a

r
X
i
v
:
1
3
0
8
.
3
1
6
2
v
1


[
c
o
n
d
-
m
a
t
.
q
u
a
n
t
-
g
a
s
]


1
4

A
u
g

2
0
1
3
Scale Invariance in Expanding Fermi Gases
E. Elliott
1,2
, J. A. Joseph
1
, and J. E.Thomas
1
1
Department of Physics, North Carolina State University, Raleigh, NC 27695, USA and
2
Department of Physics, Duke University, Durham, NC 27708, USA
(Dated: August 15, 2013)
We precisely test scale invariance in the hydrodynamic expansion of a Fermi gas of atoms with
resonant interactions, using a model-independent method. We observe that the three-dimensional
mean square cloud radius expands ballistically and nd that the bulk viscosity is consistent with
zero, as predicted for a scale-invariant system. At resonance, the expanding cloud obeys the universal
equation of state for a non-relativistic gas, p
2
3
E = 0, where p is the pressure and E is the energy
density. Tuning away from resonance, we show that the observed deviation from scale invariant ow
can be explained by the change in the equation of state p
2
3
E = 0 (conformal symmetry breaking
parameter) and a nite bulk viscosity, which is compared to predictions.
PACS numbers: 03.75.Ss
The identication and comparison of scale-invariant
physical systems, dened as those without an intrinsic
length scale, has enabled signicant advances connect-
ing diverse elds of physics. Of recent interest is a deep
connection between scale-invariant strongly interacting
systems and their weakly interacting counterparts. An
important example is the Anti de Sitter-Conformal Field
Theory correspondence, which connects a broad class of
strongly interacting quantum elds to weakly interacting
gravitational elds in ve dimensions, enabling theoreti-
cal techniques that span high energy physics, condensed
matter, and ultracold atoms [1]. These techniques have
been used to predict a universal lower bound for the ra-
tio of shear viscosity to entropy density [2], connecting
quark-gluon-plasmas [3, 4] to ultra-cold Fermi gases [5
7].
An ultra-cold Fermi gas is a paradigm for scale-
invariant quantum uids, with the unique trait that in-
teractions between spin-up and spin-down atoms are tun-
able between two scale-invariant regimes, from an ideal,
noninteracting gas to the most strongly interacting, non-
relativistic quantum system known [8]. A central con-
nection between these two regimes is that in both cases
the pressure p and the energy density E are related by
p =
2
3
E [9, 10]. However, an obvious distinction be-
tween the ideal and strongly-interacting regimes was rst
demonstrated by observing the aspect ratio of a Fermi gas
after release from the trapping potential [11]. The ideal
gas was shown to expand ballistically with an isotropic
momentum distribution, whereas the strongly interact-
ing gas was found to expand hydrodynamically and to
exhibit anisotropic elliptic ow [3, 11]. Despite this
dierence, we demonstrate both theoretically and exper-
imentally that scale-invariance requires the mean square
cloud sizes to expand identically.
We report the observation and study of scale invari-
ance in an expanding Fermi gas by measuring the mean
square cloud size r
2
= x
2
+ y
2
+ z
2
as a function of
time after release. In the resonantly-interacting, scale in-
variant regime, where p
2
3
E = 0 and the bulk viscosity
vanishes, we predict and observe that r
2
expands bal-
listically and ts a universal curve for all initial energies.
In contrast, the aspect ratios exhibit energy-dependent
elliptic ow. Tuning away from resonance, where the
scattering length a is nite and scale invariance is bro-
ken, we show that the observed deviation of r
2
from
scale invariant ballistic expansion can explained by the
change in the equation of state and a nite bulk viscosity,
which is estimated and compared to predictions.
In the experiments [12], we employ an optically-
trapped cloud of
6
Li atoms in a 50-50 mixture of the
two lowest hyperne states, which is cooled by evapora-
tion [11]. The initial energy per particle E is measured
from the trapped cloud prole [12]. To observe the ex-
pansion dynamics, the cloud is released from the trap
and the cloud prole is measured as a function of time
after release in all three dimensions using two CCD cam-
eras, which simultaneously image dierent spin states to
avoid cross-saturation. As discussed in more detail be-
low, the cigar-shaped optical trap has a 2.6:1 transverse
aspect ratio, which enables an observation of transverse
elliptic ow on a short time scale with very good signal
to background ratio, and a precise measurement of the
static shear viscosity.
To characterize the degree to which the expansion
is scale-invariant, we measure the mean-square three-
dimensional cloud radius r
2
and compare to predic-
tions. Using the hydrodynamic equation for the velocity
eld v (including pressure and viscous forces), the conser-
vation equation for the density n, and conservation of to-
tal (kinetic, internal, and potential) energy, we obtain an
exact model-independent evolution equation for r
2
as a
function of time t after release of the cloud. Without any
simplifying assumptions, a single-component uid com-
prising N atoms of mass m obeys [12],
d
2
dt
2
mr
2

2
= x U
opt

0
+
3
N
_
d
3
x[(p) (p)
0
]

3
N
_
d
3
x
B
v, (1)
where the subscript (
0
) denotes the condition at t = 0,
2
0.0 0.5 1.0 1.5
0.4
0.6
0.8
1.0
1.2
1.4
1.6
t ms
A
s
p
e
c
t
R
a
t
i
o

y
FIG. 1: Transverse aspect ratio x/y versus time after re-
lease showing elliptic hydrodynamic ow: Top to bottom,
resonantly interacting gas, E = 0.66 EF, E = 0.89 EF,
E = 1.17 EF, E = 1.46 EF , ballistic gas E = 1.78 EF. Top
four solid curves: Hydrodynamic theory with the shear vis-
cosity as the only t parameter; Lower solid curve: Ballistic
theory with no free parameters. Error bars denote statistical
uctuations in the aspect ratio.
just after the optical trap is extinguished. The depar-
ture from scale-invariance is determined by the conformal
symmetry breaking parameter p/p, where p p
2
3
E
is nonzero in general, but vanishes in the scale-invariant
resonant regime, p =
2
3
E. The bulk viscosity
B
is
nonzero in general, but is predicted to vanish in the scale-
invariant regime [1316]. Further, the bulk viscosity fre-
quency sum rule vanishes when p = 0 [17]. For brevity,
we include only the optical trap potential U
opt
in Eq. 1,
which need not be harmonic. However, for our preci-
sion measurements, as described below, it is important
to include also the small potential energy arising from
the nite curvature of the bias magnetic eld [12]. Given
the time-dependent volume integrals of p and
B
, Eq. 1
can be integrated to determine r
2
at time t from the
measured initial value of r
2

