Programmable Simulations of Molecules and Materials With Reconfigurable Quantum Processors

Programmable Simulations of Molecules and Materials with Reconfigurable Quantum Processors
Nishad Maskara,1, 2 Stefan Ostermann,1 James Shee,3, 4 Marcin Kalinowski,1 Abigail McClain Gomez,1 Rodrigo Araiza
Bravo,1 Derek S. Wang,5 Anna I. Krylov,6 Norman Y. Yao,1 Martin Head-Gordon,2, 3 Mikhail D. Lukin,1 and Susanne F. Yelin1
1
Department of Physics, Harvard University, Cambridge, MA 02138, USA
2
Lawrence Berkeley National Lab, Berkeley, CA 94720, USA
3
College of Chemistry, University of California Berkeley, Berkeley, CA 94720, USA
4
Department of Chemistry, Rice University, Houston, TX 77005, USA
5
Harvard John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA 02138, USA
6
Department of Chemistry, University of Southern California, Los Angeles, CA 90089, USA
(Dated: December 6, 2023)
Simulations of quantum chemistry and quantum materials are believed to be among the most important poten-
tial applications of quantum information processors, but realizing practical quantum advantage for such prob-
arXiv:2312.02265v1 [quant-ph] 4 Dec 2023
lems is challenging. Here, we introduce a simulation framework for strongly correlated quantum systems that
can be represented by model spin Hamiltonians. Our approach leverages reconfigurable qubit architectures
to programmably simulate real-time dynamics and introduces an algorithm for extracting chemically relevant
spectral properties via classical co-processing of quantum measurement results. We develop a digital-analog
simulation toolbox for efficient Hamiltonian time evolution utilizing digital Floquet engineering and hardware-
optimized multi-qubit operations to accurately realize complex spin-spin interactions, and as an example present
an implementation proposal based on Rydberg atom arrays. Then, we show how detailed spectral information
can be extracted from these dynamics through snapshot measurements and single-ancilla control, enabling the
evaluation of excitation energies and finite-temperature susceptibilities from a single-dataset. To illustrate the
approach, we show how this method can be used to compute key properties of a polynuclear transition-metal
catalyst and 2D magnetic materials.
A major thrust of quantum chemistry and material sci- energy states often exhibit a large degree of entanglement. In
ence involves the quantitative prediction of electronic struc- this Article, we focus on the programmable quantum simula-
ture properties of molecules and materials. While powerful tions of spin models. These correspond to a class of Hamilto-
computational techniques have been developed over the past nians that describe compounds where unpaired electrons be-
decades, especially for weakly correlated systems [1–4], the come localized at low-temperatures and can therefore be rep-
development of tools for understanding and predicting the resented as effective local spins with S ≥ 1/2. These include
properties of materials that feature strongly correlated elec- many polynuclear transition metal compounds and materials
trons remains a challenge [5–7]. Quantum computing is a containing d- and f-block elements, which play a central role
promising route to efficiently capturing such quantum cor- in chemical catalysis and magnetism [20, 28–34].
relations [8–10], and algorithms for Hamiltonian simulation Recent advances in quantum simulation [35, 36] have en-
and energy estimation [11] with good asymptotic scaling have abled the study of paradigmatic model Hamiltonians with
been developed. However, existing methods for simulating local connectivity. In particular, experiments have probed
large-scale electronic structure problems are prohibitively ex- non-equilibrium quantum dynamics [37–39], exotic forms of
pensive to run on near-term quantum hardware [12], highlight- emergent magnetism [40–43], and long-range entangled topo-
ing the need for more efficient approaches. logical matter [44, 45], in regimes that push the limits of
One approach to capturing the complexity of strongly cor- state-of-the-art classical simulations [46]. The model Hamil-
related systems utilizes model Hamiltonians [13], such as the tonians describing realistic molecules and materials however
generalized Ising, Heisenberg, and Hubbard models, which often contain more complex features, including anisotropy,
describe the interactions between the active degrees of free- non-locality, and higher-order interactions [29], demanding
dom at low temperatures. Like other coarse-graining or ef- a higher degree of programmability [18]. While universal
fective Hamiltonian approaches [14], model parameters can quantum computers can in principle simulate such systems,
be computed from an ab initio electronic structure problem standard implementations based on local two-qubit gates re-
using a number of classical techniques [15–20]. Further- quire large circuit depths [23] to realize complex interactions
more, model Hamiltonians exhibit features such as low-degree and long-range connectivity. Thus, for optimal performance
connectivity that simplify implementation, making them par- in devices with limited coherence, it is essential to utilize
ticularly promising candidates for quantum simulation [21– hardware-efficient capabilities to simulate such systems.
26]. Though approximate, simplified model Hamiltonians In what follows we introduce a framework to simulate
have proved valuable in analyzing strongly correlated prob- model spin Hamiltonians (Fig. 1) on reconfigurable quantum
lems [27–29] for small system sizes, where accurate but costly devices. The approach combines two elements. First, we de-
classical methods can be applied. However, as the system size scribe a hybrid digital-analog simulation toolbox for realizing
increases, classical numerical methods struggle to reliably complex spin interactions, which combines the programma-
solve strongly correlated model systems, as the relevant low- bility of digital simulation with the efficiency of hardware-
2
(a) (b) 3 Active Electrons (c)

= Spin 3/2
O Mn
Ca O
3d
O Mn O
4s Paired Electrons
Mn O Mn
Strongly Correlated Programmable Quantum Simulation

Matter
Spectral Weights
Model Spin
Hamiltonian
Evolution Time Frequency
FIG. 1. Model Hamiltonian approach to quantum simulation of strongly-correlated matter. (a) The procedure starts with a description
of the target molecule or material structure, whose electronic structure problem is reduced using classical computational chemistry techniques
to a simpler effective Hamiltonian that captures the relevant low-energy behaviors. (b) Here we study problems which are modeled by
spin Hamiltonians with potentially non-local connectivity and generic on-site spin S ≥ 1/2, where each spin is comprised of localized,
unpaired electrons in the original molecule. The key simplification comes from capturing charge fluctuations perturbatively, which is a good
approximation in certain contexts. (c) Programmable quantum simulation is then used to calculate properties of the model Hamiltonian. We
develop a simulation framework, based on encoding spins into clusters of qubits, that can be readily implemented on existing hardware. The
toolbox enables efficient generation of complex spin interactions by leveraging dynamical reconfigurability and hardware-optimized multi-
qubit gates (Figs. 2 and 3). The quantum simulator performs time-evolution under the spin Hamiltonian for various simulation times t, and
each qubit is projectively measured to produce an set of snapshots. Subsequent classical processing extracts properties like the low-lying
excitation spectrum and magnetic susceptibilities, all from the same dataset (see Figs. 4-6).
optimized multi-qubit analog operations. Then, we intro- pounds, including super-exchange, spin-orbit coupling, ring-
duce an algorithm, dubbed “many-body spectroscopy” that exchange, and more [29]. In our approach, spin-S variables
leverages time-dynamics and snapshot measurements to ex- are encoded into the collective spin of 2S qubits, such that the
tract detailed spectral information of the model Hamiltonian i-th spin in (1) is rewritten as
in a resource-efficient way [47]. We describe in detail how
these methods can be implemented using Rydberg atom ar-
2Si
rays [48–52], and discuss its applicability to other emerging X
platforms which can support multi-qubit control and dynamic, Ŝiα = ŝα
i,a , (2)
a=1
programmable connectivity [53], such as the ion QCCD archi-
tecture [54, 55]. Finally, we illustrate potential applications of where ŝαi,a are the spin-1/2 operators of the a-th qubit in the
the framework on model Hamiltonians describing a prototyp- i-th spin. Valid spin-S states live in the symmetric subspace
ical biochemical catalyst and 2D materials. with maximum total spin per site ⟨(Ŝi )2 ⟩ = Si (Si +1). While
several alternate approaches to encoding spins with hardware-
native qudits have been proposed recently [58–61], the clus-
Engineering spin Hamiltonians
ter approach introduced here uses the same controls as qubit-
based computations ensuring compatibility with existing se-
The general Hamiltonian we consider is tups [57], and naturally supports simulation of models with
mixed on-site spin.
X X αβ α β The core of our protocol involves applying a K step
H= Biα Ŝiα + Jij Ŝi Ŝj sequential P evolution under simpler interaction Hamiltoni-
i,α ij,αβ
ans Hi = g∈Gi hi,g , i = 1, ..., K, acting on disconnected
αβγ α β γ
X
+ Kijk Ŝi Ŝj Ŝk + higher order, (1) groups Gi of a few qubits each. The combined sequence
ijk,αβγ realizes an effective Floquet Hamiltonian HF which ap-
proximates (1). To controllably generate effective Hamil-
where Ŝiα , α = x, y, z are spin-S operators (S ≥ 1/2) acting tonians, we use the average Hamiltonian approach. In the
αβ αβ limit of small step-sizes τ ≪ 1/K, the evolution is well-
on the i-th spin, and the interaction coefficients (Jij , Kijk ,
(0) 1
PK
etc.) are potentially long-range. Hamiltonians of this form can approximated by HF = K i=1 Hi to leading order, and
capture the effects of many processes arising in physical com- contributions from higher-order terms are bounded [62].
3
(a) (b)
Tgate (2π/Ω)
US
102
UP
101
Reconfigure 100
2 4 6 8
Qubits per Cluster
(c)
Two Qubit Gates

104
US
UP
102
2 4 6 8
Qubits per Cluster
FIG. 2. Hardware-efficient implementation with neutral-atom tweezer arrays. (a) Our protocol for programmable simulation of generic
spin Hamiltonian is based on applying sequences of interactions between non-overlapping few-qubit groups. Here, we illustrate an implemen-
tation of a complex spin model using the dynamical projection approach. Spin-Si variables are encoded in the collective spin of a cluster of
2Si qubits. Then, interactions between spins are generated by evolving pairs of qubits from each cluster under an interaction Hamiltonian HI .
Second-order interactions Jij act on two qubits, while higher-order interactions act on multiple qubits, such as the fourth-order coefficient
Liijk of Ŝiα Ŝiβ Ŝj Ŝk . Then, interactions are dynamically projected into the symmetric encoding space within each cluster by evolving under
Hamiltonian HP . The target large spin Hamiltonian H is simulated by alternately evolving under HI and HP . This protocol can be realized
in any reconfigurable quantum processor. Here, we present an implementation for Rydberg atom arrays, which has two long-lived qubit states
|0⟩, |1⟩, and an excited Rydberg state |r⟩ with strong interactions. The interaction connectivity is dynamically changed by moving optical
tweezers [56], and interactions are generated using all-to-all Rydberg blockade interactions within each cluster and simultaneous global driv-
ing of the qubit Ωq (t) and Rydberg Ωr (t) transitions. (b) Fast and efficient multi-qubit spin operations US , UP are identified using optimal
control to optimize pulse sequences. Gates times are measured in units of the Rydberg driving frequency ΩT , where the two-qubit CZ gate
from Ref. [57] takes ΩT /2π ≈ 1.2. The alternating ansatz (solid lines) decomposes the target operations into symmetric diagonal gates and
single qubit rotations, which can be are individually optimized (see Methods). Simultaneous (dual) driving of both transitions (dotted lines)
enables even faster realization of approximate US and UP gates with ideal error rates below ∼ 10−3 . (c) We compare against decomposition
of US and UP into a two-qubit gate set composed of CPhase gates and single-qubit rotations. Such a decomposition rapidly becomes very
costly as the cluster size grows, in contrast to the optimized hardware-native operation.
In general, the performance of this approach will be limited While this generates the target interactions between spin
by simulation errors, characterized by the difference between clusters, it also moves encoded states out of the symmetric
HF and the target Hamiltonian H, and gate errors, determined subspace. Therefore, to prevent evolution under HI from de-
by the hardware overhead required to implement individual stroying the encoding, we alternately apply evolution under
evolutions e−ihi,g τ . To mitigate the leading sources of error,
we next develop Hamiltonian engineering protocols that lever- X
age multi-qubit spin operations to realize (1) with short peri- HP = −λ P [(Ŝi )], (3)
odic sequences. i
composed of projectors P [(Ŝi )] onto the symmetric spin

