica1simulations [20]and/or experiments
on mesoscopicspin-glasses
[21] arereally
needed to understand how theergodic
lime t$~~ is related to the number ofspins
contained in asubsystem.
b)
As shown infigure
2, the number of metastable states of their~sulating sample (Crin)
tends to increase when the temperature is lowered
(it roughly
doubles when trie temperature is144 JOURNAL DE PHYSIQUE I N°1
15
14
ll ,
l0 °
0.5 o-G 0.7 0.8 o-g TIT
Fig.
2. Evolution of the "complexity" f1 % ~logio ~~~~o
with reduced temperature(
for bothAgmn and CrIn. Note that the complexity does trot seem to go to zero continuously ai Tg.g Insert:
Evolution of the complexity per spin versus reduced temperature for the ~-spin model, with, from bottom to top: p = 2
(SK model),
3,4,5 and 6 ~p = ce is the Random EnergyModel).
Note that f1 isdiscontinuous ai Tg as soon as p > 3. From reference [24].
lowered
by
1Kelvin). However,
in contrast to trie SK model [22], this nitmber ares net seem ta vanishcontinitoitsly
ai thespin-glass
transitionTg.
Such a behaviour was found in "Pottsglas"
models [23] and in models with
p(> 2)- spin
interactions [16],[24] seefigure
2(insert).
Trie number of metastable states of trie metallicsample (Agmn)
seems to beroughly
temperatureindeper~dent, although
trie ratherlarge
errer bars leave room for trie sonne behaviour as inCrIn.
c)
The parameter x grows towards 1 when T - Tg(at
least forAgMr~,
but other experi-ments on CrIn showed the same
tendency [2]).
This behaviour was also found in a very clearway in reference
[25],
where the initialdecay
of themagnetisation
was also fittedby
a stretchedexponer~t1al.
Dur results(fl(Tg)
finite and~(Tg)
m1)
suggest that thespir~-glass
transitionis,
in these
samples,
much doser to the "RandomEnergy
Model" scenario [16], than to trie "SK"scenario where
fl(Tg)
= 0 and~(Tg)
= 0 1?i. We however note that ~significar~tly departs
from trie REM linear behaviour ~ =~,
which would hold if the energy barriers were temperature Tg
independent. Furthermore,
trie REM"replica
symmetrybreaking"
scheme iscertainly
net ricinenough
to accour~t for morecomplicated aging protocoles
[Si or "secondspectrum" experimer~ts
[21].d) Finally,
one can see that triesimplifying hypothesis
~fI "'fII isquite
accurate. Infact,
adoser look at equations
(8-10)
of reference[loi
reveals that trie ratio '~~~ ir~creases with ~land
'fi
diverges
for x -1),
which is consistent with trieexperimental
trend(see
Tab. 1).4. Conclusion.
Let us summarize our
findings: quantifying
trie idea thataging
stops when trieexploration
ofphase
space iscompleted,
we baveproposed
a way to rescale relaxations obtair~ed for differentwaiting
times on aunique
"Master" curve, ar~d to understand thephysical meaning
of thephenomenological
"scaling.
This allows to extract from theexperimer~ts
a verylong,
twtemperature
dependent,
time scale t~rgbeyond
whichaging
ceases and trie nature of relaxationchanges.
This isinteresting
from twostandpoints:
from a fundamental point of view, it allows to ourknowledge,
for the first time to estimate the number of metastable states in acomplex
system, and its temperaturedependence. Perhaps
moreimportantly,
itallows,
from apractical point
ofview,
to extract fromrelatively
short time measurements anergodic
timebeyond
which trie relaxationchanges qualitatively:
we bave in mind here trie agingexperiments
onpolymer glasses,
where trie mechanicalproperties
are very similar to those discussed here[3],[4].
It could be ofimportance
to know when these mechanicalproperties
wiIIappreciably change.
Trie ideas
developed
heremight perhaps
be useful to discussphysical agir~g
in trie context of other(molecular
ormeta1Iic) glasses.
Acknowledgments.
We wish to express warm
gratitude
towards M.Ocio,
who gave us access to bis data onAgmn,
and for
enlightening
critical discussions. We aise thank K.Biljakovic,
F. Lefloch and M.Weissmann for
interesting
conversations.References
[1] Lundgren L., Svedlindh P., Nordblad P., Beckmann O., Phys. Rev. Lent 51
(1983)
911;Nordblad P., Lundgren L., Svedlindh P., Sandlund L., Phys. Rev. B 33
(1988)
645.[2]
a)
Alba M., Ocio M., Hammann J., Europhys. Lent. 2(1986)
45; J. Phys. Lent. 46(1985)
L-l101;Alba M., Hammann J., Ocio M., Refregier Ph., J. Appt. Phys. 61
(1987)
3683.b) Vincent E., Hammam J., Ocio M., Recent Progress in Random Magnets, D.H. Ryan Ed
(World
Scientific Pub. Co. Pte. Ltd, Singapore, 1992).
