You are on page 1of 4

Density of States of Colloidal Glasses

Antina Ghosh,
1
Vijayakumar K. Chikkadi,
1
Peter Schall,
1
Jorge Kurchan,
2
and Daniel Bonn
1,3
1
Van der Waals Zeeman Institute, University of Amsterdam, Valckenierstraat 67, 1018 XE Amsterdam The Netherlands
2
PMMH, ESPCI, 10 rue Vauquelin, CNRS UMR 7636, Paris, France 75005
3
LPS de l’ENS, CNRS UMR 8550, 24 Rue Lhomond Paris, France 75005
(Received 12 October 2009; revised manuscript received 20 February 2010; published 18 June 2010)
Glasses are structurally liquidlike, but mechanically solidlike. Most attempts to understand glasses start
from liquid state theory. Here we take the opposite point of view, and use concepts from solid state
physics. We determine the vibrational modes of a colloidal glass experimentally, and find soft low-
frequency modes that are very different in nature from the usual acoustic vibrations of ordinary solids.
These modes extend over surprisingly large length scales.
DOI: 10.1103/PhysRevLett.104.248305 PACS numbers: 82.70.Dd, 63.50.Àx, 64.70.pv
The glass transition is perhaps the greatest unsolved
problem in condensed matter physics [1]: the main ques-
tion is how to reconcile the liquidlike structure with solid-
like mechanical properties. In solids, structure and
mechanics are related directly through the vibrational den-
sity of states (DOS) of the material [2]. To see whether
similar ideas could be used for glasses, a first step is to
determine the density of states of a glass that is solid
mechanically, but has no crystalline ordering.
Recent theory shows that randomly packed, frictionless
hard spheres form a system with no redundant mechanical
constraints. Because of this, a small perturbation, like, for
example, breaking a particle contact, may induce re-
arrangements at all scales [3–5]: the system is ‘‘critical’’
in this sense [6,7]. It has been argued that the marginality
of the system has dramatic consequences for the vibra-
tional spectrum [8]: in stark contrast to the well-known
phonon modes in solids [9–12], a broad band of floppy
modes emerges in the vibrational density of states, with a
gap at low frequencies that disappears as the pressure
becomes infinite.
In this Letter we measure, for the first time, vibrational
eigenstates of colloidal glasses and supercooled liquids.
We consider colloidal hard spheres that are subject to
thermal agitation, allowing us to follow the ‘‘vibrational’’
motion of particles that are trapped in cages constituted by
their neighbors. This is done at different random dense
packing configurations around the glass transition. We
obtain the density of modes from a normal-mode analysis
of particle displacements measured using confocal micros-
copy. We find that the vibrational spectrum has many soft
low-frequency modes [8], more abundant and different in
nature from the usual acoustic vibrations of ordinary sol-
ids. This results in an anomalous low-frequency peak in the
density of states which approaches zero frequency as one
goes deeper into the glass phase. The observed soft modes
are collective ‘‘swirling’’ particle motions that extend over
surprisingly long length scales.
Hard-sphere colloidal systems exhibit a glass transition.
Although the exact localization of glass transition is un-
certain and remains a point of discussion, the fact is that the
dynamics becomes very slow around a volume fraction of
ç
glass
¼ 0.58 [13]. We use poly-methylmethacrylate
(PMMA) particles, of diameter, a ¼ 1.3 µm, with a poly-
dispersity in size of about 5% that prevents crystallization
within the time scale of our experiments. Particles are
sterically stabilized to prevent aggregation and are dyed
with rhodamine which makes them visible using (confocal)
fluorescence miscoscopy. We use a mixture of cyclohexyl-
bromide and decalin as our solvent, to match the density
and index of refraction of the particles. The organic salt
TBAB (tetrabutylammoniumbromide) is used to screen
possible residual charges. To verify the hard-sphere behav-
ior we allow the system to crystallize by waiting weeks or
months, and measure the crystallization density, a very
sensitive measure for deviations from the hard-sphere be-
havior; it agrees to within a fraction of a percent to that of
true hard spheres.