states, by applying multi-qubit gates within spin clusters. This
Dynamical projection with digital Floquet engineering Hamiltonian creates an energy gap λ between symmetric and
non-symmetric states. For λ ≫ Jij , the static Hamiltonian
Our Hamiltonian engineering approach is based on the clus- HI + HP is effectively projected into the ground-space of HP
ter encoding (2). The key idea is to first generate interactions at low-energies, realizing H. However, accurate Trotter simu-
in the full Hilbert space using a spin-1/2 (qubit) version HI lation in this regime, by alternating HI and HP , requires two
of (1), and then dynamically project back onto the encoding separations of time-scale between the step-size and the target
−1
space. We select HI such that projection recovers the target interactions τ ≪ λ−1 ≪ Jij , leading to a large number of
Hamiltonian, by mapping each n-body large spin interaction gates.
in (1) onto an equivalent n-body qubit interaction acting on Instead, we develop an approach which enables projection
representatives from the n-clusters (see Methods and Fig. 2a). of HI onto the ground-state of HP with significantly fewer
4
gates, using ideas inspired by dynamical decoupling [63]. Similar tools can also be used realize a large class of spin
This is achieved by using large-angle rotations e−iτ HP , where (i)
circuits, by generating different evolutions HF during each
τ λ ∼ 1, to generate a time-dependent phase on the parts cycle. This effectively implements discrete time-dependent
of HI that couple encoded states to non-symmetric states. evolution
These phases cancel out on average, leaving the symmetric
part which commutes with HP . More precisely, the combined
T
evolution is described by Y (i)
Ucirc = e−iτi HF . (6)
i=1
Np D
Y Y Variational optimization can be further used to engineer
UF = e−iθp HP e−iτ HI,j
higher-order terms in HF , enabling generation of more com-
p=1 j=1
plex spin gates at no additional cost (see Extended Data Fig.
Np D
Y Y 9). Although the classical variational optimization procedure
= e−iτ HI,j (k) = e−iKτ HF , (4) is limited to small circuits, Hamiltonian learning protocols
p=1 j=1
could be used to perform larger-scale optimization on a quan-
where HI has been decomposed into D non-overlapping tum device directly [65, 66]. Such circuits can be used for
groups HI,j , K = Np D is the full length of the sequence, operations besides Hamiltonian simulation, including state-
and θp , τ parameterize the evolution times. In the second preparation [21].
line, we transform into an interaction picture, such that in-
termediateP terms are evolved by HP with a cumulative phase
Hardware-efficient implementation
Θk = λ−1 k′ <k θk ; for the rotating frame to be periodic, we
require ΘK (mod 2π) = 0.
Then the transformed interactions can be written as (see The digital Hamiltonian engineering sequence can, in prin-
Methods): ciple, be realized on any universal quantum processor, but it is
especially well suited for reconfigurable processors with na-
tive multi-qubit interactions [53, 55, 56, 69]. Neutral atom ar-
HI,j (k) = e−iΘk HP HI,j eiΘk HP rays are a particularly promising candidate for realizing these
nX
! techniques, for which we develop a detailed implementation
max
(0) (n) proposal. In this platform, two long-lived atomic states en-
= HI,j + e−inΘk HI,j + h.c. . (5)
code the qubit degree of freedom {|0⟩,|1⟩}, which can be in-
|{z} n=1
symmetric | {z } dividually manipulated with fidelity above 99.99% [70, 71].
symmetry violating
Strong interactions between qubits are realized by coupling
(n)
to a Rydberg state |r⟩ [72, 73], which enables parallel multi-
Here, HI captures unwanted terms that change the total spin qubit operations [73], with state-of-the-art two-qubit gate fi-
on n sites, and the symmetric part is the target Hamiltonian delities exceeding 99.5% [57]. Further, qubits can be trans-
(0)
HI = HT . To cancel out symmetry-violating terms at lead- ported with high fidelity by moving optical tweezers [56], to
ing order in τ , we choose
P an appropriate sequence of cumu- realize arbitrary groupings Gi . By placing atoms sufficiently
lative phases Θk = λ k′ <k θk , such that k e−iΘk = 0.
P
close together, atoms within a group can undergo strong all-
When nmax = 2, i. e., up to two-spin interactions, the se- to-all interactions, while interactions between groups can be
quence Θ = (0, 2π/3, 4π/3) satisfies this condition. This se- made negligible by placing them far apart.
quence, being p-independent, produces a Floquet Hamiltonian The key ingredient required for efficiently implementing
which commutes with HP in a dressed frame [64], and there- HI and HP are hardware-efficient multi-qubit spin opera-
fore preserves the encoding up to exponentially long times in tions. We show how these can be realized by using pulse
τ −1 . engineering to transform the native Rydberg-blockade inter-
This approach provides significant benefits in simulating action [72, 74] into the desired form. We illustrate this on two
complex spin-models, where the number of overlapping terms families of representative spin operations
in H grows rapidly with parameters such as the spin-size
S, number of interactions per spin d, and interaction weight 2 2
nmax . In this regime, a Trotter decomposition of H into non- US (θ) = e−iθ(Ŝ /2S)
, UP (θ) = e−iθP [Ŝ ]
(7)
overlapping terms hi,g would require a sequence of length
K ≥ d(2S)nmax −1 , where d measures the number of inter- where Ŝ2 is the total-spin operator for a cluster of 2S atoms,
actions each spin Si is involved in. Instead, the projection and P [Ŝ2 ] are the projectors appearing in (3).
approach produces decompositions
d of H into sequences of One approach to engineering these operations is based on
length K = (nmax + 1) 2S (see Methods), leading to per- an ansatz which naturally extends Ref. [57], where US and
formance improvements of several orders of magnitude for UP are found by optimizing an alternating sequence of di-
models with large spins and higher-order interactions. agonal phase gates and single-qubit rotations. As in prior
5
(a) Two-qubit Multi-qubit (i) tion Ωq (t) in addition to the usual Rydberg transition Ωr (t).
Trotter Enhanced
We find that this dual driving enables significantly faster re-
Sim. Coherence 6
Tsc × ||HT ||local
alizations of US and UP . After optimizing with GrAPE to
identify approximate gates with ideal fidelities above 99.9%,
4
we find total gate times below ΩTgate /2π = 6.0 up to clus-
2 (ii) ter sizes n = 8 and nearly constant scaling with n (see
Fig. 2b). These gates generate complex interactions in a very
0 hardware-efficient way, making them ideal for accelerating
(i) (ii)
spin-Hamiltonian simulations. Finally, we develop optimized
Hamiltonian HT
decompositions of target spin operations into two-qubit gates
(b)
(see SM), and find they are still orders of magnitude more
Sim. Coherence
Tsc × ||HT ||local
1.0 (iii) (iv) costly than the hardware-efficient implementation (Fig. 2c).
In Fig. 3, we illustrate the performance of this method
.1 on four representative examples that lie within the family of
Hamiltonians (1). To quantify the performance of the simula-
.01 tion, we present heuristic estimates of the accessible coherent
(iii) (iv) simulation time, measured in units of the target Hamiltonian’s
Hamiltonian HT local energy scale. We leverage access to multi-qubit spin
P n
operations of the form e−i n θn S , and estimate gate errors
FIG. 3. Efficiency of Hamiltonian simulation framework. Esti- based on the physical evolution time necessary to realize the
mates of the quantum simulation’s coherence time Tsc , in the tar- target operation. The step-size τ is chosen to to maximize
get Hamiltonian’s units ||HT ||local for various models. We con-
sider Hamiltonian simulation implemented using the dual driving
coherent simulation time, balancing simulation and gate er-
gates from Fig. 2b, and assume an error proportional to the gate rors (see Methods). In the representative examples, we find
time, such that ΩT /2π = 1 incurs an error of 0.1% [57, 67, 68]. the combination of dynamical projection and optimized multi-
In all cases, we compare against an implementation using two-qubit qubit gates outperforms a similarly constructed implementa-
CPhase gates with fidelity 99.9% and perfect single-qubit rotations tion based on Trotterized interactions and two-qubit gate de-
(see Methods for detailed descriptions of the estimation procedure). composition. Our approach significantly extends the available
(a) The first two two models are (i) the spin-1/2 Kagome Heisen- simulation time (up to ∼two orders of magnitude), and en-
berg model, and (ii) two interacting spin-5/2’s with Heisenberg and
Dzyaloshinskii–Moriya (DM) terms, both of which are composed of
ables much more efficient generation of complex spin Hamil-
only two-qubit interactions. In (i), a significant speedup is achieved tonians.
2
by utilizing three-qubit multi-qubit gates e−iτ S , which more effi-
ciently generates Heisenberg interactions and reduces the period of
the Floquet cycle from K = 4 to K = 2. In (ii) improvement is Spectral information from dynamics
achieved using dynamical projection, which reduces K from 2S to
2 but at the cost of additional multi-qubit gates. (b) Two complex Having described a toolbox for implementing time-
spin-models which include spin-interactions up to (iii) bi-quadratic evolution and state-preparation, we next present a general-
interactions J1 (Si · Sj ) + J2 (Si · Sj )2 and (iv) bi-quartic inter-
purpose algorithm for calculating chemically relevant infor-
actions (Si · Sj )4 . These correspond to four-body and eight-body
qubit interactions respectively. In (iii), the dramatic speedup orig- mation. The approach, dubbed “many-body spectroscopy”,
inates from using dynamical projection to significantly reduce the leverages dynamical snapshot measurements and classical co-
Floquet period, as well as the hardware-efficiency of a native four- processing [47] to compute a wide variety of spectral quanti-
qubit gate. The individual contribution to the speedup from both ties including low-lying states and finite-temperature proper-
sources is also analyzed in Methods. In (iv), the speedup arises fully ties. The procedure, combining insights from statistical phase
from the hardware-efficiency of native eight-qubit operations. estimation [80–83] and shadow tomography [84], is noise-
resilient and sample-efficient, making it especially promising
for near-term experiments.
works [57, 75, 76], the pulse profiles generating symmetric Specifically, the quantity we extract is an operator-resolved
diagonal operations can be obtained with numerical optimiza- density of states
tion via GrAPE [77]. For this alternating ansatz, we find a
roughly linear scaling of total gate time Tgate with size of X
the cluster (see Fig. 2b). Similar gates can also be imple- DA (ω) = ⟨n|A|n⟩δ(ω − ϵn ), (8)
n
mented in ion-trap architectures, where coupling to collec-
tive motional modes can be used to implement diagonal phase where ϵn and |n⟩ are the energies and eigenstates of the evo-
gates [78, 79]. lution Hamiltonian H, and A denotes an arbitrary operator.
However, Rydberg atom arrays offer additional control, Spectral functions like equation (8) can be used to access de-
which allows us to go beyond the alternating ansatz. Specif- tailed information about the properties of H: the location of
ically, we consider simultaneously driving the qubit transi- peaks provides information about energies [80, 82, 83], and
6
(a) (b) (c)

8 0.10
Susceptibility χ(T )
A=I Samples
P3 10k
6 0.08
P2 50k
DA (ω)
P1 0.06 Exact
4 P0
0.04
2
0.02
0
0.00
−30 −20 −10 0 10 20 100 101 102 103
Energy ω/J Temperature T /J
FIG. 4. Many-body spectroscopy of model Hamiltonians. (a) Schematic quantum circuit diagram for the algorithm. The first step is to apply
a state-preparation circuit S to prepare a reference state |S⟩ = S|0⟩, followed by an ancilla-controlled perturbation R preparing a superposition
of |S⟩ and the probe state |R⟩ = R|S⟩. This superposition is time-evolved by the target Hamiltonian, and then each qubit is projectively
measured to produce a snapshot. By repeating the procedure for various evolution times t, different perturbations R, and potentially different
measurement bases, this setup provides access to detailed information about the spectrum of H. (b) Consider a spin Hamiltonian H2 with two
anti-ferromagnetically coupled spin-3/2’s. We simulate 20,000 snapshot measurements and classical post processing to calculate the density-
of-states D1 (ω) (black line), and total-spin resolved versions DPS (ω) (colored lines). Vertical dashed lines correspond to exact energies and
colored regions represent 95% confidence intervals. Peaks are broadened due to finite (coherent) simulation time JTsim = 0.26, which sets
the spectral resolution. Hardware-efficient simulation schemes, which extend the simulation time (e.g. Fig. 3), are favorable because they
improve spectral resolution. Many-body spectroscopy further improves the effective spectral resolution, by leveraging multiple observables to
distinguish overlapping peaks. Here, we see that spin-resolution significantly sparsifies the signal, enabling accurate peak detection and energy
estimation, while the bare spectrum D1 (ω) is too broad to resolve all states. (c) The magnetic susceptibility χ, can also be computed from
snapshot measurements using S z -resolved density-of-states (here JTsim = 1.04). For these calculations, it is important to prevent exponential
amplification of shot-noise. We therefore use a simple empirical truncation procedure which introduces a small amount of bias (see Methods),
but enables rapid convergence with number of snapshots to the ideal value (black line).
properties of eigenstates can be computed by choosing A ap-

propriately [81]. For example,
P we can compute total-spin of
Z X
an eigenstate using A = ( i Ŝi )2 , or the local spin polariza- DA (ω) = ẼR∼R dt eiωt ⟨S|R† A Os e−iHt |S⟩
tion with A = Ŝi . | {z }| s
{z }
circuit average
classical processing
iHt −iHt
× ⟨S|e Os e R|S⟩ . (9)
| {z }
quantum evolution (Fig. 4)
The output of the quantum computation are projective mea-
surements (snapshots), produced by the circuit depicted in For this to be formally equivalent to (8), the distribution over
Fig. 4a. First, we initialize a system of N qubits and ap- perturbations R and ensemble of observables {Os } must cou-
ply a state-preparation procedure to prepare a reference state ple uniformly to all eigenstates, to ensure unbiased estimation
|S⟩ = S|0N ⟩. Next, we use a single ancilla qubit to apply a (see Methods). For example, given a polarized reference state
controlled perturbation, preparing a superposition of |S⟩ and |0⟩⊗N , X-basis measurements are sufficient.
a probe state |R⟩ = R|S⟩, which is then evolved under H for To realize the distributional average and time-integral, the
time t. Finally each qubit is projectively measured, producing quantum circuit has to be executed (Fig. 4a) for randomly
a sequence of N + 1 bits — a snapshot. By measuring the an- sampled perturbations R and evolution times t. In order to ef-
cilla in the X or Y basis, this circuit effectively performs an ficiently evaluate the classical part of (9), we require efficient
interferometry experiment between the reference and probe classical representations of |R⟩ and e−iHt |S⟩. A good choice
states. The resulting snapshot measurements enable paral- is to select |S⟩ to be a known eigenstate of H, such that the
lel estimation of two-time correlation functions of the form time-evolution is trivial, and preferentially sample |R⟩ to max-
CO,R (t) = ⟨S|eiHt Oe−iHt R|S⟩ for all 2N operators O that imize the overlap with relevant target states. In contrast, the
are diagonal in the measurement basis. quantum part of (9) includes time-evolution of |R⟩, which has
overlap with unknown eigenstates, and often includes large
amounts of entanglement. Therefore, this is estimated from
snapshot measurements produced by quantum simulation (see
Methods).
To estimate the spectral function DA (ω), we use a hy- As an example, consider two interacting spin-3/2 particles,
brid quantum-classical computation based on the following described by H2 = JS1 ·S2 . We prepare a polarized reference
expression, eigenstate |S⟩ = |0⟩⊗6 , and sample perturbations R from an
7
ensemble of random single-spin rotations. This corresponds (a) (b)

20
to a trivial state-preparation circuit and a simple controlled-
D1 (ω)
perturbation composed of two-qubit gates in Fig. 4. Next, we 10
measure the system in the X-basis, which provides access to
the full spectrum for this choice of |S⟩ (Methods). Finally, 0
during classical processing, the bare density of states is ob-
tained by choosing A = 1, and evaluating (9). The result −400 −300 −200 −100 0
contains peaks at frequencies ω associated with eigenstates of S2H-1b S2H-2b Energy ω (cm−1 )
H2 (see Fig. 4b). Further, we can isolate individual contribu- (c) (d)
tions of total spin sectors by instead choosing A = PS , the 15 40
0.3
7/2 7/2 5/2
∆E (cm−1 )
projectors onto S = 0, 1, 2, and 3. This not only allows us to 0.0
5/2 5/2 7/2
DA (ω)
10 13/2
identify the total spin of the eigenstates, but also increases the 20
9/2
effective spectral resolution in the presence of noise as it spar- 5 11/2 11/2
sifies the signal. Finite-temperature response functions [85], 9/2

