Generalized nonreciprocity in an optomechanical circuit via synthetic magnetism and

reservoir engineering
Kejie Fang,1, 2 Jie Luo,1, 2 Anja Metelmann,3, 4 Matthew H. Matheny,1, 5
Florian Marquardt,6, 7 Aashish A. Clerk,3 and Oskar Painter1, 2
1
Kavli Nanoscience Institute, California Institute of Technology, Pasadena, California 91125, USA
2
Institute for Quantum Information and Matter and Thomas J. Watson, Sr.,
Laboratory of Applied Physics, California Institute of Technology, Pasadena, California 91125, USA
3
Department of Physics, McGill University, 3600 rue University, Montréal, Quebec H3A 2T8, Canada
4
Department of Electrical Engineering, Princeton University, Princeton, New Jersey 08544, USA
5
Department of Physics, California Institute of Technology, Pasadena, California 91125, USA
6
Max Planck Institute for the Science of Light, Günther-Scharowsky-Straße 1/Bau 24, 91058 Erlangen, Germany
arXiv:1608.03620v1 [physics.optics] 11 Aug 2016

7
Institute for Theoretical Physics, Department of Physics, Universität Erlangen-Nürnberg, 91058 Erlangen
Synthetic magnetism has been used to control charge neutral excitations for applications ranging
from classical beam steering to quantum simulation. In optomechanics, radiation-pressure-induced
parametric coupling between optical (photon) and mechanical (phonon) excitations may be used to
break time-reversal symmetry, providing the prerequisite for synthetic magnetism. Here we design
and fabricate a silicon optomechanical circuit with both optical and mechanical connectivity between
two optomechanical cavities. Driving the two cavities with phase-correlated laser light results in
a synthetic magnetic flux, which in combination with dissipative coupling to the mechanical bath,
leads to nonreciprocal transport of photons with 35 dB of isolation. Additionally, optical pumping
with blue-detuned light manifests as a particle non-conserving interaction between photons and
phonons, resulting in directional optical amplification of 12 dB in the isolator through direction.
These results indicate the feasibility of utilizing optomechanical circuits to create a more general
class of nonreciprocal optical devices, and further, to enable novel topological phases for both light
and sound on a microchip.

Synthetic magnetism involving charge neutral elements tomechanical plaquettes between the photon and phonon
such as atoms [2], polaritons [3–5], and photons [6– lattices (see Fig. 1). To achieve nonreciprocal transmis-
10] is an area of active theoretical and experimental sion of intensity in the two-port device of this work –
research, driven by the potential to simulate quantum i.e., bonafide phonon or photon transport effects, not just
many-body phenomena [11], reveal new topological wave nonreciprocal transmission phase – one can combine this
effects [12, 13], and create defect-immune devices for in- synthetic field with dissipation to implement the general
formation communication [7, 10]. Optomechanical sys- reservoir engineering strategy outlined in Ref. [25]. This
tems [14], involving the coupling of light intensity to approach requires one to balance coherent and dissipa-
mechanical motion via radiation pressure, are a partic- tive couplings between optical cavities. In our system the
ularly promising venue for studying synthetic fields, as combination of the optical drives and mechanical dissipa-
they can be used to create the requisite large optical tion provide the “engineered reservoir” which is needed
nonlinearities [15]. By applying external optical driving to mediate the required dissipative coupling.
fields time-reversal symmetry may be explicitly broken To highlight the flexibility of our approach, we use it
in these systems. It was predicted that this could enable to implement a novel kind of nonreciprocal device ex-
optically tunable nonreciprocal propagation in few-port hibiting gain [26, 27]. By using an optical pump which is
devices [16–19], or in the case of a lattice of optomechan- tuned to the upper motional sideband of the optical cav-
ical cavities, topological phases of light and sound [1, 20]. ities, we realize a two-mode squeezing interaction which
Here we demonstrate a generalized form of optical non- creates and destroys photon and phonon excitations in
reciprocity in a silicon optomechanical crystal circuit [21] pairs. These particle non-conserving interactions can be
that goes beyond simple directional propagation; this is used to break time-reversal symmetry in a manner that is
achieved using a combination of synthetic magnetism, distinct from a standard synthetic gauge field. In a lattice
reservoir engineering, and parametric squeezing. system, this can enable unusual topological phases and
Distinct from recent demonstrations of optomechan- surprising behavior such as protected chiral edge states
ical nonreciprocity in degenerate whispering-gallery res- involving inelastic scattering [28] and amplification [29].
onators with inherent nontrivial topology [22–24], we em- Here, we use these interactions along with our reservoir-
ploy a scheme similar to that proposed in Refs. [1, 18] in engineering approach to create a cavity-based optical di-
which a synthetic magnetic field is generated via optical rectional amplifier: backward propagating signals and
pumping of the effective lattice formed by coupled op- noise are extinguished by 35 dB relative to forward prop-
tomechanical cavities. In such a scenario, the resulting agating waves which are amplified with an internal gain
synthetic field amplitude is set by the spatial variation of 12 dB (1 dB port-to-port).
of the pump field phase and the field lines thread op- The optomechanical system considered in this work is
 2

a J vacuum optomechanical coupling rate [14] (here we have

OL OR omitted the cavity labeling).
To enhance the effective photon-phonon interaction
i
L
e L
R
e
i R
strength each cavity is driven by an optical pump field
B with frequency relatively detuned from the optical cavity
resonance by the mechanical frequency (∆ ≡ ωp − ωc ≈
Energy ±ωm ), with a resulting intra-cavity optical field ampli-
tude |α|eiφ . In the good-cavity limit, where ωm  κ
(κ being the optical cavity linewidth), spectral filter-
ML MR ing by the optical cavity preferentially selects resonant
i V i
photon-phonon scattering, leading to a linearized Hamil-
b tonian with either a two-mode squeezing form Ĥent =
Forward Backward ~G(eiφ dˆ† b̂† +e−iφ dˆb̂) (blue detuned pumping) or a beam-
splitter form Ĥex = ~G(eiφ dˆ† b̂ + e−iφ dˆb̂† ) (red detuned
OL OR OL OR pumping). Here G = g0 |α| is the parametrically en-
hanced optomechanical coupling rate and dˆ = â − α con-
i i i
e L
e R
e
i L
e R
tains the small signal sidebands of the pump. For both
cases the phase of the resulting coupling coefficient is
ML MR ML MR nonreciprocal in terms of the generation and annihilation
of photon-phonon excitations. As has been pointed out
before, such a nonreciprocal phase resembles the Peierls
FIG. 1. Synthetic magnetic field in an optomechani- phase that a charged particle accumulates in a mag-
cal cavity system. a, In this scheme consisting of only two netic vector potential [30]. Crucially, the relative phase
optomechanical cavities, a two-dimensional plaquette can be ΦB = φL − φR is gauge independent (i.e. independent of
formed from the synthetic dimension [1] created by radia- local redefinitions of the â and b̂ cavity amplitudes), im-
tion pressure coupling from the optical modes to the mechan- plying it should have an observable effect. In the simple
ical modes. Photon hopping at rate J and phonon hopping case of ∆ = −ωm , ΦB is formally equivalent to having a
at rate V occurs between the optical and mechanical cavi-
synthetic magnetic flux threading the plaquette formed
ties, respectively, with J and V real for appropriate choice
of gauge. Pumping of the optomechancial cavities with phase by the four coupled optomechanical modes (two optical
correlated laser light (|αL |eiφL for the left cavity and (|αR |eiφR and two mechanical)[1, 8, 18]. For ∆ = +ωm , a non-zero
for the right cavity) results in a synthetic flux ΦB = φL − φR ΦB still results in the breaking of time-reversal symmetry,
threading the 4-mode plaquette. b, Scheme for detecting the though the lack of particle number conservation means
synthetic flux through nonreciprocal power transmission of that it is not simply equivalent to a synthetic gauge field.
an optical probe laser field. For forward (L → R) prop- Nonetheless, we will refer to it as a flux in what follows
agation, constructive interference set by the flux-dependent for simplicity.
phase ΦB ≈ π/2 of the dissipative phonon coupling path re-
To detect the presence of the effective flux ΦB , con-
sults in efficient optical power transmission. The accumulated
phase in the phonon coupling path is reversed for the back- sider the transmission of an optical probe signal, on res-
ward (R → L) propagation direction resulting in destructive onance with the optical cavity resonances and coupled in
interference and reduced optical power transmission in the from either the left or the right side via external optical
left output waveguide. The power in this case is sunk into coupling waveguides as depicted in Fig. 1b. The probe
the mechanical baths. light can propagate via two different paths simultane-
ously: (i) direct photon hopping between cavities via the
connecting optical waveguide, and (ii) photon-phonon
conversion in conjunction with intervening phonon hop-
ping via the mechanical waveguide between the cavities.
shown schematically in Fig. 1a and consists of two in-
As in the Aharonov-Bohm effect for electrons [32], the
teracting optomechanical cavities, labeled L (left) and
synthetic magnetic flux set up by the phase-correlated
R (right), with each cavity supporting one optical mode
optical pump beams in the two cavities causes a flux-
OL(R) and one mechanical mode ML(R) . Both the optical
dependent interference between the two paths. We de-
and mechanical modes of each cavity are coupled together
fine the forward (backward) transmission amplitude as
via a photon-phonon waveguide, resulting in optical and
TR→L(L→R) ≡ dout,L(R) /din,R(L) , where dout(in) is the am-
mechanical inter-cavity hopping rates of J and V , re-
plitude of the outgoing (incoming) electromagnetic signal
spectively (here we choose a local definition of the cavity
field in the corresponding coupling waveguide in units of
amplitudes so both are real). The radiation pressure in-
square root of photon flux. The optical transmission am-
teraction between the co-localized optical and mechanical
plitude in the forward direction has the general form
modes of a single cavity can be described by a Hamilto-
nian Ĥ = ~g0 ↠â(b̂ + b̂† ), where â(b̂) is the annihilation
TL→R [ω; ∆ = ±ωm ] = A± [ω] J − Γ± [ω]e−iΦB ,


operator of the optical (mechanical) mode and g0 is the (1)
 3

a b cavity L photon+phonon waveguide cavity R

optical coupler L optical coupler R 2.5 μm

c 1 d waveguide modes
OL 0
-1
1
ML
0

e 1
f -127
0.9 M+
optical fiber-to-chip couplers M

Normalized reflection
0.8 -128
L R

PSD (dBm/Hz)
0.7
0.6 -129
0.5
10 μm 0.4 -130
0.3 MW
0.2 -131
0.1
0 -132
1534.4 1534.45 1534.5 1534.55 1534.6 5.75 5.77 5.79 5.81 5.83
Wavelength (nm) Frequency (GHz)