0
and the initial condition,
dr
2

0
/dt = 0. As r
2
is a scalar, the contribution of
the shear viscosity pressure tensor vanishes, since it is
traceless. Hence, Eq. 1 is independent of the shear vis-
cosity, enabling a very sensitive measurement of the bulk
viscosity, where the divergence of the velocity eld v is
easily determined as described below. By tuning away
from resonance, we nd that observed deviations from
scale invariance can be explained by the pressure change
p, which is an odd function of 1/a, and a small con-
tribution, proportional to 1/a
2
, which we attribute to a
nonzero bulk viscosity.
For a scale-invariant hydrodynamically expanding
cloud, where both p and
B
are zero, Eq. 1 yields
r
2
= r
2

0
+
t
2
m
x U
opt

0
, (2)
which corresponds to ballistic expansion of the mean
square cloud size, even though the individual cloud radii
0.6 0.8 1.0 1.2 1.4 1.6
0
1
2
3
EE
F
V
i
s
c
o
s
i
t
y
C
o
e
f
f
i
c
i
e
n
t
s
FIG. 2: Measurement of bulk and shear viscosity for
a scale-invariant Fermi gas: Blue (top): Trap-averaged
shear viscosity coecient

d
3
x/(Nh) S versus energy
E/EF. Red (bottom): Trap-averaged bulk viscosity coe-
cient

d
3
xB/(Nh) B versus energy. The weighted aver-
age value of B = 0.005(0.016) is consistent with zero. (Dot-
ted curves added to guide the eye.)
expand hydrodynamically and exhibit elliptic ow, as
studied previously [11] and shown in Fig. 1 for the trans-
verse aspect ratio. Indeed, numerical modeling of hy-
drodynamic expansion of the cloud in the scale-invariant
regime shows that Eq. 2 holds even for large shear viscos-
ity, provided that both dissipative forces and heating are
properly included in the hydrodynamic equations [6, 7]
to assure that the total energy is conserved [12].
The aspect ratio
x
/
y
of the cloud is measured at
the Feshbach resonance (834 G) as a function of time
after release to establish that the ow is hydrodynamic
and to determine the shear viscosity. Fig. 1 shows data
for E/E
F
= 0.66, 0.89, 1.17, 1.46, where E x U
0
and E
F

h
2
k
2
FI
2m
is the measured Fermi energy of an
ideal gas at the trap center with k
FI
the corresponding
wavevector [12]. In contrast to the ballistic expansion
data (E/E
F
= 1.78), where the aspect ratio saturates
to unity, for the resonantly interacting cloud,
x
/
y
in-
creases to approximately 1.5 over the time range shown,
clearly demonstrating that the cloud expands hydrody-
namically. The rate at which the aspect ratio increases
with time substantially decreases as the energy is in-
creased, due to the increase in the shear viscosity.
The aspect ratio in the x y plane is very sensitive to
the shear viscosity, which slows the ow in the rapidly ex-
panding x-direction and increases the speed in the more
slowly expanding y-direction. To measure the shear vis-
cosity at 834 G, we t the aspect ratio data of Fig. 1 using
a general, energy-conserving hydrodynamic model [6, 7],
valid in the scale-invariant regime where p = 0. At the
Feshbach resonance, the shear viscosity takes the form
=
S
hn, where n is the total density of atoms and

S
is a dimensionless function of the local reduced tem-
perature. The trap-averaged shear viscosity coecient

S

_
d
3
x/( hN) is used as the only free parameter,
3
0.0 0.5 1.0 1.5 2.0
0
1
2
3
4
t ms

m
s
2

FIG. 3: Scale invariant expansion of a resonantly interacting

Fermi gas. Experimental values of
2
(t) m[r
2
r
2
0]/x
U0 versus time t after release, for the same data as in
Fig. 1 (including ballistic data) collapse onto a single curve,
demonstrating universal t
2
scaling. Dashed curve
2
(t) =
t
2
, as predicted by Eq. 2. Note that the small correction
r
2
Mag arising from the curvature in the bias magnetic
eld is subtracted from the r
2
data to obtain the
2
(t) data
that is shown in the gure [12].
initially neglecting the bulk viscosity, which is expected
to be much smaller. For the shear viscosity in the scale-
invariant regime,
S
is an adiabatic invariant, which is
therefore temporally constant in the adiabatic approxi-
mation [6, 7]. The ts to the aspect ratio obtained in
this way are shown in Fig. 1 as solid lines. We nd that
the measured shear viscosity coecients, Fig. 2, are sig-
nicantly smaller than those obtained from our previ-
ous measurements in the same energy range, based on
collective mode damping [6, 7]. We note that collective
mode measurements determine the shear viscosity at the
collective mode frequency and may be subject to excess
damping arising near the cloud edges, where the gas is
ballistic. As the frequency scale for expansion is essen-
tially the inverse of the expansion time, the expansion
measurements correspond to a low frequency, and hence
measure the static shear viscosity on a time scale where
the gas is likely to maintain thermal equilibrium and the
stress tensor has relaxed to the Navier-Stokes form [18].
Scale invariance of the expanding gas is now directly
demonstrated by measuring
2
(t) m[r
2
r
2

0
]/x
U
opt

0
, which should obey
2
(t) = t
2
for a scale-
invariant system, according to Eq. 2 [12]. Fig. 3 shows
the experimental values of
2
(t) versus t for the same
data as used in Fig. 1. In contrast to the aspect ratio
versus time data of Fig. 1, which vary substantially with
energy due to the shear viscosity, the experimental values
of
2
(t) fall on a single t
2
curve.
We estimate the bulk viscosity at resonance by assum-
ing that p = p
2
3
E = 0 in Eq. 1, which follows from the
universal hypothesis [9, 10]. This equation of state for a
resonantly interacting Fermi gas has been veried experi-
mentally to high precision [19]. We therefore assume that
0.0 0.5 1.0 1.5 2.0
0
1
2
3
4
t ms

