Real-time two-axis control of a spin qubit

Fabrizio Berritta,1, ∗ Torbjørn Rasmussen,1 Jan A. Krzywda,2 Joost van der Heijden,3
Federico Fedele,1 Saeed Fallahi,4, 5 Geoffrey C. Gardner,5 Michael J. Manfra,4, 5, 6, 7
Evert van Nieuwenburg,2 Jeroen Danon,8 Anasua Chatterjee,1, † and Ferdinand Kuemmeth1, 3, ‡
1
Center for Quantum Devices, Niels Bohr Institute,
University of Copenhagen, 2100 Copenhagen, Denmark
2
Lorentz Institute and Leiden Institute of Advanced Computer Science,
Leiden University, P.O. Box 9506, 2300 RA Leiden, The Netherlands
3
QDevil, Quantum Machines, 2750 Ballerup, Denmark
4
Department of Physics and Astronomy, Purdue University, West Lafayette, Indiana 47907, USA
arXiv:2308.02012v2 [cond-mat.mes-hall] 26 Feb 2024

5
Birck Nanotechnology Center, Purdue University, West Lafayette, Indiana 47907, USA
6
Elmore Family School of Electrical and Computer Engineering,
Purdue University, West Lafayette, Indiana 47907, USA
7
School of Materials Engineering, Purdue University, West Lafayette, Indiana 47907, USA
8
Department of Physics, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway
(Dated: February 5, 2024)
Optimal control of qubits requires the ability to adapt continuously to their ever-changing envi-
ronment. We demonstrate a real-time control protocol for a two-electron singlet-triplet qubit with
two fluctuating Hamiltonian parameters. Our approach leverages single-shot readout classification
and dynamic waveform generation, allowing full Hamiltonian estimation to dynamically stabilize
and optimize the qubit performance. Powered by a field-programmable gate array (FPGA), the
quantum control electronics estimates the Overhauser field gradient between the two electrons in
real time, enabling controlled Overhauser-driven spin rotations and thus bypassing the need for
micromagnets or nuclear polarization protocols. It also estimates the exchange interaction between
the two electrons and adjusts their detuning, resulting in extended coherence of Hadamard rotations
when correcting for fluctuations of both qubit axes. Our study highlights the role of feedback in
enhancing the performance and stability of quantum devices affected by quasistatic noise.

INTRODUCTION Achieving precise control of gated qubits can be chal-

lenging due to their sensitivity to environmental fluctu-
Feedback is essential for stabilizing quantum devices ations, making feedback necessary to stabilize and op-
and improving their performance. Real-time monitoring timize their performance. Because feedback-based cor-
and control of quantum systems allows for precise ma- rections must be performed within the correlation time
nipulation of their quantum states [1, 2]. In this way, it of the relevant fluctuations, real-time control is essential.
can help mitigate the effects of quantum decoherence and Continuous feedback then allows to calibrate the qubit
extend the lifetime of quantum systems for quantum com- environment and to tune the qubit in real time to main-
puting and quantum sensing applications [3], for example tain high-fidelity gates and improved coherence, for in-
in superconducting qubits [4–8], spins in diamond [9–14], stance by suppressing low-frequency noise and improving
trapped atoms [15, 16], and other platforms [17–22]. π-flip gate fidelity [31]. An active reset of a silicon spin
qubit using feedback control was demonstrated based on
Among the various quantum-information processing
quantum non-demolition readout [32]. Real-time opera-
platforms, semiconductor spin qubits [23, 24] are promis-
tion of a charge sensor in a feedback loop [33] maintained
ing for quantum computing because of their long coher-
the sensor sensitivity for fast charge sensing in a Si/SiGe
ence times [25] and foundry compatibility [26]. Focusing
double quantum dot, compensating for disturbances due
on spin qubits hosted in gate-controlled quantum dots
to gate-voltage variation and 1/f charge fluctuations.
(QDs), two-qubit gate fidelities of 99.5% and single-qubit
A quantum state with higher confidence than what is
gate fidelities of 99.8% have recently been achieved in sil-
achievable through traditional thermal methods was ini-
icon [27]. In germanium, a four-qubit quantum processor
tialized by real-time monitoring and negative-result mea-
based on hole spins enabled all-electric qubit logic and the
surements [34].
generation of a four-qubit Greenberger-Horne-Zeilinger
state [28]. In gallium arsenide, simultaneous coherent ex- This study implements real-time two-axis control of a
change rotations and four-qubit measurements in a 2×2 qubit with two fluctuating Hamiltonian parameters that
array of singlet-triplets were demonstrated without feed- couple to the qubit along different directions on its Bloch
back, revealing site-specific fluctuations of nuclear spin sphere. The protocol involves two key steps: first, rapid
polarizations [29]. In silicon, a six-qubit processor was estimation of the instantaneous magnitude of one of the
operated with high fidelities enabling universal operation, fluctuating fields (nuclear field gradient) effectively cre-
reliable state preparation and measurement [30]. ates one qubit control axis. This control axis is then ex-
 2

ploited to probe in real time the qubit frequency (Heisen- a c 100 nm

ΩL
berg exchange coupling) across different operating points
(detuning voltages). Our procedure allows for counter- ΩH

energy
T0
acting fluctuations along both axes, resulting in an im- S
SD
proved quality factor of coherent qubit rotations.
Our protocol integrates a singlet-triplet (ST0 ) spin
ε
qubit implemented in a gallium arsenide double quan-
detuning ε 0
tum dot (DQD) [29] with Bayesian Hamiltonian estima- π S 0
b d
tion [35–39]. Specifically, an FPGA-powered quantum
orchestration platform (OPX [40]) repeatedly separates
Ω φ
singlet-correlated electron pairs using voltage pulses and J

energy
J φ
performs single-shot readout classifications to estimate |ΔBZ| ΔBZ
on-the-fly the fluctuating nuclear field gradient within Y
the double dot [41]. Knowledge of the field gradient in X

turn enables the OPX to coherently rotate the qubit be- 0 0

detuning ε 0 T0 1
tween S and T0 by arbitrary, user-defined target angles.
Differently from previous works, we let the gradient freely e
fluctuate, without pumping the nuclear field [42], and in-
stead program the OPX to adjust the baseband control 40
pulses accordingly.
An adaptive second-axis estimation is performed to 20
also probe the exchange interaction between the two elec- ΩL ΩH (MHz)
trons. This exchange interaction estimation scheme is 0 1 2 3 4
not simply an independent repetition of the single-axis laboratory time (s)
estimation protocol [35–39]: the design of the exchange-
based free induction decay (FID) pulse sequence depends FIG. 1. A singlet-triplet (ST0 ) qubit with two fluctu-
on the outcome of the first-axis estimation and needs to ating control axes. a The dots’ electrical detuning ε tunes
be computed on the fly. Finally, fluctuations along both from a regime of low qubit frequency, ΩL , to a regime of high
axes are measured and corrected, enabling the stable co- frequency, ΩH . States outside the computational space are
herent rotation of the qubit around a symmetric axis, not plotted. b In the first (second) regime, the Overhauser
gradient |∆Bz | (the exchange coupling J) dominates the qubit
essential for performing the Hadamard gate.
frequency ΩL (ΩH ) and the polar angle φ of the qubit rotation
Our work introduces a versatile method for enhancing axis. c SEM image of the GaAs device [29], implementing a
coherent control and stability of spin qubits by harnessing two-electron double quantum dot (black circles) next to its
low-frequency environmental fluctuations coupling to the sensor dot (SD) for qubit readout. d J and ∆Bz drive ro-
system. As such, it is not limited to the operation of tations of the qubit around two orthogonal axes, providing
ST0 qubits in GaAs. Our implementation of real-time universal qubit control, as depicted in the Bloch sphere. e
reaction to fluctuating Hamiltonian parameters can find Uncontrolled fluctuations of the Larmor frequencies ΩL and
ΩH , estimated in real time on the OPX and plotted with a
application in other materials and qubit encodings, as it 30 ms moving average.
is not necessarily limited to nuclear noise.

potentials of the two dots, which in turn sets the qubit’s

RESULTS spectrum as shown in Fig. 1a. We do not plot the fully
spin-polarized triplet states, which are independent of ε
Device and Bayesian estimation and detuned in energy by the applied magnetic field. We
define ε = 0 at the measurement point close to the in-
We use a top-gated GaAs DQD array from [29] and terdot (1,1)-(0,2) transition, with negative ε in the (1,1)
tune up one of its ST0 qubits using the gate electrodes region. In the ST0 basis, we model the time-dependent
shown in Fig. 1c, at 200 mT in-plane magnetic field in Hamiltonian by
a dilution refrigerator with a mixing-chamber plate be-
σz σx
low 30 mK. Radio-frequency reflectometry off the sen- H(t) = J(ε(t)) + g ∗ µB ∆Bz (t) , (1)
2 2
sor dot’s ohmic contact distinguishes the relevant charge
configurations of the DQD [43]. which depends on the detuning ε that controls the ex-
The qubit operates in the (1,1) and (0,2) charge con- change interaction between the two electrons, J(ε(t)),
figuration of the DQD. (Integers indicate the number of and the component of the Overhauser gradient parallel to
electrons in the left and right dot.) The electrical de- the applied magnetic field between the two dots, ∆Bz (t).
tuning ε quantifies the difference in the electrochemical σi are the Pauli operators, g ∗ is the effective g-factor, and
 3