FIG. 2. Silicon optomechanical crystal circuit. a, Scanning electron microscopy (SEM) image of the optomechanical
crystal circuit studied in this work. The circuit is fabricated from a silicon-on-insulator microchip (see App. A). b, SEM of the
main part of the circuit, which consists of a left and a right nanobeam optomechanical crystal cavity with a central unpatterned
nanobeam waveguide connecting the two cavities. A left and right optical coupler, which are each fed by an adiabatic fiber-to-
chip coupler [31], are used to evanescently couple light into either of the two optical cavities. c, FEM simulated electrical field Ey
and magnitude of the displacement field for the localized optical and mechanical cavity modes, respectively, of the nanobeam.
d, FEM simulated section of the corresponding optical and mechanical modes of the connecting waveguide. e, Optical reflection
spectrum of the left (blue) and right (orange) optical cavities. f, Optically transduced mechanical power spectral density (PSD)
measured from the left (blue) and right (orange) optical cavities. M± are the two hybridized mechanical cavity modes with
frequency ωM+(−) /2π = 5788.4 (5779.1) MHz and MW is a mechanical waveguide mode with frequency ωMW /2π = 5818.3 MHz.

where ω ≡ ωs − ωp and ωs is the frequency of the probe crucial in achieving any non-reciprocity in the magnitude
light. Γ± is the amplitude of the effective mechanically- of the optical transmission amplitudes.
mediated coupling between the two optical cavities, and The general reservoir-engineering approach to nonre-
is given by ciprocity introduced in Ref. [25] provides a framework
for both understanding and exploiting the above obser-
V G L GR vation. It demonstrates that nonreciprocity is generically
Γ± [ω] = γiL . γiR
(−i(ω ± ωmL ) + 2 )(−i(ω ± +V2 ωmR ) + 2 ) achieved by balancing a direct (Hamiltonian) coupling
(2) between two cavities against a dissipative coupling of the
The prefactor A± [ω] in Eq. (1) accounts for reflection cavities; such a dissipative coupling can arise when both
and loss at the optical cavity couplers, as well as the cavities couple to the same dissipative reservoir. The
mechanically-induced back-action on the optical cavities. balancing requires both a tuning of the magnitude of the
This prefactor is independent of the transmission direc- coupling to the bath, as well as a relative phase which
tion, and for the reverse transmission amplitude TR→L , plays a role akin to the flux ΦB . In our case, the damped
only the sign in front of ΦB changes. mechanical modes can play the role of the needed reser-
The directional nature of the optical probe transmis- voir, with the optical drives controlling how the optical
sion may be studied via the frequency-dependent ratio cavities couple to this effective reservoir. One finds that
at any given frequency ω, the mechanical modes induce
J − Γ± [ω]e−iΦB both an additional coherent coupling between the two
 
TL→R
[ω; ∆ = ±ωm ] = . (3) cavities (equivalent to an additional coupling term in the
TR→L J − Γ± [ω]e+iΦB
Hamiltonian) as well as a dissipative coupling (which is
Although the presence of the synthetic flux breaks time- not describable by a Hamiltonian). As is shown explic-
reversal symmetry, it does not in and of itself result itly in App, B, in the present setting these correspond
in nonreciprocal photon transmission magnitudes upon directly to the real and imaginary parts of Γ± [ω]. Hence,
swapping input and output ports [25, 33]. In our sys- the requirement of having Im Γ[ω] 6= 0 is equivalent to
tem, if one takes the limit of zero intrinsic mechanical requiring a non-zero mechanically-mediated dissipative
damping (i.e. γik = 0), the mechanically-mediated cou- coupling between the cavities.
pling amplitude Γ± [ω] is real at all frequencies. This Achieving directionality requires working at a fre-
implies |TL→R | = |TR→L |, irrespective of the value of ΦB . quency where the dissipative coupling has the correct
We thus find that non-zero mechanical dissipation will be magnitude to balance the coherent coupling J, and a tun-
 4

ing of the flux ΦB . For |Γ± [ω]| = J and arg(Γ± ) = −ΦB using an atomic force microscope to oxidize nanoscale
(6= 0, π), one obtains purely uni-directional transport regions of the cavity. After nano-oxidation tuning,
where the right optical cavity is driven by the left optical the left (right) cavity has optical resonance wavelength
cavity but not vice versa. One finds from Eq. (3) that λL(R) = 1534.502 (1534.499) nm, total loaded damping
the mechanically-mediated dissipative coupling between rate κL(R) /2π = 1.03 (0.75) GHz, and intrinsic cavity
the cavities is maximized at frequencies near the mechan- damping rate κiL(R) /2π = 0.29 (0.31) GHz (c.f. Fig. 2e).
ical normal mode frequencies ω ≈ −ωm ± V ; to achieve Note that hybridization of the optical cavity resonances
the correct magnitude of coupling, the optical pumping is too weak to be observable in the measured left and
needs to realize a many-photon optomechanical coupling right cavity spectra due to the fact that the optical cav-
Gk ≈ (Jγik )1/2 (see App. B for details). Note that our ity linewidths are much larger than the designed cavity
discussion applies to both the choices of red-detuned and coupling J. The thermal mechanical spectra, as mea-
blue-detuned pumping. While the basic recipe for direc- sured from the two cavities using a blue-detuned pump
tionality is the same, in the blue-detuned pump case the laser (see App. A), are shown in Fig. 2f where one can
effective reservoir seen by the cavity modes can give rise see hybridized resonances M± which are mixtures of
to negative damping, with the result that the forward the localized mechanical cavity modes ML and MR . A
transmission magnitude can be larger than one. We ex- nearby phonon waveguide mode (MW ) is also observable
plore this more in what follows. in both left and right cavity spectra. The optomechanical
In order to realize the optomechanical circuit depicted coupling rate and mechanical dissipation rate of ML(R)
in Fig. 1 we employ the device architecture of optome- were measured before nano-oxidation tuning, yielding
chanical crystals [34–36], which allows for the realization g0,L(R) /2π = 0.76 (0.84) MHz and γiL(R) /2π = 4.3
of integrated cavity-optomechanical circuits with versa- (5.9) MHz.
tile connectivity and cavity coupling rates [21, 37]. Fig- The experimental apparatus used to drive and probe
ure 2a shows the optomechanical crystal circuit fabri- the optomechanical circuit is shown schematically in
cated on a silicon-on-insulator microchip. The main sec- Fig. 3a. As indicated, an optical pump field for the
tion of the circuit, shown zoomed-in in Fig. 2b, con- left and right cavities is generated from a common diode
tains two optomechanical crystal nanobeam cavities, laser. The phase difference of the pump fields at the in-
each of which has an optical resonance of wavelength put to the cavities, and thus the synthetic magnetic flux,
λ ≈ 1530 nm and a mechanical resonance of frequency is tuned by a stretchable fiber phase shifter and stabilized
ωm /2π ≈ 6 GHz. The two optical cavities can be ex- by locking the interference intensity of the reflected pump
cited through two separate optical coupling paths, one signals from the cavities. To highlight the unique kinds of
for coupling to the left cavity and one for the right cav- nonreciprocal transport possible in our setup, we present
ity. Both the left and right optical coupling paths con- results for an experiment performed with blue-detuned
sist of an adiabatic fiber-to-chip coupler which couples pump fields with frequency ωp ≈ ωc + ωm ; as discussed,
light from an optical fiber to a silicon waveguide, and this will enable non-reciprocal transport with gain. An
a near-field waveguide-to-cavity reflective coupler. This input optical probe signal is generated from either of the
allows separate optical pumping of each cavity and opti- left or right cavity pump beams by sending them through
cal transmission measurements to be carried out in either an electro-optic modulator (EOM). A vector network an-
direction. The two nanobeam cavities are physically con- alyzer (VNA) is used to drive the EOMs at modulation
nected together via a central silicon beam section which frequency ωmod and detect the photocurrent generated by
is designed to act as both an optical waveguide and an the beating of the transmitted probe and reflected pump
acoustic waveguide. The central beam thus mediates signals, thus providing amplitude and phase information
both photon hopping and phonon hopping between the of the transmitted probe signal. Owing to the spectral
two cavities even though the cavities are separated by a filtering of the cavities, only the generated lower sideband
distance much larger than the cavity mode size [21, 38]. of the blue-detuned pump at ω = −ωmod is transmitted
The numerically simulated mode profiles for the local- through the circuit as a probe signal.
ized cavities and the connecting waveguide are shown in Figure 3b shows the ratio of the forward and back-
Fig. 2c and 2d, respectively. The hopping rate for pho- ward optical power transmission coefficients of the probe
tons and phonons can be engineered by adjusting the light (|TL→R /TR→L |2 ) for several magnetic flux values be-
number and shape of the holes in the mirror section of tween ΦB = 0 and π. For these measurements the pump
the optomechanical crystal cavity along with the free- powers at the input to the left and right cavity were set
spectral range of the connecting waveguide section [21]. to PpL = −14.2 dBm and PpR = −10.8 dBm, respec-
Here we aim for a design with J/2π ≈ 100 MHz and tively, corresponding to intra-cavity photon numbers of
V /2π ≈ 3 MHz so that nonreciprocity can be realized at ncL = 1000 and ncR = 1420. So as to remove differences
low optical pump power, yet still with high transmission in the forward and reverse transmission paths external
efficiency. to the optomechanical circuit, here the |TL→R /TR→L |2
As will be presented elsewhere [39], the optical and me- ratio is normalized to 0 dB for a modulation frequency
chanical frequencies of the optomechanical cavities are ωmod /2π ≈ 5.74 GHz, detuned far from mechanical res-
independently trimmed into alignment post-fabrication onance in a frequency range where reciprocal transmis-
 5

a cL cR b = 0.18 = 0.26 = 0.34 d

dB
i i B B B 2
| L
|e L
| R
|e R 40
1.8 30

Transmission ratio (dB)