Two-dimensional images were acquired using a fast
confocal microscope (Zeiss LSM live) in a field of view
of 100 µm  100 µm. Following around 2000 particles in
real time allows us to reconstruct all particle trajectories
(Fig. 1). The 2D slice was taken at a distance of 25–30 µm
away from the coverslip, deep enough to avoid the effects
of confinement [14]. To ensure reproducibility, each sam-
ple is measured a few times, and for each volume fraction 2
to 3 samples are measured. The time interval between each
image is 0.05 sec, which is approximately 1,15
th
of the
Brownian time, r
B
¼ ,J
3
,k
B
T % 0.75 sec , where , is
the solvent viscosity. The mean-square displacement per
particle hðurÞ
2
i as a function of time shows a plateau when
plotted on a log-lin scale, indicating that in our experimen-
tal time-window, each particle moves in a ‘‘cage’’ consti-
tuted by neighboring particles. The plateau value decreases
with increasing volume fraction of the suspension, because
the particle motion is more restricted (Fig. 2). Clearly,
measurement times have to be long enough for soft modes
(if present) to be visible, yet short enough that the system
stays within a basin. The very existence of a plateau in our
measured mean-squared displacement shows that there is a
PRL 104, 248305 (2010)
P HYS I CAL RE VI E W L E T T E RS
week ending
18 JUNE 2010
0031-9007,10,104(24),248305(4) 248305-1 Ó 2010 The American Physical Society
range of time scales for which local equilibrium can be
justified and the soft modes can be measured, without
many activated events taking place.
As shown in [15], if there is a proliferation of soft modes
above a low-frequency gap, the value of the plateau in the
mean-squared displacement is dominated by the soft
modes just above the gap, leading to an anomalous scaling
of the plateau with pressure. Although each individual high
frequency vibrational mode will have a squared amplitude
that scales as ¡
À2
, the dominant effect here should be that
the new modes appear as the gap closes with increasing
pressure. An estimate of the pressure dependence on the
gap then leads [15,16] to the prediction that the plateau
value of the mean-squared displacement scales with pres-
sure as ¡
À3,2
. This scaling can be tested directly in our
experiments by plotting the plateau value of the mean-
square displacement vs the pressure. The latter can be
computed from the volume fraction using an empirical
equation of state for the colloidal hard spheres, for example
[17], ¡ ¼ Àðk
B
T,¡Þç
2 J

ln½ððç
max
,çÞ
ð1,3Þ
À 1Þ
3
Š, with
¡ the volume per particle. Fitting the data, it is clear that
they are indeed compatible with the anomalous scaling
proposed by [15,16] (Fig. 2). This is the first indication
that soft modes are present in the system.
It is worthwhile noting here that if we make the reason-
able assumption that the system is in local equilibrium
during the time interval associated with the plateau, then
the values of the squared-displacements are independent of
the dynamics, which may go from overdamped to purely
ballistic, and even contain hydrodynamic interactions.
What does depend on the nature of the dynamics is the
actual time dependence of the displacements, but not their
statistical distribution. As an analogy: over- and under-
damped Langevin dynamics in a harmonic well are very
different, but their statistical distribution of displacements
is the same.
We now turn to the experimental determination of the
DOS of our colloids. There are two standard methods to
obtain the density of modes of such a system. Spatial
Fourier analysis of displacements has been successfully
applied to colloidal crystals [9]. However, it is inapplicable
here, since it relies on the periodicity of the lattice.
Secondly, one can calculate the Fourier transform of the
velocity autocorrelation function, provided one uses a
method with high time resolution [18,19], higher than we
can achieve here.
Here we use a new method in which we directly analyze
the normal modes of the vibrations in the cages from the
displacement correlations. If the system were harmonic,
undamped and without hydrodynamic interactions, the
DOS obtained from velocity autocorrelations and our
method would coincide. However, in the former method
all the spatial information is lost. The strength of the
present analysis using confocal measurements lies in di-
rectly visualizing the collective modes at low frequencies
and therefore gaining information about the nature of such
modes. But the price to pay is the exclusion of the com-
bined effects of damping, anharmonicity and hydrody-
namic interactions which results in a frequency scale that
is somewhat different from the real frequencies of a
damped anharmonic system. In our method, the DOS is
obtained as follows. Denoting u
o
ðiÞ the components of the
particle displacements from the average position along the
confocal plane u
o
ðiÞ ¼ fðx
|
ðiÞ À hx
|
iÞ, ðy
|
ðiÞ À hy
|
iÞg, we
obtain the displacement correlation [20–22] matrix (of
dimension twice the number of observed particles) as,
D
ol
¼ hu
o
ðiÞu
l
ðiÞi, (1)
where hÁ Á Ái denotes the average over the period of mea-
surement, which is about 175 sec. Diagonalizing D
ol
we
obtain the normal modes, and the corresponding eigen-
values ¬
o
. We express the results in terms of the relevant
quantities,
0 50 100 150 200
t (sec)
0.01
0.1
δ

r
2
(
µ
m
2
)
0.54
0.57
0.58
0.59
0.60
0.1 1 10 100
t (sec)
0.001
0.1
δ
r 2
( µ
m
2
)
5 10 25
< δr
2
>
-2/3
( µm
-4/3
)
10
25
30
P
FIG. 2 (color online). Left: Mean-square displacement versus
time at different volume fractions ç ¼ 0.54, 0.57, 0.58, 0.59,
0.60 shows a plateau over which the soft modes in the system
have been measured. The same plot in log-log scale is shown in
the inset, where the plateau is not so evident. Right: pressure
from the equation of state as a function of the plateau value of
the mean-square displacement. The full line is a P
À3,2
fit.