13/2
0 0
like the z-component of the zero-field magnetic susceptibil-
−380 −370 −360 −350
ex
S2
S2
ity χ(T ) = Z1 Tr[ T1 (Sz )2 e−H/T ], can also be computed from
H
H
p.
−1
-2
-1
Energy ω (cm )
b
b
the same dataset, by integrating the Sz -projected density-of-
states DSz (ω) (see Methods). To illustrate this, the magnetic
FIG. 5. Application to the oxygen-evolving complex (OEC). Our
susceptibility is extracted from the same dataset and shown in programmable quantum simulation framework can be used to com-
Fig. 4c. The algorithm is especially promising for near-term pute detailed model spin Hamiltonian properties. (a) Here, we il-
devices, having favorable resource requirements quantified by lustrate the procedure on the OEC, an organometallic catalyst with
the number of snapshots (sample complexity) and maximum strong spin correlations. In particular, we simulate model spin
evolution time (coherence) required for accurate spectral com- Hamiltonians for two structures S2 H-1b and S2 H-2b, which have
putation (see Methods for further discussion). three spin-3/2 and one spin-2 Mn (purple) active sites (reproduced
from Ref. [28] with permission from the Royal Society of Chem-
istry). Model Heisenberg coefficients for both hypothetical struc-
tures have been computed from broken-symmetry DFT [28]. (b)
Application to transition metal clusters and magnetic solids A density of states D1 (ω) calculation is simulated for the S2 H-1b
model spin Hamiltonian. Here, we use a polarized reference state
As an illustration of a relevant computation in chemical |S⟩ = |0⟩⊗13 a probe states |R⟩ generated by random single-site
catalysis, we consider the Mn4 O5 Ca core of the oxygen- rotations, and a evolution times t. We select 50,000 circuits with in-
dependently chosen |R⟩, t pairs, and draw 10 snapshots from each
evolving complex (OEC), a transition metal catalyst central
circuit. (c) Focusing on the lowest-lying states, we see three distinct
to photosynthesis which is still not fully understood [86, 87]. peaks in D1 (ω). However, by evaluating spin-resolved quantities
Classical chemistry calculations have been used to fit model DPs (ω) on the same set of measurements, we identify three addi-
Heisenberg Hamiltonians, containing three spin-3/2 sites and tional peaks, whose energies and total-spin match exact diagonaliza-
one spin-2 site [27, 28, 88] (Fig. 5a). While this spin represen- tion results (vertical dotted lines). (d) This information is known as
tation cannot directly capture chemical reactions, it can cap- the spin-ladder, and can be computed using many-body spectroscopy
ture the ground and low-lying spin-states, i. e., the spin-ladder, for both the 1b and 2b states. Importantly, the spin-ladder can also
be measured experimentally, and therefore can be used to help deter-
which are important in catalysis because reaction pathways
mine which structure appears in nature. In this example, experimen-
depend critically on the spin multiplicity [89]. We simulate tal measurements indicate a spin-5/2 ground state and spin-7/2 first
our framework applied to the S2 H-1b structural model from excited state. However, the ordering of low-energy states is flipped
Ref. [28], by first computing the bare density-of-states D1 (A) in the 2b configuration, indicating the S2H-1b hypothesis is more
(Fig. 5b). Then, we identify a low-lying cluster of eigenstates, likely [28]. We note that quantities beyond total-spin can also be
and compute spin-projected densities DPs (A) to resolve the readily evaluated in low-lying eigenstates by inserting different op-
spin ladder (Fig. 5c). The spin ladder can also be measured erators A (see Methods).
experimentally, providing a way to evaluate candidate models
of reaction intermediates. To highlight this, we simulate the
Heisenberg model for an alternate pathway (S2 H-2b) [28], and
observe the modification reverses the ordering of the spin lad-
der, indicating S2 H-1b is more consistent with measurements
(see Fig. 5d).
The framework can also be applied to study low-energy in a noise-resiliant manner via local controlled-perturbations
properties of extended systems, including strongly correlated R. Then, local Green’s functions — two-point operators at
materials. We illustrate this on the ferromagnetic, square different positions and times — can be measured to access
lattice Heisenberg model (Fig. 6). For such large systems, properties of low-lying quasi-particle excitations, such as the
we envision utilizing an approximate ground-state preparation dispersion relation of single-particle excitations (see Methods,
method for |S⟩, so that low-energy properties can be accessed 2D Heisenberg, for details).
8
(a) tries [97, 98].

The efficiency of the Hamiltonian engineering ap-
proach originates from co-designing Floquet engineering and
hardware-specific multi-qubit gates. Extending this approach
to larger classes of strongly-correlated model Hamiltonians is
an outstanding and exciting frontier. In particular, it would
be especially interesting to further develop the toolbox to in-
(b) (c) corporate charge transport and electron-phonon interactions.
This could enable simulation of more complex model Hamil-
2 tonians, such as the t-J, Hubbard, and Hubbard-Holstein
Frequency ω
models, and expand the class of accessible chemistry prob-