30 1.6 20
p p p p 1.4
L R 20 1.2
10

/
0
10

B
FPC FPC 0.8 -10
0 0.6
-20
-10 0.4
0.2 -30

EDFA -20 0
5.77 5.78 5.79 5.8 5.77 5.78 5.79 5.8 5.77 5.78 5.79 5.8 5.775 5.78 5.785 5.79
Frequency (GHz) Frequency (GHz)
PD c e 10 0
= 1.18 = 1.26 = 1.34
φ-shifter 40
B B B
VNA

Transmission ratio (dB)

Amplification (dB)
30

Attenuation (dB)
EOM EOM 20
5 -10
FPC FPC 10
0
PD
λ-meter -10
0 -20
-20 -26 -24 -22 -20 -18 -16 -14
VOA 5.77 5.78 5.79 5.8 5.77 5.78 5.79 5.8 5.77 5.78 5.79 5.8
Frequency (GHz) Pp (dBm)

FIG. 3. Measurement of optical nonreciprocity. a, Experiment set-up. Red (blue) lines are optical (electronic) wiring.
Blue-detuned pump light from a tunable diode laser is split into two paths and fed into the two cavities (red arrows). Part of
the reflected pump laser light from the cavities (purple arrows) is collected by a photodetector (PD) and fed into a stretchable
fiber phase shifter (φ-shifter) to tune and lock the phase difference of the optical pumps. Each optical path can be modulated
by an electro-optic modulator (EOM) to generate an optical sideband which we use as the optical probe signal. The microwave
modulation signal with frequency ωmod is generated by port 1 of a vector network analyzer (VNA). After optical amplification
and photodetection, the transmitted optical probe signal through the optomechanical circuit is sent back to port 2 of the VNA
to measure the phase and amplitude of the optical probe transmission coefficient. EDFA: Erbium doped fiber amplifier, FPC:
fiber polarization controller, λ-meter: wavelength meter. b, The ratio of optical power transmission coefficients for right- and
left-propagation versus modulation frequency (ωmod = −ω = ωp − ωs ), for three different synthetic flux values ΦB /π = 0.18,
0.26, and 0.34. The blue curves correspond to the fit of the theoretical model (c.f. Eq. 3) to the measured spectra. c, The
power transmission coefficient ratio for ΦB with an additional π flux relative to those in b. d, Theoretical calculation of the
power transmission coefficient ratio for 0 ≤ ΦB ≤ 2π, where the six grey lines correspond to the six measured ΦB values in b
and c. e, Peak forward signal amplification above background level (blue squares)
p and corresponding signal attenuation in the
reverse direction (red circles) versus average optical pump power (P̄p = PpL PpR ) for fixed flux value of ΦB = 0.28π. The
solid curves are theoretical calculations based upon the theoretical model (c.f. Eq. 3 and SI) fit to the data in b and c.

sion is expected. Closer to mechanical resonance, strong with our design parameters. Figure 3d shows the theo-
nonreciprocity in the optically transmitted power is ob- retical calculation of |TL→R /TR→L |2 for a full 2π range of
served, with a peak and a dip in |TL→R /TR→L |2 oc- ΦB with the measured and fit optomechanical circuit pa-
curring roughly at the resonance frequencies of the hy- rameters. The pattern is seen to be odd symmetric with
bridized mechanical modes M+ and M− , respectively respect to ΦB = π. Inserting an additional magnetic flux
(c.f. Fig. 2c). The maximum contrast ratio between π into the measurements performed in Fig. 3b yields the
forward and backward probe transmission – the isola- spectra shown in Fig. 3c which displays a switch in the
tion level – is measured to be 35 dB for ΦB = 0.34π isolation direction as predicted by the model. The pump
near the M+ resonance. The forward transmission is power dependence of the peak (in frequency) forward sig-
also amplified in this configuration (blue-detuned pump, nal amplification and the corresponding backward signal
∆ = +ωm ), with a measured peak probe signal amplifi- attenuation relative to the background level far from me-
cation of 12 dB above the background level set by photon chanical resonance are shown in Fig. 3e for a fixed mag-
hopping alone (J/|Γ± |  1). The corresponding port-to- netic flux of ΦB = 0.28π. Good correspondence with the
port net gain is only 1 dB due to impedance mismatching theoretical power dependence (solid curves) is observed,
(J 6= κ/2) and intrinsic optical cavity losses (see SI for with nonreciprocal amplification vanishing at low pump
details). power.

From a two-parameter fit to the measured optical One can also obtain nonreciprocal optical power trans-
power transmission ratio spectra using Eq. 3 (see blue mission utilizing an even simpler system involving a sin-
curves in Figs. 3b and 3c), we obtain a waveguide- gle mechanical cavity. This is the situation we have for
mediated optical and mechanical hopping rate of J/2π = the Fabry-Perot-like mechanical resonances that exist in
110 MHz and V /2π = 2.8 MHz, respectively, consistent the central coupling waveguide (see MW resonance of
 6

b B
= 0.06 B
= 0.31 B
= 0.56 finds that for the system of Fig. 2, the spectrally-resolved
5
photon noise flux shows high directionality, but that the
2.5
sign of this directionality changes as a function of fre-
a

dB
0
J quency (analogous to what happens in the transmission
OL OR -2.5 amplitudes). In contrast, in the single-mechanical mode
B -5
5.81 5.82 5.81 5.82 5.81 5.82
setup of Fig. 4 the sign of the directionality is constant
GwL ei L
GwR ei R
Frequency (GHz) with frequency, and thus the total (frequency-integrated)
c = 1.06 = 1.31 = 1.56 noise photon flux is directional depending upon the flux
MW 5
B B B

magnitude. The laser pump fields can thus effectively

i 2.5
act as a heat pump, creating a temperature difference
0
dB

between the left and right waveguide output fields. The

-2.5 corresponding directional flow of quantum noise is espe-
-5 cially useful for quantum information applications, as it
5.81 5.82 5.81 5.82 5.81 5.82
Frequency (GHz) can suppress noise-induced damage of a delicate signal
source like a qubit [25, 27].
FIG. 4. Synthetic magnetic field with a single me- The device studied in this work highlights the poten-
chanical cavity. a, Physical configuration for generation of tial for optomechanics to realize synthetic gauge fields
a synthetic magnetic field and optical nonreciprocity with two and novel forms of nonreciprocity enabled by harness-
optical modes parametrically coupled with a common dissi- ing mechanical dissipation. Using just a few modes, it
pative mechanical waveguide mode. b,c The ratio of optical was possible to go beyond simply mimicking the physics
power transmission coefficients for right and left propagation of an isolator and realize a directional optical amplifier.
versus modulation frequency ωmod around the frequency of
By adding modes, an even greater variety of behaviours
the waveguide mode MW for various ΦB . The blue curves
correspond to a fit of the theoretical model (see App. B) to
could be achieved. For example, the simple addition of
the measured data. a third optical cavity mode, tunnel-coupled to the first
two cavities but with no mechanical coupling, would real-
ize a photon circulator similar to the phonon circulators
considered in Ref. [18]. Scaling the synthetic gauge field
Fig. 2c). As depicted in Fig. 4a, the mode configura- mechanism realized here to a full lattice of optomechani-
tion in this case consists of two optical cavity modes (OL cal cavities would allow the study of topological phenom-
and OR ) coupled together via the optical waveguide, one ena in the propgation of both light and sound. Predicted
mechanical waveguide mode MW which is parametrically effects include the formation of back-scattering immune
coupled to each of the optical cavity modes, and the syn- photonic [1] and phononic [20] chiral edge states, topo-
thetic magnetic flux ΦB = φL − φR due to the relative logically nontrivial phases of hybrid photon-phonon exci-
phases of the optical pump fields threading the triangular tations [20], dynamical gauge fields [40], and, in the case
mode space. In Fig. 4b and 4c we show the measurement of non-particle-conserving interactions enabled by blue-
of |TL→R /TR→L |2 for a series of different flux values ΦB detuned optical pumping, topologically protected inelas-
with blue-detuned pumping (∆ ≈ +ωMW ) at levels of tic scattering of photons [28] and even protected ampli-
ncL = 770 and ncR = 1090. In this single mechanical fying edge states [29].
mode case the direction of the signal propagation is de-
termined by the magnitude of the flux; ΦB ≤ π leads
to backward propagation and ΦB ≥ π to forward prop- ACKNOWLEDGMENTS
agation. The lower contrast ratio observed is a result of
the weaker coupling between the localized optical cavity The authors would like to thank Michael Roukes for
modes and the external waveguide mode, which for the the use of his atomic force microscope in the nanoox-
modest pump power levels used here (. 100 µW) does idation tuning of the cavities. This work was sup-
not allow us to reach the parametric coupling required ported by the AFOSR-MURI Quantum Photonic Mat-
for strong directional transmission. ter, the ARO-MURI Quantum Opto-Mechanics with
While our focus has been on the propagation of in- Atoms and Nanostructured Diamond (grant N00014-15-
jected coherent signals through the optomechanical cir- 1-2761), the University of Chicago Quantum Engineer-
cuit, it is also interesting to consider the flow of noise. ing Program (AAC,AM), the ERC Starting Grant OP-
As might be expected, the induced directionality of our TOMECH (FM), the Institute for Quantum Information
system also applies to noise photons generated by the up- and Matter, an NSF Physics Frontiers Center with sup-
conversion of both thermal and quantum fluctuations of port of the Gordon and Betty Moore Foundation, and
the mechanics; see App. C for detailed calculations. One the Kavli Nanoscience Institute at Caltech.