µB is the Bohr magneton. In the following, we drop the probability of obtaining mi given a value of Ω:
time dependence of the Hamiltonian parameters for ease 1
of notation. On the Bloch sphere of the qubit (Fig. 1d), P (mi |Ω) = [1 + ri (α + β cos (2πΩti ))] , (2)
2
eigenstates of the exchange interaction, |S⟩ and |T0 ⟩, are
oriented along Z, while ∆Bz enables rotations along X. where ri takes a value of 1 (−1) if mi = |S⟩ (|T0 ⟩), and
The qubit is manipulated by voltage pulses applied to α and β are determined based on the measurement error
the plunger gates of the DQD, and measured near the in- and axis of rotation on the Bloch sphere.
terdot (1,1)-(0,2) transition by projecting the unknown Applying Bayes’ rule to estimate Ω based on the ob-
spin state of (1,1) onto either the (1,1) charge state (|T0 ⟩) served measurements mN , . . . m1 , which are assumed to
or the (0,2) charge state (|S⟩). Each single-shot readout be independent of each other, yields the posterior prob-
of the DQD charge configuration involves generation, de- ability distribution P (Ω |mN , . . . m1 ) in terms of a prior
modulation, and thresholding of a few-microsecond-long uniform distribution P0 (Ω) and a normalization constant
radio-frequency burst on the OPX (see Supplementary N:
Fig. 1). P (Ω |mN , . . . m1 ) =P0 (Ω) N
The OPX allows for real-time
p calculation of the qubit N
Y
Larmor frequency Ω(ε) = ∆Bz2 + J(ε)2 at different de- × [1 + ri (α + β cos (2πΩti ))] .
tunings, based on real-time estimates of ∆Bz and J(ε). i=1
(3)
Inspecting the exchange coupling in a simplified Fermi-
Based on previous works [35, 38, 39], we fix α = 0.25 and
Hubbard hopping model [23] and inserting J(ε) into
β = ±0.5, with the latter value positive when estimating
equation 1 suggests two physically distinct regimes
ΩL and negative when estimating ΩH . The expectation
[Fig. 1b]: At low detuning, in the (1,1) charge state con-
value ⟨Ω⟩, calculated over the posterior distribution after
figuration, the Overhauser gradient dominates the qubit
all N measurements, is then taken as the final estimate
dynamics.
p In this regime, the qubit frequency reads
2 , where we have added a small phe- of Ω.
ΩL ≡ ∆Bz2 + Jres
nomenological term Jres to account for a constant resid-
ual exchange between the two electrons at low detuning. Controlled Overhauser gradient driven rotations
Such a term may become relevant when precise knowl-
edge of ∆Bz is required, for example for the Hadamard
We first implement qubit control using one ran-
protocol at the end of this study. At high detuning, close
domly fluctuating Hamiltonian parameter, through rapid
to the (1,1)-(0,2) interdot charge transition, exchange in-
Bayesian estimation of ΩL and demonstration of con-
teraction between the two electrons p dominates, and the
trolled rotations of a ST0 qubit driven by the prevail-
qubit frequency becomes ΩH (ε) ≡ ∆Bz2 + J(ε)2 . As
ing Overhauser gradient. Notably, this allows coherent
shown in Fig. 1b, the detuning affects both the Larmor
control without a micromagnet [44, 45] or nuclear spin
frequency Ω and the polar angle φ of the qubit rotation
pumping [42].
axis ω̂, with φ approaching 0 in the limit J(ε) ≫ ∆Bz
ΩL is estimated from the pulse sequence shown in
and π/2 if J(ε) ≪ ∆Bz .
Fig. 2a: for each repetition a singlet pair is initialized
Without the possibility of turning off either J or ∆Bz , in (0,2) and subsequently detuned deep in the (1, 1) re-
the rotation axes of the singlet-triplet qubit are tilted, gion (εL ≈ −40 mV) for N =101 linearly spaced separa-
meaning that pure X- and Z-rotations are unavailable. tion times ti up to 100 ns. After each separation, the
In their absence, the estimation of the qubit frequency qubit state, |S⟩ or |T0 ⟩, is assigned by thresholding the
at different operating points is crucial for navigating the demodulated reflectometry signal Vrf near the (1,1)-(0,2)
whole Bloch sphere of the qubit. Figure 1e tracks Larmor interdot transition and updating the Bayesian probabil-
frequencies ΩH and ΩL , both fluctuating over tens of MHz ity distribution of ΩL according to the outcome of the
over a period of several seconds, using a real-time pro- measurement. After measurement mN , the initially uni-
tocol as explained later. The presence of low-frequency form distribution has narrowed [inset of Fig. 2b, with
variations in time traces of ΩH and ΩL suggests that qubit white and black indicating low and high probability], al-
coherence can be extended by monitoring these uncon- lowing the extraction of ⟨ΩL ⟩ as the estimate for ΩL . For
trolled fluctuations in real time and appropriately com- illustrative purposes, we plot in Fig. 2a the N single-shot
pensating qubit manipulation pulses on-the-fly. measurements mi for 10,000 repetitions of this protocol,
To estimate the frequency of the fluctuating Hamilto- which span a period of about 20 s, and in Fig. 2b the
nian parameters on the OPX, we employ a Bayesian esti- associated probability distribution P (ΩL ) of each repe-
mation approach based on a series of free-induction-decay tition. The quality of the estimation seems to be lower
experiments [35]. Using mi to represent the outcome (|S⟩ around a laboratory time of 6 seconds, coinciding with a
or |T0 ⟩) of the i-th measurement after an evolution time reduced visibility of the oscillations in panel 2a. We at-
ti , the conditional probability P (mi |Ω) is defined as the tribute this to an enhanced relaxation of the triplet state
 4

a probe ΩL b estimate ΩL d controlled rotations

tN
20 μs 5 μs N M

S ti S
~
ti mi BE tj mj
ε
t
t1
ΩL
S T P(ΩL) 0 0.6 S T
10000

20 7690
8000

7226

laboratory time (s)

15
6000
repetition

6262
10
4000
3526

2000 5
3232

Q≈1 25 50 75 Q≳7
0.8 ΩL (MHz)
0.8
p(S)

p(S)
0.6 c compile waveforms
0.6
M
0.4
ΩL >50 MHz? ~ θj
0 50 100 tj 0 5
true ΩL θj (2π)
ti (ns)

false

FIG. 2. Controlled Overhauser gradient driven rotations of a ST0 qubit by real-time Bayesian estimation. One
loop (solid arrows) represents one repetition of the protocol. a For each repetition, the OPX estimates ΩL by separating a
singlet pair for N linearly spaced probe times ti and updating the Bayesian estimate (BE) distribution after each measurement,
as shown in the inset of b for one representative repetition. For illustrative purposes, each single-shot measurements mi is
plotted as a white/black pixel, here for N =101 ΩL probe cycles, and the fraction of singlet outcomes in each column is shown
as a red dot. b Probability distribution P (ΩL ) after completion of each repetition in a. Extraction of the expected value ⟨ΩL ⟩
from each row completes ΩL estimation. c For each repetition, unless ⟨ΩL ⟩ falls below a user-defined minimum (here 50 MHz),
the OPX adjusts the separation times t̃j , using its real-time knowledge of ⟨ΩL ⟩, to rotate the qubit by user-defined target angles
θj = t̃j ⟨ΩL ⟩. d To illustrate the increased coherence of Overhauser gradient driven rotations, we task the OPX to perform
M =80 evenly spaced θj rotations. Single-shot measurements mj are plotted as white/black pixels, and the fraction of singlet
outcomes in each column is shown as a red dot.

during readout due to the relatively high |∆Bz | gradi- and 8π. In our notation, the tilde in a symbol x̃ indicates
ent during those repetitions [46]. The visibility could that the waveform parameter x is computed dynamically
be improved by a latched or shelved read-out [47, 48] or on the OPX. To reduce the FPGA memory required for
energy-selective tunneling-based readout [38]. preparing waveforms with nanosecond resolution, we per-
form controlled rotations only if the expected ΩL is larger
Even though the rotation speed around ω̂ L at low de-
than an arbitrarily chosen minimum of 50 MHz. (The
tuning is randomly fluctuating in time, knowledge of ⟨ΩL ⟩
associated if statement and waveform compilation then
allows controlled rotations by user-defined target angles.
takes about 40 µs on the FPGA.) This reduces the num-
To show this, we task the OPX in Fig. 2d to adjust the
ber of precomputed waveforms needed for the execution
separation times t̃j in the pulse sequence to rotate the
of pulses with nanosecond-scale granularity, for which we
qubit by M =80 different angles θj = t̃j ⟨ΩL ⟩ between 0
 5