1 lems [9, 99]. Application to other settings, including lattice
gauge theories [100, 101] and quantum optimization prob-
0 lems [102] are also of interest. Incorporation of error miti-
−2 0 2
gation and correction into the Hamiltonian simulation should
be considered; specifically, the present method can potentially
Wavevector kx
be generalized to control logical, encoded degrees of freedom
in a hardware-efficient way [103].
FIG. 6. Application to 2D magnetic materials. (a) Properties of
extended materials can also be investigated using our quantum simu- Finally, characterization and development of the model
lation framework. As an example, we study the square-lattice
P ferro- Hamiltonian approach itself is an interesting and challeng-
magnetic (J > 0) Heisenberg model H2D = −J ⟨ij⟩ ŝi · ŝj . By ing problem. Key challenges include development of effi-
preparing the polarized ground state |S⟩ = |0⟩⊗N , applying a single- cient schemes to compute parameters for higher-order inter-
site perturbation R = X0 on the central site, and measuring the sys- actions [91], estimation of corrections arising from coupling
tem in the X-basis after time evolution, we can estimate the single- to states outside the model space [104], and validation of
particle Green’s function G(r, t) = ⟨S|Xr (t)X0 |S⟩. Therefore we
the model Hamiltonian approximation [19]. Feedback be-
select O = Xr for various positions r as the observables in (9), all
of which are diagonal in the measurement basis. We visualize the tween the classical and quantum parts of the computation
real-part of G(r, t), where the plotted intensity and color denotes the is an important part of these developments [25, 105, 106].
magnitude and sign at Jt = 0, 0.5, and 1.0. (b) The structure of For these reasons, the large-scale simulation of model Hamil-
excited states is extracted by classical post-processing of these mea- tonians on quantum processors will be invaluable for test-
surements. Even though the spectrum is continuous, additional struc- ing approximations by comparing simulation outputs with
ture can be identified by computing the momentum-resolved density experimental measurements. Hence, the approach proposed
of states DPk (ω), where Pk is a projector onto plane-wave states
in this work facilitates exciting directions in computational
(see Methods). Restricting to ky = 0 and evolving to maximum
time JTmax = 8.0, we see DPk (ω) forms a band-like structure, chemistry and quantum simulation, aiming towards construct-
from which a peak ω can be estimated for each k (black line). (c) ing a novel ab initio simulation pipeline that utilizes hybrid
This peak extraction allows us to directly estimate the single-particle quantum-classical resources.
dispersion ω(k) across the 2D Brillouin zone.
Outlook
[1] A. Szabo and N. S. Ostlund, Modern quantum chemistry: in-
troduction to advanced electronic structure theory (Courier
These considerations indicate that reconfigurable quantum Corporation, 2012).
processors enable a powerful, hardware-efficient framework [2] R. J. Bartlett and M. Musiał, Coupled-cluster theory in quan-
for quantum simulation of problems from chemistry and ma- tum chemistry, Rev. Mod. Phys. 79, 291 (2007).
terials science, illustrating potential directions for the search [3] N. Mardirossian and M. Head-Gordon, Thirty years of density
functional theory in computational chemistry: an overview
for useful quantum advantage. Specifically, in addition to the and extensive assessment of 200 density functionals, Mol.
OEC, other organometallic catalysts could be studied with this Phys. 115, 2315 (2017).
approach, including iron-sulfur clusters [23, 90], for which [4] K. Burke, Perspective on density functional theory, J. Chem.
bi-quadratic terms appear in the model Hamiltonian to cap- Phys. 136 (2012).
ture higher-order perturbative charge fluctuation effects [91]. [5] Y. A. Aoto, A. P. de Lima Batista, A. Kohn, and A. G.
Another promising direction involves quantum simulation of de Oliveira-Filho, How to arrive at accurate benchmark values
for transition metal compounds: Computation or experiment?,
low-energy properties of 2D and 3D frustrated spin systems,
J. Chem. Theory Comput. 13, 5291 (2017).
including model Hamiltonians for Kitaev materials [92–94] [6] D. Hait, N. M. Tubman, D. S. Levine, K. B. Whaley, and
and molecular magnets [95, 96]. The ability to realize non- M. Head-Gordon, What levels of coupled cluster theory are
local interactions further opens the door to simulation of spin- appropriate for transition metal systems? a study using near-
Hamiltonians defined on non-Euclidean interaction geome- exact quantum chemical values for 3d transition metal binary
9
compounds, J. Chem. Theory Comput. 15, 5370 (2019). G. K.-L. Chan, Simulating Models of Challenging Correlated
[7] C. Ahn, A. Cavalleri, A. Georges, S. Ismail-Beigi, A. J. Millis, Molecules and Materials on the Sycamore Quantum Proces-
and J.-M. Triscone, Designing and controlling the properties sor, PRX Quantum 3, 040318 (2022).
of transition metal oxide quantum materials, Nat. Mater. 20, [24] D. Wecker, M. B. Hastings, N. Wiebe, B. K. Clark, C. Nayak,
1462 (2021). and M. Troyer, Solving strongly correlated electron models on
[8] Y. Alexeev, D. Bacon, K. R. Brown, R. Calderbank, L. D. Carr, a quantum computer, Physical Review A 92, 062318 (2015).
F. T. Chong, B. DeMarco, D. Englund, E. Farhi, B. Fefferman, [25] B. Bauer, D. Wecker, A. J. Millis, M. B. Hastings, and
A. V. Gorshkov, A. Houck, J. Kim, S. Kimmel, M. Lange, M. Troyer, Hybrid Quantum-Classical Approach to Correlated
S. Lloyd, M. D. Lukin, D. Maslov, P. Maunz, C. Monroe, Materials, Physical Review X 6, 031045 (2016).
J. Preskill, M. Roetteler, M. J. Savage, and J. Thompson, [26] H. Ma, M. Govoni, and G. Galli, Quantum simulations of ma-
Quantum Computer Systems for Scientific Discovery, PRX terials on near-term quantum computers, npj Computational
Quantum 2, 017001 (2021). Materials 6, 85 (2020).
[9] B. Bauer, S. Bravyi, M. Motta, and G. K.-L. Chan, Quantum [27] V. Krewald, F. Neese, and D. A. Pantazis, On the Magnetic and
algorithms for quantum chemistry and quantum materials sci- Spectroscopic Properties of High-Valent Mn 3 CaO 4 Cubanes
ence, Chem. Rev. 120, 12685 (2020). as Structural Units of Natural and Artificial Water-Oxidizing
[10] S. McArdle, S. Endo, A. Aspuru-Guzik, S. C. Benjamin, and Catalysts, J. Am. Chem. Soc. 135, 5726 (2013).
X. Yuan, Quantum computational chemistry, Rev. Mod. Phys. [28] V. Krewald, M. Retegan, N. Cox, J. Messinger, W. Lubitz,
92, 015003 (2020). S. DeBeer, F. Neese, and D. A. Pantazis, Metal oxidation states
[11] L. Lin, Lecture notes on quantum algorithms for scientific in biological water splitting, Chem. Sci. 6, 1676 (2015).
computation, arXiv:2201.08309 [quant-ph] (2022). [29] J. P. Malrieu, R. Caballol, C. J. Calzado, C. de Graaf, and
[12] M. E. Beverland, P. Murali, M. Troyer, K. M. Svore, T. Hoe- N. Guihéry, Magnetic Interactions in Molecules and Highly
fler, V. Kliuchnikov, G. H. Low, M. Soeken, A. Sundaram, Correlated Materials: Physical Content, Analytical Deriva-
and A. Vaschillo, Assessing requirements to scale to practical tion, and Rigorous Extraction of Magnetic Hamiltonians,
quantum advantage, arXiv:2211.07629 [quant-ph] (2022). Chem. Rev. 114, 429 (2014).
[13] A. Auerbach, Interacting electrons and quantum magnetism [30] J. Shee, M. Loipersberger, D. Hait, J. Lee, and M. Head-
(Springer Science & Business Media, 1998). Gordon, Revealing the nature of electron correlation in tran-
[14] N. P. Bauman, G. H. Low, and K. Kowalski, Quantum simula- sition metal complexes with symmetry breaking and chemical
tions of excited states with active-space downfolded Hamilto- intuition, J. Chem. Phys. 154, 194109 (2021).
nians, J. Chem. Phys. 151, 234114 (2019). [31] D. A. Pantazis, M. Orio, T. Petrenko, S. Zein, W. Lub-
[15] H. Bolvin, From ab Initio Calculations to Model Hamiltoni- itz, J. Messinger, and F. Neese, Structure of the oxygen-
ans: The Effective Hamiltonian Technique as an Efficient Tool evolving complex of photosystem II: information on the S2
to Describe Mixed-Valence Molecules, J. Phys. Chem. A 107, state through quantum chemical calculation of its magnetic
5071 (2003). properties, Phys. Chem. Chem. Phys. 11, 6788 (2009).
[16] N. J. Mayhall and M. Head-Gordon, Computational quantum [32] C. J. Calzado, J. M. Clemente-Juan, E. Coronado, A. Gaita-
chemistry for single Heisenberg spin couplings made simple: Arino, and N. Suaud, Role of the Electron Transfer and Mag-
Just one spin flip required, J. Chem. Phys. 141, 134111 (2014). netic Exchange Interactions in the Magnetic Properties of
[17] N. J. Mayhall and M. Head-Gordon, Computational Quantum Mixed-Valence Polyoxovanadate Complexes, Inorg. Chem.
Chemistry for Multiple-Site Heisenberg Spin Couplings Made 47, 5889 (2008).
Simple: Still Only One Spin–Flip Required, J. Phys. Chem. [33] X. Li, H. Yu, F. Lou, J. Feng, M.-H. Whangbo, and H. Xiang,
Lett. 6, 1982 (2015). Spin Hamiltonians in Magnets: Theories and Computations,
[18] P. Pokhilko and A. I. Krylov, Effective Hamiltonians de- Molecules 26, 803 (2021).
rived from equation-of-motion coupled-cluster wave func- [34] P. Pokhilko, D. S. Bezrukov, and A. I. Krylov, Is Solid Copper
tions: Theory and application to the Hubbard and Heisenberg Oxalate a Spin Chain or a Mixture of Entangled Spin Pairs?,
Hamiltonians, J. Chem. Phys. 152, 094108 (2020). J. Phys. Chem. C 125, 7502 (2021).
[19] S. Kotaru, S. Kähler, M. Alessio, and A. I. Krylov, Magnetic [35] C. Monroe, W. Campbell, L.-M. Duan, Z.-X. Gong, A. Gor-
exchange interactions in binuclear and tetranuclear iron (iii) shkov, P. Hess, R. Islam, K. Kim, N. Linke, G. Pagano,
complexes described by spin-flip dft and heisenberg effective P. Richerme, C. Senko, and N. Yao, Programmable quan-
hamiltonians, J. Comput. Chem. 44, 367 (2023). tum simulations of spin systems with trapped ions, Rev. Mod.
[20] D.-T. Chen, P. Helms, A. R. Hale, M. Lee, C. Li, J. Gray, Phys. 93, 025001 (2021).
G. Christou, V. S. Zapf, G. K.-L. Chan, and H.-P. Cheng, [36] A. J. Daley, I. Bloch, C. Kokail, S. Flannigan, N. Pearson,
Using Hyperoptimized Tensor Networks and First-Principles M. Troyer, and P. Zoller, Practical quantum advantage in quan-
Electronic Structure to Simulate the Experimental Properties tum simulation, Nature 607, 667 (2022).
of the Giant {Mn84} Torus, J. Phys. Chem. Lett. 13, 2365 [37] H. Bernien, S. Schwartz, A. Keesling, H. Levine, A. Omran,
(2022). H. Pichler, S. Choi, A. S. Zibrov, M. Endres, M. Greiner,
[21] A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, V. Vuletić, and M. D. Lukin, Probing many-body dynamics
J. M. Chow, and J. M. Gambetta, Hardware-efficient varia- on a 51-atom quantum simulator, Nature 551, 579 (2017).
tional quantum eigensolver for small molecules and quantum [38] A. Keesling, A. Omran, H. Levine, H. Bernien, H. Pich-
magnets, Nature 549, 242 (2017). ler, S. Choi, R. Samajdar, S. Schwartz, P. Silvi, S. Sachdev,
[22] A. Chiesa, F. Tacchino, M. Grossi, P. Santini, I. Tavernelli, P. Zoller, M. Endres, M. Greiner, V. Vuletić, and M. D. Lukin,
D. Gerace, and S. Carretta, Quantum hardware simulating Quantum Kibble–Zurek mechanism and critical dynamics on
four-dimensional inelastic neutron scattering, Nat. Phys. 15, a programmable Rydberg simulator, Nature 568, 207 (2019).
455 (2019). [39] D. Bluvstein, A. Omran, H. Levine, A. Keesling, G. Semegh-
[23] R. N. Tazhigulov, S.-N. Sun, R. Haghshenas, H. Zhai, ini, S. Ebadi, T. T. Wang, A. A. Michailidis, N. Maskara,
A. T. Tan, N. C. Rubin, R. Babbush, A. J. Minnich, and W. W. Ho, S. Choi, M. Serbyn, M. Greiner, V. Vuletić, and
10
M. D. Lukin, Controlling quantum many-body dynamics in Thompson, Universal Gate Operations on Nuclear Spin Qubits
driven Rydberg atom arrays, Science 371, 1355 (2021). in an Optical Tweezer Array of 171 Yb Atoms, Phys. Rev. X
[40] H. Labuhn, D. Barredo, S. Ravets, S. de Léséleuc, T. Macrı̀, 12, 021028 (2022).
T. Lahaye, and A. Browaeys, Tunable two-dimensional arrays [51] K. Singh, S. Anand, A. Pocklington, J. T. Kemp, and
of single Rydberg atoms for realizing quantum Ising models, H. Bernien, Dual-Element, Two-Dimensional Atom Array
Nature 534, 667 (2016). with Continuous-Mode Operation, Phys. Rev. X 12, 011040
[41] S. Ebadi, T. T. Wang, H. Levine, A. Keesling, G. Semeghini, (2022).
A. Omran, D. Bluvstein, R. Samajdar, H. Pichler, W. W. Ho, [52] A. Jenkins, J. W. Lis, A. Senoo, W. F. McGrew, and
S. Choi, S. Sachdev, M. Greiner, V. Vuletić, and M. D. Lukin, A. M. Kaufman, Ytterbium Nuclear-Spin Qubits in an Opti-
Quantum phases of matter on a 256-atom programmable quan- cal Tweezer Array, Physical Review X 12, 021027 (2022).
tum simulator, Nature 595, 227 (2021). [53] O. Katz, L. Feng, A. Risinger, C. Monroe, and M. Cetina,
[42] P. Scholl, M. Schuler, H. J. Williams, A. A. Eberharter, Demonstration of three- and four-body interactions between
D. Barredo, K.-N. Schymik, V. Lienhard, L.-P. Henry, T. C. trapped-ion spins, Nat. Phys. 19, 1452 (2023).
Lang, T. Lahaye, A. M. Läuchli, and A. Browaeys, Quantum [54] H. Häffner, C. F. Roos, and R. Blatt, Quantum computing with
simulation of 2D antiferromagnets with hundreds of Rydberg trapped ions, Phys. Rep. 469, 155 (2008).
atoms, Nature 595, 233 (2021). [55] S. A. Moses, C. H. Baldwin, M. S. Allman, R. Ancona, L. As-
[43] C. Chen, G. Bornet, M. Bintz, G. Emperauger, L. Leclerc, carrunz, C. Barnes, J. Bartolotta, B. Bjork, P. Blanchard,
V. S. Liu, P. Scholl, D. Barredo, J. Hauschild, S. Chatterjee, M. Bohn, J. G. Bohnet, N. C. Brown, N. Q. Burdick, W. C.
M. Schuler, A. M. Läuchli, M. P. Zaletel, T. Lahaye, N. Y. Burton, S. L. Campbell, J. P. C. I. au2, C. Carron, J. Chambers,
Yao, and A. Browaeys, Continuous symmetry breaking in a J. W. Chan, Y. H. Chen, A. Chernoguzov, E. Chertkov, J. Col-
two-dimensional Rydberg array, Nature 616, 691 (2023). ina, J. P. Curtis, R. Daniel, M. DeCross, D. Deen, C. Delaney,
[44] G. Semeghini, H. Levine, A. Keesling, S. Ebadi, T. T. J. M. Dreiling, C. T. Ertsgaard, J. Esposito, B. Estey, M. Fab-
Wang, D. Bluvstein, R. Verresen, H. Pichler, M. Kali- rikant, C. Figgatt, C. Foltz, M. Foss-Feig, D. Francois, J. P.
nowski, R. Samajdar, A. Omran, S. Sachdev, A. Vishwanath, Gaebler, T. M. Gatterman, C. N. Gilbreth, J. Giles, E. Glynn,
M. Greiner, V. Vuletić, and M. D. Lukin, Probing topological A. Hall, A. M. Hankin, A. Hansen, D. Hayes, B. Higashi,
spin liquids on a programmable quantum simulator, Science I. M. Hoffman, B. Horning, J. J. Hout, R. Jacobs, J. Johansen,
374, 1242 (2021). L. Jones, J. Karcz, T. Klein, P. Lauria, P. Lee, D. Liefer, C. Ly-
[45] K. J. Satzinger, Y.-J. Liu, A. Smith, C. Knapp, M. New- tle, S. T. Lu, D. Lucchetti, A. Malm, M. Matheny, B. Mathew-
man, C. Jones, Z. Chen, C. Quintana, X. Mi, A. Dunsworth, son, K. Mayer, D. B. Miller, M. Mills, B. Neyenhuis, L. Nu-
C. Gidney, I. Aleiner, F. Arute, K. Arya, J. Atalaya, R. Bab- gent, S. Olson, J. Parks, G. N. Price, Z. Price, M. Pugh,
bush, J. C. Bardin, R. Barends, J. Basso, A. Bengtsson, A. Ransford, A. P. Reed, C. Roman, M. Rowe, C. Ryan-
A. Bilmes, M. Broughton, B. B. Buckley, D. A. Buell, B. Bur- Anderson, S. Sanders, J. Sedlacek, P. Shevchuk, P. Siegfried,
kett, N. Bushnell, B. Chiaro, R. Collins, W. Courtney, S. De- T. Skripka, B. Spaun, R. T. Sprenkle, R. P. Stutz, M. Swallows,
mura, A. R. Derk, D. Eppens, C. Erickson, L. Faoro, E. Farhi, R. I. Tobey, A. Tran, T. Tran, E. Vogt, C. Volin, J. Walker,
A. G. Fowler, B. Foxen, M. Giustina, A. Greene, J. A. Gross, A. M. Zolot, and J. M. Pino, A race track trapped-ion quan-
M. P. Harrigan, S. D. Harrington, J. Hilton, S. Hong, T. Huang, tum processor, arXiv:2305.03828 [quant-ph] (2023).
W. J. Huggins, L. B. Ioffe, S. V. Isakov, E. Jeffrey, Z. Jiang, [56] D. Bluvstein, H. Levine, G. Semeghini, T. T. Wang, S. Ebadi,
D. Kafri, K. Kechedzhi, T. Khattar, S. Kim, P. V. Klimov, A. N. M. Kalinowski, A. Keesling, N. Maskara, H. Pichler,
Korotkov, F. Kostritsa, D. Landhuis, P. Laptev, A. Locharla, M. Greiner, V. Vuletić, and M. D. Lukin, A quantum processor
E. Lucero, O. Martin, J. R. McClean, M. McEwen, K. C. based on coherent transport of entangled atom arrays, Nature
Miao, M. Mohseni, S. Montazeri, W. Mruczkiewicz, J. Mutus, 604, 451 (2022).
O. Naaman, M. Neeley, C. Neill, M. Y. Niu, T. E. O’Brien, [57] S. J. Evered, D. Bluvstein, M. Kalinowski, S. Ebadi,
A. Opremcak, B. Pató, A. Petukhov, N. C. Rubin, D. Sank, T. Manovitz, H. Zhou, S. H. Li, A. A. Geim, T. T. Wang,
V. Shvarts, D. Strain, M. Szalay, B. Villalonga, T. C. White, N. Maskara, H. Levine, G. Semeghini, M. Greiner, V. Vuletić,
Z. Yao, P. Yeh, J. Yoo, A. Zalcman, H. Neven, S. Boixo, and M. D. Lukin, High-fidelity parallel entangling gates on a
A. Megrant, Y. Chen, J. Kelly, V. Smelyanskiy, A. Kitaev, neutral-atom quantum computer, Nature 622, 268 (2023).
M. Knap, F. Pollmann, and P. Roushan, Realizing topolog- [58] A. Kruckenhauser, R. v. Bijnen, T. V. Zache, M. D. Liberto,
ically ordered states on a quantum processor, Science 374, and P. Zoller, High-dimensional SO(4)-symmetric Rydberg
1237 (2021). manifolds for quantum simulation, Quantum Sci. Technol. 8,
[46] Y. Kim, A. Eddins, S. Anand, K. X. Wei, E. Van Den Berg, 015020 (2022).
S. Rosenblatt, H. Nayfeh, Y. Wu, M. Zaletel, K. Temme, and [59] Y. Chi, J. Huang, Z. Zhang, J. Mao, Z. Zhou, X. Chen, C. Zhai,
A. Kandala, Evidence for the utility of quantum computing J. Bao, T. Dai, H. Yuan, M. Zhang, D. Dai, B. Tang, Y. Yang,
before fault tolerance, Nature 618, 500 (2023). Z. Li, Y. Ding, L. K. Oxenløwe, M. G. Thompson, J. L.
[47] H. H. S. Chan, R. Meister, M. L. Goh, and B. Koczor, Al- O’Brien, Y. Li, Q. Gong, and J. Wang, A programmable qudit-
gorithmic shadow spectroscopy, arXiv:2212.11036 [quant-ph] based quantum processor, Nat. Commun. 13, 1166 (2022).
(2023). [60] P. Hrmo, B. Wilhelm, L. Gerster, M. W. van Mourik, M. Hu-
[48] D. Barredo, S. De Léséleuc, V. Lienhard, T. Lahaye, and ber, R. Blatt, P. Schindler, T. Monz, and M. Ringbauer, Native
A. Browaeys, An atom-by-atom assembler of defect-free ar- qudit entanglement in a trapped ion quantum processor, Nat.
bitrary two-dimensional atomic arrays, Science 354, 1021 Commun. 14, 2242 (2023).
(2016). [61] D. González-Cuadra, T. V. Zache, J. Carrasco, B. Kraus, and
[49] A. Cooper, J. P. Covey, I. S. Madjarov, S. G. Porsev, M. S. P. Zoller, Hardware Efficient Quantum Simulation of Non-
Safronova, and M. Endres, Alkaline-Earth Atoms in Optical Abelian Gauge Theories with Qudits on Rydberg Platforms,
Tweezers, Phys. Rev. X 8, 041055 (2018). Phys. Rev. Lett. 129, 160501 (2022).
[50] S. Ma, A. P. Burgers, G. Liu, J. Wilson, B. Zhang, and J. D. [62] D. A. Abanin, W. De Roeck, W. W. Ho, and F. Huveneers, Ef-
11
fective Hamiltonians, prethermalization, and slow energy ab- [81] S. Lu, M. C. Bañuls, and J. I. Cirac, Algorithms for quantum
sorption in periodically driven many-body systems, Phys. Rev. simulation at finite energies, PRX Quantum 2, 020321 (2021).
B 95, 014112 (2017). [82] R. D. Somma, Quantum eigenvalue estimation via time series
[63] J. Choi, H. Zhou, H. S. Knowles, R. Landig, S. Choi, and analysis, New J. Phys. 21, 123025 (2019).
M. D. Lukin, Robust Dynamic Hamiltonian Engineering of [83] T. E. O’Brien, B. Tarasinski, and B. M. Terhal, Quantum phase
Many-Body Spin Systems, Phys. Rev. X 10, 031002 (2020). estimation of multiple eigenvalues for small-scale (noisy) ex-
[64] D. V. Else, B. Bauer, and C. Nayak, Prethermal Phases of Mat- periments, New J. Phys. 21, 023022 (2019).
ter Protected by Time-Translation Symmetry, Phys. Rev. X 7, [84] H.-Y. Huang, R. Kueng, and J. Preskill, Predicting many prop-
011026 (2017). erties of a quantum system from very few measurements, Nat.
[65] J. Carrasco, A. Elben, C. Kokail, B. Kraus, and P. Zoller, The- Phys. 16, 1050 (2020).
oretical and Experimental Perspectives of Quantum Verifica- [85] M. Alessio and A. I. Krylov, Equation-of-Motion Coupled-
tion, PRX Quantum 2, 010102 (2021). Cluster Protocol for Calculating Magnetic Properties: Theory
[66] M. Benedetti, M. Fiorentini, and M. Lubasch, Hardware- and Applications to Single-Molecule Magnets, J. Chem. The-
efficient variational quantum algorithms for time evo- ory Comput. 17, 4225 (2021).
lution, Physical Review Research 3, 033083 (2021), [86] M. Askerka, G. W. Brudvig, and V. S. Batista, The o2-evolving
arXiv:2009.12361 [quant-ph]. complex of photosystem ii: Recent insights from quantum
[67] P. Scholl, A. L. Shaw, R. B.-S. Tsai, R. Finkelstein, J. Choi, mechanics/molecular mechanics (qm/mm), extended x-ray ab-
and M. Endres, Erasure conversion in a high-fidelity Rydberg sorption fine structure (exafs), and femtosecond x-ray crystal-
quantum simulator, Nature 622, 273 (2023). lography data, Acc. Chem. Res. 50, 41 (2017).
[68] S. Ma, G. Liu, P. Peng, B. Zhang, S. Jandura, J. Claes, A. P. [87] S. Paul, F. Neese, and D. A. Pantazis, Structural models of the
Burgers, G. Pupillo, S. Puri, and J. D. Thompson, High-fidelity biological oxygen-evolving complex: achievements, insights,
gates and mid-circuit erasure conversion in an atomic qubit, and challenges for biomimicry, Green Chem. 19, 2309 (2017).
Nature 622, 279 (2023). [88] V. Krewald, M. Retegan, F. Neese, W. Lubitz, D. A. Pantazis,
[69] J. W. Lis, A. Senoo, W. F. McGrew, F. Rönchen, A. Jenkins, and N. Cox, Spin State as a Marker for the Structural Evo-
and A. M. Kaufman, Mid-circuit operations using the omg- lution of Nature’s Water-Splitting Catalyst, Inorg. Chem. 55,
architecture in neutral atom arrays (2023), arXiv:2305.19266 488 (2016).
[cond-mat, physics:physics, physics:quant-ph]. [89] S. Shaik, H. Hirao, and D. Kumar, Reactivity of high-valent
[70] C. Sheng, X. He, P. Xu, R. Guo, K. Wang, Z. Xiong, M. Liu, iron–oxo species in enzymes and synthetic reagents: a tale of
J. Wang, and M. Zhan, High-fidelity single-qubit gates on neu- many states, Acc. Chem. Res. 40, 532 (2007).
tral atoms in a two-dimensional magic-intensity optical dipole [90] Z. Li, S. Guo, Q. Sun, and G. K.-L. Chan, Electronic landscape
trap array, Phys. Rev. Lett. 121, 240501 (2018). of the P-cluster of nitrogenase as revealed through many-
[71] H. Levine, D. Bluvstein, A. Keesling, T. T. Wang, S. Ebadi, electron quantum wavefunction simulations, Nat. Chem. 11,
G. Semeghini, A. Omran, M. Greiner, V. Vuletić, and M. D. 1026 (2019).
Lukin, Dispersive optical systems for scalable raman driving [91] S. Sharma, K. Sivalingam, F. Neese, and G. K.-L. Chan, Low-
of hyperfine qubits, Phys. Rev. A 105, 032618 (2022). energy spectrum of iron–sulfur clusters directly from many-
[72] D. Jaksch, J. I. Cirac, P. Zoller, S. L. Rolston, R. Côté, and particle quantum mechanics, Nat. Chem. 6, 927 (2014).
M. D. Lukin, Fast Quantum Gates for Neutral Atoms, Phys. [92] H.-K. Jin, W. M. H. Natori, F. Pollmann, and J. Knolle, Unveil-
Rev. Lett. 85, 2208 (2000). ing the S=3/2 Kitaev honeycomb spin liquids, Nat. Commun.
[73] H. Levine, A. Keesling, G. Semeghini, A. Omran, T. T. Wang, 13, 3813 (2022).
S. Ebadi, H. Bernien, M. Greiner, V. Vuletić, H. Pichler, and [93] C. Xu, J. Feng, M. Kawamura, Y. Yamaji, Y. Nahas,
M. D. Lukin, Parallel implementation of high-fidelity multi- S. Prokhorenko, Y. Qi, H. Xiang, and L. Bellaiche, Possible
qubit gates with neutral atoms, Phys. Rev. Lett. 123, 170503 Kitaev Quantum Spin Liquid State in 2D Materials with S = 3
(2019). / 2, Phys. Rev. Lett. 124, 087205 (2020).
[74] E. Urban, T. A. Johnson, T. Henage, L. Isenhower, D. D. [94] H. Takagi, T. Takayama, G. Jackeli, G. Khaliullin, and S. E.
Yavuz, T. G. Walker, and M. Saffman, Observation of Ryd- Nagler, Concept and realization of Kitaev quantum spin liq-
berg blockade between two atoms, Nat. Phys. 5, 110 (2009). uids, Nat. Rev. Phys. 1, 264 (2019).
[75] S. Jandura and G. Pupillo, Time-Optimal Two- and Three- [95] L. Bogani and W. Wernsdorfer, Molecular spintronics using
Qubit Gates for Rydberg Atoms, Quantum 6, 712 (2022). single-molecule magnets, Nature Materials 7, 179 (2008).
[76] A. Pagano, S. Weber, D. Jaschke, T. Pfau, F. Meinert, S. Mon- [96] D. N. Woodruff, R. E. P. Winpenny, and R. A. Lay, Lanthanide
tangero, and H. P. Büchler, Error budgeting for a controlled- Single-Molecule Magnets, Chem. Rev. 113, 5110 (2013).
phase gate with strontium-88 rydberg atoms, Phys. Rev. Res. [97] A. Anshu, N. P. Breuckmann, and C. Nirkhe, NLTS Hamilto-
4, 033019 (2022). nians from good quantum codes, in Proceedings of the 55th
[77] N. Khaneja, T. Reiss, C. Kehlet, T. Schulte-Herbrüggen, and Annual ACM Symposium on Theory of Computing (2023) pp.
S. J. Glaser, Optimal control of coupled spin dynamics: de- 1090–1096.
sign of NMR pulse sequences by gradient ascent algorithms, [98] T. Rakovszky and V. Khemani, The physics of (good) ldpc
J. Magn. Reson. 172, 296 (2005). codes i. gauging and dualities, arXiv:2310.16032 [quant-ph]
[78] A. Sørensen and K. Mølmer, Quantum computation with ions (2023).
in thermal motion, Phys. Rev. Lett. 82, 1971 (1999). [99] M. Motta and J. E. Rice, Emerging quantum computing algo-
[79] O. Katz, M. Cetina, and C. Monroe, N-body interactions be- rithms for quantum chemistry, Wiley Interdiscip. Rev. Com-
tween trapped ion qubits via spin-dependent squeezing, Phys. put. Mol. Sci. 12, e1580 (2022).
Rev. Lett. 129, 063603 (2022). [100] A. Celi, B. Vermersch, O. Viyuela, H. Pichler, M. D. Lukin,
[80] L. Lin and Y. Tong, Heisenberg-limited ground state energy and P. Zoller, Emerging Two-Dimensional Gauge Theories
estimation for early fault-tolerant quantum computers, PRX in Rydberg Configurable Arrays, Phys. Rev. X 10, 021057
Quantum 3, 010318 (2022), arXiv:2102.11340 [quant-ph]. (2020).
12
[101] T. V. Zache, D. González-Cuadra, and P. Zoller, Fermion-qudit [cond-mat, physics:quant-ph].