[1] M. Schmidt, S. Keßler, V. Peano, O. Painter, and F. Mar- photons on a lattice,” Optica, vol. 2, pp. 635–641, 2015.
quardt, “Optomechanical creation of magnetic fields for
 7

[2] J. Dalibard, F. Gerbier, G. Juzelinas, and P. Öhberg, cavity-optomechanical circuits,” Nature Photon., vol. 10,
“Colloquium: Artificial gauge potentials for neutral p. 489496, Jun. 2016.
atoms,” Rev. Mod. Phys., vol. 83, p. 1523, 2011. [22] J. Kim, M. C. Kuzyk, K. Han, H. Wang, and
[3] J. Koch, A. Houck, K. L. Hur, and S. Girvin, “Time- G. Bahl, “Non-reciprocal brillouin scattering induced
reversal-symmetry breaking in circuit-qed-based photon transparency,” Nat. Phys., vol. 11, pp. 275–280, Jan.
lattices,” Phys. Rev. A, vol. 82, p. 043811, 2010. 2015.
[4] P. Roushan, C. Neill, A. Megrant, Y. Chen, R. Bab- [23] Z. Shen, Y.-L. Zhang, Y. Chen, C.-L. Zou, Y.-F. Xiao,
bush, etal, and J. Martinis, “Chiral groundstate currents X.-B. Zou, F.-W. Sun, G.-C. Guo, and C.-H. Dong, “Ex-
of interacting photons in a synthetic magnetic field,” p. perimental realization of optomechanically induced non-
arXiv:1606.00077, 2016. reciprocity,” arXiv:1604.02297, 2016.
[5] K. Sliwa, M. Hatridge, A. Narla, S. Shankar, L. Frun- [24] F. Ruesink, M.-A. Miri, A. Alú, and E. Verhagen, “Non-
zio, R. Schoelkopf, and M. Devoret, “Reconfigurable reciprocity and magnetic-free isolation based on optome-
josephson circulator/directional amplifier,” Phys. Rev. chanical interactions,” arXiv:1607.07180, Jul. 2016.
X, vol. 5, p. 041020, 2015. [25] A. Metelmann and A. Clerk, “Nonreciprocal photon
[6] R. Umucalılar and I. Carusotto, “Artificial gauge field for transmission and amplification via reservoir engineer-
photons in coupled cavity arrays,” Phys. Rev. A, vol. 84, ing,” Phys. Rev. X, vol. 5, p. 021025, 2015.
p. 043804, 2011. [26] B. Abdo, K. Sliwa, L. Frunzio, and M. Devoret, “Direc-
[7] M. Hafezi, E. Demler, M. Lukin, and J. Taylor, “Ro- tional amplification with a josephson circuit,” Phys. Rev.
bust optical delay lines with topological protection,” Nat. X, vol. 3, p. 031001, Jul. 2013.
Phys., vol. 7, p. 907, 2011. [27] B. Abdo, K. Sliwa, S. Shankar, M. Hatridge, L. Frunzio,
[8] K. Fang, Z. Yu, and S. Fan, “Realizing effective mag- R. Schoelkopf, and M. Devoret, “Josephson directional
netic field for photons by controlling the phase of dy- amplifier for quantum measurement of superconducting
namic modulation,” Nature Photon., vol. 6, pp. 782–787, circuits,” Phys. Rev. Lett., vol. 112, p. 167701, Apr. 2014.
2012. [28] V. Peano, M. Houde, C. Brendel, F. Marquardt, and
[9] M. Rechtsman, J. Zeuner, Y. Plotnik, Y. Lumer, A. Clerk, “Topological phase transitions and chiral in-
D. Podolsky, F. Dreisow, S. Nolte, M. Segev, and A. Sza- elastic transport induced by the squeezing of light,” Nat.
meit, “Photonic floquet topological insulators,” Nature, Commun., vol. 7, p. 10779, 2016.
vol. 496, p. 196, 2013. [29] V. Peano, M. Houde, F. Marquardt, and A. A. Clerk,
[10] L. Tzuang, K. Fang, P. Nussenzveig, S. Fan, and M. Lip- “Topological quantum fluctuations and travelling wave
son, “Nonreciprocal phase shift induced by an effective amplifiers,” arXiv:1604.04179, Apr. 2016.
magnetic flux for light,” Nature Photons, vol. 8, p. 701, [30] K. Fang, Z. Yu, and S. Fan, “Photonic aharonov-bohm
2014. effect based on dynamic modulation,” Phys. Rev. Lett.,
[11] Y.-J. Lin, R. Compton, K. Jimnez-Garca, J. Porto, and vol. 108, p. 153901, 2012.
I. Spielman, “Synthetic magnetic fields for ultracold neu- [31] S. Gröblacher, J. T. Hill, A. H. Safavi-Naeini, J. Chan,
tral atoms,” Nature, vol. 462, p. 628, 2009. and O. Painter, “Highly efficient coupling from an optical
[12] M. W. Ray, E. Ruokokoski, S. Kandel, M. Möttönen, and fiber to a nanoscale silicon optomechanical cavity,” App.
D. S. Hall, “Observation of dirac monopoles in a synthetic Phys. Lett., vol. 108, p. 181104, 2013.
magnetic field,” Nature, vol. 505, p. 657, 2014. [32] Y. Aharanov and D. Bohm, “Signigicance of electromag-
[13] L. Lu, J. Joannopoulos, and M. Soljai, “Topological pho- netic potentials in the quantum theory,” Phys. Rev., vol.
tonics,” Nat. Photons., vol. 8, p. 821, 2014. 115, pp. 485–491, 1959.
[14] M. Aspelmeyer, T. J. Kippenberg, and F. Marquardt, [33] L. Deák and T. Fülöp, “Reciprocity in quantum, electro-
“Cavity optomechanics,” Rev. Mod. Phys., vol. 86, pp. magnetic and other wave scattering,” Ann. Phys., vol.
1391–1452, 2014. 327, pp. 1050–1077, 2012.
[15] J. Rosenberg, Q. Lin, and O. Painter, “Static and dy- [34] M. Eichenfield, J. Chan, R. M. Camacho, K. J. Vahala,
namic wavelength routing via the gradient optical force,” and O. Painter, “Optomechanical crystals,” Nature, vol.
Nature Photon., vol. 3, pp. 478–483, Aug. 2009. 462, pp. 78–82, 2009.
[16] S. Manipatruni, J. T. Robinson, and M. Lipson, “Optical [35] S. Sadat-Saleh, S. Benchabane, F. I. Baida, M.-P. Bernal,
nonreciprocity in optomechanical structures,” Phys. Rev. and V. Laude, “Tailoring simultaneous photonic and
Lett., vol. 102, p. 213903, May 2009. phononic band gaps,” J. Appl. Phys., vol. 106, p. 074912,
[17] M. Hafezi and P. Rabl, “Optomechanically induced 2009.
non-reciprocity in microring resonators,” Opt. Express, [36] J. Gomis-Bresco, D. Navarro-Urrios, M. Oudich, S. El-
vol. 20, no. 7, pp. 7672–7684, Mar. 2012. Jallal, A. Griol, D. Puerto, E. Chavez, Y. Pennec,
[18] S. J. M. Habraken, K. Stannigel, M. D. Lukin, P. Zoller, B. Djafari-Rouhani, F. Alzina, A. Martı́nez, and C. S.
and P. Rabl, “Continuous mode cooling and phonon Torres, “A one-dimensional optomechanical crystal with
routers for phononic quantum networks,” New J. Phys., a complete phononic band gap,” Nat. Commun., vol. 5,
vol. 14, p. 115004, 2012. p. 4452, 2014.
[19] Z. Wang, L. Shi, Y. Liu, X. Xu, and X. Zhang, “Optical [37] A. H. Safavi-Naeini and O. Painter, “Design of op-
nonreciprocity in asymmetric optomechanical couplers,” tomechanical cavities and waveguides on a simultaneous
Scientific Reports, vol. 5, p. 8657, Mar. 2015. bandgap phononic-photonic crystal slab,” Opt. Express,
[20] V. Peano, C. Brendel, M. Schmidt, and F. Marquardt, vol. 18, pp. 14 926–14 943, 2010.
“Topological phases of sound and light,” Phys. Rev. X, [38] Y. Sato, Y. Tanaka, J. Upham, Y. Takahashi, T. Asano,
vol. 5, p. 031011, 2015. and S. Noda, “Strong coupling between distant photonic
[21] K. Fang, M. H. Matheny, X. Luan, and O. Painter, “Op- nanocavities and its dynamic control,” Nature Photon.,
tical transduction and routing of microwave phonons in vol. 6, pp. 56–61, 2012.
 8

[39] K. Fang, X. Luan, M. H. Matheny, M. L. Roukes, and R. Schoelkopf, “Introduction to quantum noise, measure-
O. Painter, in preparation. ment, and amplification,” Rev. Mod. Phys., vol. 82, p.
[40] S. Walter and F. Marquardt, “Dynamical gauge fields in 1155, 2010.
optomechanics,” arXiv:1510.06754, Oct. 2015.
[41] A. Clerk, M. Devoret, S. Girvin, F., Marquardt, and

Appendix A: Device Fabrication and Methods

1. Device fabrication and atomic force microscope nano-oxidation tuning

The devices were fabricated from a silicon-on-insulator wafer with a silicon device layer thickness of 220 nm and
buried-oxide layer thickness of 2 µm. The device geometry was defined by electron-beam lithography followed by
inductively coupled plasma reactive ion etching to transfer the pattern through the 220 nm silicon device layer. The
devices were then undercut using an HF:H2 O solution to remove the buried oxide layer and cleaned using a piranha
etch.
After device fabrication, we used an atomic force microscope to draw nanoscale oxide patterns on the silicon device
surface. This process modifies the optical and mechanical cavity frequencies in a controllable and independent way
with the appropriate choice of oxide pattern. The nano-oxidation process was carried out using an Asylum MFP-3D
atomic force microscope and conductive diamond tips (NaDiaProbes) in an environment with relative humidity of
48%. The tip was biased at a voltage of −11.5 V, scanned with a velocity of 100 nm/s, and run in tapping mode with
an amplitude of 10 nm. The unpassivated silicon device surface was grounded.

