An atom interferometer with a shaken optical lattice

C.A. Weidner,1 Hoon Yu,1, Ronnie Kosloff,2 and Dana Z. Anderson1

arXiv:1609.00416v1 [quant-ph] 1 Sep 2016

1
Department of Physics and JILA, University of Colorado, Boulder, Colorado, 80309-0440, USA
Fritz Haber Research Centre and The Department of Physical Chemistry, Hebrew University, Jerusalem 91904, Israel
(Dated: September 5, 2016)

We introduce shaken lattice interferometry with atoms trapped in a one-dimensional optical lattice. The atoms undergo an interferometer sequence of splitting, propagation, reflection, and recombination by phase modulation of the lattice through a sequence of shaking functions. Each function
in the sequence is determined by a learning procedure that is implemented with a genetic algorithm.
Numerical simulations determine the momentum state of the atoms, which is experimentally accessible with time-of-flight imaging. The shaking function is then optimized to achieve the desired
state transitions. The sensitivity of the interferometer to perturbations such as those introduced
by inertial forces scales the same way as for conventional matter wave interferometers. The shaken
lattice interferometer may be optimized to sense signals of interest while rejecting others, such as
the measurement of an AC signal while rejecting a DC bias.
PACS numbers: 37.25.+k,37.10.Jk, 03.75.Dg

I.

INTRODUCTION

This work introduces an approach to atom interferometry using a shaken optical lattice. Consider a onedimensional standing light field produced by retro reflecting an incoming laser beam with a mirror. By shaken
we mean that the longitudinal positions of the lattice
nodes are modulated, for example, by moving the reflecting mirror back and forth. The pioneering work of
P
otting et al. established that it is possible to transform
an initial atomic wave function into another desired wave
function by shaking the lattice in a prescribed way [1].
By now the shaken lattice concept has been broadly applied to study atom tunneling and transport in lattices
[24] and ferromagnetism [5]. Atoms held in a shaken
lattice have also been used to measure gravity [6, 7]. In
their original work, P
otting et al. sought a specific redistribution of the momentum state of atoms in the lattice.
They achieved the desired wavefunction transformation
through the use of a genetic learning algorithm (GA)
[1, 8, 9].
In one extension of the shaken lattice concept we carry
out a set of transformations that reproduce the ordered
sequence of operations associated with a Michelson interferometer [10], namely splitting, propagation, reflection,
reverse-propagation, and recombination of the atomic
wavefunction. In all five distinct cases, the protocol
needed to execute each operation is developed through
a learning algorithm. For the Michelson case we show
that the sensitivity for acceleration detection increases
as T 2 , where T is the atom propagation time. This is the
same result as for a free-space atom interferometer [11
15] (in both cases, the times associated with splitting,
reflection, and recombination are assumed to be negligibly small compared with the propagation times).

Current address: Hanwha Corporation Defence R&D center,

Daejeon, Republic of Korea

One aspect of shaken lattice interferometry (SLI) that

is distinguishable from free space interferometry is that
SLI affords a means to control the transfer function of the
system. As a simple example consider a scenario in which
one wishes to eliminate the sensitivity of the device to the
constant force of gravity. By optimizing the Michelson interferometry sequence in the presence of gravity, this DC
signal may be eliminated. In a reciprocal system where
the atoms propagate for a full period in the lattice, a DC
signal may be eliminated in favor of sensitivity to an AC
signal. Indeed one can optimize sensitivity of interferometry to a given signal class by carrying out learning to
enhance desired signals in this class and reject undesired
signals outside of the target sensitivity range. While we
use a GA in this paper to optimize the shaking, the interferometer sequence may also be found through use of
an optimal control algorithm such as the Krotov [16, 17]
or CRAB methods [18]. Similar methods have been used
to solve problems in other atomic systems, such as matter wave pulse shaping [19] or state inversion of a BEC
[20]. Optimization techniques have also been used to find
efficient light pulse schemes for atom interferometry [21].
This work proceeds as follows: In section II, the simulation and optimization processes are detailed. Section III
shows how a reciprocal scheme may be used to enhance
an AC acceleration signal while suppressing a DC signal.
Section IV analyzes the robustness of the results in the
presence of variations in parameters, and section V gives
a detailed analysis of the sensitivity of the shaken lattice
interferometer as well as the errors that could affect measurements made in an experimental system. Section VI
concludes.

II.

OPTIMIZATION OF THE SHAKEN

LATTICE INTERFEROMETER

Consider atoms trapped in a red-detuned optical lattice potential with a time varying phase (t). The po-

FIG. 1. Quantized momentum population of atoms in the

ground Bloch state of an infinite lattice with V0 = 10ER . This
ground state is the initial wavefunction of the atoms at the
start of the interferometer sequence and the final state of the
atoms after recombination.