m
s
2

FIG. 4: Breaking of scale invariance in the expansion for a

Fermi gas near a Feshbach resonance. The data are the ex-
perimental values of
2
(t) m[r
2
r
2
0]/x U0 for
E/EF 1.0, versus time t after release. Solid curves are
the predictions using Eq. 1 with B = 0, where the pressure
change p is approximated using the second virial coecient
without any free parameters [12]. Top: 1/(kFIa) = 0.59;
Center: 1/(kFI a) = 0; Bottom: 1/(kFIa) = +0.61.
at resonance the bulk viscosity term in Eq. 1 produces
the only deviation from scale-invariance in the evolution
of r
2
. Analogous to the shear viscosity, we take the
bulk viscosity to be of the form
B
=
B
hn, where
B
is dimensionless, and consider rst a large nite scatter-
ing length. Since the bulk viscosity must be positive, the
leading contribution in powers of the inverse scattering
length must be of the form
B
= f
B
() h n/(k
F
a)
2
, where
k
F
= (3
2
n)
1/3
is the local Fermi wavevector. Here
f
B
() is a dimensionless function of the reduced temper-
ature , which is an adiabatic invariant, and hence time-
independent in the adiabatic approximation that we use
for the small bulk viscosity contribution. As the cloud
expands, the density decreases as n 1/ in the scaling
approximation, where the ts to the aspect ratio data
in all three dimensions accurately determine the volume
scale factor (t) [12]. Since 1/(k
F
a)
2

2/3
, the trap-
averaged bulk viscosity coecient,
B

_
d
3
x
B
/( hN)
is time-dependent and scales as

B
(t) =
B
(0)
2/3
(t). (3)
With the scaling approximation v =

/,
the bulk viscosity term then takes the simple form
3 h
B
(0)

/
1/3
. We determine
B
(0) with high pre-
cision, by using a least-squares t of Eq. 1 to the mea-
sured r
2
data [12]. In contrast to the shear viscosity
coecient, which increases with increasing energy, the
bulk viscosity coecient at resonance remains nearly zero
over the entire energy range, Fig. 2. We nd that the
weighted-average
B
(0) = 0.005(0.016), which is consis-
tent with zero, as predicted for a scale-invariant cloud[13
16].
We investigate the departure from scale invariance and
the thermodynamics of the expanding gas at nite scat-
4
0 0.5 1 1.5 2
1.5
1
0.5
0
0.5
1
1.5
p scaling

B
s
c
a
l
i
n
g
1. 1.2 1.4 1.6 1.8
0.8
0.9
1.0
1.1
1.2
EE
F

c
1

c
0

c
1

c
0

8
3
4
FIG. 5: Pressure change p and bulk viscosity B contribu-
tions to conformal symmetry breaking as a function of en-
ergy. The expansion data are t with r
2
= c0 + c1 t
2
.
The ratio (c1/c0)/(c1/c0)834 is shown for the resonantly in-
teracting gas 1/(kFI a) = 0 (black dots, black line-theory), for
1/(kFI a) = 0.59 (top, red dots) and for 1/(kFIa) = +0.61
(bottom, blue dots). The ratios are compared to predictions
using two scaling parameters, B for B of Ref. [16] and p for
p based on the second virial coecient approximation. The
contour plot shows
2
as a function of B and p. Solid curves
top and bottom show the best t, where p = 1.08(0.32) and
B = 0.16(0.61). The dashed (dotted) curves show the pre-
dictions for p = 1 and B = 0 (B = 1), to illustrate the
eect of the bulk viscosity.
tering length by tuning the bias magnetic eld above and
below the Feshbach resonance. We measure r
2
r
2

0
for E 1.0 E
F
, as shown in Fig. 4. Compared to the res-
onant case, we see qualitatively that the cloud expands
more rapidly when the scattering length is negative and
more slowly when the scattering length is positive. This
behavior is a signature of the [(p)(p)
0
] term in Eq. 1,
where |p(t)| |p(0)| for any time t after release and
p has the same sign as the scattering length.
To obtain a model-independent quantitative compari-
son with predictions for nite scattering length, we pa-
rameterize the data using a
2
t with r
2
= c
0
+ c
1
t
2
,
where c
1
is the eective coecient of t
2
[12]. This method
avoids utilizing model-dependent expansion factors for
the cloud radii in the data analysis. We determine c
0
and c
1
parameters for both the on-resonance and o-
resonance data at several dierent energies. We correct
all of the c
1
/c
0
ratios for the eective potential arising
from magnetic eld curvature [12], and nally determine
the ratio (c
1
/c
0
)/(c
1
/c
0
)
834
, Fig. 5. If p and
B
were
zero at all magnetic elds, then this ratio would be unity
everywhere, corresponding to the black line (black dots)
in the gure. The systematic deviation from unity arises
from nite p and
B
, where the red dots (top) show
data at 986 G where a < 0 and the blue dots (bottom)
show data for 760 G, where a > 0.
We t the data shown in Fig. 5, by modeling the c
1
/c
0
ratio as a function of energy using two parameters, a
scale factor
p
for p calculated to second order in fu-
gacity (high temperature limit) [12, 20] and a scale fac-
tor
B
for the high temperature bulk viscosity predicted
in Ref. [16]. We assume that all contributions to p
that require three-body and higher order interactions to
maintain equilibrium are frozen over the expansion time
scale, and therefore do not aect (p) (p)
0
in Eq. 1.
In particular, the molecular contribution to the second
order virial coecient can be neglected. Retaining only
the translational degrees of freedom in the two-body scat-
tering contribution, p is evaluated using an adiabatic
approximation for the translational temperature in this
small correction, so that p is then a known function of
time and is odd in 1/a [12].
The bulk viscosity at nite scattering length [16] makes
the only 1/a
2
correction in Eq. 1, since the 1/a
2
part of
p is generally time-independent and does not contribute
to (p) (p)
0
[12]. Dusling and Sch afer [16] point out
that the leading order bulk viscosity must be second or-
der in the fugacity z n
3
T
/2 in the high-temperature
limit. They obtain the form
B
=
1
24