use the OPX baked waveforms capability. Accordingly, knowledge of the Overhauser field gradient, the qubit
the number of rows in Fig. 2d (1,450) is smaller than in would traditionally be initialized near the equator by
panel a, and we only label a few selected rows with their adiabatically reducing detuning from (0,2) to the (1,1)
repetition number. charge configuration, and a reverse ramp for readout.
To show the increased rotation-angle coherence of con- Such adiabatic ramps usually last several microseconds
trolled |∆Bz |-driven rotations, we plot the average of all each, while our t̃π/2 pulses typically take less than 10 ns,
1,450 repetitions of Fig. 2d and compare the associated thereby significantly shortening each probe cycle.
quality factor, Q > ∼ 7, with that of uncontrolled oscilla- For the estimate of ΩH , the Bayesian probability distri-
tions, Q ∼ 1 (we define the quality factor as the number bution of ΩH is updated after each of the M =101 single-
of oscillations until the amplitude is 1/e of its original shot measurement mj , each corresponding to a separa-
value). The average of the uncontrolled S-T0 oscillations tion time tj that is evenly stepped from 0 to 100 ns. The
in Fig. 2a can be fit by a decay with Gaussian envelope Bayesian probability distributions of both ΩL and ΩH are
(solid line), yielding an inhomogeneous dephasing time shown in Fig. 3c and d, respectively, with the latter being
T2∗ ≈ 30 ns typical for ST0 qubits in GaAs [49]. We asso- conditioned on 20 MHz < ⟨ΩL ⟩ < 40 MHz to reduce the
ciate the relatively smaller amplitude of stabilized qubit required FPGA memory.
oscillations with the low-visibility region around 6 sec- This section demonstrated a real-time baseband con-
onds in Fig. 2d, discussed earlier. Excluding such regions trol protocol that enables manipulation of a spin qubit
by post selection increases the visibility and quality fac- on the entire Bloch sphere.
tor of oscillations (see Supplementary Fig. 4). Overall,
the results presented in this section exemplify how adap-
tive baseband control pulses can operate a qubit reliably, Controlled exchange-driven rotations
out of slowly fluctuating environments.
Using Bayesian inference to estimate control axes in
real-time offers new possibilities for studying and mitigat-
Real-time two-axis estimation ing qubit noise at all detunings. Figure 4a describes the
real-time controlled exchange-driven rotations protocol
In addition to nuclear spin noise, ST0 qubits are ex- aimed at stabilizing frequency fluctuations of the qubit
posed to electrical noise in their environment, which af- at higher detunings. Following the approach of Fig. 3,
fects the qubit splitting in particular at higher detunings. we first estimate ΩL and ΩH using real-time Bayesian es-
It is therefore important to examine and mitigate low- timation. We then use our knowledge of ΩH to increase
frequency noise at different operating points of the qubit. the rotation angle coherence of the qubit where the ex-
In the previous section, the qubit frequency ΩL was es- change coupling is comparable with the Overhauser field
timated entirely at low detuning where the Overhauser gradient.
field gradient dominates over the exchange interaction. As illustrated in Fig. 4a, the qubit control pulses now
In order to probe and stabilize also the second control respond in real time to both qubit frequencies ΩL and
axis, namely J-driven rotations corresponding to small ΩH . Similar to the previous section, after determining
φ in Fig. 1d, we probe the qubit frequency ΩH at higher ⟨ΩL ⟩ and confirming that 30 MHz < ⟨ΩL ⟩ < 50 MHz is
detunings, using a similar protocol with a modified qubit fulfilled, the qubit is initialized near the equator of the
initialization. Bloch sphere by fast diabatic ΩL (π/2) pulses, followed
Free evolution of the initial state |S⟩ around ω̂ L would by an exchange-based FID that probes ΩH . Based on
result in low-visibility exchange-driven oscillations be- the resulting ⟨ΩH ⟩, the OPX adjusts the separation times
cause of the low value of φ. To circumvent this prob- t̃l to rotate the qubit by user-defined target angles θl =
lem, we precede the ΩH estimation by one repetition of t̃l ⟨ΩH ⟩.
ΩL estimation, as shown in Fig. 3a. This way, real-time To show the resulting improvement of coherent ex-
knowledge of ⟨ΩL ⟩ allows the initial state |S⟩ to be ro- change oscillations, we plot in Fig. 4c the interleaved
tated to a state near the equator of the Bloch sphere, K=101 measurements ml and compare them in Fig. 4b to
before it evolves freely for probing ΩH . This rotation is the M =101 measurements mj . Fitting the average of the
implemented by a diabatic detuning pulse from (0,2) to uncontrolled rotations by an oscillatory fit with Gaussian
εL (diabatic compared to the interdot tunnel coupling) envelope decay yields Tel∗ ≈ 60 ns and Q ≈ 3, presumably
for time t̃π/2 , corresponding to a rotation of the qubit limited by electrical noise [49], while the quality factor of
around ω̂ L by an angle ΩL t̃π/2 = π/2. After evolution the controlled rotations is enhanced by a factor of two,
for time tj under finite exchange, another π/2 rotation Q ≈ 6.
around ω̂ L rotates the qubit to achieve a high readout The online control of exchange-driven rotations using
contrast in the ST0 basis, as illustrated on the Bloch Bayesian inference stabilizes fluctuations of the qubit fre-
sphere in Fig. 3b. quency at higher detunings, where fluctuations are more
As a side note, we mention that in the absence of sensitive to detuning noise. Indeed, we attribute the
 6

a ΩL estimation ΩH estimation b
S 0
30 μs 5 μs
N M
Compute
S S
~ -1 tj
ti mi BE tπ/2 4 ΩL mj BE
~ ~
tπ/2 tπ/2
T0 1

c d
0.0 0.6 0.0 0.4

60
40

ΩH (MHz)
ΩL (MHz)

40
20

1 2 3 1 2 3
laboratory time (s) laboratory time (s)

FIG. 3. Real-time Bayesian estimation of two control axes. a One repetition of the two-axis estimation protocol. After
estimating ΩL from N =101 ti probe cycles (Fig. 2a), the OPX computes on-the-fly the pulse duration t̃π/2 required to initialize
the qubit near the equator of the Bloch sphere by a diabatic ΩL (π/2) pulse. After the ΩL (π/2) pulse, the qubit evolves for
time tj under exchange interaction before another ΩL (π/2) pulse initiates readout. After each single-shot measurement mj , the
OPX updates the BE distribution of ΩH . Similar to ti in the ΩL estimation, tj is spaced evenly between 0 and 100 ns across
M = 101 exchange probe cycles. b Qubit evolution on the Bloch sphere during one exchange probe cycle. c Each column plots
P (ΩL ) after completion of the ΩL estimation in each protocol repetition. d Each column plots P (ΩH ) after completion of the
ΩH estimation in each protocol repetition.

slightly smaller quality factor, relative to Overhauser- resonant-driving approach of previous works [35, 38, 39].
driven rotations in Fig. 2d, to an increased sensitivity In the resonant implementation, constrained to the op-
to charge noise at larger detuning, which, owing to its erating regime |∆Bz | ≫ J, low-frequency fluctuations of
high-frequency component, is more likely to fluctuate on J result in transverse noise that causes dephasing and
the estimation timescales [36]. phase shifts of the Rabi rotations [51, 52].
This section established for the first time stabilization In previous sections, we have shown how to probe the
of two rotation axes of a spin qubit. This advancement qubit Larmor frequencies ΩH and ΩL at different detun-
should allow for stabilized control over the entire Bloch ings in real time and correct for their fluctuations. Now,
sphere, which we demonstrate in the next section. we simultaneously counteract fluctuations in J and |∆Bz |
on the OPX in order to perform the Hadamard gate. As
we do not measure the sign of ∆Bz , we identify the polar
Hadamard rotations angle of ω̂ Had as either φ = π/4 or −π/4.
In other words, starting from the singlet state, the
In this experiment, we demonstrate universal ST0 con- qubit rotates towards +X on the Bloch sphere for one
trol that corrects for fluctuations in all Hamiltonian pa- sign of ∆Bz , and towards −X for the other sign. The
rameters. We execute controlled Hadamard rotations sign of the gradient may change over long time scales due
around ω̂ Had , as depicted by the trajectory on the Bloch to nuclear spin diffusion (on the order of many seconds
sphere of Fig. 5d, by selecting the detuning εHad in real [41]), but the measurement outcomes of our protocol are
time such that J(εHad ) = |∆Bz |. To achieve this, we do expected to be independent of the sign.
not assume that ∆Bz = ΩL (i.e. we allow contributions The relative sign of Overhauser gradients becomes rel-
of Jres to ΩL ) or that J = ΩL (i.e. we allow contribu- evant for multi-qubit experiments [53], and could be de-
tions of ∆Bz to ΩH ). The full protocol is detailed in termined following [54] by comparing the relaxation time
Supplementary Discussion. of the ground state (e.g. |↑↓⟩) of ∆Bz with its excited
Real-time knowledge of both ∆Bz and J would po- state (|↓↑⟩). Such diagnostic sign-probing cycles on the
tentially benefit two-qubit gate fidelities [50] and the FPGA should not require more than a few milliseconds,
 7

a controlled ΩH rotations tuning up J(ε) = |∆Bz |. We also provide the OPX

K with a value for the residual exchange at low detuning,
ΩL & ΩH
estimation S Jres ≈ 20 MHz, determined offline as described in Sup-
~
tl ml plementary Fig. 5.
~ ~ As illustrated in Fig. 5a, the Hadamard rotation pro-
tπ/2 tπ/2
tocol starts by estimating |∆Bz | from ΩL , taking into
account a constant residual exchange by solving ∆Bz2 =
b S T c Ω2L − Jres
2
.
1000 Next, an initial value of εHad is chosen based on the
linear offline model to fulfill J(εHad ) = |∆Bz | [feed-
12.5 back ○ 1 in panel (a,c)]. To detect any deviations of the
800 prevailing J from the offline model, an exchange-driven
FID is performed at εHad to estimate J from ΩH , using
10.0 J 2 = Ω2H − ∆Bz2 .
Any deviation of ⟨J⟩ from the target value |∆Bz | is

laboratory time (s)

600
subsequently corrected for by updating εHad based on
7.5
repetition

the linearized J(ε) model [feedback ○ 2 in panels (a,c)].