FIG. 2. (color online) The full interferometer sequence. Blue

clouds represent atom wavepackets trapped in the shaken reddetuned lattice. Atoms begin in the ground Bloch state of the
lattice (Fig. 1) and are split into two oppositely propagating
wavepackets. The atoms are then reversed, propagated, and
recombined back into their initial ground state (in the absence
of any accumulated phase difference), thus completing the
interferometer sequence.

tential is given by
V (x, t) =

V0
cos {2kL x + (t)},
2

(1)

where the lattice wavenumber is kL = 2/L for lattice

wavelength L and the lattice depth is V0 , usually given
in units of the lattice recoil energy ER = h
2 kL2 /2m.
The results presented in this paper are based on numerical simulation of the one-dimensional time-dependent
Schr
odinger equation (TDSE). The TDSE is simulated
using a symmetrized split-step operator method [22] with
the potential in Eq. (1). In this work we take the lattice depth to be V0 = 10ER . In neglecting atom-atom
interactions we assume that the lattice is sparsely populated, e.g. though the use of a 3-D lattice where in
the non-shaken dimensions the lattice is so deep that few
atoms populate each site. This results in an array of 1D low atom number interferometers largely immune to
the effects of phase diffusion [23]. If a few hundred 1D tubes can be loaded with about 100 atoms in each
dimension then total atom numbers can reach 106 , thus
lowering shot noise to levels comparable with state-ofthe-art light-pulse atom interferometers based on Bosecondensed atoms [24]. The infinite lattice approximation
used here is reasonable if the lattice beam has a Rayleigh
range is much larger than the lattice wavelength. This
is true for lattice beams with near-infrared wavelengths
and waists on the order of tens of microns.
Given an initial wavefunction 0 and a desired wavefunction d , the GA finds a (t) that transforms the
wavefunction from 0 to d to within a specified precision. For example, when the algorithm learns to split the
atom wavefunction, the initial state is the ground Bloch
state of the lattice. The momentum population of the
ground state is shown in Fig. 1 [25]. The final split state
is a pair of equal-amplitude waves moving with equal and
opposite momenta.

As the split states are not eigenstates of the lattice

we find a second modulation protocol that maintains
the purity of the momentum states as they propagate.
Three additional shaking protocols are needed to reflect,
reverse-propagate, and recombine the atoms. Once the
five phase-modulation protocols are learned, interferometric measurements may be performed by repetition of
their ordered sequence. The full interferometer sequence
is illustrated in Fig. 2.
During the optimization process the GA modifies (t)
and changes how the lattice is shaken. The algorithm is
based on the work in references [1, 8] and relevant details
are included here. A block diagram of the GA procedure is shown in Fig. 3. First, the algorithm produces
an initial population of A individuals. This is known
as the first generation, denoted G = 1. For the results
presented here, A = 20. Each individual is produced
by generating a random vector of normally-distributed
Fourier amplitudes up to a certain bandwidth set by the
user. To get a time-varying shaking function, the Fourier
transform of this vector is taken. To maintain smooth
turn-on and turn-off this shaking function is multiplied
by an envelope function that goes smoothly to zero at
its endpoints, resulting in a shaking function (t) for
each individual. As an example the optimal shaking protocol for splitting is shown in Fig. 4. The shaking time
T is a parameter chosen by the user and is set here to
approximately 0.5 ms.
Once the initial individuals are generated the simulation takes the initial wavefunction 0 and propagates
the TDSE with (t), producing a final wavefunction
. Because experimental measurements can provide
momentum state populations via time-of-flight imaging,
we represent a simulated result for the Fourier transform
of the final wavefunction (k) = F{ (x)} as a vector P~ ( ) with components Pn, representing the rela-

FIG. 3. A block diagram illustrating the steps taken in the genetic algorithm. Given the initial and desired states, the first
generation G = 1 is randomly generated. The simulation then runs, and after the jth run of the simulation a generation
G = j + 1 results from the mixing of the previous generations individuals (see Fig. 5). Once the convergence criterion is met,
the simulation stops.

FIG. 4. (color online) (red, left axis) Shaking protocol for

the optimal splitting shown in Table I. The envelope function
ensures slow turn-on and turn-off at the endpoints. (blue,
right axis) The magnitude of the effective acceleration felt by
the atoms. The mean acceleration (amean = 22.5g) is marked
with a black dotted line. (inset) Fourier transform of the
bandwidth-limited shaking function.

tive atom population in each momentum state 2nhkL for

n = [, ]. A similar vector P~d (d ) is constructed
for the desired final state. In this work n is truncated
at n = [5, 5] since higher-order momentum states are
negligibly populated. While it is possible to consider final states with non-integer momenta, these states remain
unpopulated in our simulations of the TDSE and are thus
not considered in the GA. This is due to the fact that in
momentum space the TDSE takes the form [1]
2 2

k
i
h (k,t)
= h2m
(k, t)
t

V0 i(t)
(k
4 [e

2kL ) + ei(t) (k + 2kL )].