quantum processors for simulating lattice gauge theories with [118] I. Affleck, T. Kennedy, E. H. Lieb, and H. Tasaki, Rigorous re-
matter, Quantum 7, 1140 (2023). sults on valence-bond ground states in antiferromagnets, Phys-
[102] S. Ebadi, A. Keesling, M. Cain, T. T. Wang, H. Levine, D. Blu- ical Review Letters 59, 799 (1987).
vstein, G. Semeghini, A. Omran, J.-G. Liu, R. Samajdar, X.- [119] F. T. Arecchi, E. Courtens, R. Gilmore, and H. Thomas,
Z. Luo, B. Nash, X. Gao, B. Barak, E. Farhi, S. Sachdev, Atomic Coherent States in Quantum Optics, Physical Review
N. Gemelke, L. Zhou, S. Choi, H. Pichler, S.-T. Wang, A 6, 2211 (1972).
M. Greiner, V. Vuletić, and M. D. Lukin, Quantum optimiza- [120] L. H. Pedersen, N. M. Møller, and K. Mølmer, Fidelity of
tion of maximum independent set using Rydberg atom arrays, quantum operations, Phys. Lett. A 367, 47 (2007).
Science 376, 1209 (2022). [121] A. Elben, S. T. Flammia, H.-Y. Huang, R. Kueng, J. Preskill,
[103] D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, B. Vermersch, and P. Zoller, The randomized measurement
T. Manovitz, S. Ebadi, M. Cain, M. Kalinowski, D. Hangleiter, toolbox, Nature Reviews Physics 5, 9 (2022).
J. P. Bonilla Ataides, N. Maskara, I. Cong, X. Gao, P. S. Acknowledgments. We thank A. Aldossary, T. Betley, D.
Rodriguez, T. Karolyshyn, G. Semeghini, M. J. Gullans, Bluvstein, M. Cain, L. Cunha, J. Feldmeier, C. Kokail, J.
M. Greiner, V. Vuletic, and M. D. Lukin, Logical quantum
Lee, N. Leitao, S. Gant, J. Haber, S. Hollerith, T. Manovitz,
processor based on reconfigurable atom arrays, Submitted to
Nature. J. Neaton, S. Sachdev, A. Schuckert, K. Seethram, and
[104] J. Günther, A. Baiardi, M. Reiher, and M. Christandl, More P. Zoller for insightful discussions. This work was sup-
quantum chemistry with fewer qubits, arXiv:2308.16873 ported by the US Department of Energy [DE-SC0021013
[quant-ph] (2023). and DOE Quantum Systems Accelerator Center (contract no.
[105] C. Vorwerk, N. Sheng, M. Govoni, B. Huang, and G. Galli, 7568717)], the Defense Advanced Research Projects Agency
Quantum embedding theories to simulate condensed systems (grant no. W911NF2010021), the National Science Founda-
on quantum computers, Nature Computational Science 2, 424
tion (via CUA PFC, Q-IDEAS HDR, PHY-2207972), the De-
(2022).
[106] W. J. Huggins, B. A. O’Gorman, N. C. Rubin, D. R. Re- partment of Defense Multidisciplinary University Research
ichman, R. Babbush, and J. Lee, Unbiasing fermionic quan- Initiative (ARO MURI, grant no. W911NF2010082), Well-
tum Monte Carlo with a quantum computer, Nature 603, 416 come Leap under the Q for Bio program, the DARPA IM-
(2022). PAQT programme (grant number HR0011-23-3-0030) and
[107] M. Kalinowski, N. Maskara, and M. D. Lukin, Non-Abelian the Harvard-MIT Center for Ultracold Atoms. N.M. ac-
Floquet Spin Liquids in a Digital Rydberg Simulator, Phys. knowledges support from the Department of Energy Compu-
Rev. X 13, 031008 (2023).
tational Science Graduate Fellowship under Award Number
[108] M. S. J. Tepaske, D. J. Luitz, and D. Hahn, Optimal
compression of constrained quantum time evolution (2023), DE-SC0021110. A.M.G and R.A.B acknowledges support
arXiv:2311.06347 [cond-mat, physics:quant-ph]. from the NSF through the Graduate Research Fellowships
[109] Qiskit contributors, Qiskit: An open-source framework for Program, and A.M.G also acknowledges support through the
quantum computing (2023). Theodore H. Ashford Fellowships in the Sciences.
[110] M. Hermele, M. P. A. Fisher, and L. Balents, Pyrochlore pho-
tons: The U ( 1 ) spin liquid in a S = 1 2 three-dimensional
frustrated magnet, Physical Review B 69, 064404 (2004). APPENDIX
[111] K. Wan, M. Berta, and E. T. Campbell, Randomized Quantum
Algorithm for Statistical Phase Estimation, Physical Review
Letters 129, 030503 (2022). Hamiltonian Engineering
[112] I. Zintchenko and N. Wiebe, Randomized gap and am-
plitude estimation, Physical Review A 93, 062306 (2016), The Hamiltonian engineering toolbox introduced here is
arXiv:1603.03672 [quant-ph]. based on the average Hamiltonian approach. This approach
[113] R. Haghshenas, J. Gray, A. C. Potter, and G. K.-L. Chan, The
Variational Power of Quantum Circuit Tensor Networks, Phys-
uses the fact that in the high-frequency limit, the effective
ical Review X 12, 011047 (2022), arXiv:2107.01307 [cond- Floquet cycle period Kτ is much smaller than the inverse lo-
mat, physics:quant-ph]. cal energy scales of H(k), and the Floquet Hamiltonian HF
[114] M. Foss-Feig, D. Hayes, J. M. Dreiling, C. Figgatt, J. P. Gae- can be well-approximated by expanding in a small parame-
bler, S. A. Moses, J. M. Pino, and A. C. Potter, Holographic Kτ
ter ||H(k)||local
[62]. The leading contribution is the average
quantum algorithms for simulating correlated spin systems, Hamiltonian
Physical Review Research 3, 033002 (2021).
[115] J. R. McClean, J. Romero, R. Babbush, and A. Aspuru-Guzik,
The theory of variational hybrid quantum-classical algorithms, K−1
New Journal of Physics 18, 023023 (2016). (0) 1 X
HF = H(k). (10)
[116] A. Schuckert, A. Bohrdt, E. Crane, and M. Knap, Prob- K
k=0
ing finite-temperature observables in quantum simulators with
short-time dynamics (2022), arXiv:2206.01756 [cond-mat, The second order term also takes a simple form,
physics:quant-ph].
[117] M. Knap, A. Kantian, T. Giamarchi, I. Bloch, M. D. Lukin,
and E. Demler, Probing real-space and time resolved correla- (1) τ X
tion functions with many-body Ramsey interferometry, Phys- HF = [H(k), H(k ′ )] . (11)
2K ′
ical Review Letters 111, 147205 (2013), arXiv:1307.0006 k<k
13
The results presented in the main text involve engineering the corresponding to the different ways to change the expectation
(0)
average Hamiltonian HF to reproduce the target (1). In this value of 2 − Pi − Pj . n-spin interactions will have parts run-
setting, the second order term generates simulation errors, and ning from h(−n) to h(+n) . Therefore, in the rotating frame,
is order O(τ ). It can be cancelled by selecting time-reversal the Hamiltonian terms transform as
symmetric sequences of length 2K, where the second half of nX
max
the pulse is defined by Θk = ΘK−1−k . This reduces sim- h(k) = h(n) eiΘk n . (19)
ulation errors to order O(τ 2 ), but might potentially alter the n=−nmax
pre-thermal properties of the Floquet Hamiltonian, which is
an interesting problem for further research. We further choose a sequence of Θk ’s such that only the
Going beyond average Hamiltonian engineering is also pos- h(0) contribution is non-zero on average. The simplest se-
sible by optimizing the Floquet sequence [66, 107, 108]. We quence which satisfies these conditions are a family of cyclic
demonstrate engineering of higher-order terms for a small sys- pulses of order P .
tem composed of two interacting spin-3/2’s, to controllably
engineer up to bi-cubic terms (Ŝi · Ŝj )3 using only two- and
2πi
three-qubit operations in Extended Data Fig. 7 and SM. Θk = k k = 0, ..., P − 1 (20)
P
which satisfy the cancellation condition as long as nmax < P .
High-spin Hamiltonian engineering with dynamical Floquet (0)
projection In the two-body case, of the terms which contribute to hij ,
(0) (0)
only ŝi,1 ·ŝj,1 acts non-trivially in the symmetric subspace, so
To compute the form of the interaction Hamiltonian in the we focus on this term. An explicit form can be computed by
rotated frame HI (k) = eiΘk HP HI e−iΘk HP as used in (5) decomposing ŝi,1 into its permutation-symmetric and orthog-
let’s consider a single spin-1/2 particle belonging to a spin-Si onal components,
cluster, and define projectors Pi = P [Ŝ2i ] onto the symmetric
space and Qi = 1 − Pi onto its complement. The spin-1/2 2S 2S
!
term can be split into four parts 1 Xi 2Si − 1 1 Xi
ŝi,1 = ŝi,a + ŝi,1 − ŝi,a′ .
2Si a=1 2Si 2Si ′
a =2
| {z } | {z }
ŝi,1 = Pi ŝi,1 Pi + Qi ŝi,1 Qi + Pi ŝi,1 Qi + Qi ŝi,1 Pi (12) symmetric non−symmetric
| {z } | {z } (21)
symmetric non−symmetric
(0)
where the first two terms preserve the on-site total spin, and Therefore, the symmetric part ŝi,1 = 2S1 i Ŝi is proportional to
the second two change the on-site total spin. Since HP acts the collective spin.
as 1 − Pi on the i-th spin, we label terms by how they change With this understanding, we can construct a spin-1/2 inter-
the expectation value of (1 − Pi ) action Hamiltonian HI that recovers (1) under projection, by
replacing each n-site high-spin interaction with an analagous
(0) spin-1/2 one. For example, for n = 2 the replacement pro-
ŝi,1 = Pi ŝi,1 Pi + Qi ŝi,1 Qi (13) ceeds as
(+1)
ŝi,1 = Qi ŝi,1 Pi (14)
(−1) αβ α β αβ α β
ŝi,1 = Pi ŝi,1 Qi (15) Jij Ŝi Ŝj → Jij ŝi,a ŝj,b
αβ αβ
This rule can be extended to higher-weight operators. For Jij = 4Si Sj Jij , (22)
example, two-spin interactions between clusters hij = ŝi,1 ·
(n) where intra-cluster indexes a, b encode which representative
ŝj,1 decompose into five parts hij , n = −2, −1, 0, 1, 2,
from spin’s i and j are used to generate the interaction. The
interaction strength is further boosted to Jij , to account for
(0) (0) (+1) (−1) (+1) (−1) 1
hij (k) = ŝi,1 · ŝj,1 + ŝi,1 · ŝj,1 + ŝi,1 · ŝj,1 (16) the 2S factor in (21). A straightforward calculation shows
that for higher-weight interactions (e.g. nmax ), the interaction
| {z }
(0)
hij
should also be boosted (e.g. Kijk = 8Si Sj Sk Kijk ) to recover
(+1) (0) (0) (+1) the target large-spin operator under projection.
+ ŝi,1 · ŝj,1 + ŝi,1 · ŝj,1 +h.c. (17)
| {z } In general, it may not be feasible to uniquely assign each
(+1)
hij spin-Si interaction in H to qubits in HI especially when a