400
Matching J to |∆Bz | in the two detuning feedback steps
5.0 each takes about 400 ns √ on the OPX. Finally, real-time
knowledge of ΩHad ≡ 2|∆Bz | is used to generate the
free evolution times t̃i , spent at the updated value ε̃Had ,
200
2.5 in order to perform Hadamard rotations by K user de-
fined target angles.
The resulting Hadamard oscillations are shown in
Fig. 5e (top panel) and fitted with an exponentially de-
Q≈3 Q≈6
0.8 caying sinusoid, indicating a quality factor Q > 5. (Ac-
cording to this naive fit, the amplitude drops to 1/e over
p(S)

approximately 40 rotations, although we have not exper-

0.6 imentally explored rotation angles beyond 9π.) In com-
parison to exchange-controlled rotations from Fig. 4c, the
0 50 100 0 2 4 6 Hadamard rotations are more stable, which we attribute
tj (ns) θl (2π)
to the additional feedback on detuning that fixes the os-
cillation axis and decreases sensitivity to charge noise.
FIG. 4. Real-time-controlled exchange-driven qubit To illustrate the crucial role of real-time estimation for
rotations. a One repetition of the exchange rotation proto- this experiment, we also performed rotation experiments
col. After estimation of ΩL and ΩH as in Fig. 3, the OPX
adjusts exchange duration times t̃l , using real-time knowl-
that do not involve any real-time estimation and feedback
edge of ⟨ΩH ⟩, to rotate the qubit by user-defined target an- cycles (dark gray data, bottom panel), as follows. Within
gles θl = t̃l ⟨ΩH ⟩. Pulse durations t̃π/2 for qubit initialization minutes after performing the controlled Hadamard rota-
and readout use real-time knowledge of ⟨ΩL ⟩. b Each row tions (purple data), we executed Hadamard rotations as-
plots measurements mj from one protocol repetition, here suming a fixed value of |∆Bz | = |∆Bz |, i.e. by pulsing to
M =101 exchange probe outcomes. c Each row plots measure- a fixed detuning value corresponding to J(εH ) = |∆Bz |
ment ml from one protocol repetition, here K=101 controlled- according to the offline model. Here, |∆Bz | ≈ 40 MHz
exchange-rotation outcomes. To illustrate the increased co-
herence of controlled exchange rotations, we also plot in b and
is the average Overhauser gradient that we observed just
c the fraction of singlet outcomes of each column. before executing the Hadamard protocol. Not surpris-
ingly, the quality factor of the resulting Hadamard-like
oscillations is low and the rotation angle deviates from
the intended target angle, likely due to the Overhauser
negligible compared to the expected time between sign gradient having drifted in time. As a side note, we men-
reversals. tion that the purple data in Fig. 5e constitutes an aver-
In preparation for our protocol, we first extract the age over 5,000 repetitions, corresponding to a total ac-
time-averaged exchange profile by performing exchange quisition time of 2 minutes including Overhauser and ex-
oscillations as a function of evolution time [Fig. 5b]. Re- change estimation cycles. In contrast, the dark gray data
moving contributions of ∆Bz to ΩH then yields J(ε) also constitutes an average over 5,000 repetitions, but
in Fig. 5c. A linear approximation in the target range only required 15 seconds because of the omission of all
40 MHz < ⟨J⟩ < 60 MHz (dashed blue line) is needed estimation and feedback cycles.
later on the OPX to allow initial detuning guesses when To verify that the enhancement in Q is not solely due to
 8

a b d S
7.25 Vrf (mV) 7.55

80
Overhauser gradient estimation
N X

exchange evolution (ns)

S 60
T0
ti
e |ΔBz| and J estimation

40 0.85

Choose J(ε) = |ΔBz| 1

20 0.80
Exchange interaction estimation
M
|ΔBz| estimation
S 0
tj -10 -5
~ detuning (mV)
ε 0.90
~ ~

p(S)
tπ/2 tπ/2 c
80 0.80
Feedback on ε 2

60 no estimation
J (MHz)

|ΔBz|
Hadamard rotations 1
K 2
40 0.85
S
~
~ ti
εHad 20 0.80

-10 -5 1 3 5 7
detuning (mV) Hadamard rotation (π)

FIG. 5. Real-time universal ST0 control demonstrated by Hadamard rotations. a Hadamard rotation protocol.
After estimating ΩL , ε is chosen in real-time such that J(ε) = |∆Bz |, based on a linearized offline model from panel c. If
40 MHz < ⟨|∆Bz |⟩ < 60 MHz, the√ detuning is adjusted to account for deviations of the prevailing J from the offline model.
Real-time knowledge of ΩHad = 2 |∆Bz | then dictates t̃i to achieve a user-defined Hadamard rotation angle. b Averaged
exchange driven FID as a function of detuning and evolution time. Here, a diabatic ΩL (π/2) pulse initializes the qubit near
the equator of the Bloch sphere, prior to free exchange evolution, and subsequently prepares it for readout. c J as a function
of ε extracted offline from b, as well as a linearized model (dashed line) used in the two feedback steps of panel a. d Hadamard
rotation depicted on the Bloch sphere. e Measurement of Hadamard rotations with |∆Bz | and J estimation (purple, top panel),
only |∆Bz | estimation (light gray, middle panel), and without the feedback shown in a (dark gray, bottom panel).

the more accurate knowledge of |∆Bz |, we also performed rescaled to the Hadamard evolution time using the fitted
Hadamard rotations only using the estimation of |∆Bz |. frequency ≈ 29 MHz.) We see that (i) a reduction of the
The FPGA was programmed to perform a measurement uncertainty in |∆Bz | from ≈ 30 MHz (r.m.s.) to ≈ 2 MHz
where the initialized singlet is pulsed to a fixed detuning (dark gray to light gray) does not yield a proportional
J(εH ) ≈ 20 MHz to perform a Hadamard rotation, only gain in Q and (ii) the improvement in Q when includ-
if the estimated |∆Bz | on the FPGA satisfies 17 MHz < ing estimation of J (light gray to purple) is much larger
⟨|∆Bz |⟩ < 23 MHz. We then post select the repetitions than can be justified solely by the slight further reduc-
where 19.5 MHz < ⟨|∆Bz |⟩ < 20.5 MHz. Fitting this tion of the uncertainty in |∆Bz | (roughly from ≈ 2 MHz
by an oscillatory fit with Gaussian envelope decay yields to ≈ 1 MHz). This demonstrates the crucial contribution
Tel∗ ≈ 70 ns, Q ≈ 2.0 and frequency ≈ 29 MHz. of the estimations along both axes in the improvement of
our Hadamard gate quality factor.
In Figure 5e we compare these data (light gray, middle
panel) with the cases where the FPGA estimated both Further evidence for the fluctuating nature of non-
|∆Bz | and J (purple, top panel) and where the micro- stabilized Hadamard rotations is discussed in Supplemen-
processor does not perform any estimation but simply tary Fig. 6.
pulses to J(εH ) to perform the rotations (dark gray, bot- The stabilized Hadamard rotations demonstrate real-
tom panel). (In the middle panel the horizontal axis was time feedback control based on Bayesian estimation of
 9

J and |∆Bz |, and suggest a significant improvement in namics [58–60], possibly via long short-term memory ar-
coherence for ST0 qubit rotations around a tilted control tificial neural networks as reported for superconducting
axis. Despite the presence of fluctuations in all Hamil- qubits [61]. While our current qubit cycle time (approx-
tonian parameters, we report effectively constant ampli- imately 30 µs) is dominated by readout and qubit ini-
tude of Hadamard oscillations, with a reduced visibility tialization, it can potentially be reduced to a few mi-
that we tentatively attribute to estimation and readout croseconds through faster qubit state classification, such
errors. as enhanced latched readout [48], and faster reset, such
as fast exchange of one electron with the reservoir [62].
Our protocol could be modified for real-time non-local
DISCUSSION noise correlations [63] or in-situ qubit tomography us-
ing fast Bayesian tomography [64] to study the underly-
Our experiments demonstrate the effectiveness of feed- ing physics of the noisy environment, thereby providing
back control in stabilizing and improving the perfor- qualitatively new insights into processes affecting qubit
mance of a singlet-triplet spin qubit. The protocols pre- coherence and multi-qubit error correction.
sented showcase two-axis control of a qubit with two Beyond ST0 qubits, our protocols uncover new per-
fluctuating Hamiltonian parameters, made possible by spectives on coherent control of quantum systems manip-
implementing online Bayesian estimation and feedback ulated by baseband pulses. This work represents a signif-
on a low-latency FPGA-powered qubit control system. icant advancement in quantum control by implementing
Real-time estimation allows control pulses to counteract an FPGA-powered technique to stabilize in real time the
fluctuations in the Overhauser gradient, enabling con- qubit frequency at different manipulation points.
trolled Overhauser-driven rotations without the need for
micromagnets or nuclear polarization protocols. Notably,
even in the absence of a deterministic component of the METHODS
Hamiltonian purely noise-driven coherent rotations of a
two-level quantum system were demonstrated. Experimental setup
The approach is extended to the real-time estimation of
the second rotation axis, dominated by exchange interac- We use an Oxford Instruments Triton 200 cryofree di-
tion, which we then combine with an adaptive feedback lution refrigerator with base temperature below 30 mK.
loop to generate and stabilize Hadamard rotations. In The experimental setup employs a Quantum Machines
particular, executing the Hadamard gate involves (i) se- OPX+ for radio-frequency (RF) reflectometry and gate
quentially executing two distinct estimation cycles, where control pulses. The RF carrier frequency is ≈ 158 MHz
the design of the second cycle relies on the outcomes of and the gate control pulses sent to the left and right
the first, (ii) correlating the detected frequencies to dis- plunger gates of the DQD are filtered with low-pass filters
tinguish independent fluctuations of the two control axes, (≈ 220 MHz) at room temperature, before being attenu-
and (iii) utilizing this correlated information to dynam- ated at different stages of the refrigerator. Low-frequency
ically construct and execute a Hadamard gate. These tuning voltages (high-frequency baseband waveforms) are
steps demand real-time adaptive estimations and signal applied by a QDAC [65] (OPX) via a QBoard high-
generations throughout the protocol, which has not been bandwidth sample holder [66].
demonstrated before. A constant Overhauser field gra-
dient, whether stemming from nuclear spin pumping or
a micromagnet, is expected to further improve the feed- Measurement details
back control. From this perspective, our work represents
a worst-case scenario, demonstrating the effectiveness of Before qubit manipulation, an additional reflectometry
our experimental technique. measurement is taken as a reference to counteract slow
Our protocols assume that ∆Bz does not depend on drifts in the sensor dot signal. At the end of each qubit
the precise dot detuning in the (1,1) configuration and re- cycle, a ≈ 1 µs long pulse is applied to discharge the bias
mains constant on the time scale of one estimation. Sim- tee. As the qubit cycle period (tens of µs) is much shorter
ilarly, stabilization of exchange rotations is only effective than the bias tee cutoff (≈ 300 Hz), we do not correct the
for electrical fluctuations that are slow compared to one pulses for the transfer function of the bias tee.
estimation. Therefore, we expect potential for further
improvements by more efficient estimation methods, for
example through adaptive schemes [55] for Bayesian esti- DATA AVAILABILITY
mation from fewer samples, or by taking into account the
statistical properties of a time-varying signal described The datasets generated and analyzed during the cur-
by a Wiener process [56] or a nuclear spin bath [57]. rent study are available from the corresponding authors
Machine learning could be used to predict the qubit dy- (F.B., A.C., and F.K.) upon request.
 10