FIG. 1 (color online). A two-dimensional image of the colloi-
dal suspension at a volume fraction 0.60 acquired by confocal
microscopy. The field of view is 100 Â 100 µm. The inset
shows a few typical particle trajectories, which show that we
obtain the DOS before large-scale rearrangements happen.
PRL 104, 248305 (2010)
P HYS I CAL RE VI E W L E T T E RS
week ending
18 JUNE 2010
248305-2
w
o
¼
ffiffiffiffiffiffiffiffiffiffiffi
1,¬
o
q
, (2)
which are the temporal frequencies the system would have
if it were harmonic and undamped.
The resulting density of states is shown in Fig. 3.
Because ‘‘hard’’ modes are expected to have eigenvalues
proportional to the pressure, we scale out this effect by
plotting the DOS in terms of scaled frequencies w,¡. The
rescaled cumulative density of states (Fig. 3) shows that, in
agreement with [15,20], the gap in the rescaled variables
closes with increasing pressure. This is also the reason for
the anomalous scaling of the plateau in the mean-squared
displacement. Hence we conclude that indeed there is a
proliferation of soft modes when the system becomes
glassy.
What do the soft modes look like? The two-dimensional
eigenvector field corresponding to a single frequency
shows the emergence of large vortexlike structures
(Fig. 4), indicating that the low-frequency soft modes
consist in fact of a large scale swirling motion of the
particles. To the contrary, for higher frequencies, the ei-
genvector field tends to look random. To quantify the order
in the low-frequency eigenvector field, we calculated the
orientational correlation function defined as the scalar
product ð ¯ ¡
|
. ¯ ¡
]
Þ, where ¯ ¡
|
and ¯ ¡
]
are the two-dimensional
eigenvector components corresponding to the |th and ]th
particle for a single eigenmode. The result for eigenmodes
in the different parts of the spectrum are shown in Fig. 5.
The low-frequency soft modes involve motion correlated
over many (tens of) interparticle distances and thus extend
over large length scales. One surprise is that the correla-
tions characterizing the soft modes are very weakly depen-
dent on volume fraction and hence on pressure (Fig. 5),
suggesting that the large correlation length stays finite in
the large-pressure limit. The negative part (anticorrelation)
of the correlation function (Fig. 5) becomes slightly more
pronounced upon increasing the volume fraction. To be
able to see the swirling motion disappearing completely
one presumably has to go to even lower volume fractions
(equilibrium fluid phase), where the present method of
analysis no longer applies.
As was pointed out to us by Henkes, Brito, and van
Saarloos [23] the finite resolution of the microscope leads
to a significant uncertainty in the high frequency tail of the
calculated DOS. Roughly, the error in the DOS becomes
important when the value of the matrix elements of the
displacement correlation matrix becomes on the order of
the square of the resolution of the microscope (estimated to
be 30 nm [14]). The corresponding upper limit of the of
frequency w
0
up to which the spectrum is reliable is
indicated in Fig. 3. The eigenvector fields presented here
all correspond to frequencies below w
0
.
One possible caveat of our analysis is the following. In
our experimental setup we measure the displacements of
particles in a small and two-dimensional slice of the real
system, and we diagonalise a submatrix of the full corre-
lation matrix. This is not only due to an experimental, but
also to a numerical limitation: diagonalizing the full cor-
relation matrix is an impossible task. The problem is not
hopeless, because our slice is typical of all others and will
give statistically equal results. In particular, the mean-
squared displacement per particle and per spatial dimen-
sion is the same when restricted to a slice and in the whole
system. One can say more: consider an eigenvector
~
V with
eigenvalue
~
¬ of the restricted covariance matrix, involving
only the measured particles and their motion along the
confocal plane. The eigenvectors of the full system fV
o
g
forma complete basis so we can expand
~
V in terms of V
o
as
0 20 40 60 80 100 120
ω (µm
-1
)
0
0.005
0.02
D
(
ω
)
0.58
0 5 10 15
ω / p (µm
1/3
)
0.0
1.0
0.54
0.57
0.58
0.59
0.60
D
(
ω
)
FIG. 3 (color online). Left: Density of states shown for the
volume fractions, ç ¼ 0.58. The frequency axis is in the unit of
µm
À1
. The dashed line shows the upper-limit of frequency w
0
until which the spectrum is reliable. Right: Closure of the gap:
cumulative density of states D
0
ðwÞ ¼
R
DðwÞJw at different
volume fractions respectively ç ¼ 0.54, 0.57, 0.58, 0.59, 0.60.