(+1) (+1)
spin-Si is involved in more than 2S interactions. In this case,
+ ŝi,1 · ŝj,1 +h.c., (18) HI can be implemented by splitting it into a sequence of non-
| {z
(+2)
} overlapping groups HI,1 , ..., HI,D that approximate HI on
hij
average. Each sequence can handle up to D(2S) interactions
14
per spin, so if d is the interaction degree then we require a se- For comparison, we estimate gate counts for a decompo-
d
quence of length D = ⌈ 2S ⌉. While manual decompositions sition of US and UP into single and two qubit gates (see
sufficed for the models studied here, automated methods to Fig. 2c). Using the Qiskit transpiler [109], we find two-qubit
determine efficient decompositions will have to be developed decompositions into single qubit rotations and CPhase gates
for more complex systems. for UP . For US , the diagonal gates D(ϕ) found numerically
in the alternating ansatz only require two-qubit interactions,
and can be decomposed into n2 two-qubit CPhase gates, out-

Multi-qubit gates with Rydberg blockade performing the Qiskit result for n ≥ 3. Finally, we show in
SM that symmetric operations like US and UP can be imple-
The Rydberg Hamiltonian governing a cluster of N atoms mented with poly(n) two-qubit gates using ancilla qubits, by
is constructing an efficient MPO representation of arbitrary n-
spin operations.
The fidelity of the multi-qubit gates is subject to errors
Ωq (t) X Ωr (t) X
Hcluster = |1⟩i ⟨0| + |r⟩i ⟨1| + h.c. like spontaneous emission and dephasing due to a finite T2∗ ,
2 i
2 i as seen in current Rydberg gates [57]. These errors can be
+
X
Vij |r⟩i ⟨r| ⊗ |r⟩j ⟨r| (23) mitigated by improving the excitation schemes. Specifically,
i<j
single-photon schemes as used in as in Refs. [67, 68], avoid-
ing intermediate-state scattering, may enhance performance
where Ωq (t), Ωr (t) are complex valued driving fields. In the for bigger clusters. This is because the rate of decay from
blockade approximation, which is valid when Vij ≫ Ωr , the Rydberg state does not depend on cluster size, since the
there is at most one atom in state |r⟩. Therefore, at lead- number of Rydberg excitations is never larger than one.
ing order in Ωr /Vij , Hcluster is approximated by projecting
into the manifold of blockade consistent states. This produces
Estimating simulation time
an interacting model with an emergent permutation symme-
try (see SM). This symmetry allows us to write Hcluster in a
low-dimensional representation of the Hilbert space scaling as The simulation time is defined as the maximum evolution
O(N NS ) for a representation including the NS largest total- time, before which the typical error-per-qubit is below some
spin sectors (see SM). target threshold ϵ. We account for both coherent Hamilto-
Figure 2b shows optimization results for an alternating nian simulation errors and incoherent gate errors. For a sym-
ansatz with separate Ωr (t) and Ωq (t) applications (solid metrized sequence, we estimate the scaling of both contribu-
lines) versus a dual driving scheme with simultaneous field tions to be
control (dashed lines). For the alternating ansatz the op-
timization process begins with finding short sequences of gT
symmetric diagonal gates D(ϕ) and global single-qubit ro- εsim = (c2 τ 2 )2 T 2 , εgate = , (24)
τ
tations Q(θ) that combined realize US and UP (see SM).
While Q(θ) uses global Ωq (t) control, D(ϕ) involves multi- with target evolution time T and step-size τ . Here, the coef-
qubit interactions, and the pulse sequences to realize D(ϕ) ficient c2 depends on the detail of the Hamiltonian simulation
can be optimized through gradient ascent pulse engineering protocol, and can be estimated from numerics or the third or-
(GrAPE) [57, 75, 77]. Numerical optimization is feasible due der term in the Magnus expansion (see SM). The coefficient
to the manageable size of the low-dimensional basis. We find g measures the gate error probability per cycle, and is deter-
that the maximum gate time T ∗ (n) for generic phases (see mined by asuming each multi-qubit gate has a fixed probabil-
SM). Gate times in Fig. 2 are computed by multiplying T ∗ (n) ity of failure Tgate g0 which scales linearly in the time of the
with the shortest sequence length determined in the first step. gate. For simplicity, we utilize the estimated Tgate for large-
The operations Q(θ) and D(ϕ) can also be promoted to con- angle unitary UP with θ = π. Both c2 and g are estimated
trolled operations (see SM), as required for the controlled per- to grow extensively in system size. Thus, we work instead
turbation in Fig. 4a. with the intensive version of these quantities, c̃2 = c2 /L and
Dual driving gate profiles are directly optimized using g̃ = g/L.
GrAPE, with an added smoothness regularization to ensure Optimizing the step-size τ to minimize error (see SM), the
driving profiles can be implemented with available classical maximum evolution time scales as
controls (see SM and Extended data Fig. 7). Interestingly the
optimized fidelity is roughly system size independent, up to 22/3 ε5/6 /L
error rates around 10−3 . We select this fidelity threshold, and Topt = (25)
55/6 (c̃2 g̃ 2 )1/3
plot the resulting gate times in Fig. 2b (dotted lines). For more
stringent thresholds, such as 10−6 , we recover an approxi- for a target error rate ε.
mately linear dependence in n, producing gate times compa- In Fig. 3, we use this formula, along with numerical es-
rable to the alternating decomposition. timates for c2 in models (i) and (ii), and heuristic estimates
15
for models (iii) and (iv). We further estimate g using the si- Finally, we consider a model including up to bi-quartic
multaneous driving gate times of Fig. 2b, and select an error terms k ≤ 4. Here, we further consider using an eight-qubit
per Rabi cycle of g0 = 10−3 . The target error we select is multi-qubit gate to directly realize the interaction between two
ε5/6 /L = 0.1, which grows approximately extensively with large-spins (iv). This interaction naturally preserves the sym-
system size. metry, so dynamical projection is not needed in this case. As
We illustrate the benefits of our approach on four example such, the speedup comes entirely from the efficiency of the
Hamiltonians (Fig. 3). First we consider the spin-1/2 Heisen- optimized eight-qubit operation, compared against the large
berg model on a Kagome lattice (i). The standard two-qubit overhead associated with a two-qubit decomposition of an
decomposition involves applying a sequence of six-steps, each eight-qubit interaction.
applying a e−iθsi ·sj gate along an edge (i, j) (solid lines in
Fig. 3). However, using native three-qubit Heisenberg gates,
we group the interactions into upwards and downwards facing Many-body Spectroscopy
triangles (dotted lines Fig. 3), reducing the sequence to only
two-steps, improving c2 . The three-qubit gate also generates To complete our simulation framework, we also develop
interactions more efficiently, further improving g. Note that tools for resource-efficient readout of Hamiltonian properties.
four-qubit version of this scheme could simulate the spin-1/2 First, we illustrate how to compute two-time correlation func-
Heisenberg model on a pyrochlore lattice using K = 2 [110]. tions of the form
The second model (ii) are two interacting high-spins with
Si = 5/2, which interact via a mix of Heisenberg and
Dzyaloshinskii–Moriya (DM) interactions, CO,R (t) = ⟨S|O(t)R(0)|S⟩, (26)
αβ
X
Jij = Jδαβ + D ϵαβγ . using the circuit in Fig. 4a. The real and imaginary parts of
γ CO,R (t) are independently accessed by measuring the ancilla
in the X and Y basis, respectively [80, 81, 83, 111]. More
The conventional decomposition splits the 25 pairwise inter- concretely, consider the state of the system right before mea-
actions into five groups of five, which are applied sequentially. surement, including both the ancilla qubit and the system,
In contrast, there is a dynamical projection scheme for this
model with K = 2 (see SM). This improvement in Floquet
engineering efficiency compensates the overhead from intro- 1
|ψf ⟩ = √ (|0⟩ ⊗ U (t) |S⟩ + |1⟩ ⊗ U (t)R |S⟩) , (27)
ducing a five-qubit gate, leading to slightly higher accessible 2
simulation times. Scaling with Si is discussed in SM.
where U is the time-evolution operator. Measuring X ⊗ O or
The final two models (iii and iv) are composed of spin-2
Y ⊗ O results in
particles on a square lattice. In model (iii), nearest-neighbor
spins
P2 interact via ak Heisenberg and bi-quadratic interaction,
k=1 Jk (Si · Sj ) . The conventional Trotter approach re- 1
⟨S| R† O(t) |S⟩ + ⟨S| O(t)R |S⟩ , (28)

⟨X ⊗ O⟩ψf =
quires realizing many four-qubit interactions per spin-2 pro- 2
ducing very long Floquet periods. In contrast, dynamical pro- i
⟨S| R† O(t) |S⟩ − ⟨S| O(t)R |S⟩ , (29)

jection can be implemented by applying a single four-qubit ⟨Y ⊗ O⟩ψf =
2
gate per edge, resulting in smaller K (see SM). The multi-
qubit implementation is also signficantly more efficient, as which together gives the full complex-valued CO,R (t) by tak-
the two-qubit decomposition of a four-qubit interaction comes ing a linear combination of the two,
with large overhead, as illustrated in Fig. 2c. To highlight the
separate contribution from Floquet projection and multi-qubit
gates, we compute the effective simulation time for four cases CO,R = ⟨(X + iY ) ⊗ O⟩ψf . (30)
in the table below.
For observables O diagonal in the measurement basis, CO,R
Tsc × ||H||local Two-qubit gates Four-qubit gates can be efficiently estimated in parallel from snapshots. During
Trotterization 0.15 1.0 the i-th run, let µ(i) = {x, y} be the randomly sampled ancilla
Dynamical Projection 0.25 2.5 measurement basis, and a(i) = {0, 1}, and |b(i) ⟩ be the ancilla
and system measurement outcomes respectively. Then, the
TABLE I. The quantum simulation’s achievable coherence time Tsc
estimator can be written as
in the target Hamiltonian’s units ||H||local (see also Fig. 3) for four
different simulation schemes applied to model (iii). This comparison
illustrates the advantages of the techniques we propose here. Hamil- M
tonian engineering based on dynamical projection in combination 1 X
CO,R (t) = 2σ(µ(i) , a(i) )⟨b(i) |O|b(i) ⟩ (31)
with efficient, hardware-optimized interactions outperforms alterna- M i=1
tive approaches based on bare Trotter evolution of two- or multi-qubit
gates, as well as dynamical projection just using two-qubit gates. where σ is a function taking on the values
16
NN si si
Ps = i=1 σi , where s is a base-four string, and σ de-
notes the four Pauli operators I, X, Y, Z; this can be accessed
σ(x, 0) = +1, σ(x, 1) = −1, with randomized measurements (see SM), and has a normal-
σ(y, 0) = +i, σ(y, 1) = −i. (32) ization factor N (Ps ) = 2N .
and |b(i) ⟩ is the measured projected state.