[21] Magrini, L. et al. Real-time optimal quantum control of

mechanical motion at room temperature. Nature 595,
373–377 (2021).
∗
fabrizio.berritta@nbi.ku.dk [22] Tebbenjohanns, F., Mattana, M. L., Rossi, M., Frimmer,
†
anasua.chatterjee@nbi.ku.dk M. & Novotny, L. Quantum control of a nanoparticle
‡
kuemmeth@nbi.dk optically levitated in cryogenic free space. Nature 595,
[1] Wiseman, H. M. Quantum theory of continuous feedback. 378–382 (2021).
Physical Review A 49, 2133–2150 (1994). [23] Burkard, G., Ladd, T. D., Pan, A., Nichol, J. M. & Petta,
[2] Wiseman, H. M. & Milburn, G. J. Quantum Measure- J. R. Semiconductor spin qubits. Rev. Mod. Phys. 95,
ment and Control (Cambridge University Press, 2009). 025003 (2023).
[3] Zhang, J., xi Liu, Y., Wu, R.-B., Jacobs, K. & Nori, F. [24] Chatterjee, A. et al. Semiconductor qubits in practice.
Quantum feedback: Theory, experiments, and applica- Nature Reviews Physics 3, 157–177 (2021).
tions. Physics Reports 679, 1–60 (2017). [25] Stano, P. & Loss, D. Review of performance metrics
[4] Vijay, R. et al. Stabilizing Rabi oscillations in a super- of spin qubits in gated semiconducting nanostructures.
conducting qubit using quantum feedback. Nature 490, Nature Reviews Physics 4, 672–688 (2022).
77–80 (2012). [26] Zwerver, A. M. J. et al. Qubits made by advanced semi-
[5] Campagne-Ibarcq, P. et al. Persistent control of a super- conductor manufacturing. Nature Electronics 5, 184–190
conducting qubit by stroboscopic measurement feedback. (2022).
Physical Review X 3, 021008 (2013). [27] Noiri, A. et al. Fast universal quantum gate above the
[6] de Lange, G. et al. Reversing quantum trajectories with fault-tolerance threshold in silicon. Nature 601, 338–342
analog feedback. Physical Review Letters 112, 080501 (2022).
(2014). [28] Hendrickx, N. W. et al. A four-qubit germanium quan-
[7] Masuyama, Y. et al. Information-to-work conversion tum processor. Nature 591, 580–585 (2021).
by Maxwell’s demon in a superconducting circuit quan- [29] Fedele, F. et al. Simultaneous operations in a two-
tum electrodynamical system. Nature Communications dimensional array of singlet-triplet qubits. PRX Quan-
9 (2018). tum 2, 040306 (2021).
[8] Vepsäläinen, A. et al. Improving qubit coherence using [30] Philips, S. G. J. et al. Universal control of a six-qubit
closed-loop feedback. Nature Communications 13 (2022). quantum processor in silicon. Nature 609, 919–924
[9] Blok, M. S. et al. Manipulating a qubit through the (2022).
backaction of sequential partial measurements and real- [31] Nakajima, T. et al. Coherence of a driven electron spin
time feedback. Nature Physics 10, 189–193 (2014). qubit actively decoupled from quasistatic noise. Physical
[10] Bonato, C. et al. Optimized quantum sensing with a sin- Review X 10, 011060 (2020).
gle electron spin using real-time adaptive measurements. [32] Kobayashi, T. et al. Feedback-based active reset of a spin
Nature Nanotechnology 11, 247–252 (2015). qubit in silicon. npj Quantum Information 9, 52 (2023).
[11] Hirose, M. & Cappellaro, P. Coherent feedback control [33] Nakajima, T. et al. Real-time feedback control of charge
of a single qubit in diamond. Nature 532, 77–80 (2016). sensing for quantum dot qubits. Phys. Rev. Appl. 15,
[12] Cramer, J. et al. Repeated quantum error correction L031003 (2021).
on a continuously encoded qubit by real-time feedback. [34] Johnson, M. A. I. et al. Beating the thermal limit of qubit
Nature Communications 7, 11526 (2016). initialization with a Bayesian Maxwell’s demon. Phys.
[13] Wu, S.-H., Turner, E. & Wang, H. Continuous real-time Rev. X 12, 041008 (2022).
sensing with a nitrogen-vacancy center via coherent pop- [35] Shulman, M. D. et al. Suppressing qubit dephasing using
ulation trapping. Physical Review A 103, 042607 (2021). real-time Hamiltonian estimation. Nature Communica-
[14] Turner, E., Wu, S.-H., Li, X. & Wang, H. Spin-based tions 5, 5156 (2014).
continuous Bayesian magnetic-field estimations aided by [36] Dial, O. E. et al. Charge noise spectroscopy using coher-
feedback control. Physical Review A 106, 052603 (2022). ent exchange oscillations in a singlet-triplet qubit. Phys-
[15] Bushev, P. et al. Feedback cooling of a single trapped ical Review Letters 110, 146804 (2013).
ion. Physical Review Letters 96, 043003 (2006). [37] Cerfontaine, P. et al. Closed-loop control of a GaAs-based
[16] Singh, K. et al. Mid-circuit correction of correlated phase singlet-triplet spin qubit with 99.5% gate fidelity and low
errors using an array of spectator qubits. Science 380, leakage. Nature Communications 11, 4144 (2020).
1265–1269 (2023). [38] Kim, J. et al. Approaching ideal visibility in singlet-
[17] Steck, D. A., Jacobs, K., Mabuchi, H., Bhattacharya, T. triplet qubit operations using energy-selective tunneling-
& Habib, S. Quantum feedback control of atomic motion based Hamiltonian estimation. Physical Review Letters
in an optical cavity. Physical Review Letters 92, 223004 129, 040501 (2022).
(2004). [39] Yun, J. et al. Probing two-qubit capacitive interactions
[18] Sayrin, C. et al. Real-time quantum feedback prepares beyond bilinear regime using dual Hamiltonian parame-
and stabilizes photon number states. Nature 477, 73–77 ter estimations. npj Quantum Information 9, 30 (2023).
(2011). [40] Quantum Machines model OPX+, www.quantum-
[19] Wilson, D. J. et al. Measurement-based control of a me- machines.co.
chanical oscillator at its thermal decoherence rate. Nature [41] Malinowski, F. K. et al. Spectrum of the nuclear envi-
524, 325–329 (2015). ronment for GaAs spin qubits. Physical Review Letters
[20] Rossi, M., Mason, D., Chen, J., Tsaturyan, Y. & 118, 177702 (2017).
Schliesser, A. Measurement-based quantum control of [42] Foletti, S., Bluhm, H., Mahalu, D., Umansky, V. & Ya-
mechanical motion. Nature 563, 53–58 (2018). coby, A. Universal quantum control of two-electron spin
quantum bits using dynamic nuclear polarization. Nature
 11