2. Optical transmission coefficient measurement

a B
= 0.34 b = 1.34
B
-35 -35
Transmission coefficient (dB)

-40
Transmission coefficient (dB)

-40
-45 -45
-50 -50
-55 -55
-60 -60
-65 -65
-70 -70
-75 -75
-80 -80
5.74 5.76 5.78 5.8 5.82 5.74 5.76 5.78 5.8 5.82
Frequency (GHz) Frequency (GHz)

FIG. 5. a Microwave signal power transmission through the optomechanical circuit for forward (right-propagation; blue) and
backward (left-propagation; blue curve) directions, with flux set to ΦB = 0.34π and cavity photon number ncL = 1000 and
ncR = 1420. b Same as a but with ΦB = 1.34π.

To measure the optical power transmission through the optomechanical circuit we used a vector network analyzer
(VNA). The VNA outputs a microwave tone from port 1 with frequency ωmod to an electro-optic modulator which
modulates the optical pump to generate an optical sideband corresponding to the optical probe. In the case of a
blue-detuned pump from the optical cavity resonance, the probe field corresponds to the lower sideband (selected by
the filtering properties of the cavity itself). Both the optical probe and pump are launched into one optomechanical
cavity in the circuit. At the other cavity, the transmitted optical probe combines with a second pump and the beating
of the two is detected by a high-speed photodetector (both the first and second pump beams are from the same
laser source, and thus phase coherent). The photocurrent signal from the photodetector is sent into port 2 of the
VNA to measure the microwave signal transmission coefficient Tµ . Fig. 5 shows |Tµ |2 for forward (right-propagating;
blue curve) and backward (left-propagating; red curve) directions through the optomechanical circuit as a function
of the modulation frequency ωmod . In Fig. 5a the synthetic flux value is locked to ΦB = 0.34π whereas in Fig. 5b
ΦB = 1.34π. In both flux settings the optical pumping levels were such that the left and right cavity photon numbers
 9

were ncL = 1000 and ncR = 1420, respectively. This is the raw transmission data corresponding to the normalized
transmission ratio shown in Figs. 3b and 3c of the main text.
While absolute optical transmission is not directly measured, the ratio of the optical transmission coefficients for
forward and backward propagation can be obtained from the normalized microwave signal transmission coefficient T̄µ ,

|TL→R /TR→L |2 = |T̄µR /T̄µL |2 , (A1)

where |T̄µ |2 is normalized using the value of |Tµ |2 away from all mechanical resonances to remove all the external
asymmetry in the experimental setup for left and right propagation paths. These external asymmetries include
modulator efficiency, cable/fiber loss, etc. In our analysis the normalization level is the average value of |Tµ |2 in the
frequency range of 5.74-5.76 GHz. To be clear, the reason this calibration is necessary is because we don’t actually
physically swap the source and detector in our measurements. Rather, for the left-to-right transmission path we have
one modulator on the left side which generates the probe tone and one detector on the right side which measures the
transmission through to the right side. When we measure right-to-left transmission we have a different modulator on
the right side to generate the probe tone and a different detector on the left side to detect the transmitted probe. If
the modulator on the left side is different from the modulator on the right side, then for the same microwave drive that
excites the modulators we would get different a different optical probe power in the sidebands of the pump. Similarly
if the left and right detectors have different efficiencies then they would produce a different photocurrent for the
same transmitted optical probe power. Since we measure in practice the ratio of the microwave drive to the detected
microwave photocurrent, this could cause an inherent asymmetry in the measured transmission for left-to-right and
right-to-left transmission even if the optical transmission was perfectly symmetric.

3. Device characterization

To determine the components of optical cavity loss (intrinsic decay rate κi , external waveguide-to-cavity coupling
κe , total cavity decay rate κ) of both the left and right optical cavities we used a pump-probe scheme similar to that
used to measure the nonreciprocity of the optomechanical circuit. The pump beam in this case, however, is set to be
very weak so as to not resonantly excite the mechanics as the probe signal is swept across the optical cavity resonance.
The cavity scans are plotted in Fig. 6a and 6b for the left and right cavities, respectively. We fit the phase response
curves and get κiL(R) /2π = 0.29 (0.31) GHz, κeL(R) /2π = 0.74 (0.44) GHz, and κL(R) /2π = 1.03 (0.75) GHz. The
intrinsic and external optical cavity rates are used to determine the intra-cavity photon number for a given optical
pump power (specified at the input to the cavity).
Thermal mechanical spectra of the two cavities are measured with a weak blue-detuned optical pump so as to avoid
back-action; a single pump is used for each of the left and right cavity measurements. The reflected pump light from
the cavity contains modulation sidebands from the thermal mechanical motion, which upon detection with a high-
speed photodetector creates a photocurrent with the thermal motion of the mechanical cavity modes imprinted on it.
Since the mechanical modes can be hybridized between left-cavity, right-cavity, and waveguide modes, a measurement
with the left-side pump produces a local measurement of the cavity modes as measured by the localized left optical
cavity mode, and similarly for the right-side pump and cavity. The intrinsic decay rate of the mechanical modes is
inferred from the linewidth of the Lorentzian mechanical spectrum.
Measurements of the mechanical mode spectra were performed both before and after the cavities were nano-oxidized
to tune their localized optical and mechanical modes into resonance. Measurements prior to nano-oxidation allowed us
to determine the local (left and right) mechanical and optical cavity mode properties (i.e., the bare, uncoupled mode
properties). Knowing the left and right cavity mode properties from independent measurements allowed us to fit with
fewer fitting parameters the measured forward and backward transmission curves of the hybridized cavities presented
in the main article text. Note that after nano-oxidation the left and right optical cavity modes were only very weakly
hybridized so as to maintain their left-cavity and right-cavity character. The mechanical modes were tuned to be
strongly hybdridized as evidenced in Fig. 2f of the main text. Figures 6c and 6d show the measured linewidth of the
mechanical cavity modes ML(R) versus optical pumping power. In Fig. 6c the left cavity was pumped with a blue
detuning ∆ = +ωmL ; in Fig. 6d the right cavigty was pumped with a blue detuning of ∆ = +ωmR . By fitting the
measured data with formula γ = γi − 4g02 nc /κ (nc corresponding to the intra-cavity photon number determined from
the OL(R) measured cavity properties), we obtain g0,L(R) /2π = 0.76 (0.84) MHz and γiL(R) /2π = 4.3 (5.9)MHz for
the left (right) localized cavity modes.
The optical (J) and mechanical (V ) hopping rates between the two optomechanical cavities via the connecting
waveguide are determined from a global fitting using Eq. (1) of the main text for the group of measured transmission
coefficient ratio curves in Figs. 3c and 3d with varying ΦB . The intra-cavity cavity photon number, optomechanical
coupling rates and intrinsic mechanical decay rates are all taken as fixed and equal to the independently measured
 10

a 0
b
180
-20 160
-40 140

Phase (degree)

Phase (degree)
-60 120
-80 100
-100 80
-120 60
-140 40
-160 20
-180 0
-20
2 4 6 8 2 4 6 8
Frequency (GHz) Frequency (GHz)
c 3.8 d 5.5

3.6 5.3

3.4 5.1

γ R /2π (MHz)
γ L/2π (MHz)

3.2 4.9

3.0 4.7

2.8 4.5

2.6 4.3
300 400 500 600 700 800 100 150 200 250 300 350 400 450
n n
cL cR
e f
1
0.9
0.8
Normalized reflection

0.7

Transmission (dB)
0.6
0.5
0.4
0.3
0.2
0.1
0
-100 -50 0 50 100 Wavelength (pm)
Wavelength (pm)

FIG. 6. a, Left optical cavity phase response as measured by scanning the probe signal across the cavity resonance with weak
blue-detuned pump. b, Same as in a for the right optical cavity. c, Measured back-action modified mechanical linewidth versus
intra-cavity pump photon number for the left optical cavity. Here only left cavity pump beam is applied, and the pump is tuned
to the upper motional sideband of the cavity (blue-detuned with ∆ = +ωmL ). d, Same as in c for the right-side cavity and
right-side pump. Measurements in a-d were performed prior to nano-oxidation tuning. e Measured (circles) and theoretical
(solid curves) optical reflection spectra using a left-cavity (blue) and right-cavity (red) optical pump. These measurements are
taken after nano-oxidation and the theoretical calculation includes the fit coupling (J/2π = 110 MHz) between the left and
right optical cavity modes and a splitting between the uncoupled modes. The wavelength origin is taken to correspond to the
right optical cavity resonance. f Calculated optical transmission power from one optical port to the other of an optical probe
signal near resonance of the coupled optical cavity modes. Here there is no pump beam, and so no coupling to phonons. The
parameters of the optical cavity modes are taken from the fit to the measured optical reflection spectra in e.

values as described above.