The frequencies along the horizontal axis are scaled by the
pressure.
FIG. 4 (color online). The two-dimensional eigenvector fields.
The volume fraction is ç ¼ 0.60. Frequencies (µm
À1
) from left
to right, top: 1.01, 3.40, bottom: 10.4, 18.5. The length scale
associated with swirling structures (top left corner) is about
34 microns. Large displacements (big arrows) in the eigenvector
field are mostly seen always at low frequencies and are often an
isolated event.
PRL 104, 248305 (2010)
P HYS I CAL RE VI E W L E T T E RS
week ending
18 JUNE 2010
248305-3
~
V ¼
X
o
c
o
V
o
with
X
o
c
2
o
¼ 1. (3)
One has
~
¬ ¼
P
o
c
2
o
¬
o
, which can also be written as
P
o
c
2
o
ð
~
¬ À ¬
o
Þ ¼ 0. It is clear from this and the fact that
eigenvalues are positive that the existence of a large eigen-
value
~
¬ for the restricted problemimplies the existence of a
large eigenvalue ¬
o
!
~
¬ for the complete problem, or,
equivalently, w
o
~ w. Each soft mode of the restricted
problem implies the existence of—and has a large projec-
tion over—soft modes of the complete system. In particu-
lar, a mode with ~ w ¼ 0 in the restricted problem implies a
mode w ¼ 0 in the true system. Although we do not expect
that the spectrum associated with a large two-dimensional
slice coincides exactly with that of the full system, we may
be certain that the soft modes we observe are indeed a
reflection of all the soft modes of the full system. Another
argument is that the vibration we observe in our slice is
induced by the global soft modes. As mentioned above,
DðwÞ has a sharp peak at a gap value w
Ã
ð¡Þ, below which
there are very few modes (w
Ã
ð¡Þ goes to zero as ¡ ! 1).
The mean-squared displacement is dominated at high pres-
sures by the peak value contribution [15] hðurÞ
2
i $
Dðw
Ã
Þ
w
Ã
.
Thus, at large pressure, essentially all the motion of each
particle is given by a combination of the soft modes just
above the gap, i.e., those having w $ w
Ã
ð¡Þ.
In conclusion, we have argued that the normal modes of
a restricted set of particles gives information on the whole
system. Thanks to this, a well-established technique such
as confocal microscopy can be used to study global vibra-
tional properties of the system, which contain very detailed
information on the geometry of the configurations. The
above strategy can be applied for much more complex
colloidal systems, e.g., with interactions other than hard-
sphere repulsions. Here, we have reported an excess of
low-frequency modes in a colloidal hard-sphere system
around the glass transition volume fraction. These modes
show large-scale correlations in the velocities of particles,
extending over many particle diameters. The existence of
soft modes had been predicted for ideal hard spheres on the
basis of theoretical considerations. Our experimental study
shows that they do exist in glassy colloidal suspensions,
probably the most studied system in the field of glassy
dynamics. These vibrational modes signal the onset of
macroscopic elasticity and give a microscopic insight in
the collectivity of the particle dynamics near the glass
transition. Following particles for longer times will further
allow us to assess the relevance of the low-frequency soft
modes to long time, activated dynamics near the glass
transition.
[1] Liquids, Freezing and the Glass Transition, edited by J. P.
Hansen, D. Levesque, and J. Zinn-Justin (North-Holland,
Amsterdam, 1991).
[2] C. Kittel, Introduction to Solid State Physics (Wiley,
New York, 1953).
[3] S. Alexander, Phys. Rep. 296, 65 (1998).
[4] G. Combe and J. N. Roux, Phys. Rev. Lett. 85, 3628
(2000).
[5] A. Widmer-Cooper, H. Perry, P. Harrowell, and D. R.
Reichman, Nature Phys. 4, 711 (2008).
[6] C. S. O’Hern, L. E. Silbert, A. J. Liu, and S. R. Nagel,
Phys. Rev. E 68, 011306 (2003).