Physics 5, 903–908 (2009). [60] Fiderer, L. J., Schuff, J. & Braun, D. Neural-network
[43] Vigneau, F. et al. Probing quantum devices with radio- heuristics for adaptive Bayesian quantum estimation.
frequency reflectometry. Applied Physics Reviews 10, PRX Quantum 2, 020303 (2021).
021305 (2023). [61] Koolstra, G. et al. Monitoring fast superconducting qubit
[44] Wu, X. et al. Two-axis control of a singlet–triplet qubit dynamics using a neural network. Phys. Rev. X 12,
with an integrated micromagnet. Proceedings of the Na- 031017 (2022).
tional Academy of Sciences 111, 11938–11942 (2014). [62] Botzem, T. et al. Tuning methods for semiconductor spin
[45] Jang, W. et al. Individual two-axis control of three qubits. Physical Review Applied 10, 054026 (2018).
singlet-triplet qubits in a micromagnet integrated quan- [63] Szańkowski, P., Trippenbach, M. & Cywiński, Ł. Spec-
tum dot array. Applied Physics Letters 117, 234001 troscopy of cross correlations of environmental noises
(2020). with two qubits. Physical Review A 94, 012109 (2016).
[46] Barthel, C. et al. Relaxation and readout visibility of [64] Evans, T. et al. Fast Bayesian tomography of a two-qubit
a singlet-triplet qubit in an Overhauser field gradient. gate set in silicon. Phys. Rev. Appl. 17, 024068 (2022).
Physical Review B 85, 035306 (2012). [65] QDevil model QDAC-II, www.quantum-machines.co.
[47] Yang, C. H. et al. Charge state hysteresis in semiconduc- [66] QDevil model QBoard-I, www.quantum-machines.co.
tor quantum dots. Applied Physics Letters 105, 183505
(2014).
[48] Harvey-Collard, P. et al. High-fidelity single-shot readout
ACKNOWLEDGMENTS
for a spin qubit via an enhanced latching mechanism.
Phys. Rev. X 8, 021046 (2018).
[49] Martins, F. et al. Noise suppression using symmetric ex- This work received funding from the European Union’s
change gates in spin qubits. Phys. Rev. Lett. 116, 116801 Horizon 2020 research and innovation programme under
(2016). grant agreements 101017733 (QuantERA II) and 951852
[50] Petit, L. et al. Design and integration of single-qubit (QLSI), from the Novo Nordisk Foundation under Chal-
rotations and two-qubit gates in silicon above one kelvin.
Communications Materials 3, 82 (2022).
lenge Programme NNF20OC0060019 (SolidQ), from the
[51] Koppens, F. H. L. et al. Universal phase shift and nonex- Inge Lehmann Programme of the Independent Research
ponential decay of driven single-spin oscillations. Physical Fund Denmark, from the Research Council of Norway
Review Letters 99, 106803 (2007). (RCN) under INTFELLES-Project No 333990, as well
[52] Ramon, G. & Cywiński, Ł. Qubit decoherence under two- as from the Dutch National Growth Fund (NGF) as part
axis coupling to low-frequency noises. Physical Review B of the Quantum Delta NL programme.
105, l041303 (2022).
[53] Cerfontaine, P., Otten, R., Wolfe, M. A., Bethke, P. &
Bluhm, H. High-fidelity gate set for exchange-coupled
AUTHOR CONTRIBUTIONS
singlet-triplet qubits. Physical Review B 101, 155311
(2020).
[54] Cai, X., Connors, E. J., Edge, L. F. & Nichol, J. M. Co- F.B. lead the measurements and data analysis, and
herent spin-valley oscillations in silicon. Nature Physics wrote the manuscript with input from all authors. F.B.,
19, 386–393 (2023). T.R., J.v.d.H., A.C. and F.K. performed the experi-
[55] Sergeevich, A., Chandran, A., Combes, J., Bartlett, S. D. ment with theoretical contributions from J.A.K., J.D.
& Wiseman, H. M. Characterization of a qubit Hamilto-
and E.v.N. F.F. fabricated the device. S.F., G.C.G. and
nian using adaptive measurements in a fixed basis. Phys-
ical Review A 84, 052315 (2011). M.J.M. supplied the heterostructures. A.C. and F.K. su-
[56] Bonato, C. & Berry, D. W. Adaptive tracking of a time- pervised the project.
varying field with a quantum sensor. Physical Review A
95, 052348 (2017).
[57] Scerri, E., Gauger, E. M. & Bonato, C. Extending qubit COMPETING INTERESTS
coherence by adaptive quantum environment learning.
New Journal of Physics 22, 035002 (2020).
The authors declare no competing interests.
[58] Mavadia, S., Frey, V., Sastrawan, J., Dona, S. & Biercuk,
M. J. Prediction and real-time compensation of qubit de-
coherence via machine learning. Nature Communications
8, 14106 (2017). ADDITIONAL INFORMATION
[59] Gupta, R. S. & Biercuk, M. J. Machine learning for
predictive estimation of qubit dynamics subject to de- Correspondence and requests for materials should be
phasing. Physical Review Applied 9, 064042 (2018). addressed to F.B., A.C., and F.K.
 Supplementary Information for “Real-time two-axis control of a spin qubit”

Fabrizio Berritta,∗ Torbjørn Rasmussen, Federico Fedele, and Anasua Chatterjee†

Center for Quantum Devices, Niels Bohr Institute,
University of Copenhagen, 2100 Copenhagen, Denmark

Jan A. Krzywda and Evert van Nieuwenburg

Lorentz Institute and Leiden Institute of Advanced Computer Science,
Leiden University, P.O. Box 9506, 2300 RA Leiden, The Netherlands

Joost van der Heijden

arXiv:2308.02012v2 [cond-mat.mes-hall] 26 Feb 2024

QDevil, Quantum Machines, 2750 Ballerup, Denmark

Saeed Fallahi
Department of Physics and Astronomy, Purdue University, West Lafayette, Indiana 47907, USA and
Birck Nanotechnology Center, Purdue University, West Lafayette, Indiana 47907, USA

Geoffrey C. Gardner
Birck Nanotechnology Center, Purdue University, West Lafayette, Indiana 47907, USA

Michael J. Manfra
Department of Physics and Astronomy, Purdue University, West Lafayette, Indiana 47907, USA
Birck Nanotechnology Center, Purdue University, West Lafayette, Indiana 47907, USA
Elmore Family School of Electrical and Computer Engineering,
Purdue University, West Lafayette, Indiana 47907, USA and
School of Materials Engineering, Purdue University, West Lafayette, Indiana 47907, USA

Jeroen Danon
Department of Physics, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway

Ferdinand Kuemmeth‡
Center for Quantum Devices, Niels Bohr Institute,
University of Copenhagen, 2100 Copenhagen, Denmark and
QDevil, Quantum Machines, 2750 Ballerup, Denmark
(Dated: January 28, 2024)

CONTENTS

Supplementary note 1: Experimental setup 2

Supplementary note 2: Relation between the quality factor and the Bayesian estimation procedure 5
Finite frequency resolution 5
Low quality estimates 6

Supplementary note 3: Extracting exchange energy and Overhauser field gradient from Larmor frequencies 8

Supplementary note 4: Controlled Hadamard rotations 10

Protocol 10
Examples of Hadamard rotations without feedback 11

References 11
 2

SUPPLEMENTARY NOTE 1: EXPERIMENTAL SETUP

Supplementary Supplementary Fig. 1 displays the experimental setup used in this study, which comprises a Triton
200 cryofree dilution refrigerator from Oxford Instruments capable of reaching a base temperature below 30 mK. A
superconducting vector magnet is thermally anchored to the 4 K plate and can generate 6 T along the main axis z of
the refrigerator and 1 T along x or y. It is used to apply an in-plane magnetic field of B ≈ 200 mT, along the double
quantum dot (DQD) axis (see device schematic in Supplementary Fig. 1). An upper bound of the electron temperature
is 100 mK, determined by attributing the observed broadening of the interdot (1,1)-(0,2) charge transition to thermal
broadening.
The Quantum Machines OPX+ includes real-time classical processing at the core of quantum control with fast
analog feedback. It enables on-the-fly pulse manipulation, which is critical for this experiment [1]. The RF carrier
frequency, approximately 158 MHz, is attenuated at room temperature by a programmable step attenuator. The
RF carrier power incident onto the PCB sample holder corresponds to between −70 and −80 dBm. Before entering
the cryostat, the signal is filtered with low-pass and high-pass filters, and a DC block reduces heating in the coaxial
lines caused by the DC component of the RF signal. The RF carrier is reflected by the surface-mounted tank circuit
wirebonded to one ohmic of the sensor dot and is amplified by a cryogenic amplifier at 4 K. The RF carrier is then
amplified again at room temperature, filtered to avoid aliasing, and digitized at 1 GS/s. As the signal is AC coupled
and does not have a 50 Ω resistance to ground, a bias tee is used to bias the input amplifier of the OPX+.
The QDAC-II [2] can trigger the OPX+, for example while tuning the device in video-mode using the RF reflectom-
etry measurements taken by the OPX+. All the ohmics of the device are grounded at the QDevil QBox, a breakout
box where low-pass filters are installed for all the DC lines for the gate voltages. To reduce noise and interference in
the signal chain, a QDevil QFilter-II [2] is installed in the cryostat for the DC lines. Additionally, the setup includes
a QDevil QBoard sample holder [2], which is a printed circuit board with surface-mounted tank circuits used for
multiplexed RF reflectometry. The bias tees have a measured cut-off frequency of ≈ 300 Hz, whereas the RC filters
have a cut-off frequency > 30 kHz.
 3

Legend
Computer QM orchestrator (server)
AC gate pulsing
RF reflectometry
OPX+ DC biasing
(pulsing and RF reflectometry)
Ext. trigger
QDevil QDAC-II
Analog output (DC bias)
PL PR RF-carrier Digitizer Sync x24
( (
Picosecond Picosecond
PSPL5915 Bias tee
PSPL5915
ZFBT-6GW+ 50 Ω

ANNE-50+
RCDAT-6000-60
-36 dB
SLP-300+

SLP-450+ Miteq AU-1447

+55 dB
Pomona BLP-1.9+

SHP-50+ SHP-50+

DCB-3538- SLP-450+
QDevil QBox
IO-SMA-02
(breakout box)
RT

55 K

LNF-LNC0.2_3A
+ 25 dB

BLK-18-S+
4K 10 dB 10 dB 10 dB

6 dB 6 dB 6 dB 0 dB
Still plate

6 dB 6 dB 3 dB 0 dB
Cold plate

20 mK 0 dB 0 dB 3 dB 0 dB

3 dB 3 dB ZFDC-20-50-S+ QDevil
QFilter II

PCB QDevil QBoard

cryogenic attenuator
1.2 kΩ

Bext
1.2 kΩ
22 nF

1.0 nF

50 kΩ 1.0 nF variable attenuator

Bias tee RC
L R SD filter
SMD DC-block
5 kΩ
750 nH

Cp ~ 1 pF
low-pass filter
22 pF 100 pF

high-pass filter

Supplementary Figure 1. See next page for caption.