(2)

Therefore, due to the presence of the applied potential,

only transitions between momentum states separated by

2
hkL are allowed.
For a given shaking protocol (t), the GA assesses
the quality of the final momentum-space population via
a fitness function f (P~ ). The fitness function will be discussed in detail later in this section. Once the fitness
of the initial population is evaluated the individuals are
ranked in terms of their fitness (lower fitness is ranked
higher because we are performing minimization) and the
genetic algorithm proceeds. In general the procedure is
identical when proceeding from one generation G = j to
the next, denoted G = j + 1. The Alive best individuals
are allowed to proceed unaltered to the next generation,
a process known as elitism that ensures that the best
(lowest) fitness value will never increase. The Adie worst
individuals are deleted entirely. In this work we choose
Alive = 2 and Adie = 4. The remaining A Alive new
individuals are generated by randomly picking parents
from the surviving population for each new individual.
These parents are then mixed using methods adapted
from reference [1] and illustrated in Fig. 5. The idea
behind the mixing in this and other GAs is to use the
components of the A Adie best individuals to produce
a new generation with better fitness. When generation
j + 1 has been generated via this method, the simulation
runs again and evaluates the new fitness values. This
procedure iterates until it is terminated, e.g. after a certain number of iterations or if the fitness has reached a
suitable level.
Here, the goal of the GA is to minimize f (P~ ) subject to the constraints on the total time T for a given
shaking prescription. As a result, the final momentum
population converges to the desired state, and the lattice
learns to control the atoms. Splitting of the ground
state wavefunction results in two traveling matter waves
of the same amplitude with momenta of 2nhkL , analogous to an optical beamsplitter. For the simplest case
modeled here the final state has momenta corresponding
to n = 1. The fitness function used to evaluate the final
split wavefunction is

FIG. 5. (color online) Each iteration, the genetic algorithm

mixes the individuals from the previous generation to create
children that populate the next generation. Each gene is a
vector of values where red corresponds to a minimum value
and purple corresponds to a maximum value. In each case,
the indices of crossover or mutation are chosen at random.
The methods presented here are adapted from reference [1].
a) One-point crossover at index 6. b) Random mutation of
the value at index 2. c) Two-point crossover at indices 5 and
9. d) Change of the value at index 5 through a small creep
of the value.

f (P~ ) = |P~d P~ | +

2
P

|Pn,d Pn, |



 P P1, 
P
. (3)
+
|Pn,d Pn, | + |P0,d P0, | +  P1,
1, +P1, 
n=

run begins with atoms in the 2nhkL state and applies the
same phase modulation function pr (t). For both runs
the desired state is identical to the initial state. The GA
then sums the fitness of both final states and optimizes
pr (t). The optimal pr (t) will thus propagate the atoms
without crosstalk between the two momentum states
when applied to the linear combination of the two states.
For reflection this simultaneous two-state optimization
proceeds in the same way, but for an initial state with
2hkL , the final state has momentum 2hkL . After reflection the atoms are again propagated, this time with
reversed momentum.
The final step of the interferometry sequence recombines the two split waves, allowing a measurement of
their relative phase . In our case the phase is deduced
from the relative populations in each momentum state.
The recombination scheme is as follows: the initial and
desired states considered in splitting are swapped, such
that the desired state is now the initial state and vice
versa. The GA is run to find a modulation sequence that
returns all of the atoms in the two split matter waves
to the ground Bloch state. It is this final state of the
lattice after interferometry that changes when a signal is
applied, as shown in Sections III and IV. Thus, a quantifiable measure of the sensitivity of the interferometer is
related to how well deviations from this ground state can
be measured, which is related to the orthogonality of the
final state relative to the ground state.
In this paper the percent difference D1,2 between two
states with momentum vectors P~1 and P~2 is defined as
D1,2 = (1 P~1 P~2 ) 100%.