2
T
a
2
h

3
T
z
2
, where

T

h

2mkBT
is the thermal wavelength. Assuming
that the temperature evolves adiabatically in this small
bulk viscosity term, the integral over the trap volume [12]
takes the same form as Eq. 3, with
B
(0) = c
B
_
EF
E
_
4
and c
B
=
9
32
1
(kFI a)
2
, which is symmetric in 1/a in con-
trast to p, which changes sign with 1/a.
Our measurements provide a test of the degree to
which equilibrium thermodynamics holds in the expan-
sion of an ultra-cold cloud. In Fig. 5, the
2
contour
plot as a function of
p
and
B
shows a maximum for

p
= 1.08(0.32) and
B
= 0.16(0.61). To show the rel-
ative scale of the bulk viscosity and the p corrections,
the predictions for
p
= 1 and
B
= 0 are shown as
dashed curves, while the dotted curves show the predic-
tions for
p
= 1 and
B
= 1. The contribution of the bulk
viscosity appears smaller than predicted. As
p
1, the
observed breaking of scale invariance is primarily due the
direct change in the pressure p = p
2
3
E, and our p
model adequately describes the data, showing that the
pressure in the expanding cloud is not far from thermal
equilibrium in the translational degrees of freedom.
This research is supported by the Physics Division of
5
the National Science Foundation (Quantum transport in
strongly interacting Fermi gases) and by the Division of
Materials Science and Engineering, the Oce of Basic
Energy Sciences, Oce of Science, U.S. Department of
Energy (Thermodynamics in strongly correlated Fermi
gases). Additional support has been provided by the
Physics Divisions of the Army Research Oce and the
Air Force Oce of Scientic Research. The authors are
pleased to acknowledge K. Dusling and T. Sch afer, North
Carolina State University, for stimulating conversations.
[1] A. Adams, L. D. Carr, T. Schafer, P. Steinberg, , and
J. E. Thomas, New J. Phys. 14, 115009 (2012).
[2] P. K. Kovtun, D. T. Son, and A. O. Starinets, Phys. Rev.
Lett. 94, 111601 (2005).
[3] P. F. Kolb and U. Heinz, Quark Gluon Plasma 3 (World
Scientic, 2003), p. 634.
[4] E. Shuryak, Prog. Part. Nucl. Phys. 53, 273 (2004).
[5] T. Schafer, Phys. Rev. A 76, 063618 (2007).
[6] C. Cao, E. Elliott, J. Joseph, H. Wu, J. Petricka,
T. Schafer, and J. E. Thomas, Science 331, 58 (2011).
[7] C. Cao, E. Elliott, H. Wu, and J. E. Thomas, New J.
Phys. 13, 075007 (2011).
[8] G. Rupak and T. Schafer, Phys. Rev. A 76, 053607
(2007).
[9] T.-L. Ho, Phys. Rev. Lett. 92, 090402 (2004).
[10] J. E. Thomas, J. Kinast, and A. Turlapov, Phys. Rev.
Lett. 95, 120402 (2005).
[11] K. M. OHara, S. L. Hemmer, M. E. Gehm, S. R.
Granade, and J. E. Thomas, Science 298, 2179 (2002).
[12] See Appendix: Supplemental Material.
[13] D. T. Son, Phys. Rev. Lett. 98, 020604 (2007).
[14] M. A. Escobedo, M. Mannarelli, and C. Manuel, Phys.
Rev. A 79, 063623 (2009).
[15] Y.-H. Hou, L. P. Pitaevskii, and S. Stringari (2013),
arXiv:1302.2258v1 [cond-mat.quant-gas].
[16] K. Dusling and T. Schafer (2013), arXiv:13054688v1
[cond-mat.quant-gas].
[17] E. Taylor and M. Randeria, Phys. Rev. A 81, 053610
(2010).
[18] This idea was suggested to us by Thomas Schafer, North
Carolina State University, private communication.
[19] M. Ku, A. T. Sommer, L. W. Cheuk, and M. W. Zwier-
lein, Science (2012).
[20] T.-L. Ho and E. Mueller, Phys. Rev. Lett. 92, 160404
(2004).
[21] L. Luo and J. E. Thomas, J. Low Temp. Phys. 154, 1
(2009).
[22] F. Werner, Phys. Rev. A 78, 025601 (2008).
Appendix A: Supplemental Material
1. Experimental Methods
In the experiments, we employ an optically-trapped
cloud of
6
Li atoms in a 50-50 mixture of the two low-
est hyperne states, which is tuned to a broad collisional
(Feshbach) resonance in a bias magnetic eld of 834 G,
and cooled by evaporation. The initial energy per par-
ticle E is measured from the trapped cloud prole, as
discussed below. A focused CO
2
laser beam forms the
cigar-shaped optical trap with a transverse aspect ratio
of 2.6:1, which enables an observation of transverse el-
liptic ow and a precise measurement of the static shear
viscosity even at low temperature, where the shear viscos-
ity is small. The transverse aspect ratio is controlled by
using two sets of cylindrical ZnSe lenses, one set just after
the acousto-optic modulator that controls the laser inten-
sity and one just before an expansion telescope, which is
used to increase the beam radii before focusing into a
high vacuum chamber, where the optical trap is loaded
from a standard magneto-optical trap. The rst set of
cylindrical lenses controls the transverse aspect ratio of
focused beam, while the second set enables matching of
the beam curvatures to achieve a common focal plane.
To observe the expansion dynamics, the cloud is re-
leased from the trap and the cloud prole is measured
as a function of time after release in all three dimen-
sions using two identical CCD cameras, Fig. 6 which si-
multaneously image dierent spin states to avoid cross-
saturation. The magnications of the imaging systems
are measured by translating the trap focus. The mea-
sured magnications yield average axial dimensions
z
that are consistent within 1%. To obtain the most pre-
cise measurements of the cloud prole, we adjust the
eective magnication of one camera so that the aver-
age axial dimensions precisely agree. In this way, the
cloud radii
i
in all three dimensions are consistently
measured, to determine the aspect ratios
x
/
y
,
x
/
z
,
and
y
/
z
, as well as the mean square cloud radius, r
2
.
Two-dimensional density distributions are t to the cloud
proles to extract the cloud radii. For fast data handling,
gaussian proles are assumed for most of the data and
a zero-temperature Thomas-Fermi prole is assumed for
the lowest energies. Both types of t proles are com-
pared to full nite-temperature Thomas-Fermi proles
to estimate multiplicative corrections to the cloud radii,
which are needed to correct for the small error arising
from the form of the t functions.
We derive an exact, model-independent evolution
equation for r
2
based on hydrodynamics and energy
conservation in A2. This enables precise characteri-
zation of scale-invariance and thermodynamics in an ex-
panding cloud. The primary result, Eq. A17, is indepen-
dent of the shear viscosity and includes the corrections
to the ow arising from the bulk viscosity and the devia-
tion p p
2
3
E of the pressure from the scale invariant
equation of state, p =
2
3
E. We also include the potential
energy arising from the nite bias magnetic eld curva-
ture, as required for our precision measurements.
The pressure change p is determined for the high
6
FIG. 6: Imaging the expanding cloud in three dimensions.
Two CCD cameras are used to measure the density prole of
the cloud. The cloud is released from an asymmetric optical
trap with a 2.6 : 1 transverse aspect ratio, enabling observa-
tion of elliptic ow in the x y plane.
temperature limit in A3. Then the method of estimat-
ing the bulk viscosity is described in A4. Finally, we
discuss the method of tting the mean square size r
2

data in the resonantly and interacting regimes A5.