 4

Supplementary Figure 1. Experimental setup (previous page). The cryostat is a Triton 200 dilution refrigerator by
Oxford Instruments with a base temperature lower than 30 mK. A Quantum Machines OPX+ is used for RF reflectometry
and fast gate control pulses with a bandwidth of less than 1 GHz. The RF carrier frequency used is approximately 158 MHz.
Additionally, the setup includes a QDevil QDAC-II for generating low-frequency analog signals. A QDevil QFilter-II is used
to suppress noise and interference in the signal chain. Finally, the setup incorporates a QDevil QBoard PCB sample holder
with surface-mount tank circuits for multiplexed RF reflectometry. Only one of the four qubits of the chip is activated for this
experiment.
 5

SUPPLEMENTARY NOTE 2: RELATION BETWEEN THE QUALITY FACTOR AND THE BAYESIAN
ESTIMATION PROCEDURE

In this section we show that most of the oscillation decay, i.e., the finite quality factor of coherent rotations in
Fig. 2 in the main text, can be explained by the estimation errors. For each estimation procedure we represent the
probability distribution of Ω on a regular grid, and update the weights based on the measurement outcome. To
convert the final probability distribution to an estimation of the frequency we use the average:
X
⟨Ω⟩ = p(Ω[n])Ω[n] (1)
n

where Ω[n] is the frequency and p(Ω[n]) is the corresponding probability. When the estimated frequency is used to
adjust the time for coherent rotations t = ϕ/(2π⟨Ω⟩) [in the main text the qubit rotation angle θ ≡ ϕ/(2π)], any
estimation error δΩ will result in a random phase evolution δϕ = 2πδΩ t that contributes to a decrease in a quality
factor.

Finite frequency resolution

The first source of errors can be associated with the finite resolution of the probability distribution. While too high
a resolution would significantly slow down an on-the-fly estimation scheme, too low a resolution would introduce an
estimation error, the scale of which can be related to a step size δΩ ≈ (Ω[n + 1] − Ω[n])/2.
To optimize the frequency resolution used in the Bayesian estimation, we perform different measurements for both
ΩL and ΩH . Supplementary Supplementary Fig. 2a shows the averaged traces after performing controlled Overhauser-
driven rotations whenever ⟨ΩL ⟩ > 30 MHz, as described in Fig. 2 of the main text. Increasing the frequency resolution
from 10 MHz to 0.3 MHz (with frequency span from 10 MHz to 70 MHz) results in an increased number of visible
oscillations. However, considering the required computational time by the FPGA and no appreciable improvements
above 0.5 MHz, we choose a resolution of 0.5 MHz throughout this work, resulting in ≈ 5 µs per qubit cycle to update
the estimate on the FPGA.
We also test the frequency resolution of ΩH in a similar way by performing exchange-driven controlled rotations, as

a b

10 MHz 10 MHz

6 MHz 6 MHz

3 MHz 3 MHz

1 MHz 1 MHz

0.3 MHz
0.3 MHz

Supplementary Figure 2. Frequency resolution of Bayesian estimation. a Each trace shows averaged Overhauser-driven
controlled rotations as Fig. 2d in the main text. For each trace we used a different frequency resolution for the probability
distribution when estimating ΩL , as indicated. The y-axis of each curve is offset for clarity. b Each trace shows exchange-
driven controlled rotations as shown in Fig. 4c in the main text. For each trace we used a different frequency resolution for the
probability distribution when estimating ΩH , keeping fixed the frequency resolution of ΩL at 1 MHz. The y-axis of each curve
is offset for clarity.
 6

shown in Fig. 4 of the main text, and estimating ΩH between 40 MHz and 90 MHz. In this case, we also set the
resolution to 0.5 MHz to keep the estimation cycle on the few µs scale.
We support the above by a simple model of uncertainty, in which for each realization of the experiment we draw
2
a random error δΩ ≈ N (0, σδΩ ). For simplicity we assume the initialization and measurement axis are perpendicular
to the rotation axis, such that the probability of measuring the initial state reads:

1 1 1 cos(2π Ωτ )
PS (τ ) = + cos(2π[Ω + δΩ]τ ) δΩ
= + W (τ ), (2)
2 2 2 2
where W (τ ) is the attenuating function related to classical averaging over repetitions of δΩ, which we denoted
as ⟨. . . ⟩δΩ . For simplicity

we assume the errors are normally distributed with characteristic width σδΩ , for which
W (τ ) = exp −2π 2 σδΩ 2
τ 2 . For the relevant case of stabilized rotations, which are obtained by adjusting evolution
time τ = 2πΩ/ϕ, we have
(  2 ) (   )
2
1 1 1 σδΩ 2 1 1 ϕ
PS (ϕ) = + cos(ϕ) exp − ϕ ≡ + cos(ϕ) exp − ∗ , (3)
2 2 2 Ω 2 2 ϕ2

where in analogy to dephasing time T2∗ we defined dephasing angle ϕ∗2 , which measures the angle of rotation at which
the amplitude falls as 1/e. For a constant estimation error it depends on the estimated frequency, since:
√ Ω
ϕ∗2 = 2 , (4)
σδΩ
and can be related to a Q-factor via the equation:
ϕ∗2 Ω
Q= = √ . (5)
2π 2πσδΩ
Using the above, we estimate that an resolution-related error of σδΩ ≈ 0.5 MHz at Ω = 20 MHz (the smallest estimated
frequency) should allow for at least Q ≥ 9. As this number is larger than any measured Q, we conclude the finite
resolution of the estimation is not the main factor responsible for observed amplitude decay.

Low quality estimates

Having confirmed sufficiently high resolution (0.5 MHz), another source of errors is associated with low-quality
estimates resulting from multi-modal or not sufficiently narrow probability distributions. As a measure of estimation
quality we take the variance of the final distribution, defined as:
X
2
2
σest = Var(Ω) ≡ Ω[n] − ⟨Ω⟩ p(Ω[n]) (6)
n

To show that low-quality estimates play an important role in the loss of oscillation amplitude we post-process the
measured data based on σest . We reject the repetitions with a probability distribution with calculated variance of
σest > σest,max , and we average over the remaining traces. We highlight that in principle, such or an even more
sophisticated procedure of quality assessment could also be performed on-the-fly, for instance as a part of the same
feedback loop and without the need of measuring the low-quality oscillations.
We test this approach on the experimental data used to generate Fig. 2d in the main text. In Supplementary
Fig. 3a, we plot the obtained probability distributions with ⟨ΩL ⟩ ≥ 50 MHz. For each distribution we compute its
variance, whose histogram over all the distribution is plotted in Supplementary Fig. 3b. By imposing an upper bound
for the variance σest,max , we can now compute an average such as shown in the bottom panel of Fig. 2d in the main
text, but now using only a selection of the best estimates.
In Supplementary Fig. 4b we plot the ratio of used data Ñ versus total repetition N = 1450, as a function of
σest,max . For each value of σest,max we plot in Supplementary Fig. 4a the change in the averaged singlet probability
PS (θj ) of stabilized oscillations, as compared to the trace shown in the main text. In particular, we focus on two
different bounds of tolerated variance σest,max = 0.8 MHz and 2.7 MHz, that correspond to rejecting 80% and 50% of
the worst estimations, respectively. We mark those values of σest,max using dashed lines in both panels. Finally, in
Fig 4c we compare the coherent oscillations obtained using the 20% (green) and 50% (yellow) best estimates against
the unfiltered result from Fig. 2d (black). We see a significant improvement in the oscillation amplitude, which
suggests that the observed decay may be associated with poor performance of the estimation scheme.
 7

a b

Supplementary Figure 3. Standard deviation of Bayesian estimations. a Probability distributions used to estimate
frequencies ⟨ΩL ⟩ > 50 MHz. b Histogram of corresponding variances σest of the probability distributions p(Ω[n]).

a c

Supplementary Figure 4. Improvement of the visibility of coherent oscillations by rejecting low-quality estimates.
a Change in oscillation amplitude after averaging only the best Ñ estimates out of N = 1450 as a function of bound estimation
uncertainty σest,max . b The fraction of used estimations Ñ /N as a function of bound estimation uncertainty σest,max . By dashed
lines we mark σest,max = 0.7 MHz (green), and 2.7 MHz (yellow), that correspond to rejecting 80% and 50% of repetitions
respectively. c The resulting averaged oscillations in comparison to unfiltered data (black) from Fig. 2d in the main text.
 8