(4)

By quantifying the difference between the final state

after a signal is applied (P~sig ) and the final ground state
after recombination (P~g ) one may quantify the degree of
orthogonality between these two states. When the two
states are completely orthogonal and Dg,sig = 100%, the
states are maximally distinguishable experimentally.

n=2

The first term in Eq. (3) quantifies the difference between the two momentum state populations for the final
state and the desired state. The second and third terms
penalize for higher-order momentum state populations,
and the fourth term penalizes for atoms in the zero momentum state. The last term penalizes for asymmetry
in the 2
hkL momentum states. Similar fitness functions are used for all shaking protocols. For each shaking
protocol, Table I shows the initial, desired, and final momentum states, as well as the optimized fitness value.
The best result of 5 runs is shown, but in all cases, the
average fitness is below 0.1.
During propagation and reflection, the initial and final
momentum populations are the same. In these cases the
TDSE solver is run twice. The first run begins with an
initial state where all atoms are in the 2nhkL state and
modulates the lattice with a function pr (t). The second

III.

OPTIMIZATION OF A RECIPROCAL
INTERFEROMETER

One of the advantages of SLI is its flexibility in terms

of the possible detectable signals. That is, it is possible to
use the learning algorithm to teach the atoms a recombination protocol, as in reference [26]. For example, the
interferometer may be set up in a reciprocal configuration as shown in Fig. 6. The difference between this and
the interferometer configuration shown in Fig. 2 is that
the atoms take a fully symmetric path. In a traditional
trapped-atom Michelson interferometer [10], this configuration is immune to any DC accelerations. Therefore,
we expect that a reciprocal shaken lattice interferometer
will exhibit the same characteristics.
To test this the interferometer was simulated in the
following ways: First, while simulating of the standard

5
TABLE I. Genetic algorithm results, best of 5 optimization runs.
Protocol
Split

Prop.

Refl.

Recomb.

a
b

State
Init.
Des.
Final
Init.
Des.
Final
Init.
Des.
Final
Init.
Des.
Final

4
hkL
0.0026
0
0
0
0
0.0006
0
0
0.0022
0
0.0026
0.0026

Momentum populationa
2
hkL
0
hkL
2
hkL
0.1345
0.7259
0.1345
0.5
0
0.5
0.4999
0.0001
0.4998
0.5
0
0.5
0.5
0
0.5
0.4992
0.0010
0.4980
0.5
0
0.5
0.5
0
0.5
0.4962
0.0021
0.4948
0.5
0
0.5
0.1345
0.7259
0.1345
0.1345
0.7258
0.1343

4
hkL
0.0026
0
0
0
0
0.0006
0
0
0.0042
0
0.0026
0.0026

Fitness

Bandwidth (kHz)b

amean /g

0.0009

35.2

22.5

0.0995

32.8

43.5

0.0374

34.0

37.5

0.0009

34.0

13.7

Normalized to 1. Momentum states higher than |p| = 4

hkL are negligibly populated.
Bandwidth is defined as the frequency where the amplitude drops below 30 dB (0.1%) of the maximum. The shaking time is 0.5 ms.

FIG. 6.
(color online) The reciprocal interferometer sequence. The reciprocal interferometer modifies the standard
Michelson interferometer sequence shown in Fig. 2 so that
the atoms travel a fully symmetric path. This configuration
is designed to be sensitive to AC accelerations and immune
to DC accelerations.

Michelson interferometer, the atom wavefunction was

split, propagated for a time 2T , reflected, reverse propagated for a time 2T , then recombined into the ground
state, as in Fig. 2. The second simulation of the reciprocal interferometer (Fig. 6) split the wavefunction, propagated for a time T , reflected the atoms, reverse propagated them for a time 2T , reflected the atoms again,
propagated them for a final time T , then recombined
them back into the ground state. In both cases the total
interrogation (propagation) time was Ttot = 4T , where
we neglect the time needed to split, reflect, and recombine the atoms. Once these simulations were completed,
the shaking function opt (t) was used to simulate propagation of the TDSE with the potential

FIG. 7. (color online) Response of the reciprocal (red, solid)

and non-reciprocal (blue, dashed) interferometers to a sinusoidal signal as in Eq. (5). The response is given in terms
of the difference between the ground state of the lattice and
the final state after shaking with the applied signal as in Eq.
(4). The total interrogation time of each interferometer is
200.4 s and the amplitude of the applied acceleration signal is a = 0.115 m/s2 . The reciprocal interferometer is 20
times more sensitive than the non-reciprocal interferometer
to a signal with f 7 kHz, showing that SLI may be modified to tailor the interferometer response to an AC signal.

V0
cos {2kL x + opt (t)} + m~a ~x sin (2f t).
2
(5)
In Eq. (5), m is the atom mass and ~a is the amplitude
of the sinusoidal acceleration with frequency f . The dot
product arises because we can only measure the component of ~a that lies along the lattice propagation direction
x. For the proof-of-principle simulations done here we set
V (x, t) =

6
values of the DC bias and AC signal acceleration.

IV.