2. Expansion of the Mean-Square Cloud Radius
We employ a hydrodynamic description for a single
component uid [6, 7], where the velocity eld v(x, t)
is determined by the scalar pressure and the viscosity
pressure tensor,
nm(
t
+v ) v
i
=
i
p +

j
(
ij
+
B

ij
)
n
i
U
total
, (A1)
where p is the scalar pressure and m is the atom mass.
U
total
is the total trapping potential energy arising from
the optical trap and the bias magnetic eld curvature.
The second term on the right describes the friction forces
arising from both shear and bulk
B
viscosities, where

ij
= v
i
/x
j
+ v
j
/x
i
2
ij
v/3 and

v.
Current conservation for the density n(x, t) requires
n
t
+ (nv) = 0. (A2)
Finally, conservation of total energy yields
d
dt
_
d
3
x
_
n
1
2
mv
2
+E + nU
total
_
= 0. (A3)
Here, the rst term is the kinetic energy arising from the
velocity eld and E is the internal energy density of the
gas. As shown below, Eq. A3 will play an important
role in determining a general evolution equation for the
volume integral of the pressure in both the scale-invariant
regime and away from scale-invariance.
To explore scale invariance for an expanding cloud,
we measure the mean-square cloud radius, r
2
, which
is a scalar quantity. In this section, we derive gener-
ally the equation of motion for r
2
, with no simplify-
ing assumptions, except that of a single component uid,
which is appropriate in the normal uid regime above
the superuid transition temperature as well as in the
superuid regime when the normal and superuid com-
ponents move together [15]. We show that in the scale-
invariant regime at a Feshbach resonance, where p =
2
3
E
and
B
= 0, conservation of total energy leads to bal-
listic expansion of r
2
for a hydrodynamic gas. Away
from resonance, the departure from scale-invariance is
determined by the change in the equation of state, char-
acterized by p p
2
3
E and nite
B
, permitting a
measurement of the bulk viscosity, given p.
We begin by noting that for each direction i = x, y, z,
the mean square size x
2
i

1
N
_
d
3
xn(x, t) x
2
i
obeys
dx
2
i

dt
=
1
N
_
d
3
x
n
t
x
2
i
=
1
N
_
d
3
x[ (nv)]x
2
i
=
1
N
_
d
3
xnv x
2
i
= 2x
i
v
i
, (A4)
where N is the total number of atoms. We have used
integration by parts and n = 0 for x
i
to obtain
the second line. Similarly,
dx
i
v
i

dt
=
1
N
_
d
3
xnx
i
v
i
t
+
1
N
_
d
3
x
n
t
x
i
v
i
=
1
N
_
d
3
xnx
i
v
i
t
+
1
N
_
d
3
xnv (x
i
v
i
)
= x
i
(
t
+v )v
i
+v
2
i
. (A5)
Combining Eq. A4 and Eq. A5, we obtain,
d
2
dt
2
x
2
i

2
= x
i
(
t
+v )v
i
+v
2
i
. (A6)
To proceed, we use Eq. A1, which yields
_
d
3
xnx
i
(
t
+ v )v
i
=
1
m
_
d
3
xx
i
(
i
p n
i
U
total
)
+
1
m

j
_
d
3
xx
i

j
(
ij
+
B

ij
)
Integrating by parts on the right hand side, assuming
that the surface terms vanish, we obtain
x
i
(
t
+v )v
i
=
1
Nm
_
d
3
xp
1
m
x
i

i
U
total

1
Nm
_
d
3
x(
ii
+
B

)(A7)
with

v. Dening the viscosity coecients

S
and
B
by
S
hn and
B

B
hn, respectively, we
7
can write,
x
i
(
t
+v )v
i
=
1
Nm
_
d
3
xp
1
m
x
i

i
U
total

h
m

S

ii
+
B

, (A8)
where

S

ii
+
B

1
N
_
d
3
xn(
S

ii
+
B

). (A9)
Using Eq. A8 in Eq. A6, we then obtain for one direction
x
i
,
d
2
dt
2
x
2
i

2
=
1
Nm
_
d
3
xp +v
2
i

1
m
x
i

i
U
total

h
m

S

ii
+
B

. (A10)
Eq. A10 determines the evolution of the mean square
cloud radii along each axis, x
2
i
, which depends on the
conservative forces arising from the scalar pressure and
the trap potential, as well as the viscous forces arising
from the shear and bulk viscosities.
Summing Eq. A10 over all three directions, the shear
viscosity term vanishes, since
ij
is traceless, yielding
d
2
dt
2
r
2

2
=
3
Nm
_
d
3
xp +v
2

1
m
x U
total

3 h
m

B
v. (A11)
At t = 0

, before release from the trap, v = 0, Eq. A11

shows that the volume integral of the pressure is,
3
N
_
d
3
xp
0
= x U
total

0
, (A12)
where the subscript ( )
0
denotes the initial condition.
Here U
total
= U
Opt
+U
Mag
is the total trapping potential,
comprising an optical component from the laser trap and
a magnetic component arising from the curvature of the
bias magnetic eld used in the experiments, as described
further below.
It will be convenient to rewrite Eq. A11 using
p =
2
3
E +
_
p
2
3
E
_
, (A13)
where the rst term denes the equation of state for
the pressure in the scale-invariant regime, and the sec-
ond term describes the departure from scale invariance.
Then,
d
2
dt
2
r
2