[7] M. Wyart, S. Nagel, and T. Witten, Europhys. Lett. 72, 486
(2005).
[8] N. Xu, M. Wyart, A. J. Liu, and S. R. Nagel, Phys. Rev.
Lett. 98, 175502 (2007); For a point of view unrelated to
isostaticity, see, e.g.: T. S. Grigera, V. Martin-Mayor, G.
Parisi, and P. Verrocchio, Nature (London) 422, 289
(2003); T. S. Grigera, V. Martin-Mayor, G. Parisi, and P.
Verrocchio, J. Phys. Condens. Matter 14, 2167 (2002).
[9] P. Keim, G. Maret, U. Herz, and H. H. von Gru¨nberg,
Phys. Rev. Lett. 92, 215504 (2004).
[10] B. J. Ackerson, and N. A. Clark, Faraday Discuss. Chem.
Soc. 76, 219 (1983).
[11] M. Hoppenbrouwers and W. van de Water, Phys. Rev. Lett.
80, 3871 (1998).
[12] Z. Cheng, J. Zhu, W. B. Russel, and P. M. Chaikin, Phys.
Rev. Lett. 85, 1460 (2000).
[13] P. Pusey and W. van Megen, Nature (London) 320, 340
(1986).
[14] E. R. Weeks et al., Science 287, 627 (2000).
[15] C. Brito and M. Wyart, J. Stat. Mech. (2007) L08003.
[16] C. Brito and M. Wyart, arXiv:0804.2439.
[17] R. D. Kamien and A. J. Liu, Phys. Rev. Lett. 99, 155501
(2007).
[18] W. Van Megen, Phys. Rev. E 73, 020503(R) (2006).
[19] Stephen R. Williams, G. Bryant, I. K. Snook, and W. van
Megen, Phys. Rev. Lett. 96, 087801 (2006).
[20] R. Mari, F. Krzakala, and J. Kurchan, Phys. Rev. Lett. 103,
025701 (2009).
[21] B. Hess, Phys. Rev. E 62, 8438 (2000).
[22] S. A. Harris and C. A. Laughton, J. Phys. Condens. Matter
19, 076103 (2007).
[23] S. Henkes, C. Brito, and W. van Saarloos (private com-
munication).
0 2.5 5
ω /P (µm
1/3
)
0
100
180
D
(
ω
)
0 25 50 75
δr
ij
(µm)
0
7.5

v
i

.

v
j


(
1
0
-
5
)
+
= 1.5 µ δ r m
s
5 25 50 75 100
δr (µm)
0
5.0
v
i


.

v
j


(
1
0
-
5
)
0.54
0.57
0.58
0.59
0.60
FIG. 5 (color online). Left: The average scalar product ¯ ¡
|
¯ ¡
]
,
for a vol. fraction 0.60. This has been computed for eigenmodes
in different parts of the spectrum, as shown in DðwÞ in the inset.
Lowest frequency modes have a long tail with a negative part
characterizing the swirling motion. The shortest interparticle
distance ur
s
¼ 1.5 µm is indicated by the arrow. Right: The
orientational correlation function for the lowest frequency modes
for different packing fractions ç ¼ 0.54, 0.57, 0.58, 0.59, 0.60.
The general shape of the modes seems only very weakly depen-
dent on the packing fraction.