FIG. 8.
(color online) The response of an interferometer
optimized in the presence of a bias acceleration aDC = 0.76
m/s2 . The interferometer response is clearly minimized in
this vicinity of aDC and increases away from this bias. This
shows that SLI can be used to reject a DC bias of a given
magnitude or measure perturbations around this bias.

T = 50.2s, a = 0.115 m/s2 , and scan the acceleration

frequency from DC to f 1/T as shown in Fig. 7. We
expect that the sensitivity of the reciprocal interferometer is maximal around f = 1/2T = 9.96 kHz. Fig. 7
shows a maximum sensitivity for the reciprocal interferometer around fmax 7 kHz, and the reciprocal interferometer shows a factor of 20 enhancement in sensitivity
over the non-reciprocal interferometer. The error in the
maximally sensitive frequency is due to the fact that the
splitting, reflection, and recombination times are on the
order of the propagation time, not negligibly small as assumed. Furthermore, during propagation there is some
error in the fidelity of the atom wavefunction over long
propagation steps.
The interferometer may be optimized with a bias so
that it rejects a DC signal of a given magnitude. For the
results shown in Fig. 8 the standard non-reciprocal interferometer was optimized in the presence of a DC acceleration aDC = 0.76 m/s2 . Then the response of the interferometer was simulated for accelerations around aDC . Fig.
8 shows that the overlap of the interferometer response
increases around this minimum value, showing that the
interferometer may be optimized and operated around a
DC bias point. Thus, by combining this DC bias and AC
sensitivity one can sense a time-varying acceleration of a
certain frequency while rejecting a background DC acceleration. For example, a seismic signal could be detected
while rejecting a DC signal due to gravity. An interferometer optimized in the presence of gravity at one spatial
location could also be used to perform gravity gradiometry. A longer interrogation time will narrow the dip
shown in Fig. 8, improving the interferometer sensitivity.
Future work will focus on increasing the interferometer
sensitivity, decreasing the error in propagation, and optimization of the reciprocal interferometer for different

ROBUSTNESS OF THE
INTERFEROMETER

The shaking function is optimized in an ideal situation

where the lattice wavelength or depth is known exactly.
In an experimental setting, however, there is uncertainty
in these values. Thus, the robustness of the shaking function to these errors is of interest. In the following section
analysis was done with respect to the optimal splitting
function shown in Fig. 4, but the general results should
hold for all ingredients of the interferometer due to the
similarities in shaking time and bandwidth.
The results of variation of lattice depth and wavelength
are shown in Fig. 9. Variations in the lattice depth of
2.8% and variations in the wavelength of 0.016% maintain the fitness below a value of 0.1. For a mechanically
stable system such lattice depth variations can be controlled by servoing the laser intensity, and as detailed in
Section V such a servo is desirable to limit heating due to
laser intensity noise. Variations in the wavelength may
be controlled by locking the laser, e.g. to an atomic hyperfine transition, providing frequency stability to better
than 1 MHz. Therefore for common atomic physics wavelengths the wavelength stability provided by laser locking
is sufficient.
Even if the lattice parameters are known exactly, in
an experimental system the shaking will have some associated noise. For example, if the shaking is done via a
piezo-electric transducer on a retro-reflecting mirror then
noise in the electronics will be added to the desired shaking function. Undesired shaking due to mechanical instability will also add noise to the optimal shaking function.
To simulate this, Gaussian white noise of varying amplitudes was added to the optimized splitting function, and
the resulting fitness was recorded. This is plotted in Fig.
10. To maintain fitness less than 0.1 the noise amplitude
must be less than 10% of the maximum shaking phase,
thus setting a limit on the allowable noise added to the
shaking function.
Finally, we calculate the robustness of the shaking
function to undesired stray reflections in the lattice system. In reference [27] the forces due to these reflections
caused undesired phase shifts and imposed a limit on the
contrast of the interferometer fringes. In our case such
undesired reflections set up secondary lattice potentials
that can affect the fitness of the final state relative to the
desired state. We simulate this by shaking the lattice
and adding a parasitic potential of the form

VP (, ) =

V0
cos (2kL x + ).
2

(6)

In Eq. (6)  is the reflection amplitude and is the

phase of the parasitic lattice relative to the initial unshaken main lattice potential. The value of is generally

FIG. 10.
(color online) Fitness of the optimized splitting
shaking function with varying noise amplitudes added. The
results shown here are the average of 5 runs with random
white Gaussian noise added to the shaking function (t). Error bars give the standard deviation of the fitness from each
run. Noise amplitude is given as a fraction of the maximum
of (t). shown in Fig. 4. To limit the fitness to < 0.1, the
allowable noise level is 10% of the maximum shaking phase.