2
=
2
Nm
_
d
3
xE +v
2
+
3
Nm
_
d
3
x
_
p
2
3
E
_

1
m
x U
total

3 h
m

B
v. (A14)
Just after release from the trap, t 0
+
, from the
optical trap, the trapping potential changes abruptly,
U
total
U
Mag
. To determine the evolution of r
2
af-
ter release, we use total energy conservation to eliminate
v
2
from Eq. A14. From Eq. A3, the nal total energy
per particle is equal to the initial total energy. Then,
1
N
_
d
3
xE +
m
2
v
2
+U
Mag
=
1
N
_
d
3
xE
0
+U
Mag

0
.
(A15)
To determine the initial internal energy,
1
N
_
d
3
xE
0
,
we use E
0
=
3
2
p
0

3
2
(p
2
3
E)
0
. With Eq. A12, this yields
1
N
_
d
3
xE
0
=
1
2
x U
total

3
2N
_
d
3
x
_
p
2
3
E
_
0
.
(A16)
We have x U
total

0
= x U
Opt

0
+ x U
Mag

0
in Eq. A16. Multiplying Eq. A16 by 2/m determines the
rst two terms in Eq. A14. Then, with x U
total

x U
Mag
for t 0
+
in Eq. A14, we obtain nally our
central result for studying scale invariance,
d
2
dt
2
mr
2

2
= x U
Opt

0
+
3
N
_
d
3
x[(p) (p)
0
]
3 h
B
v + U
Mag
, (A17)
where
p p
2
3
E (A18)
describes the departure of the pressure from the scale-
invariant regime and
U
Mag
2U
Mag

0
+xU
Mag

0
2U
Mag
xU
Mag

(A19)
corrects for the small potential energy arising from the
bias magnetic eld curvature. As the bias coils are ori-
ented along the x direction, the eective potential is re-
pulsive along the x axis and twice the magnitude of the
attractive potential along y and z,
U
Mag
(x) =
1
2
m
2
Mag
_
y
2
+ z
2
2x
2
_
, (A20)
where
Mag
= 2 21.5(0.25) Hz is measured by from
the oscillation frequency of the cloud in the y z plane.
The rst three terms in Eq. A17 reproduce the Eq. 1
of the main text, where the small magnetic contribution
was neglected for brevity. The potential energy arising
from the magnetic eld curvature depends on the mean
square cloud radii, x
2
i
, which are determined as a func-
tion of time by tting the aspect ratio data using a scaling
approximation for the density prole. x U
Opt

0
is de-
termined by parametric resonance measurements of the
trap frequencies and measurements of the cloud radii,
as discussed in more detail below. The cloud radii for
t = 0
+
are dominated by z
2

0
, the longest direction of
the cigar-shaped cloud in the trap. This is consistently
measured both by in-situ imaging and by measurements
after expansion, using the calculated expansion factor,
which is close to unity.
8
Eq. A17 is easily integrated using the initial conditions
r
2

0
and
t
r
2

0
= 0. To clearly demonstrate scale-
invariant expansion after the optical trap is extinguished,
we integrate Eq. A17 in two steps. First, we integrate the
magnetic contribution, r
2

Mag
determined from
d
2
dt
2
mr
2

Mag
2
=
1
2
m
2
Mag
_
y
2

0
b
2
y
(t) +z
2

0
b
2
z
(t) 2x
2

0
b
2
x
(t)

. (A21)
We employ homogeneous initial conditions for the mag-
netic contribution, r
2

0 Mag
= 0 (which does not
change r
2

0
) and
t
r
2

0 Mag
= 0. The time depen-
dent expansion factors b
i
(t) are determined by tting
the measured cloud radii as a function of time after re-
lease, using the hydrodynamic equations, Eq. A10 in a
scaling approximation [6, 7]. As the magnetic contribu-
tion to r
2
arising from Eq. A21 is only a few percent,
the expansion factors are readily determined with su-
cient precision. After integration, the quantity r
2

Mag
is subtracted from the measured r
2
data to determine
the eective r
2
, which then expands according to the
remaining terms in Eq. A17,
d
2
dt
2
mr
2

2
= x U
Opt

0
+
3
N
_
d
3
x[(p) (p)
0
]
3 h
B
v. (A22)
The rst term is the optical trap contribution, which is
dominant and leads to a t
2
scaling for the mean square
cloud radius of a resonantly interacting gas or for a bal-
listic gas, with the initial condition r
2

0
. The remaining
small pressure change and bulk viscosity terms can be in-
tegrated separately, using the scale factors b
i
(t) and the
same homogeneous initial conditions as for the magnetic
contribution. These contributions are described in more
detail below.
3. High Temperature Approximation to p
We study the regime away from the Feshbach reso-
nance by using in Eq. A17 a nonzero correction to the
pressure p. As a consequence of energy conservation,
only the dierence between the pressure at time t and
at time t = 0, i.e., (p) (p)
0
, appears in Eq. A17.
Hence, any static contribution to p has no eect.
In the high-temperature limit, we can evaluate p us-
ing a virial expansion [20]. To determine p, we make
the assumption that three-body and higher order contri-
butions are frozen at their equilibrium values over the
time scale of the expansion, and do not contribute to
(p) (p)
0
. In particular, the molecular contribution
to the second virial coecient [20] requires three-body
collisions to populate and depopulate the molecular state
as the gas expands and cools in the translational degrees
of freedom. Therefore, the molecular contribution is ex-
pected to be negligible.
With the frozen equilibrium hypothesis, we neglect
three-body and higher order processes, and evaluate p
in the high temperature limit to second order in the fu-
gacity [20] , where p = p
2
3
E is given by,
p = p
2
3
E =

2
3
nk
B
T
_
T
b
2
T
_
(n
3
T
). (A23)
Here
T
h/

2mk
B
T is the thermal wavelength and
b
2
is the part of the second virial coecient that describes
two-body interactions. Ignoring the molecular contribu-
tion, which is frozen on the short time of the expansion
as discussed above, we take
b
2
(x) =
sgn[a]
2
e
x
2
erfc(x), (A24)
where erfc(x) = 1
2

_
x
0
dx

e
x
2
and x =
T
|a|

2
, with
a the s-wave scattering length.
As p causes only a small perturbation to the ow, we
make an adiabatic approximation for the temperature,
T = T
0

2/3
, where T
0
is the initial temperature of the
trapped cloud and = b
x
b
y
b
z
is the volume scale factor,
i.e., the density n of the expanding gas scales as 1/.
We determine by tting the aspect ratio data with a
scaling approximation to the hydrodynamics [6, 7], using
the shear viscosity as the only free parameter. Then, x =
x
0