With the fit value of J from forward and reverse transmission measurements versus ΦB , and the measured cavity
coupling rates (κ, κi ) from the left and rigth optical cavity modes prior to nano-oxidation tuning, we fit the measured
optical reflection spectra of the two weakly coupled optical cavity modes after nano-oxidation. This allows us to
determine the uncoupled left and right optical cavity mode frequencies. The measured and fit spectra as measured
from the left and right cavities are shown in Fig. 6e. As noted earlier, the measured spectra after nano-oxidation
are still predominantly given by uncoupled left and right cavity modes. Based on the theoretical fit to the measured
optical reflection spectra, we also calculate the transmission of an optical probe signal through the optomechanical
 11

circuit in the absence of a pump beam (i.e., no coupling to phonons, just pure optical transmission)
√
J κeL κeR
η= 2 . (A2)
|J + κL κR /4 − (ω − ωcL )(ω − ωcR ) − iκL (ω − ωcL )/2 − iκR (ω − ωcR )/2|

Fig. 6f shows the numerical result, and the minimum insertion loss for transmission from one port to the other port
is found to be about 11 dB for a probe signal frequency in between the two cavity resonances. This is the estimated
port-to-port optical transmission effiency in absence of optomechanical amplification.

Appendix B: Theory of optical nonreciprocity

1. Input-output formula

We provide theoretical analysis of optical nonreciprocity in the coupled optomechanical cavity system. We first
consider the case with two optical and two mechanical cavity modes. The Hamiltonian of this system can thus be
written down as follows,

~ωck â†k âk + J(â†L âR + âL â†R ) + ~ωmk b̂†k b̂k + V (b̂†L b̂R + b̂L b̂†R )
X X
Ĥ = (B1)
k=L,R k=L,R
√
~g0k (b̂†k b̂k )â†k âk i~ κek αpk e−iωpk t−iφk (âk − â†k ),
X X
+ + +
k=L,R k=L,R

where J and V are the waveguide mediated optical and mechanical coupling strength (we gauged out the phase of J
and V and take both of them to be real), and the last two terms are the optical driving fields (pumps) which have
the same frequency and correlated phases. We consider the situation where the optical cavities are nearly degenerate,
i.e., ωcL ' ωcR ≡ ωc and both optomechanical systems are driven with a blue-detuned laser (ωpk = ωc + ωmk ).
We perform a displacement transformation âk = αk + dˆk , separating the classical steady state amplitude of the local
optical cavity field from its fluctuations. With this we can linearize the optomechanical interaction in the Hamiltonian
of Eq. B1 in the usual manner. Assuming the good cavity limit (sideband resolved, ωmk  κk ), we apply a rotating
wave approximation and obtain for the equations of motions (~ = 1)

d ˆ  κL  ˆ √ √
dL = i∆L − dL − κeL dˆL,in − κiL ξˆL,in − iJ dˆR − iGL b̂†L eiφL ,
dt 2
d ˆ  κR  ˆ √ √
dR = i∆R − dR − κeR dˆR,in − κiR ξˆR,in − iJ dˆL − iGR b̂†R eiφR ,
dt 2
d  γiL  √
b̂L = − iωmL + b̂L − γiL b̂L,in − iV b̂R − iGL dˆ†L eiφL ,
dt 2
d  γiR  √
b̂R = − iωmR + b̂R − γiR b̂R,in − iV b̂L − iGR dˆ†R eiφR , (B2)
dt 2
with the total damping rates κk = κek + κik , the detunings ∆k = ωp − ωck and the many-photon optomechanical
couplings Gk = g0k αk . The latter contains the steady state amplitude of the local optical cavity field αk eiφk , which
is related to the pump amplitudes through
√ √
iφL(R)
(i∆R(L) − κR(L) /2) κeL(R) αpL(R) e−iϕL(R) + iJ κeR(L) αpR(L) e−iϕR(L)
αL(R) e = . (B3)
(i∆L − κL /2)(i∆R − κR /2) + J 2
√
We find the steady state amplitude is approximately κek αpk e−iϕk /i∆k under the condition ∆k ≈ ωmk  κk , J,
which means each cavity is effectively only driven by its own optical pump. Thus, each pump-enhanced optomechanical
coupling and its phase can be independently controlled. The intrinsic noise operators ξˆk,in and b̂k,in in the coupled
mode equations B2 describe thermal and vacuum fluctuations impinging on the the cavities and the mechanical modes
respectively. The associated noise of a possible input signal is described via dˆk,in .
 12

2. Mechanically-mediated coupling

−iωt
We perform a Fourier transform (b̂[ω] ≡ dt b̂(t)e+iωt ; b̂(t) ≡ dω
R R
2π b̂[ω]e ) of the coupled mode equations Eqs. B2
†
and insert the resulting solution for b̂L,R [ω] into the equations of the cavity operators. Ignoring the intrinsic noise
terms ξˆin,k and b̂in,k for the moment, we obtain for the cavity operators in frequency space (ΦB = φL − φR )

ˆ √
e−1 κeL dˆL,in [ω] − i J − Γ+ [ω]e+iΦB dˆR [ω],


χ L,+ [ω]dL [ω] = −
ˆ √
e−1 κeR dˆR,in [ω] − i J − Γ+ [ω]e−iΦB dˆL [ω],


χR,+ [ω]dR [ω] = − (B4)

e−1 κk


with the modified susceptibility χ k,+ [ω] = −i(ω + ∆k ) + 2 + iΣk,+ [ω] . The frequency dependent coupling Γ+ [ω]
and the self-energy Σk,+ [ω] are defined as

VG G iGk h γ i
Γ+ [ω] =  γiL
 R L γiR
 , Σk,+ [ω] = −i(ω + ωmk̄ ) + ik̄ Γ+ [ω], (B5)
−i(ω + ωmL ) + (−i(ω + ωmR ) + +V 2 V Gk̄ 2
2 2

here the coupling Γ+ [ω] coincides with Eq. (2) of the main text. After eliminating the mechanical degrees of freedom,
one finds both a ”local” modification of each cavity (described by the self energy Σk,+ [ω]) and an induced coupling
between the cavities. The self-energies lead to damping (or anti-damping) of each cavity resonance as well as a
frequency shift of the resonance. Here the subscript + indicates blue-detuning (∆k = ωpk − ωc ≈ +ωmk ). The poles
of the self energy read
s  2
i 1 1 i
ω± = − (γiL + γiR ) − (ωmL + ωmR ) ± V 2 − (γiL − γiR ) − (ωmL − ωmR ) . (B6)
4 2 4 2

The induced coupling has a coherent and a dissipative aspect. To illustrate this we separate the coupling into real
and imaginary parts Γ+ [ω] ≡ ΓRe [ω] + iΓ= [ω]. The real and imaginary parts of this frequency-dependent coupling
have completely different physical interpretations. We see this, by considering again the coupling terms in Eq. (B4).
We have
h i
dˆL [ω] ∼ −i J − ΓRe [ω]e+iΦB − Γ= [ω]e+iΦB dˆR [ω] ≡ −iJe [ω] − Γ= [ω]e+iΦB dˆR [ω],


h i
dˆR [ω] ∼ −i J − ΓRe [ω]e−iΦB − Γ= [ω]e−iΦB dˆL [ω] ≡ −iJe∗ [ω] − Γ= [ω]e−iΦB dˆL [ω].


(B7)

For the given frequency of interest, we see that the real part of the induced coupling is completely equivalent to having a
Hamiltonian, coherent tunneling term between the cavities; we can absorb it into a redefinition of the coherent hopping
strength J, i.e., J → J[ω].
e In contrast, the coupling mediated by the imaginary part Γ= [ω] is not equivalent to some
effective coherent tunneling interaction between the cavities, i.e., the Γ= [ω] terms in dˆL and dˆR Eqs.(B7) cannot be
incorporated into a definition of J. The terms involving Γ= [ω] instead represent a dissipative coupling between the
two cavities mediated by the mechanics. Such dissipative interactions (if we ignore their frequency hdependence) can i
be obtained in a master equation formalism via an effective Lindblad dissipator of the form 2Γ= L d†L + e−i∆φ d†R ,
where L[ô]ρ̂ = ôρ̂ô† − 1/2ô† ôρ̂ − 1/2ρ̂ô† ô is the standard Lindblad superoperator.

3. Directionality by balancing coherent and dissipative interactions

The dissipative coupling is crucial for directionality: by balancing the dissipative interaction against the coherent
interaction we obtain a nonreciprocal system (following the general recipe outlined in Ref.[25]). For example, if we
aim for a directional transport from the left to the right cavity, we want to decouple the left cavity from the right
cavity (while still having the right cavity influenced by the left cavity). This is accomplished by balancing coherent
and dissipative interactions, i.e.,
!
J[ω]
e = iΓ= [ω]eiΦB , (B8)

in which case the coupling from the left to right cavity vanishes, cf. Eq. (B7), and we obtain a unidirectional coupling
where the right cavity is driven by the left cavity but not vice versa. Crucially, this would not be possible without
 13

the dissipative interaction, i.e., we need Γ= [ω] 6= 0. Note, for the situation that Γ= [ω] = 0, i.e., γik = 0, but finite
ΦB , we still obtain a directional dependent phase. However, to use this as the basic for nonreciprocal transmission
additional interference processes have to implemented.
The directionality condition Eq. (B8) can be reformulated in terms of the original J and the phase difference ΦB
as used in Eq. (B4). This translates to the condition