FIG. 9.
(color online) Fitness of the optimized splitting
shaking function shown in Fig. 4 after variations of a) the
simulated lattice depth and b) the simulated lattice wavelength. The red lines guide the eye by connecting the results
of simulations denoted by the black points. A fitness value of
0.1 is marked in each plot with a blue dashed line.

unknown in a real experiment. The simulation results

are shown in Fig. 11 for varying values of  from 0.1%
(i.e. reflection from an anti-reflection coated window)
to 4% (reflection from uncoated glass). The results of
these simulations demonstrate the importance of using
AR coatings on the windows to the science chamber or
tilting optics so that parasitic lattices do not interfere
with the main lattice.

V.

INTERFEROMETER SENSITIVITY AND

ERROR

FIG. 11.
(color online) Fitness of the optimized splitting
shaking function after the addition of spurious lattice potentials with varying phase due to unwanted reflections with reflection amplitudes of 0.1% (red), 1% (blue, dashed), and 4%
(cyan, dash-dot). A gray dotted line marks a fitness of 0.1.
The lines guide the eye between the black simulation points.

celeration can be calculated as in references [28, 29]. For

an interrogation time T , the sensitivity is
= 4nkL aT 2 .

In order to quantify the interferometer performance we

must calculate its sensitivity and evaluate the effects of
systematic and other errors on the measured results. If
we assume that the splitting, reflection, and recombination times are short relative to the propagation time and
the atoms momenta do not change appreciably during
interrogation, then the interferometers sensitivity to ac-

(7)

In general an applied acceleration will change the

atoms momenta as they propagate. Any potential term
linear in x that is added to a lattice potential may be unitarily transformed into phase factor modifying the lattice
potential (and vice versa) [30]. Thus, from Eq. (2), we
expect that the momenta remain quantized even in the

8
presence of an applied force, and this is verified by our
simulations. Therefore, the phase difference may be written as

= 4kL a

N
X
n=N

Z
n

dt c+,n (t) c,n (t) t.

(8)

In Eq. (8) the c,n (t) denote the relative population

in the 2n
hkL states for the arm with initial momentum
of 2l
hkL , respectively, that is, c,n (0) = n,l . In practice, the c,n (t) are found by solving the TDSE. As expected, this interferometer will maintain T 2 sensitivity.
Because the envelope function allows for smooth turnon and turn-off of the phase function, we can stitch
together successive propagation steps to increase the interrogation time. Therefore, the dynamic range of the
interferometer can be controlled by changing the total
propagation time. The GA can be used to force the atom
wavefunction to maintain its state over longer propagation times, correcting for any errors that arise as the
propagation protocol is repeatedly applied.
We consider the Michelson interferometer scheme in
this work as it is a natural and familiar starting point.
However, it is possible to optimize the interferometer to
obtain higher sensitivities by changing the shaking protocol. For example, higher momentum transfer generally
provides another means to increase sensitivity [31, 32].
Furthermore, as shown in Section III shaking protocols
can be synthesized to optimize the interferometer response to a signal of interest.
In practice an optimized shaking function can be found
computationally and adjustments may be made by running an experiment with the learning algorithm in a
closed-loop scheme, which has been done before in cold
and ultracold atom experiments [3335]. This closedloop scheme may be used to optimize in the presence
of inevitable systematics due to nonlinearities in shaking, laser wavefront errors, atom-atom interactions, and
finite lattice effects. For example, a closed-loop system
can correct deviations from optimal fitness due to the
parasitic lattice reflections discussed in the previous section as long as the deleterious effects are constant from
shot-to-shot. Uncertainties in lattice parameters such as
the lattice depth or wavelength due to imperfect lattice
alignment or the Gouy phase [27, 36] may also be corrected in a closed-loop system.
Other sources of error must be considered when evaluating the shaken lattice interferometer. Stray constant
fields result in an offset that does not affect the atoms,
as the interferometer maintains the atoms internal state.
If the atoms are pumped into a magnetically insensitive
state, errors due to magnetic field gradients are secondorder effects. For a spin unpolarized atomic sample these
effects will average to zero. Stray electric field gradients
can arise from speckle caused by imperfect optical elements such as a dusty window. If the effects due to
speckle are constant then the closed-loop scheme discussed above can correct for the undesired forces that

arise. However, if the speckle effects change in time (for

example, due to the accumulation of dust on the lattice
optics) then the forces caused by the gradients in the
light field can mimic unwanted signals. Following reference [27], the undesired acceleration aS due to a speckle
pattern with intensity V0 /2 and correlation length along
the lattice direction Lx is given by

aS =

V0
.
2mLx

(9)