1/3
, where x
0
=
T
0
|a|

2
. Using the high temperature
approximation for the energy per particle, E = 3 k
B
T
0
and E
F
=
h
2
k
2
FI
2m
= (3N)
1/3
h , the Fermi energy of an
ideal gas at the cloud center, we have
x = x
0

1/3
x
0
=

6
k
FI
|a|
_
E
F
E
_
1/2
, (A25)
where k
FI
=
_
2mE
F
/h
2
is the Fermi wavevector of an
ideal gas at the trap center. Note that E
F
is measured
using the total atom number and the oscillation frequen-
cies in the trap, (
x

z
)
1/3
, where the
i
are given
in A5.
Now, T
b2
T
= xb

2
/2, where b

2
(x) sgn[a] f

2
(x) with
f

2
(x)
_
1

xe
x
2
erfc(x)
_
. (A26)
Integrating over the trap volume, and using the adiabatic
approximation for the temperature and a scaling approx-
imation for the density, we obtain
1
N
_
d
3
x
_
p
2
3
E
_
=

2
3
k
B
T
0

2/3
z [xb

2
(x)], (A27)
where the trap-averaged fugacity z is an adiabatic invari-
ant. For a gaussian density prole,
z
1
N
_
d
3
xn
_
n
3
T0
2
_
=
9
4

2
_
E
F
E
_
3
. (A28)
9
In Eq. A25 and Eq. A28, we have made the harmonic ap-
proximation
2
x
x
2

0
=
2
y
y
2

0
=
2
z
x
2

0
and we have
used the high temperature approximation E = 3 k
B
T
0
,
We determine the energy E and temperature T
0
, using
E = 3 k
B
T
0
= 3
_
z
U
total
z
_
0
= x U
total

0
, (A29)
which follows from force balance in the trapping potential
and p = n(0)k
B
T
0
in the high temperature limit. This
approximation is adequate for evaluating p in the sec-
ond virial approximation. We discuss the measurement
of z
z
U
total

0
using the axial cloud prole in A5.
To use Eq. A17 to determine r
2
as a function of time,
the volume integral Eq. A27 is written as,
1
N
_
d
3
xp =
x U
Trap

6
4
_
E
F
E
_
7/2

1/3
k
FI
a
f

2
(x),
(A30)
where the time dependence of x is determined by Eq. A25
and f

2
(x) is given by Eq. A26. We note that the total
dependence of p on the Fermi energy is E
7/2
F
/k
FI
E
3
F
which is proportional to N, the total atom number, as it
should be. We have used sgn[a]/|a| = 1/a to explicitly
show that the volume integral of p changes sign with
the scattering length as the bias magnetic eld is tuned
across the Feshbach resonance.
As discussed above, the net pressure correction in
Eq. A17 is (p)(p)
0
. Hence, we also evaluate Eq. A30
in the limit t = 0, where 1 and x x
0
. As
|(p)| |(p)
0
| for all t, the net pressure correction is
positive (negative) when p is negative (positive). Then,
compared to the resonant case, where 1/(k
FI
a) = 0,
the cloud is expected to expand more rapidly when the
scattering length 1/(k
FI
a) < 0 and more slowly when
1/(k
FI
a) > 0, as observed in the experiments (see the
main text).
4. Estimating the Bulk Viscosity
By tuning the bias magnetic eld away from the Fesh-
bach resonance, we explore the departure from the scale-
invariant regime, where the correction to the pressure
p p
2
3
E is nonzero, as described above. In this
case, the bulk viscosity can be nonzero. We estimate the
bulk viscosity by measuring r
2
and tting the data us-
ing Eq. A17. This measurement exploits the fact that
the bulk viscosity provides the only contribution propor-
tional to
1
a
2
, while the
1
a
2
contribution to the volume
integral of (p) (p)
0
generally vanishes, as we now
show.
The bulk viscosity is positive for nite a and must
vanish in the scale-invariant regime, where |a| .
Hence, to leading order in 1/a, the bulk viscosity must be
quadratic in
1
kF a
. Using dimensional analysis, the bulk
viscosity then takes the general form

B
= hn
f
B
()
(k
F
a)
2
, (A31)
where hn is the natural unit of viscosity, k
F
n
1/3
is
the local Fermi wavevector and f
B
() is a dimensionless
function of the reduced temperature T/T
F
(n), where
k
B
T
F
(n)
F
(n) n
2/3
is the local Fermi energy. As
discussed in the main text, by using an adiabatic approxi-
mation for , one obtains the trap-averaged bulk viscosity
coecient, which takes the form

B
(t) =
B
(0)
2/3
(t), (A32)
where the
2/3
factor arises from the 1/k
2
F
scaling.
Again using dimensional analysis, the most general
1/(k
F
a)
2
contribution to p, which we dene as p
2
must be of the form
p
2
= n
F
(n)
f
p
()
(k
F
a)
2
,
where
F
(n) is the local Fermi energy and f
p
() is a di-
mensionless function of the reduced temperature. As
F (n)
(kF a)
2
=
h
2
2ma
2
,
1
N
_
d
3
xp =
h
2
2ma
2
f
p
() is time-
independent, since the number of atoms in a vol-
ume element is conserved during the expansion and
1
N
_
d
3
xnf
p
() is constant in the adiabatic approxima-
tion. Hence, the
1
a
2
part of (p) (p)
0
in Eq. A17
vanishes and the bulk viscosity provides the only time-
dependent
1
a
2
contribution, which increases as
2/3
, ac-
cording to Eq. A32.
As described in the main text, the estimation of the
bulk viscosity is accomplished using a
2
t to the data
for the mean square cloud radii as a function of time t
after release of the cloud with the t function
r
2
= c
0
+ c
1
t
2
. (A33)
Eq. A33 is exactly valid for a resonantly interacting gas
and for a ballistic gas. For the o-resonance experiments,
the c
1
determined from the t function is an average
value, which is perturbed by p and
B
(0).
We compare the tted
B
(0) to the recent prediction
of Dusling and Sch afer, Ref. [16] for the bulk viscosity in
the high temperature limit, which is second order in the
fugacity z,