[3] Struik L.C.E., Physical
Aging
in Amorphous Polymers and Other Materials(Elsevier,
Houston,1978)
[4] Perez J., Physique et Mecanique des Polymères Amorphes
(Technique
et Documentation Lavoisier, Paris, 1992).[si Hammann J., Lederman M., Ocio M., Orbach R., Vincent E., Physica A 185
(1992)
278; Lefloch F., Hammam J., Ocio M., Vincent E., Europhys. Lent. 18(1992)
647.[GI Binder K., Young A.P., Rev. Med. Phys. 58
(1986)
801.[7] Mézard M., Parisi G., Virasoro M.A., Spin Glass Theory and Beyond
(World
Scientific, Singapore1987)
[8] Fisher D.S, Huse D.A., Phys. Rev. Lent 56
(1986)
1601; Phys. Rev. B 38(1988)
373.[9] Dotsenko V.S., Feigelmann M. V., Ioffe L.B., Spin-Glasses and related problems, Soviet Scientific Reviews, vol. 15
(Harwood, 1990);
Feigelmann M. V., Vinokur V., J. Phys. France 49
(1988)
1731.[10] Bouchaud J-P-, J. Phys. i France 2
(1992)
1705.[Il]
Bouchaud J. P., Georges A., Phys. Rep. 195(1990)
127.[12] Some of these predictions have recently been confirmed by theoretical and numerical calculations:
see [20]
Cughandolo L.F., Kurchan J., Phys. Rev. Lent. 71
(1993)
173;Cugliandolo L.F., Kurchan J. and Ritort F., prepnnt
(lune
93);Marinan E., Pansi G., prepnnt
(August 93).
146 JOURNAL DE PHYSIQUE I N°1
[13] Schins A.G., Dons E-M-, Arts A.F.M., de Wijn H-W-, Vincent E., Leylekian L. and Hammann J., to appear in Phys. Rev. B
(Dec.
1~~, 1993).[14] Lee A., Mc Kenna G-B-, Polymer 29
(1988)
1812;Mc Kenna G-B-, Santore M., Lee A., Duran R. S., J. Non.Gryst. Sol. 131
(1991)
497.[15] Biljakovic K., Lasjaunias J-C-, Monceau P., Lévy F., Phys. Rev. Lent. 67
(1991)
1902;Biljakovic K., Lasjaunias J-C-, Monceau P., Phys. Rev. Lent 62
(1989)
1512.[16] Derrida B., Phys. Rev. B 24
(1981)
2613.[17] Another physical situation where a non-normalisable probability distribution occurs is presented
in: Bardou F., Bouchaud J-P-, Cohen-Tannouji C., Emile O., Subrecoil laser cooling and Lévy flights, preprint ENS
(July 1993),
submitted to Phys. Rev. Lent.[18] Simple numerical simulations of models based on equation (1) indeed confirm that aging occurs
as soon as x < 1 without any further physical ingredient: E. Vincent
(unpublished)
[19] Since t$rg turns out to be ci 10~ seconds,
one may proceed "backwards" as follows. The "youngest"
relaxation
curve is such that the correction term in equation (3)is negligible: this allows to identify
the Master curve
fi
Now, plotting the ratio of -say- the oldest relaxation curve to the youngest twversus the variable
(t
+ tw)~ determines the functionF(~).
In the interval of ~ which is probed,a fit of the form
F(~)
= ~ + B~~ is
very satisfactory.
[20] Rieger H., preprint, J. Phys. A 26
(1993)
L615.[21] Alers G. B., Weissmann M. B., Isrealoff N. E., Phys. Rev. B 46
(1992)
507;Weissmann M. B., Isrealoff N.E., Alers G. B., J. Magn. Magn. Mater. l14
(1992)
87; Weissmann M. B., Rev. Med. Phys.(July,1993,
to appear).[22] Bray A. J., Moore M. A., J. Phys. G13
(1980)
L469.[23] Thirumalai D., Kirkpatrick T. R., Phys. Rev. B 36
(1987)
5388;Kirkpatrick T. R., Wolynes P. G.,Phys. Rev. B 36
(1987)
8552.[24] Rieger H., Phys. Rev. B 46
(1992)
14655.[25] Hoogerbeets R., Luo W. L., Orbach R., Phys. Rev. B 34