For  = 1% and correlation lengths on the order of

1 mm [27], we get aS /g = 7.4 104 , which can be
improved by the use of high quality optical elements in
the lattice beamline and protecting these optics from dust
accumulation.
As in light-pulse atom interferometry the effects of unwanted inertial signals, e.g. spurious rotations, can be
subtracted out with the use of two interferometers operating in differential mode [37]. The common-mode signal
can then be recovered by subtracting the two interferometer measurements.
The effects of decoherence from lattice shaking have
been studied extensively elsewhere [24, 6]. These effects occur for a certain range of shaking amplitudes and
frequencies, depending on the lattice depth and applied
acceleration. If the desired dynamic range is known then
undesired frequencies and amplitudes can be filtered out
accordingly in the learning algorithm.
The shaking imparts an effective acceleration

aeff = (t)/2k
L on the atoms, which tilts the lattice [30].
This effective acceleration can be compared to a typical
acceleration measured in a practical environment, as
shown in Fig. 4. For all optimized shaking protocols
given here the mean effective acceleration amean is
an order of magnitude above g. Therefore, in many
cases the inertial forces of interest may be considered
a perturbation on the acceleration due to shaking. If
the tilt due to shaking becomes so large that the lattice
potential has no minima, the atoms will be untrapped.
Our calculations show that for a lattice depth of
V0 = 10ER the atoms are untrapped at accelerations
above approximately 72g. Thus, the mean acceleration
should be kept high enough so that the inertial forces
are perturbative but not so high that the atoms become
untrapped.
Finally, the fundamental limit of lattice experiments is
defined by photon scattering. In other optical lattice experiments long atom lifetimes in the lattice are enabled
by servo systems that lower laser noise [38]. If the lattice light is sufficiently far off of resonance the limit due
to photon scattering is reduced (although more power is
needed). Therefore, interrogation times on the order of
tens of seconds are possible in shaken lattice systems.
In the retro-reflecting lattice scheme, laser phase noise
is irrelevant, but unwanted motion of the retro-reflecting
mirror will cause unwanted shaking and give rise to spurious signals. Thus any noise in the retro mirror motion

9
must be stabilized via a servo system.

VI.

CONCLUSION

an increase in propagation time and optimization of the

interferometer sensitivity to a signal of interest. Finally,
an experimental implementation will require a means of
closed-loop learning to overcome inevitable experimental
noise and systematic errors.

In conclusion we introduce an atom interferometer

based on a shaken lattice, where the shaking functions
are found via a learning algorithm. This scheme provides the basis for a robust trapped-atom interferometer
wherein the forces of interest are small perturbations on
the shaking forces. Several other aspects of shaken lattice interferometry will be studied in future work, such as

C.W, H.Y, and D.Z.A. would like to acknowledge funding from the NSF PFC and Northrop Grumman Corporation.

[1] S. P
otting, M. Cramer, and P. Meystre, Phys. Rev. A
64, 063613 (2001).
[2] A. Alberti, V. V. Ivanov, G. M. Tino, and G. Ferrari,
Nat. Phys. Lett. 5, 547 (2009).
[3] H. Lignier, C. Sias, D. Ciampini, Y. Singh, A. Zenesini,
O. Morsch, and E. Arimondo, Phys. Rev. Lett. 99,
220403 (2007).
[4] A. Zenesini, H. Lignier, D. Ciampini, O. Morsch, and
E. Arimondo, Phys. Rev. Lett. 102, 100403 (2009).
[5] C. Parker, L.-C.Ha, and C. Chin, Nat. Phys. Lett. 9,
769 (2013).
[6] V. V. Ivanov, A. Alberti, M. Schioppo, G. Ferrari, M. Artoni, M. L. Chiofalo, and G. M. Tino, Phys. Rev. Lett.
100, 043602 (2008).
[7] M. G. Tarallo, A. Alberti, N. Poli, M. L. Chiofalo, F.-Y.
Wang, and G. M. Tino, Phys. Rev. A 86, 033615 (2012).
[8] R. S. Judson and H. Rabitz, Phys. Rev. Lett. 68, 1500
(1992).
[9] R. L. Haupt and S. E. Haupt, Practical Genetic Algorithms (John Wiley and Sons, Inc., 2004).
[10] Y.-J. Wang, D. Z. Anderson, V. M. Bright, E. A. Cornell,
Q. Diot, T. Kishimoto, M. Prentiss, R. A. Saravanan,
S. R. Segal, and S. Wu, Phys. Rev. Lett. 94, 090405
(2005).
[11] M. Kasevich and S. Chu, Phys. Rev. Lett. 67, 181 (1991).
[12] T. L. Gustavson, P. Bouyer, and M. A. Kasevich, Phys.
Rev. Lett. 78, 2046 (1997).
[13] J. M. McGuirk, M. J. Snadden, and M. A. Kasevich,
Phys. Rev. Lett. 85, 4498 (2000).
[14] B. Canuel, F. Leduc, D. Holleville, A. Gauguet, J. Fils,
A. Virdis, A. Clairon, N. Dimarcq, C. J. Borde, A. Landragin, and P. Bouyer, Phys. Rev. Lett. 97, 010402
(2006).
[15] S. M. Dickerson, J. M. Hogan, A. Sugarbaker, D. M. S.
Johnson, and M. A. Kasevich, Phys. Rev. Lett. 111,
083001 (2013).
[16] J. P. Palao and R. Kosloff, Phys. Rev. A 68, 062308
(2003).
[17] J. P. Palao, R. Kosloff, and C. P. Koch, Phys. Rev. A
77, 063412 (2008).
[18] T. Caneva, T. Calarco, and S. Montangero, Phys. Rev.
A 84, 022326 (2011).
[19] M. Nest, Y. Japha, R. Folman, and R. Kosloff, Phys.
Rev. A 81, 043632 (2010).
[20] R. B
ucker, T. Berrada, S. van Frank, J.-F. Schaff,
T. Schumm, J. Schmiedmayer, G. J
ager, J. Grond, and