PRL 104, 248305 (2010)
P HYS I CAL RE VI E W L E T T E RS
week ending
18 JUNE 2010
248305-4

If the system were harmonic.58.60 0 50 100 150 200 P t (sec) 10 5 10 25 t (sec) < δr > 2 -2/3 ( µm -4/3 ) computed from the volume fraction using an empirical equation of state for the colloidal hard spheres. 0.01 0. 2). the dominant effect here should be that the new modes appear as the gap closes with increasing pressure. it is inapplicable here. where the plateau is not so evident. Here we use a new method in which we directly analyze the normal modes of the vibrations in the cages from the displacement correlations. Diagonalizing Dab we obtain the normal modes. We express the results in terms of the relevant quantities. Dab ¼ hua ðtÞub ðtÞi. and even contain hydrodynamic interactions.57 0.16] to the prediction that the plateau value of the mean-squared displacement scales with pressure as pÀ3=2 . An estimate of the pressure dependence on the gap then leads [15.PRL 104. with v the volume per particle. However. ðyi ðtÞ À hyi iÞg. Right: pressure from the equation of state as a function of the plateau value of the mean-square displacement. The inset shows a few typical particle trajectories. which may go from overdamped to purely ballistic.59 0. We now turn to the experimental determination of the DOS of our colloids. undamped and without hydrodynamic interactions. The latter can be 0. since it relies on the periodicity of the lattice.001 0. 2 δr (µm ) 2 where hÁ Á Ái denotes the average over the period of measurement. which is about 175 sec. As an analogy: over.1 2 0. In our method. It is worthwhile noting here that if we make the reasonable assumption that the system is in local equilibrium during the time interval associated with the plateau. which show that we obtain the DOS before large-scale rearrangements happen. 0. Secondly. Spatial Fourier analysis of displacements has been successfully applied to colloidal crystals [9]. The full line is a PÀ3=2 fit. p ¼ ÀðkB T=vÞ2 d ln½ððmax =Þð1=3Þ À 1Þ3 Š. Left: Mean-square displacement versus time at different volume fractions  ¼ 0:54.59.54 0. the value of the plateau in the mean-squared displacement is dominated by the soft modes just above the gap. it is clear that they are indeed compatible with the anomalous scaling proposed by [15. we obtain the displacement correlation [20–22] matrix (of dimension twice the number of observed particles) as. The strength of the present analysis using confocal measurements lies in directly visualizing the collective modes at low frequencies and therefore gaining information about the nature of such modes. However.19]. This is the first indication that soft modes are present in the system. 0. Fitting the data. Denoting ua ðtÞ the components of the particle displacements from the average position along the confocal plane ua ðtÞ ¼ fðxi ðtÞ À hxi iÞ. What does depend on the nature of the dynamics is the actual time dependence of the displacements.60 acquired by confocal microscopy. 2 (color online). A two-dimensional image of the colloidal suspension at a volume fraction 0. without many activated events taking place. for example d [17].16] (Fig.57. range of time scales for which local equilibrium can be justified and the soft modes can be measured. The field of view is 100  100 m. This scaling can be tested directly in our experiments by plotting the plateau value of the meansquare displacement vs the pressure. one can calculate the Fourier transform of the velocity autocorrelation function. but their statistical distribution of displacements is the same. 248305 (2010) PHYSICAL REVIEW LETTERS week ending 18 JUNE 2010 FIG. and the corresponding eigenvalues a . but not their statistical distribution. higher than we can achieve here. 248305-2 . the DOS is obtained as follows. Although each individual high frequency vibrational mode will have a squared amplitude that scales as pÀ2 . 0. the DOS obtained from velocity autocorrelations and our method would coincide. As shown in [15]. (1) δ r (µm ) FIG.1 1 10 100 2 0. if there is a proliferation of soft modes above a low-frequency gap.58 0.60 shows a plateau over which the soft modes in the system have been measured. anharmonicity and hydrodynamic interactions which results in a frequency scale that is somewhat different from the real frequencies of a damped anharmonic system. But the price to pay is the exclusion of the combined effects of damping.1 30 25 0. The same plot in log-log scale is shown in the inset. There are two standard methods to obtain the density of modes of such a system. in the former method all the spatial information is lost.and underdamped Langevin dynamics in a harmonic well are very different. 1 (color online). leading to an anomalous scaling of the plateau with pressure. provided one uses a method with high time resolution [18. then the values of the squared-displacements are independent of the dynamics.