A
These measurements can be used to compute the operator- Efficient estimation of DR (t) from snapshots
resolved density of states (8), which can be rewritten as fol-
lows One of the key advantages of our approach, is the ability
A
to compute many estimators DR (t) from one dataset, includ-
Z X ing complex operators such as projectors onto spin-sectors
A
D (ω) = dt eiωt Tr[A|n⟩e−iϵn t ⟨n|] A = PS . This enhances the sample efficiency and reduces
Z n simulation time requirements, which are crucial limited re-
iωt sources in quantum computation.
= dt e Tr[AU (t)], (33)
To estimate density-of-states in a sample efficient way, we
introduce a classical co-processing algorithm which utilizes
where we have replaced δ(ω − ϵn ) → dt ei(ω−ϵn )t . In prac-
R
knowledge of the prepared reference state. In particular, we
tice, we will sample evolution times t from a probability dis- discuss the estimation procedure for
tribution p(t),Rsuch that
R ∞the integral is normalized to one when
ω = ϵn , i. e., dt = −∞ p(t)dt = 1. To arrive at (9), we can X
replace the trace with an average over probe states [112]
A
DR (t) = ⟨R|AOs U (t)|S⟩ ⟨S|U † (t)Os U (t)|R⟩
s
| {z }| {z }
classical quantum
X
Tr[AU (t)] = ẼR∼R ⟨R|AU (t)|R⟩, (34) = e−iES t ⟨R|AOs |S⟩⟨S|Os (t)|R⟩ (37)
s
where ẼR∼R = Tr[1]ER∼R is a normalized expectation which becomes (9) after averaging over times t and perturba-
value, and Tr[1] is the dimensionality of the Hilbert space. tions R. In the second line, we assumed |S⟩ is a known eigen-
This is valid, as long as the ensemble forms a 2-design, state, so U (t)|S⟩ = e−iES t |S⟩, and the classical part becomes
equivalent to a zero-time correlation function. One challenge
1 is that the sum in (37) involves exponentially many observ-
ER∼R |R⟩⟨R| = , (35) ables. However, by using the estimator (31) this reduces to a
Tr[1]
sum over M snapshots. In particular, for the polarized refer-
Observables such as (34) can in principle be computed via ence state, and Pauli-X measurements, a simple calculation
a modified Hadamard test by applying controlled-time evolu- (see SM) shows that
tion (see Ref. [81]). Since time-evolution is generally the most
costly step, we avoid the overhead associated with controlled- M
evolution and instead utilize a reference state |S⟩ with simple 1 X −iES t N/2
A (t) =
DR e 2 σ(µ(i) , a(i) )⟨R|A|b(i) ⟩ (38)
time-evolution. In particular, we select an ensemble of ob- M i=1
servables Os such that
is an unbiased estimator. The variance of this estimator is in-
dependent of system size√ for appropriately chosen ensembles
1 ∗
X R, and scales as O(1/ M ). While CO (0) could instead
Os U (t)|S⟩⟨S|U † (t)Os = 1. (36) s ,AR
N (Os ) s be estimated by measuring the unevolved state, this would
produce an estimator that converges very slowly, leading to
The normalization factor N (Os ) depends on the choice of en- a sample complexity that is exponential in N . However, clas-
∗
semble. Then, we can insert this resolution of the identity into sical methods can compute CO s ,AR
(0) with no error, making
(34) to get (9). In Fig. 4 and Fig. 5, we consider the polar- the procedure much more sample-efficient. In the SM, we dis-
ized reference state |S⟩ = |0⟩⊗N which is an exact eigenstate cuss how such calculations can be efficiently performed for
of the Heisenberg Hamiltonian, NNand thesi ensemble of Pauli- any pair of reference and probe states with an efficient matrix
X operators Os = Xs = i=1 (Xi ) where s is an N - product state (MPS) description, and any observable A with a
bit string. This satisfies the condition, and has N (Xs ) = 1 matrix product operator (MPO) description. Further, we show
since Xs |0⟩⊗N is an orthonormal basis. For generic refer- that projectors PS onto spin-sectors can be written as MPOS
ence states |S⟩ prepared by applying S to |0⟩⊗N , the ensem- with poly(N ) bond-dimension, making evaluation of DPS (ω)
ble Os = SXs S † satisfies the condition, and can be mea- classically efficient and scalable. We further discuss how en-
sured by applying the inverse preparation circuit S † before tangled measurements between two systems could be used to
measuring in the X-basis. Lastly, an ensemble which is in- estimate these correlators efficiently, even when |S⟩ does not
dependent of the reference state, is the set of Pauli strings have a known classical description.
17
Thermal expectation values that for a spin-1 AFM chain a simple bond-dimension two
MPS can outperform product states by orders-of-magnitude
The operator-resolved density of states can be used to com- in ground-state estimation. While bond-dimension two states
pute thermal expectation values via [81] can be efficiently prepared with simple circuits of two qubit
gates, more general ansatze can also be efficiently realized us-
R −βω A ing the simulation techniques described here [113, 114]. Op-
e D (ω)dω timized ansatze could be further combined with importance
⟨A⟩β = R −βω 1 . (39)
e D (ω)dω sampling [115], to improve the sample-efficiency of comput-
ing finite-temperature or excited state properties (see SM).
For example, to compute the magnetic susceptibility, we sim- The simulation time Tmax will depend on the required spec-
ply select the operator A = β(S z )2 , where β = 1/T is the tral resolution, which does not scale with system size for a
inverse temperature. Interestingly, this method of estimating gapped eigenstate. However, the rate of spectral broadening
thermal expectation values is insensitive to uniform spectral depends sensitively on the weight of measured observables
broadening of each peak, due to a cancellation between the nu- (Fig. 8). When the reference state is high in energy, such as
merator and denominator (see SM). However, it is highly sen- the polarized state for an AFM chain, the relevant observables
sitive to noise at low ω, which is exponentially amplified by typically have extensive weight, requiring Tmax ∼ N to main-
e−βω . To address this, we estimate the SNR for each DA (ω) tain constant spectral resolution. In contrast, preparing a low-
independently, and zero-out all points with SNR below three energy reference state, such as the ground-state |S⟩ = |GS⟩,
times the average SNR. This potentially introduces some bias allows coupling to other low-energy states using low-weight
by eliminating peaks with low-signal, but ensures the effects operators. This results in a noise-resilient and system-size in-
of shot-noise are well controlled. dependent procedure (Fig. 9). We further note that ground-
state preparation can be approximate, which would result in
additional spectral broadening in the computation of DA (ω).
Noise Modelling While the spectral resolution requirements should also grow
as the gap shrinks, we have illustrated that operator-resolution
To quantify the effect of noise on the engineered time- can mitigate this in certain settings (e.g. Fig. 5 and Fig. 6). As
dynamics, we simulate a microscopic error model by apply- such, understanding the general capabilities of this approach
ing a local depolarizing channel with an error probability p is an interesting direction for continued research.
at each gate. This results in a decay of the obtained signals
A
for the correlator DR (t). The rate of the exponential decay
grows roughly linearly with the weight of the measured oper-
ators (see Extended Data Fig. 8). This scaling with operator OEC Hamiltonians
weight can be captured by instead applying a single depolar-
izing channel at the end of the time-evolution, with a per-site The two candidates for the closed S2 state of the oxygen
error probability of γt with an effective noise rate γ. This ef- evolving complexP(OEC) are parameterized with Heisenberg
fective γ also scales roughly linear as a function of the single- models H = − ij Jij Ŝi · Ŝj [27, 88]. The coupling con-
qubit error rate per gate p (see Extended Data Fig. 8). stants for the two cases used to generate the data presented
in Fig. 5 are summarized in Table II. The corresponding spin
sizes of the different sites are S1 = 3/2, S2 = 3/2, S3 = 3/2,
Scaling the approach S4 = 2.
Quantum simulations are constrained by the required num- J12 J13 J14 J23 J24 J34
ber of samples and the simulation time needed to reach a cer- S2 H-1b 30.5 12.9 4.5 36.5 1.3 -7.3
tain target accuracy. These factors are crucial for determining S2 H-2b 32.6 11.7 4.0 37.3 1.5 -2.6
the size of Hamiltonians which can be accessed for particular
quantum hardware. TABLE II. Exchange coupling constants in units of wavenumbers
(cm)−1 for the two different candidates for the S2 stage of the OEC
Focusing on a single gapped eigenstate we determine the
used in Fig. 5.
number of snapshots CM needed to distinguish a spectral peak
from noise (Fig. 10). The signal arises from the overlap of the
probe states with the target eigenstate. The noise is given by Additional information about eigenstates can be calculated
the variance of the estimator (38), and decays as ϵ ∼ M −1/2 . by choosing the operator A in the operator resolved density of
For certain ensembles of probe states, the variance can be states appropriately, and multiplying by a narrow “band-pass”
made system size independent (see SM). However, a random filter in Fourier space to isolate a small set of frequencies [81,
probe state will have exponentially vanishing overlap with any 116]. For example, we investigate the total spin of the cubane
specific eigenstate. One approach to mitigate this, is to ini- sub-unit, i. e., the three magnetic sites supported on opposite
tialize probe states with higher overlap. In Fig. 9b we show vertices of the cube, using A = (Ŝ1 + Ŝ2 + Ŝ3 )2 , and compute
18
DA (ω)
A(ω) = (40)
D1 (ω)
evaluated at energies ωn of the eigenstates, which can be

extract from peaks in the spectral functions (Extended data
fig. 11). We further use spin-projection to improve the esti-
mate in the presence of broadening. For example, if the peak
at ωn occurs in spin-sector PS , we insert the projector PS in
APS
the numerator and denominator, A(ωn ) = DDPS (ω (ωn )
n)
.
2D Heisenberg calculations
The square lattice Heisenberg calculation was performed on

a large (L × L) system, with Hamiltonian
X
H2D = −J Ŝr · Ŝr+x̂ + Ŝr · Ŝr+ŷ (41)
r
We measure the Green’s function G(r, t) = ⟨S|Xr (t)X0 |S⟩

2
from a polarized reference |S⟩⊗L [117]. Since S z is con-
served under the dynamics G(r, t) is classically simulated by
restricting it to the space containing the |S⟩ and single spin-
flip states Xr |S⟩, which has dimension L2 + 1. We evolve
under equally spaced times up to JTmax = 8, and select
L = 23 which is large enough such that G(r, t) vanishes
far from the boundaries. Therefore, by letting G(r, t) = 0
outside the simulated region, this provides a good approxi-
mation for the L → ∞ limit. As such, we define projectors
onto (unnormalized) plane-wave states Pk = |k⟩⟨k|, where
⟨k|Xr |S⟩ = e−ik·r . Then, DPk (ω) can be written as
Z X
Pk
D (ω) = dtei(ω−ω0 )t ⟨S|X0 Pk Xr |S⟩ ⟨S|Xr (t)X0 |S⟩
r
| {z }| {z }
eik·(r−0) G(r,t)
(42)
which reduces to the Fourier transform G(k, ω) when the en-

ergy of the polarized state ω0 is set to zero. Plotting this for
a continuous set of k and ω produces the spectral weight de-
picted in Fig. 6. This further shows that finite-size systems
are sufficient to simulate extended systems at finite evolution
times.
By computing the peak value of ω for each k in the vicin-
ity of the single-particle excitations, we estimate the disper-
sion relation ω(k) associated with a single spin-flip excitation.
Computations of the many-particle Green’s function could be
performed similarly by applying multi-site perturbations, to
extract finite-temperature properties and characterize the in-
teractions between the quasi-particles.
19
(a) 100
1
φ/π
10−3 0
Gate Error E
−1
−6
10 0.5
θ/π
0
−0.5
10−9
0 1 2 3 3.65
Time (2π/Ω)
10−12
0 5 10 15 20
Tgate (2π/Ω) ×10−3
(b) (c) 6
0.03
4
0.02
J2 coefficient
0.01 2
0.00 0
−0.01 −2
−0.02
−4
−0.03
−6
−0.4 −0.2 0.0 0.2 0.4
J1 coefficient
Extended Data – FIG. 7. Gate optimization procedure and variational Hamiltonian Engineering. (a) To find smooth gates we perform
the GrAPE optimization procedure in two steps. In the first step, we penalize rapid changes in the pulse profile by introducing an extra term
in the cost-function. In this case, the resulting relationship between ideal gate-error vs. time (yellow circles) saturates around 10−6 . For
the second step we initialize the search with the smooth gates found in the previous step, which are re-optimized by removing the smooth
penalty. This significantly reduces the ideal gate error (orange). The data are shown for a N = 4 qubit cluster. Since the first step already
confined the problem into a subspace of the search space with smooth gates, the resultant pulses also remain smooth after the second step.
On the right we show an example smoothened pulse profile for the hyperfine angle θ and the Rydberg phase ϕ (see Methods) with an ideal
gate error rate E = 10−3 . (b) Higher order interaction terms can be controllably engineered via a simple modification of the K = 6 Floquet
projection sequence. For example, reducing the second time-step and increasing the fourth time-step by the same amount, δ preserves the target
Hamiltonian at leading order. (c) By tuning τ and δ, for a system of two interacting spin-3/2’s, a very large family of coefficients J1 , J2 , J3 in
the general Hamiltonian HT = J1 (S1 · S2 ) + J2 (S1 · S2 )2 + J3 (S1 · S2 )3 can be engineered. In particular, the roughly horizontal gray lines
correspond to constant δ grid lines, and roughly vertical gray lines correspond to values with constant τ . We see that changing δ primarily
modifies J2 and J3 , while changing τ primarily modifies J1 , consistent with our analysis that δ picks out certain higher-order terms. We
further note that the especially interesting AKLT family, with J2 /J1 = 116/243 and J3 /J1 = 16/243 lies among the family of efficiently
engineerable interactions.
20
(a) (d) 4.0