U. Hohenester, J. Phys. B: At. Mol. Opt. Phys. 46,

104012 (2013).
S. S. Szigeti, J. E. Debs, J. J. Hope, N. P. Robins, and
J. D. Close, New J. Phys. 14, 023009 (2012).
M. D. Feit, J. A. Fleck, and A. Steiger, J. Comp. Phys.
47, 412 (1982).
J. Javanainen and M. Wilkens, Phys. Rev. Lett. 78, 4675
(1997).
T. Kovachy, P. Asenbaum, C. Overstreet, C. A. Donnelly,
S. M. Dickerson, A. Sugarbaker, J. M. Hogan, and M. A.
Kasevich., Nature 528, 530 (2015).
J. Hecker-Denschlag, J. E. S. H. H
affner, C. McKenzie,
A. Browaeys, D. Cho, K. Helmerson, S. L. Rolston, and
W. D. Phillips, J. Phys. B: At. Mol. Opt. Phys. 35, 3095
(2002).
S. van Frank, A. Negretti, T. Berrada, R. B
ucker, S. Montangero, J.-F. Schaff, T. Schumm, T. Calarco, and
J. Schmiedmayer, Nat. Comm. 5, 4009 (2014).
R. Charri`ere, M. Cadoret, N. Zahzam, Y. Bidel, and
A. Bresson, Phys. Rev. A 85, 013639 (2012).
G. D. McDonald, C. C. N. Kuhn, S. Bennetts, J. E. Debs,
J. D. Close, and N. P. Robins, Eur. Phys. Lett. 105,
63001 (2014).
G. D. McDonald and C. C. N. Kuhn, arXiv:quantph/1312-2713v1.
Q. Thommen, J. C. Garreau, and V. Zehnle, Phys. Rev.
A 65, 053406 (2002).
G. D. McDonald, C. C. N. Kuhn, S. Bennetts, J. E. Debs,
K. S. Hardman, M. Johnsson, J. D. Close, and N. P.
Robins, Phys. Rev. A 88, 053620 (2013).
S.-W. Chiow, T. Kovachy, H.-C. Chien, and M. A. Kasevich, Phys. Rev. Lett. 107, 130403 (2011).
W. Rohringer, R. B
ucker, S. Manz, T. Betz, C. Koller,
M. G
obel, A. Perrin, J. Schmiedmayer, and T. Schumm,
Appl. Phys. Lett. 93, 264101 (2008).
P. B. Wigley et al., arXiv:quant-ph/1507.04964v2.
M. Krenn, M. Malik, R. Fickler, R. Lapkiewicz, and
A. Zeilinger, Phys. Rev. Lett. 116, 090405 (2016).
P. Hamilton, M. Jaffe, J. Brown, L. Maisenbacher,
B. Estey, and H. M
uller, Phys. Rev. Lett. 114, 200405
(2015).
T. M
uller, M. Gilowski, M. Zaiser, P. Berg, C. Schubert,
T. Wendrich, W. Ertmer, and E. M. Rasel, Eur. Phys.
J. D 53, 273 (2009).
S. Blatt, A. Mazurenko, M. F. Parsons, C. S. Chiu, F. Huber, and M. Greiner, Phys. Rev. A 92, 021402(R) (2015).

ACKNOWLEDGMENTS

[21]
[22]
[23]
[24]

[25]

[26]

[27]
[28]

[29]
[30]
[31]

[32]
[33]

[34]
[35]
[36]

[37]

[38]