The corresponding upper limit of the of frequency !0 up to which the spectrum is reliable is indicated in Fig.005 0 0 20 40 60 -1 D(ω) D(ω) 0. 4 (color online).02 1. In particular. we calculated the orientational correlation function defined as the scalar ~ ~ ~ ~ product ðvi :vj Þ. for higher frequencies. One surprise is that the correlations characterizing the soft modes are very weakly dependent on volume fraction and hence on pressure (Fig.60 0 5 10 1/3 80 100 120 15 ω (µm ) ω / p (µm ) FIG. 3) shows that.60. This is not only due to an experimental. 3. The result for eigenmodes in the different parts of the spectrum are shown in Fig. 0. The frequencies along the horizontal axis are scaled by the pressure. and we diagonalise a submatrix of the full correlation matrix. the meansquared displacement per particle and per spatial dimension is the same when restricted to a slice and in the whole ~ system. To the contrary.  ¼ 0:58.4. 248305-3 . 3. Left: Density of states shown for the volume fractions. The eigenvector fields presented here all correspond to frequencies below !0 . The negative part (anticorrelation) of the correlation function (Fig. top: 1. Hence we conclude that indeed there is a proliferation of soft modes when the system becomes glassy.59. The rescaled cumulative density of states (Fig. because our slice is typical of all others and will give statistically equal results. Right: Closure of the gap: R cumulative density of states D0 ð!Þ ¼ Dð!Þd! at different volume fractions respectively  ¼ 0:54.58 0. The length scale associated with swirling structures (top left corner) is about 34 microns. Brito. the error in the DOS becomes important when the value of the matrix elements of the displacement correlation matrix becomes on the order of the square of the resolution of the microscope (estimated to be 30 nm [14]).20]. we scale out this effect by plotting the DOS in terms of scaled frequencies !=p. FIG. 4). indicating that the low-frequency soft modes consist in fact of a large scale swirling motion of the particles. To be able to see the swirling motion disappearing completely one presumably has to go to even lower volume fractions 0. The low-frequency soft modes involve motion correlated over many (tens of) interparticle distances and thus extend over large length scales. the gap in the rescaled variables closes with increasing pressure. PHYSICAL REVIEW LETTERS (2) week ending 18 JUNE 2010 which are the temporal frequencies the system would have if it were harmonic and undamped. As was pointed out to us by Henkes. 5) becomes slightly more pronounced upon increasing the volume fraction.57. and van Saarloos [23] the finite resolution of the microscope leads to a significant uncertainty in the high frequency tail of the calculated DOS.58. 0. One possible caveat of our analysis is the following. 3 (color online). The two-dimensional eigenvector fields. 3. involving only the measured particles and their motion along the confocal plane. suggesting that the large correlation length stays finite in the large-pressure limit. where the present method of analysis no longer applies.54 0. 5). 248305 (2010) qffiffiffiffiffiffiffiffiffiffiffi !a ¼ 1=a .01.5. in agreement with [15. 18. Because ‘‘hard’’ modes are expected to have eigenvalues proportional to the pressure. bottom: 10.PRL 104. where vi and vj are the two-dimensional eigenvector components corresponding to the ith and jth particle for a single eigenmode. Roughly. The volume fraction is  ¼ 0:60.40. Large displacements (big arrows) in the eigenvector field are mostly seen always at low frequencies and are often an isolated event. The dashed line shows the upper-limit of frequency !0 until which the spectrum is reliable.0 0. One can say more: consider an eigenvector V with ~ eigenvalue  of the restricted covariance matrix.58 0. The frequency axis is in the unit of mÀ1 .0 (equilibrium fluid phase). In our experimental setup we measure the displacements of particles in a small and two-dimensional slice of the real system. 0. the eigenvector field tends to look random.57 0. The problem is not hopeless. 5. To quantify the order in the low-frequency eigenvector field. What do the soft modes look like? The two-dimensional eigenvector field corresponding to a single frequency shows the emergence of large vortexlike structures (Fig. but also to a numerical limitation: diagonalizing the full correlation matrix is an impossible task. The resulting density of states is shown in Fig. The eigenvectors of the full system fVa g ~ form a complete basis so we can expand V in terms of Va as 0.59 0. 0. Frequencies (mÀ1 ) from left to right. This is also the reason for the anomalous scaling of the plateau in the mean-squared displacement.