SUPPLEMENTARY NOTE 3: EXTRACTING EXCHANGE ENERGY AND OVERHAUSER FIELD

GRADIENT FROM LARMOR FREQUENCIES

In this section, we seek to extract time-dependent knowledge about the exchange energy, J(t), and the Overhauser
field gradient,
p ∆Bz (t), from the ΩL (t) data p and ΩH (t) data shown in Fig. 3(c,d). To achieve this, we assume
ΩL (t) ≡ ∆Bz2 (t) + Jres 2 (t) and Ω (t) ≡
H ∆Bz2 (t) + J 2 (εH , t), and combine this with statistical methods as the
problem is not analytically solvable: at each time t, we have two known quantities [ΩL (t) and ΩH (t)] and three
unknown ones [∆Bz2 (t), Jres2
(t), and J 2 (εH , t)].
As ΩH (t) is probed only when 20 MHz < ⟨ΩL (t)⟩ < 40 MHz, we downsample ⟨ΩL (t)⟩ to the same number of points
as we have for ⟨ΩH (t)⟩, by choosing for each value of ⟨ΩH (t)⟩ the one of ⟨ΩL (t)⟩ that is closest in time. We then remove
any outliers in the data by considering a window of size w = 15 around each data point and reject any data point
that is more than the standard deviation of its 14 neighboring data points away from the average of those points. In
such a case, we replace the rejected data point with that average. We then finally apply a running average with a
window size of w = 5 to smoothen the data and remove high-frequency noise. In Supplementary Fig. 5a we show the
resulting ⟨ΩL,H (t)⟩, cf. Fig. 3(c,d) in the main text.
We assume the fluctuations of |∆Bz | to dominate on this timepscale, and thus we square ΩL and ΩH and determine
the shift κ2opt that results in the best overlap of ⟨ΩL (t)⟩ and ⟨ΩH (t)⟩2 − κ2 , using a least squares method. The
result is plotted in Supplementary Fig. 5b, with κopt ≈ 36 MHz, suggesting that on average J(εH )2 ≈ Jres 2
+ κ2opt .
The clear correlation between the two traces confirms the assumption that the fluctuations of |∆Bz | dominate over
the fluctuations of J. The slightly worse overlap seen in some regions, for instance at about 1 s, are possibly due to
fluctuations of Jres and J(εH ).
Nevertheless, in order to extract direct knowledge about ∆Bz (t) we still need to determine Jres (t) [or J(εH , t)]
independently. To obtain an average value of Jres (t), we combine together longer measurements of ⟨ΩL (t)⟩ and
⟨ΩH (t)⟩ taken at different times at the same tuning and cooldown over ≈ 36 s and ≈ 6 s, respectively. We then
consider histograms of the measured frequencies without filtering [we show in Supplementary Fig. 5c the one obtained

a c

b d

Supplementary Figure 5. Extracting fluctuations of the exchange energy and Overhauser field gradient from two
different estimated Larmor frequencies. a Fluctuations of the Larmor frequencies Ωq
L and ΩH after removal of estimation

outliers (see text). b The estimated Larmor frequency ⟨ΩL (t)⟩ and the shifted frequency ⟨ΩH (t)⟩2 − κ2opt , based on the same
data as shown in Fig. 3 of the main text, where κopt provides the best overlap using a least-squares error. c Histogram of a
≈ 36 s long measurement of ⟨ΩL ⟩. From a fit to equation (7) (red solid line) the average Jres ≈ 20.2 MHz is found. d |∆Bz (t)|
and J(εH , t) extracted from the Larmor frequencies ΩL (t) and ΩH (t) of main Fig. 3, assuming constant Jres = 20.2 MHz.
 9

for ⟨ΩL ⟩], which we fit to the distribution function

Z Ω    
1 Ω (J − µJ )2 Ω2 − J 2
P(Ω) = dJ √ exp − exp − , (7)
−Ω πσJ σB Ω2 − J 2 2σJ2 2σB2

√
that gives the probability density for Ω = J 2 + B 2 when J and B are normally distributed variables with means µJ
and zero and variances σJ2 and σB 2
, respectively. The red solid line in Supplementary Fig. 5c shows the least-square
fit of the data, yielding σJ ≈ 4.63 MHz, σB ≈ 22.0 MHz, µJ ≈ 20.2 MHz for the ≈ 36 s long trace of ⟨ΩL (t)⟩.
The results of this fit confirm that σB > σJ , i.e., that the fluctuations in the Overhauser gradient dominate those
in Jres . Indeed, first-order detuning fluctuations of the residual exchange interaction are expected to be zero, as the
qubit is tuned close to the symmetry point in the (1, 1) charge state when measuring ΩL [3, 4]. To get a rough picture
of the time dependence of |∆Bz (t)| we thus replace thepresidual exchange term Jres (t) p by its constant average value
µJ as extracted from the fit and we extract |∆Bz (t)| = Ω2L (t) − µ2J and J(εH , t) = Ω2H (t) − ∆Bz2 (t). The result is
plotted in Supplementary Fig. 5d.
 10

SUPPLEMENTARY NOTE 4: CONTROLLED HADAMARD ROTATIONS

Protocol

The aim of this protocol is to perform an adaptive Hadamard gate in a singlet-triplet qubit in GaAs by Bayesian
estimation of two Larmor frequencies at different detunings. We define the following quantities:
p
• Ω(ε) ≡ ∆Bz2 + J(ε)2 is the Larmor frequency at a given detuning ε
• Jres is the residual exchange deep in (1,1) (ε ≈ −40 mV)
p
• ΩL ≡ ∆Bz2 + Jres 2 is the Larmor frequency deep in the (1,1) charge state.

• At the detuning where J ≈ |∆Bz | (with |∆Bz | ≈ [40, 60] MHz), we approximate J(ε) ≈ J(ε0 ) + α ∆ε, where α
is ≈ 10 MHz/mV and ε0 ≈ −16 mV
p
• ΩH ≡ ∆Bz2 + J(ε0 )2 is the Larmor frequency at the detuning point defined above
• εHad is the detuning at which J = |∆Bz |, and in general εHad ̸= ε0 because J fluctuates.
This protocol assumes we have prior knowledge of the residual exchange (assumed constant), an offline model
of J(ε), and |∆Bz | does not depend on detuning ε and it fluctuates sufficiently slowly to be considered constant
throughout the protocol. Typical values for Jres ≈ [10, 20] MHz.
1. Estimate ΩL .
2
2. Calculate ∆Bz2 = Ω2L − Jres . If 40 MHz < |∆Bz | < 60 MHz, go to the next point. Otherwise repeat 1.
3. Adjust detuning such that J = |∆Bz |, based on the offline knowledge of J(ε) = J(ε0 ) + α ∆ε.
4. To learn the prevailing J, perform exchange-based FID around the field found at 3, i.e. where J = ∆Bz ,
interleaved with ΩL (π/2) pulses for initialization and readout, from which we estimate ΩH as in Fig. 3 of the
main text.
5. Estimate J 2 = Ω2H − ∆Bz2 and adjust detuning again such that J = |∆Bz | to account for fluctuations of J from
the offline model.
6. Perform Hadamard with rotation angle calculated based on |∆Bz | by directly jumping to εHad .
7. Back to 1.
Details
1. Estimate ΩL by 101 single shots linearly spaced between 0 ns and 100 ns.
2. Calculate ∆Bz2 = Ω2L − Jres
2
= (ΩL + Jres )(ΩL − Jres ).
3. Adjust detuning such that J = |∆Bz |. In practice this condition is equivalent to Ω2H = 2(Ω2L − Jres
2
). Since
Ω2H (ε) ≈ (J(ε0 ) + α∆ϵ)2 + Ω2L − Jres
2
, (8)
∗
by solving we get the required detuning shift ∆ε = (∆Bz − J(ε0 ))/α.
4. Perform FID with interleaved ΩL (π/2) pulses, from which we measure ΩH,meas .
5. Adjust detuning again such that J = |∆Bz |. We have
Ω2H,meas (ε) ≈ (J(ε0 ) + α∆ϵmeas )2 + Ω2L − Jres
2
, (9)
where we have assumed Ω2L , Jres
2
and J(ε0 ) have not changed. We look for the fluctuation in detuning by looking
at Ω2H,meas − Ω2H and find
Ω2H,meas − Ω2H
∆εmeas ≈ + ∆ε∗ , (10)
2J(ε0 )α
having neglected second order terms in ∆ε. So we shift the detuning of the qubit by the amount ∆ε∗ − ∆εmeas .
√
6. Perform Hadamard with rotation angle calculated based on ΩHad = 2|∆Bz | by directly jumping to εHad =
ε0 + ∆ε∗ − ∆εmeas .
7. Back to 1.
 11

a c

b d

Supplementary Figure 6. Comparing Hadamard rotations with and without real-time stabilization. a Measurement
of Hadamard rotations with feedback, same data as shown in Fig. 5e of the main text. b-c-d Additional measurements of
Hadamard rotations without feedback, not shown in the main text.

Examples of Hadamard rotations without feedback

In Supplementary Fig. 6 we compare the quality of Hadamard rotations with feedback (reproduced from Fig. 5e
of the main text) with naive Hadamard rotations without feedback, as explained in the main text. In the absence of
feedback, the resulting oscillation curves fluctuate randomly in time, even though the cycles of qubit control pulses
were nominally identical for panel b, c and d. This is likely due to the Overhauser gradient drifting over time.
(All panels were taken within minutes of each other.) For example, data in panel b and d suggests a slight under-
and over-rotation relative to the target rotation angles, while in c we have picked a data set that happens to show
approximately the correct rotation angles. For all nominally identical uncontrolled Hadamard experiments that we
acquired, we observe a quality factor that is much lower compared to the stabilized oscillations in panel a.

∗
fabrizio.berritta@nbi.ku.dk
†
anasua.chatterjee@nbi.ku.dk
‡
kuemmeth@nbi.dk
[1] Quantum Machines, www.quantum-machines.co.
[2] QDevil, www.qdevil.com.
[3] Martins, F. et al. Noise suppression using symmetric exchange gates in spin qubits. Phys. Rev. Lett. 116, 116801 (2016).
[4] Reed, M. D. et al. Reduced sensitivity to charge noise in semiconductor spin qubits via symmetric operation. Phys. Rev.
Lett. 116, 110402 (2016).