J = |Γ+ [ω]| , ΦB = − arg(Γ+ [ω]), (B9)

where we still aim for unidirectional behavior from left to right. For the case of a purely real coupling Γ+ [ω] = ΓRe [ω]
these conditions could still be satisfied, i.e., for ΦB = 0 and ΓRe [ω] = J. However, this means that there is effectively
no coupling between the cavities, and thus no forward transport either. Note, that a sign change in arg(Γ+ [ω]) would
lead to the opposite situation, where the propagation direction would be from right to left.
In general, the directionality balancing condition obtained here is frequency dependent, for the simple reason that
the mechanically-mediated cavity-cavity coupling is frequency-dependent. If we could somehow fulfill the directionality
condition in Eq. (B9) at every frequency, the cavity output field operators would be given by (using the standard
√
input-output relation dˆk,out = dˆk,out + κek dˆk )

dˆL,out [ω] = [1 − κeL χ

eL,+ [ω]] dˆL,in [ω],
√  
dˆR,out [ω] = [1 − κeR χ
eR,+ [ω]] dˆR,in [ω] − i κeR κeL χ χL,+ [ω] |Γ+ [ω]| ei2 arg(Γ+ [ω]) − 1 dˆL,in [ω],
eR,+ [ω]e (B10)

where we neglected the noise contributions originating from the mechanical modes, i.e., the coupling to b̂n,in in
Eq. (B4), and the intrinsic cavity noise ξin,k for simplicity. Here, we see again that the dissipative interaction is
crucial as we need arg(Γ+ [ω]) 6= nπ, n ∈ Z, i.e., we need a finite imaginary part of Γ+ [ω].
The experimentally relevant situation is where dissipative and coherent interactions are only balanced at a single
frequency (by appropriate tuning of phase and J). Achieving this condition close to the normal modes resonance
frequencies is favorable given the resonantly-enhanced transmission. Enforcing directionality at ω = −ωm ± V for
equal mechanical resonance frequencies, results in the directionality conditions

2V (γiL + γiR ) V GR GL
ωmL = ωmR : ΦB = ∓ arctan , J=q , (B11)
γiL γiR 1 2 2 2 γ2
γiL
4V (γiL + γiR ) + iR
16

where the upper (lower) sign in the phase difference ΦB realizes directionality at ω = −ωm +V (−ωm −V ). Directionality
here means that an input signal injected on the left cavity is transmitted to the right cavity, whereas the backward
propagation path, i.e., from right to left, is blocked.
On the other side, if we assume identical bare mechanical damping of the mechanical modes (γiL = γiR = γi ), but
q (ωmL 6= ωmR ), then we find that at the frequencies of the hybridized mechanical
unequal bare mechanical frequencies
modes Ω± = − 21 (ωmL + ωmR ) ± V 2 + 14 (ωmL − ωmR )2 the directionality condition is modified to
q
4 V 2 + 14 (ωmL − ωmR )2 V GL GR
γiL = γiR : ΦB = ∓ arctan , J= q . (B12)
γ γ2
γ V2+ 16 + 14 (ωmL − ωmR )2

where the upper (lower) sign in the phase difference ΦB realizes directionality at ω = Ω+(−) . The directionality
conditions for a perfectly symmetric device, i.e., for equal mechanical resonance frequencies (ωm ) and decay rates (γ),
can simply be read off from either Eq. B12 or Eq. B11.

4. Nonreciprocal optical transmission: two blue-detuned pumps

From the equations for the cavity operators in Eqs. B4 we can calulate the transmission coefficients via input/output
theory. Note, that although Eqs. B4 are formulated on the basis of noise operators, they as well describe the
dynamics of the cavity field amplitudes dk around their steady state solution. The right transmission coefficient
 14

TL→R ≡ dR,out /dL,in and left transmission coefficient TR→L ≡ dL,out /dR,in are given by
√
i κeL κeR J − Γ+ [ω]e∓iΦB
 
∓iΦB
 
TRL [ω] = −1 −1 ≡ A + [ω] J − Γ + [ω]e , (B13)
χ
eL [ω]eχR [ω] + [Γ+ [ω]2 + J 2 − 2Γ+ [ω]J cos(φ)]

with the modified susceptibilities χ

ek [ω] as defined after Eq. (B4). The prefactor A+ [ω] is the same for both transmission
amplitudes, it accounts for the mechanically-induced back-action on the optical cavities, cf. main text after Eq. (2).
Note, that the corresponding prefactor for two red-detuned pumps is simply A− [ω] = −A∗+ [−ω].
We now assume a completely symmetric pair of mechanical cavities (ωmL = ωmR = ωm and γiL = γiR = γi ) and
apply the corresponding directionality direction for symmetric parameters, cf. Eq. B12 or Eq. B11. The transmission
coefficient for the through direction (→) under these conditions of perfect nonreciprocity is given by,
s
γ √
κeL κeR 1 ± i 4Vi
r
8i CL CR
T→ [−ωm ± V ] = γi h  
i h
 
i ,
κR κL 1 ∓ i 4V C 1 ± i γi − 1 ∓ i 2V


2 ± i γi C 1 ± i γi − 1 ∓ i 2V 2 ± i γi
L 2V κL 2V R 2V κR 2V
(B14)

introducing the single cavity cooperativity Ck ≡ 4G2k /γi κk . Considering as well symmetric optical cavities (κeL =
κeR = κe ; κL = κR = κ) with symmetric optical pumping (GL = GR = G) the transmission coefficient simplifies to

V κ 8iC κκe
T→ [−ωm ± V ] '  γi
2 , (B15)
2 − C ± i 2V (1 − C)

with C ≡ 4G2 /γi κ and under the realistic assumption that the hopping rate V is much lower than the cavity decay
rate κ. Here we work with blue-detuned pumping of both optical cavities (∆ ≈ +ωm ), which results in parametric
amplification of each of the left and right mechanical modes and leads to amplification of the optical probe signal.
This becomes apparent for the situation that the mechanical hopping rate is much faster than the intrinsic mechanical
decay rate (V /γi  1). In this case the gain diverges for C → 2 (this is twice as large as for a single cavity instability
because the mechanical modes are hybridized
√ and thus the effective optomechanical coupling from the left or right
optical cavity is reduced by a factor of 2, hence the cooperativity by a factor of 2). Note, for the situation V /γi  1,
the directionality conditions at the hybridized mechanical modes ω = −ωm ± V simplifies to J ' GL GR /γi and
ΦB → ∓π/2.

5. Nonreciprocal optical transmission: two red-detuned pumps

The analysis for the the case of two red detuned pumps is similar to the blue-detuned case. The cavity operators
in Eq.(B2) couple now to the mechanical lowering operators b̂k and vice versa, while the detuning between the cavity
resonances and the external pump tones yields ∆k = −ωmk . The ratio of transmission coefficients is found to be given
by the following expression

J − −i(ω−ω )+ γiLV G−i(ω−ω

L GR
e−iΦB
TL→R J − Γ− [ω]e−iΦB [ mL 2 ][
γiR
mR )+ 2 ]+V
2
= = (B16)
TR→L J − Γ− [ω]e+iΦB J − −i(ω−ω )+ γiLV G−i(ω−ω
L GR
γiR e+iΦB
[ mL 2 ][ mR )+ 2 ]+V
2

where we have Γ− [ω] = Γ∗+ [−ω], thus the ratio |TL→R /TR→L | is the same for blue and red detuned pumps evaluated at
corresponding frequencies. The reason for this is that the transmission is either amplified or suppressed simultaneously
for both directions and thus their ratio stay unchanged. Comparing to the blue detuned case, the perfect nonreciprocity
condition remains the same in the red detuned case, while the transmission coefficient for the through direction the
hybridized mechanical modes Ω± = ωm ± V is given by (assuming ωmL = ωmR , γiL = γiR = γi and V  κk )
r √
κeL κeR 8i CL CR
T→ [ωm ± V ] '  γi
 γi
. (B17)
κR κL CL + 2 ± i 2V (CL + 1) CR + 2 ± i 2V (CR + 1)
p
From Eq. B17, we note in general an attenuated transmission for the red detuned case as T→ ≤ κeL κeR /(κL κR ) < 1.
For the case of a fast hopping rate V /γi  1 equality is achieved when Ck = 2 and/or κk /2 = GLγG
i
R
. Comparing the
latter to Eq. B11 we see the maximal through transmission efficiency is achieved when the optical cavity loss rate
 15

κk /2 is matched to the inter-cavity photon hopping rate J for both cavities (impedance matching condition).

6. Nonreciprocity associated with a single mechanical waveguide mode

In our optomechanical circuits, we also observed optical nonreciprocity with a single mechanical waveguide mode. In
this case, the Hamiltonian describing the interaction between two optical cavity modes and one mechanical waveguide
mode is given by,

~ωc,k â†k âk + J(â†L âR + âL â†R ) + ~ωMW b̂†W b̂W
X
Ĥ = (B18)
k=L,R
  √
b̂†W â†k âk + i~ κek αpk e−iωp t−iφk (âk − â†k ).
X X
∗
+ ~ g0,Wk b̂W + g0,Wk
k=L,R k=L,R

Going through a similar calculation using coupled mode equations, we find that the ratio of right and left optical
transmission coefficients is
|GWL GWR | −i(ΦB ±ΦW )
J ± i −i ω±ω γiW e
TL→R ( MW )+ 2
= |GWL GWR |
, (B19)
TR→L J ± i γiW e
+i(ΦB ±ΦW )
−i(ω±ωMW )+ 2

where the upper (lower) sign corresponds to the blue (red) detuned case and ΦW = arg(G∗WL GWR ). The corresponding
conditions for perfect directionality from left to right and at ω = ∓ωMW are

2|GWL GWR | π
J= , ΦB = ± ∓ ΦW . (B20)
γiW 2
This in turn leads to the transmission coefficients
r √
κeL κeR 4i CW L CW R
T→ [∓ωMW ] = . (B21)
κL κR (CW L ∓ 1)(CW R ∓ 1)