Phys. 0. Phys. A. 087801 (2006). extending over many particle diameters. Phys. Lett. with interactions other than hardsphere repulsions. [23] S. van Megen. 627 (2000). E. Kurchan. [21] B. R. J. Martin-Mayor. Clark. Lett. fraction 0. M. Left: The average scalar product vi  vj . Van Megen. van Saarloos (private communication). K. Rep. equivalently.59 0. [22] S. 96. Chaikin. Williams.PRL 104. Nature (London) 422. Alexander. at large pressure. Maret. Stat. J. 2167 (2002). T. In conclusion. 85. The existence of soft modes had been predicted for ideal hard spheres on the basis of theoretical considerations. G. [12] Z. 175502 (2007).: T.2439. 1953). E 62. and N. A. Hansen. [18] W. 020503(R) (2006). below which there are very few modes (!à ðpÞ goes to zero as p ! 1). Lett. A. P. Brito. S. Zinn-Justin (North-Holland. and S. [16] C. as shown in Dð!Þ in the inset. Snook. [2] C. J. Perry. J. 0. G. Rev. Lett. essentially all the motion of each particle is given by a combination of the soft modes just above the gap. Henkes. 711 (2008).60. E 68. which can also be written as P 2 ~ a ca ð À a Þ ¼ 0. C. !a !. D. Phys. Verrocchio. and T. I. Phys. 296. Rev. Following particles for longer times will further allow us to assess the relevance of the low-frequency soft modes to long time. e. E 73. Right: The orientational correlation function for the lowest frequency modes for different packing fractions  ¼ 0:54. Each soft mode of the restricted ~ problem implies the existence of—and has a large projection over—soft modes of the complete system. R. D. V.. a mode with ! ¼ 0 in the restricted problem implies a ~ mode ! ¼ 0 in the true system. Phys. activated dynamics near the glass transition. 0. Lett. Matter 19. Wyart. 248305 (2010) 7. Laughton. and D. ~ ~ for a vol. 99.57. These modes [1] Liquids. In particular.5 D(ω) 180 PHYSICAL REVIEW LETTERS 5. Phys. Russel. H. Reichman. [5] A. 85. Lowest frequency modes have a long tail with a negative part characterizing the swirling motion. Phys. J. Silbert. Wyart. Combe and J. Phys. S. and W. Matter 14. 3628 (2000). 92. ¨ [9] P. J. Nagel. see. The mean-squared displacement is dominated at high presà sures by the peak value contribution [15] hðrÞ2 i $ Dð! Þ .54 0. 103. 80. and J. !à Thus. L. [17] R. e. [15] C. Levesque. This has been computed for eigenmodes in different parts of the spectrum. Here. Phys. These vibrational modes signal the onset of macroscopic elasticity and give a microscopic insight in the collectivity of the particle dynamics near the glass transition. Rev. Martin-Mayor. Phys. we may be certain that the soft modes we observe are indeed a reflection of all the soft modes of the full system. Lett. Zhu. vj (10 ) 100 0 0 2. or..5 5 ω /P (µm ) 1/3 0. F. Phys. Kittel. Rev. 8438 (2000). van de Water. Bryant. [4] G. 011306 (2003). and P. B. S. G. van Megen. Science 287. Dð!Þ has a sharp peak at a gap value !à ðpÞ. show large-scale correlations in the velocities of particles. 025701 (2009). Another argument is that the vibration we observe in our slice is induced by the global soft modes. and P. P. 76. 340 (1986). Rev. S. Rev. 4. Freezing and the Glass Transition. Liu. Parisi. 486 (2005). Herz. and P. Faraday Discuss. Widmer-Cooper. New York. For a point of view unrelated to isostaticity. Cheng. Condens. Lett. -5 -5 ~ V¼ X a ca Va with X c2 ¼ 1: a a (3) ~ P a One has  ¼ a c2 a . we have argued that the normal modes of a restricted set of particles gives information on the whole system. Introduction to Solid State Physics (Wiley. 1991). 65 (1998). The above strategy can be applied for much more complex colloidal systems. Brito and M. 219 (1983). [6] C. Kamien and A. Liu. 076103 (2007). 3871 (1998). 289 (2003). Amsterdam. Wyart. 215504 (2004). Keim. (2007) L08003. 248305-4 . Harris and C. R. and W.g. The general shape of the modes seems only very weakly dependent on the packing fraction. which contain very detailed information on the geometry of the configurations. Rev. Condens. Nagel. Krzakala. 5 (color online).57 0. Rev. Grigera. Parisi. Mech.58 0. Hess. we have reported an excess of low-frequency modes in a colloidal hard-sphere system around the glass transition volume fraction. V. arXiv:0804. Lett. Mari. and H. A. probably the most studied system in the field of glassy dynamics.e. Weeks et al. Verrocchio. those having ! $ !à ðpÞ. As mentioned above. Brito and M. H. 1460 (2000). Roux. R. [19] Stephen R. Witten. Harrowell.60 0 δ rs = 1. Rev. Hoppenbrouwers and W. Lett. von Grunberg.. [7] M. [8] N. Although we do not expect that the spectrum associated with a large two-dimensional slice coincides exactly with that of the full system. [3] S.g. and S. Chem. Pusey and W. [14] E.60. [13] P. Phys. W. Europhys. O’Hern. 0. [11] M. Xu. Nature Phys. [10] B.58. J. 98. Nagel. Our experimental study shows that they do exist in glassy colloidal suspensions.59. It is clear from this and the fact that eigenvalues are positive that the existence of a large eigen~ value  for the restricted problem implies the existence of a ~ large eigenvalue a !  for the complete problem. [20] R. The shortest interparticle distance rs ¼ 1:5 m is indicated by the arrow. Wyart. Grigera. Liu. Soc. and J.0 vi .5 µ m 0 0 25 δrij (µm) 50 75 5 25 50 75 100 δr (µm) FIG. M. a well-established technique such as confocal microscopy can be used to study global vibrational properties of the system. Ackerson. Rev. edited by J. Phys. G. vj (10 ) week ending 18 JUNE 2010 + vi . Thanks to this. 72. N. Nature (London) 320. Rev. 155501 (2007). J. U. i. A.