B
= c
B
_

T
a
_
2
h

3
T
z
2
. (A34)
Here we can approximate z = n
3
T
/2 and Dusling and
Sch afer give c
B
=
1
24

2
.
Integrating over the trap volume, we obtain
B
(t) =

B
(0)
2/3
(t) as in Eq. A32 with

B
(0) = c
B
27
2

2
1
(k
FI
a)
2
_
E
F
E
_
4
c
B
_
E
F
E
_
4
.
(A35)
Here, the value of c
B
based on the prediction is
c
B
=
9
32
1
(k
FI
a)
2
, (A36)
which is measurable in the experiments.
10
5. Fitting r
2
= c0 + c1 t
2
for Resonant and Ballistic
Gases.
We begin by noting that the energy scale

E in the
experiments is twice the total potential energy in the
harmonic oscillator approximation. In terms of the mea-
sured mean square cloud size along the axial z-direction,
we have

E = 3 m(
2
z opt
+
2
Mag
)z
2

0
, (A37)
where
z opt
is the axial oscillation frequency for atoms
in the optical trap, in the harmonic approximation, and

Mag
is the corresponding oscillation frequency arising
from the bias magnetic eld curvature. For the reso-
nantly interacting gas, where the virial theorem holds,

E is precisely the energy of the cloud for a harmonic

trapping potential. For an anharmonic trap, the virial
theorem gives the total energy for the resonantly inter-
acting gas in terms of the trapping potential [21], but
at nite scattering length, the relation between the total
energy and the trapping potential energy is scattering
length dependent [22]. By using

E to characterize the
energy, we avoid the scattering length dependence.
As discussed in A3, for evaluating p and
B
in the
high temperature limit, we approximate the energy per
particle as E = 3 k
B
T
0
and take
E = x U
total

0
. (A38)
In the experiments, we determine the harmonic oscil-
lation frequencies,
i
by parametric resonance methods.
We subtract o the contribution from the magnetic po-
tential and extrapolate to the harmonic values of the
optical frequencies by correcting for trap anharmonicity.
We obtain
x
= 2 2210(4) Hz,
y
= 2 830(2),
and
z opt
= 2 60.7(0.1). The corresponding trap
depth is U
0
= 60.3(0.2) K. The Fermi energy of an
ideal gas at the trap center is E
F
= (3N)
1/3
h , where
(
x

z
)
1/3
. With a typical total number of atoms
N 2.5 10
5
, E
F
k
B
2.0 K.
The quantity c
1
= x U
optical
(x)
0
/m is crucial for
determining the time evolution of

r
2
_
0
. Here, we recall
that the total trapping potential takes the form
U
total
(x) = U
opt
(x) + U
Mag
(x), (A39)
where U
opt
arises from the optical trap and U
Mag
arises
from the magnetic eld curvature, Eq. A20.
For precise measurements, it is necessary to determine
both the harmonic value and the anharmonic corrections
to x U
total
(x)
0
. To accomplish this, we note that
for a scalar pressure p, force balance for the trapped
cloud requires
_
d
3
xp = x
x
U
total

0
= y
y
U
total

0
=
z
z
U
total

0
. Then,
x U
total
(x)
0
= 3
_
z
U
total
z
_
0
. (A40)
After subtraction of the magnetic bowl correction, as
described above, the atoms eectively expand according
to the initial optical potential, where
x U
opt
(x)
0
= 3
_
z
U
total
z
_
0
x U
Mag
(x)
0
.
(A41)
Since x
2

0
and y
2

0
are small compared to z
2

0
, the
last term is just m
2
Mag
z
2

0
. Using Eq. A39 in Eq. A41,
we then have
xU
opt
(x)
0
= 3
_
z
U
opt
z
_
0
+2 m
2
Mag
z
2

0
. (A42)
The harmonic oscillator frequency for atoms in the op-
tical trap is generally energy dependent, because the trap
is less conning at higher energy, causing the frequency
to decrease. We model this by writing
_
z
U
opt
z
_
0
= m
2
z opt
z
2

0
h
A
(z
2

0
), (A43)
where m
2
z opt
z
2

0
arises from the harmonic trapping
potential. Here h
A
is the anharmonic correction factor,
h
A
(z
2

0
) 1 +
1
z
2

0
, (A44)
Note that at low energy, where the cloud size is small
compared to the width of the trapping potential, we take
h
A
1. Finally, we write the optical trap contribution
as
c
1

x U
opt
(x)
0
m
= 3
2
z opt
z
2

0
h
A
(z
2

0
) + 2
2
Mag
z
2

0
.(A45)
For p = 0 and
B
= 0, Eq. A17 shows that the eec-
tive mean square cloud radius (with the magnetic contri-
bution subtracted as described above), expands ballisti-
cally according to
r
2
= c
0
+ c
1
t
2
. (A46)
where
c
0
= r
2

0
= z
2

0
_
1 +

2
z

2
x
+

2
z

2
y
_
, (A47)
and
2
z

2
z opt
+
2
Mag
.
We t Eq. A46 to the mean square radius of the res-
onantly interacting clouds at a bias magnetic eld of
B = 834 G to determine c
1
and z
2

0
. For dierent
initial cloud sizes, we determine the quantity
h
A
[B, z
2

0
] =
c1
z
2
0
2
2
Mag
3
2
z opt
= 1 +
1
z
2

0
, (A48)
where
Mag
B. We nd that h
A
[834, z
2

0
] decreases
with increasing z
2

0
(
1
< 0), as expected for a correc-
tion arising from trap anharmonicity, where the quartic
11
terms in the optical trapping potential decrease the av-
erage oscillation frequency.
The optical frequency
z opt
in Eq. A48 is most pre-
cisely determined by demanding h
A
= 1 for energies close
to the ground state, where the anharmonic correction is
small and the resonantly interacting gas is nearly a pure
superuid, with a negligible bulk viscosity.
The slope
1
is determined by measuring the ballistic
expansion of the noninteracting Fermi gas at 528 G as a
function of initial cloud size, using the same method as
for the resonantly interacting gas. In the experiments,
we nd that the
1
obtained from ballistic expansion at
528 G is within 10% of that obtained from the h
A
data
for the highest energies of the resonantly interacting gas.
We use the linear anharmonic correction h
A
and
the
z opt
determined in this manner to predict x
U
opt
(x)
0
for the resonantly interacting gas at all initial
energies and nd that the corresponding bulk viscosity is
then very small, consistent with zero, as described in the
main text.
For the o-resonant studies, we measure the quantity
c
1
/z
2

0
2
2
Mag
at a bias eld B and divide by the tted
resonant value, 3
2
z opt
h
A
[834, z
2

0
].