In the case of blue detuned tones an input signal is amplified and the corresponding gain increases for CW k → 1.
Note in Eq. B20 we included the phase of the product G∗WL GWR . This addition comes from the fact that we
have already chosen definitions for the local cavity mode amplitudes (aL,R and bL,R ) such that the phase of the
optomechanical couplings of the localized cavity modes – GL ≡ |αL |g0,L and GR ≡ |αR |g0,R – are both zero. With
these same definitions for amplitudes aL and aR we are not then free to set the phases of both GWL and GWL to be
zero; not at least for the same set of pump phases φL and φR chosen for the localized cavity mode coupling. A simple
example helps to illustrate this. The mode MW can be viewed as a hybridization between the localized left and right
cavity modes and a delocalized waveguide mode [21]. Using perturbation theory, we have for the mechanical mode
amplitude of the hybridized mode MW ,
tL tR
bW = bW0 + bL + bR , (B22)
0 − ωmL
ωMW 0 − ωmR
ωMW

where bW0 is the unperturbed delocalized waveguide mode amplitude and ωMW 0 is the unperturbed frequency of the
delocalized waveguide mode. tL(R) is the coupling coefficient between the delocalized waveguide mode and the localized
cavity mode ML(R) . The phases of tL and tR are determined by the field distribution of the hybridized mode MW
in the left and right cavities, respectively, and cannot be (both) chosen arbitrarily . Using the mode decomposition
∗
of Eq. B22, we have that arg(g0,WL g0,WR ) = arg(t∗L tR ) as we have already chosen a local cavity mode amplitude
basis such that arg(g0,L ) = arg(g0,R ) = 0 and ωMW 0 > ωmL , ωmR (this assumes of course that the left (right) optical
cavity mode only couples to the portion of bW which is due to bL (bR ), which is a good approximation due to the fact
that the optical cavities are in the far field of each other). Thus, by simultaneously measuring the flux-dependent
transmission near the resonance of the localized mechanical cavity modes and the hybridized mechanical waveguide
∗
mode we can determine the arg(g0,WL g0,WR ) in this mode basis (see Fig. 8 for example). For the MW mode in our
∗
experiment, we find arg(g0,WL g0,WR ) ≈ π, which means for this hybridized mode and chosen localized cavity mode
basis the mechanical motion in the left cavity as seen by the left cavity optical mode is approximately 180 degrees
out of phase with the motion in the right cavity as seen by the right cavity optical mode.
 16

Appendix C: Directional flow of quantum and thermal noise

Besides the nonreciprocal optical signal transmission, the flow of quantum and thermal noise in the optomechanical
circuit is directional. This is a natural consequence of the system’s scattering matrix having a directional form; the
scattering matrix determines both the transmission of coherent signals, as well as noise properties. To show this, we
calculate the symmetrized output noise spectral density via
Z
1 dΩ Dn ˆ oE
S̄k,out [ω] = dk,out [ω], dˆ†k,out [Ω] , (C1)
2 2π

defined in the standard manner [41]. The mechanical and optical noise operators introduced in Eqs. B2 have zero
mean and satisfy the canonical correlation relations:

hôk,in [ω]ô†k0 ,in [Ω]i = ho†k,in [ω]ok0 ,in [Ω]i + δk,k0 δ(ω + Ω) = nth ôk,in = dˆk,in , ξˆk,in , b̂k,in . (C2)


ok + 1 δk,k0 δ(ω + Ω),

where nth ok is the thermal occupation of each bath. In what follows, we assume that we have no thermal occupation
of the optical field. This is justified as we work with a very high optical frequency.
Figure 7a-d depicts the output spectra for the situation that both pumps are blue detuned from the cavity by ωm .
Here we assumed equal mechanical frequencies ωmL = ωmR = ωm and work in a rotating frame where the uncoupled
mechanical resonance frequencies are shifted to zero. The remaining parameters are as used in the experiment,
i.e., we take γiL /2π = 4.3 MHz, γiR /2π = 5.9 MHz, κL /2π = 1.03 GHz, κR /2π = 0.75 GHz, κiL /2π = 0.29 GHz,
κiR /2π = 0.31 GHz, V /2π = 2.8 MHz, J/2π = 110 MHz. The multiphoton couplings GL = GR used in the calculation
are determined from Eq. B11.
Figure 7a shows the result for zero temperature mechanical baths and a finite phase ΦB = 0.36π (determined from
Eq. B11). As expected, the L and R output spectra are not identical: while each has a double-peaked structure
(corresponding to the two normal mode resonances), the right output spectra S̄R,out [ω] has the upper-frequency peak
larger than the lower-frequency peak, while the situation is reversed for the left output spectra. This does not lead
to any asymmetry in the total output photon number fluxes (i.e., intergrated over all frequencies). It does however
lead to an asymmetry in the energy fluxes (i.e., as the higher energy peak is bigger for the right output spectrum,
and the low energy peak is bigger for the left spectrum). Thus, the ”quantum heating” of zero-point fluctuations
preferentially cause an energy flow to the right (rather than to the left) for this choice of phase.
It is also worth noting that if all dissipative rates are equal for the R and L cavities, then the L output spectrum
is just the frequency-mirrored R output spectrum. The latter is visible in Fig. 7(c), where we plotted the output
spectra for symmetric parameters, i.e., we set γiR /2π = γiL /2π = 4.3 MHz, κR /2π = κL /2π = 1.03 GHz, κiR /2π =
κiL /2π = 0.31 GHz and ΦB = 0.38π (determined from Eq. B11 for the new γiR ). However, having unequal decay
rates, i.e., γR 6= γL and κR 6= κL , leads to a slight asymmetry even if the phase is set to zero, i.e., ΦB = 0, as visible
in Fig.7b. In Fig. 7g we plot the asymmetry S̄L,out [ω] − S̄R,out [ω] for all the four cases corresponding to Fig. 7a-d.
For finite temperature, we find that the output spectrum has a roughly linear dependence on the mechanical bath
temperature: S̄k,out (T ) = ck nth + S̄k,out (0) (assuming nth th th
bL = nbR ≡ n ). This linear dependence is visible if we
compare Fig. 7c,d and Fig. 7e,f, where the latter show the output noise spectra for nth = 10 with symmetric cavity
parameters. Additionally, we also calculate the added noise quanta to the transmitted signal

S̄k,out [ω] 1
n̄k,add [ω] ≡ − , (C3)
|Tk [ω]|2 2

where 12 is the half quanta noise of the vacuum optical fields injected from the coupler. Fig.7h shows the added noise
for left-right propagation with ΦB = 0.36π (and asymmetric experimental cavity parameters). The mechanical baths
nth are varied as denoted in each graph. Even if the cavities and the mechanics are only driven by vacuum noise the
standard quantum limit (SQL) of half a quanta is not achieved. This is due to the limited amount of gain achieved
in the experiment, i.e., the transmission coefficient is not high enough to suppress the noise contributions. Moreover,
even in the large gain limit the added noise would be roughly one quanta due to the finite amount of intrinsic optical
cavity loss.

Appendix D: Reciprocal device

Realizing optical nonreciprocity in the optomechanical circuits studied in this work is not simple or easy as just
creating a circuit with optical and mechanical coupling between two optomechanical cavities. One is limited by the
 17

a b 4
5

4 3

3
2
2
1
1

0 0
-0.01 -0.005 0 0.005 0.01 -0.01 -0.005 0 0.005 0.01

c 4 d 3

3
2

1
1

0 0
-0.01 -0.005 0 0.005 0.01 -0.01 -0.005 0 0.005 0.01

e f 25
30
20

20 15

10
10
5

0 0
-0.01 -0.005 0 0.005 0.01 -0.01 -0.005 0 0.005 0.01

g 3 h 12

2 10

1 8
nth =2
0 6
nth =1
-1 4
nth =0
-2 2

-3 0
-0.01 -0.005 0 0.005 0.01 -0.01 -0.005 0 0.005 0.01

FIG. 7. a-d Output noise spectra at zero temperature for a set of parameters given in the text. e-f Output noise spectra at
finite temperature with thermal phonon occupation of nth = 10. g Difference of the left and right output spectra for a-d. f
Added noise for right-propagation signal.
 18

a b
-127 6

Normalized transmission (dB)

5
-128
4

PSD (dBm/Hz)
-129 3
2
-130
1

-131 0
-1
-132
5.77 5.78 5.79 5.8 5.77 5.78 5.79 5.8
Frquency (GHz) Frequency (GHz)

FIG. 8. Optical reciprocity in a circuit with large optical cavity coupling, J. a Mechanical spectra measured from the left
(red) and right (blue) optical cavities. b Normalized optical signal power transmission coefficient for forward (red) and reverse
(blue) optical signal propagation.

practical realities of device power handling capability, finite optical and mechanical Q-factors, etc. As such, not all the
circuits that were tested exhibited nonreciprocal transmission and amplification; the effects were too weak to observe
in some circuits. This, however, was a useful test of our set-up as nonreciprocity could be effectively turned on and
off by looking at different circuits with only slightly different parameters.
Eq. B12 sets the desired circuit parameters in order to achieve significant nonreciprocity, which for the optome-
chanical coupling, optical and mechanical Q-factors, and the power handling capabilities of the nanobeam cavities
requires optical hopping rate between cavities to be less than J/2π ≈ 500 MHz. Devices with larger coupling rates can
simply not be pumped hard enough to satisfy Gk ≈ (Jγik )1/2 . To confirm this, here we show another optomechanical
crystal circuit with bare cavity wavelengths of λL(R) = 1535.051 (1535.060) nm and inter-cavity photon hopping
rate of J/2π = 1.4 GHz (more than ten times larger than the device studied in the main text). The mechanical
spectra of this device as measured from both the left and right optical cavities is shown in Fig. 8a. Figure 8b shows
the normalized transmission coefficient for forward and reverse optical signal propagation for a blue-detuned pump
wavelength of λp = 1534.99 nm and synthetic flux of ΦB = π/2. Even at the largest pump powers (Pp ≈ 100 µW;
nc ≈ 1.5 × 103 ) this device does not satisfy the condition of Eq. B12 due to the large J, resulting in nearly perfect
reciprocity in the optical signal transmitted power. These measurements were performed on the exact same set-up as
the circuit studied in the main text.