0.02
1 Spin Flip Sz Spin Flip Sector
Phenomenological Noise Rate γ/|J|

1 4
A
DR
3.5
0.00 2 6
3.0 3 8
(b)
2.5
0.2 4 Spin Flips
A
DR
2.0
0.0
1.5
−0.2
(c)
0.004 8 Spin Flips A
DR,Floquet
1.0
A
DR,Floquet × exp(−γt)
A
A
DR,noise
DR
0.5
0.002
0.0
0 1 2 3 4 0.000 0.002 0.004 0.006 0.008 0.010
Time t Gate Error Probability p
A
Extended Data – FIG. 8. Gate-level noise simulations. (a)-(c) Time evolution of the correlations DR (t) for a polarized reference state
z
|S⟩ = |0⟩ and A chosen to be projectors onto different S sectors. This involves measuring operators of fixed weight. We see that the
simulation of gate-level noise modelled as a depolarizing channel with a gate error probability p = 0.001 (orange dash dotted curve) matches
well with the dynamics obtained by adding an additional phenomenological noise ∝ exp(−γt) with rate γ = (0.128, 0.251, 0.36) for
A
Nspin flip = (1, 4, 8) (solid yellow lines) to the loss-less time evolution DR,Floquet (green dashed). (d) Relationship between the optimal
phenomenological noise rate γ and the gate error probability p for different spin flip sectors. We see a roughly linear-relationship of the decay
rate γ with the gate error rate p and operator weight Sz . All data are obtained using time-steps Jτ = 0.05.
(a) (b)
|S = | ↑N
10−1 D=1
Ground-state Overlap
nflips = 1 0.6
P (ω)
nflips = 2 D=2
nflips = 6 spin-1
10−3
0.4
|S = |GS
10−1 nflips = 1
0.2
P (ω)
nflips = 2
nflips = 6
10−3
0.0
−15 −10 −5 0 5 10 2 8 14 20 26 32
Energy ω (cm −1
) Chain Length
Extended Data – FIG. 9. Noise susceptibility and eigenstate overlaps. (a) Density-of-states for a spin-1 AFM chain, computed from a
polarized reference state |S⟩ = |0⟩⊗N (top), and ground state reference |S⟩ = |GS⟩ (bottom). The spectrum is separated into sectors
distinguished by their operator weight from |S⟩. For the polarized state, these correspond to sectors with well-defined S z . For the ground-
state, each sector is orthogonalized with respect to lower-weight sectors. Each sector is phenomenologically broadened by e−γtnflips to
simulate the operator-weight dependence of decoherence. When computing low-energy properties, the ground-state reference is more robust
to noise, since the low-energy eigenstates can be reached with lower-weight operators. (b) The amplitude of the corresponding spectral peak
is determined by the eigenstate overlap. We analyze the ground-state overlap for an AFM spin-1 chain performing DMRG for low bond
dimensions (D = 1, 2) and find that it is significantly larger for bond-dimension D = 2 matrix product states (red diamonds) compared to
bond-dimension D = 1, i. e., mean-field states (green circles). Interestingly, the ground state overlap decays slower with the chain length
for D = 2 indicating that the fidelity density is large. This feature makes low bond dimension states a promising direction for efficient state
preparation within our scheme since they can efficiently be decomposed into short circuits.
21
(a) 4 (b)
1 1
0.4
P0 P3
Average σ(D(ω))
3 P3
0.2
D(ω)
2
0.0
1 −0.2
0 −0.4 2
3 4 5 6 10 103 104 105
Chain Length Number of Samples
Extended Data – FIG. 10. Shot noise scaling with system size and convergence with number of snapshots. (a) Numerically computed
standard deviations of the estimator (9) for the density of states DA (ω) of a spin-1 AFM chain, for different chain lengths and observables.
Here, we consider a polarized reference state |S⟩ and random single-qubit rotations R for the controlled-perturbation. For the bare density
of states the standard deviation slowly scales with the chain length (blue circles). For the projectors into the zero (gray diamonds) and three
(orange circles) spin-flip sectors the standard deviation is consistent with being independent of system size. (b) Convergence of estimator with
number of samples for √ two different observables. Shaded regions correspond to 2σ error bars around the mean, and decrease with the number
of measurements as 1/ M . Dark lines are running averages for a specific sampled dataset. The sample complexity is defined as the number
of samples needed such that the error bars around the mean do not include zero. After this point, spectral peaks can be reliably distinguished
from noise.
(a) (b) 4.6

DOS
Open cubane 6
Proj. DOS Closed cubane 4 4.5
Cubane Spin
Cubane Spin
ED
3 4.4
4 −370 −360
3.75
2
2 DOS 3.50
Proj. DOS
1 ED
3.25
−200 −180
−400 −200 0 200 400 −250 0 250
Energy ω (cm−1 )
−1 −1
Energy ω (cm ) Energy ω (cm )
2
Extended Data – FIG. 11. Additional observables calculated for OEC. By choosing A to be the local cubane spin S123 , and calculating
A
D (ω) at each of the peaks in the density of states, we can infer the local structure of spin correlations within each eigenstate. (a) The cubane
spin for each eigenstate for the open cubane configuration (S2 H-1a). This model has coefficients: J12 = −15.8, J13 = 1.9, J14 , J23 =
23.1, J24 = 1.9, J34 = −13.9 in cm−1 , and spin-sizes S1 = 2, S2 = S3 = S4 = 3/2. The triangles, circles and crosses are obtained using
three different methods. The crosses indicate the exact value obtained via exact diagonalization (ED), the green circles indicate the values
obtained via the spin resolved (projected) density of states (DOS), and the orange triangles are obtained by the bare DOS. The projected DOS,
which is able to resolve individual eigenstates, matches well with the ED result. In contrast the bare DOS, which has overlapping peaks (see
Fig. 5), does not capture the expectation values accurately. (b) Same as in (a) but for the closed cubane configuration S2 H-1b. This model has
S1 = S2 = S3 = 3/2 and S4 = 2. We see that for the lowest-lying cluster, all spins in the cubane subunit are approximately polarized with
a total spin value S = 9/2 = 4.5. In contrast, the second cluster of eigenstates seems to differ by a single spin flip as indicated by the value
S = 7/2 = 3.5.
1
Supplemental Material
HAMILTONIAN ENGINEERING
Variational spin-gate optimization.
An exciting route to efficiently generating more complex interactions is to go beyond average Hamiltonian engineering. In
particular, we introduce a p-dependent evolution time τp , and consider tuning the parameters τp and θp in the pulse sequence (4)
to control certain higher-order terms, with no additional gate overhead. As a concrete example (see also Extended data Fig. 7),
we consider two spin-3/2 particles, evolved by HI = 9J 4 ŝ1,1 · ŝ2,1 and a K = 6 Floquet sequence with parameters
Θ = (0, 2π/3, 4π/3, 4π/3, 2π/3, 0) (S1)

τ = (τ, τ − δ, τ, τ, τ + δ, τ ). (S2)
We find that this sequence naturally engineers an SU(2) symmetric Hamiltonian with Heisenberg interactions, but also bi-
quadratic and bi-cubic terms
H2,extended = J1 (Ŝ1 · Ŝ2 ) + J2 (Ŝ1 · Ŝ2 )2 + J3 (Ŝ1 · Ŝ2 )3 . (S3)
To characterize the values of J1 , J2 , and J3 which can be accurately realized, we simulate the action of UF on the state
ρsym which is maximally mixed in the encoded subspace, for various pairs of values τ, δ. Then, as long as long as the average
population leaving the subspace is below 10−4 , we project UF into the encoded subspace, compute the effective Hamiltonian
HF = −i log(UF ) by exact diagonalization, and fit the coefficients J1 , J2 , J3 by solving a set of linear equations. The resulting
family of engineerably coefficients forms a 2D surface in a 3D space, and is visualized in Extended data Fig. 7b. Interestingly,
the AKLT family of Hamiltonians [S118]
3 k
⃗i · S
⃗i+1 ,
XX
HAKLT,3/2 = JkAKLT S (S4)
i k=1
lies within the space of engineerable Hamiltonians. Here, the coefficients JkAKLT are chosen such that HAKLT,3/2 is equivalent
to the projection Hamiltonian HP on six qubits. In other words the groundstate of HAKLT,3/2 is the total spin-3 subspace, and
is separated by a constant gap to the other lower spin subspaces.
The procedure outlined above could also be applied to anisotropic Hamiltonians HI , to try and engineer the full-space of
two-spin operations. Extensions beyond pairwise interactions could also be considered, for example to efficiently realize three-
spin interactions as in Ref. [S107]. We note that, higher-order interactions in our approach emerge similarly to those in realistic
materials. In our case they arise from virtual total-spin fluctuations, whereas in a material they arise from virtual charge-
fluctuations.
Procedures for Multi-Qubit Gate Optimization
The Hamiltonian engineering approach described in this work stands out for its hardware efficiency, owing to the strategic
use of efficient multi-qubit gates. To perform optimization efficiently, we first introduce a low-dimensional basis for each
cluster, leveraging the permutation symmetry of the blockade interaction. Then, we illustrate two complementary approaches
to determing pulse profiles which realize the target gate. First, we show how to use the GrAPE algorithm to find smooth off-
diagonal gates where the pulses Ωq (t) and Ωr (t) are applied simultaneously. Second we describe the two-stage optimization
used to identify alternating sequences of diagonal multi-qubit interaction gates and single-qubit gates to generate the target
off-diagonal operations.
Low-dimensional construction of blockade cluster basis
To efficiently perform numerical optimization of the pulse sequences Ωq (t) and Ωr (t) (see below), we use a low-dimensional
representation of the Hilbert space, leveraging the permutation-symmetry of the time-dependent Rydberg Hamiltonian (23). In
the blockade approximation, states with more than one Rydberg excitation within the cluster are disallowed. When the Rydberg
Hamiltonian is projected into this manifold, it becomes an all-to-all interacting “PXP” model, where a spin-flip is only allowed
2
if all other
Qatoms are in the qubit (|0⟩, |1⟩) states. Define Pi = |1⟩i ⟨1| + |0⟩i ⟨0| to be the projector onto the qubit state at site i,
n
and P = i=1 Pi to be the global projector. Then, the constrained Hamiltonian can be written as
Ωq (t) + Ωr (t) +
Hproj = Σq + Σr P + h.c. (S5)
2 2
where we have defined collective raising and lowering operators Σ+ +
P P
q = i |0⟩i ⟨1| and Σr = i |r⟩i ⟨0|. This is a good approxi-
mation for the dynamics at leading order in Ωr /V , and has an emergent permutation symmetry. In particular, notice that Hproj
is invariant under exchange of any two sites. In general, the Hilbert space of an N -atom cluster has 3N states, and the blockade
projection further reduces this to an O(2N ) scaling. However, to efficiently perform numerical optimization for larger clusters,
we construct a basis whose dimension grows quadratically in N by leveraging the permutation symmetry of Hproj .
We start by defining a class of symmetric states which are fully specificed their occupation numbers |n1 , n0 , nr ⟩. Each state
is constructed by taking a symmetric superposition of all configurations with the corresponding occupations. For example, the
N = 2 triplet state is given by
√
|1, 1, 0⟩ = (|01⟩ + |10⟩)/ 2.
More generally, the states spanned by |n1 , n0 , 0⟩ form the degenerate ground space of Ŝ2 , and have total spin N/2 (where
N = n1 + n0 ). Excited states of S 2 live outside the symmetric space, and have total spin <N/2; thus, to form the lowest-lying
excited states, i.e., states with total spin N/2 − 1, we need to expand our basis.
In general, there are N − 1 independent irreducible representations (irrep) with S = N/2 − 1 [S119]. One irrep can be formed
by constructing states that are symmetric under permutation of spins i = 1, ..., N − 1, but orthogonal to the fully symmetric
state. A simple way to construct a basis spanning this S = N/2 − 1 irrep and the symmetric S = N/2 irrep is to take a tensor
product of symmetric spaces on N − 1 atoms and 1 atom. To do this, we split the atoms into two groups, A and B, and label
states by six numbers |nA A A B B B
1 , n0 , nr , n1 , n0 , nr ⟩, associated with the occupation of each group. Therefore, states must satisfy
n1 + n0 + nr = N − 1 and n1 + n0 + nr = 1. The blockade constraint is imposed by further requiring nA
A A A B B B B
r + nr ≤ 1. Then,
matrix elements of Hproj can be efficiently constructed in this manifold. More generally, we can construct a basis which spans
irreps of spin S = N/2 − k,S = N/2 − k + 1, to S = N/2, by splitting the system into N − k atoms and k atoms. Simulations
are still efficient, since the dimensionality scales as Θ(N 2 ) when k = ⌊N/2⌋.
Simultaneous drive – Smooth gates with GrAPE
The primary tool we utilize to identify fast, hardware-optimized gate operations is gradient ascent pulse engineering
(GrAPE) [S77], which has recently been utilized to develop high-fidelity diagonal gates based on Rydberg driving [S57, S75].
Here, we extend the optimization results to include off-diagonal gates based on simultaneous Rydberg Ωr (t) and hyperfine Ωq (t)
driving. The core of the algorithm involves approximating time-dependent evolution under H(t) with discrete, Trotterized time
evolution. The time-dependent profile we consider has two parameters, a phase on the Rydberg drive Ωr (t) = Ωr e−iϕ(t) , and
a hyperfine rotation angle θ(t) defined by Ωq (t) = ∂t θ(t). The action of the continuous pulse is approximated by discretizing
evolution of nT steps with step-size τ each, resulting in a unitary of the form
nT
+ − iϕ(kτ ) −
Σ+ +
eiθ(kτ )(Σq +Σq ) e−iτ (e r +h.c.) −iθ(kτ )(Σq +Σq )
Y
UF (θ, ϕ) = e . (S6)
k=1
Optimization is performed by fixing T = nT τ , and tuning discrete set of variational parameters θ(kτ ), ϕ(kτ ) by computing
the distance between UF and the target unitary Utarget . Since the initial and final states should be fully supported in the qubit
manifold, we evaluate the distance between Utarget and UF within this manifold. Then, we can measure the gate fidelity using
the expression
†
UF (θ, ϕ)|q⟩|2
P
Nq + | q ⟨q|Utarget
F({θ, ϕ}) = (S7)
Nq (Nq + 1)
where |q⟩ is an orthonormal basis for the qubit states and Nq is the dimensionality of the associated subspace [S120]. We note
that in our implementation we use the reduced symmetric basis described above, with k = ⌊N/2⌋, so the gate fidelity differs
slightly from a calculation in the full qubit space. In particular, it weights the symmetric subspace more heavily than lower-spin
subspaces. Nevertheless, the basis is complete, so if F = 1 in the reduced subspace, it also is unity in the full qubit subspace.
When we optimize dual-driving gates, we further add a regularization to the cost function to bias the solution towards smooth
gates. Unlike Refs. [S75], we initialize the optimization with random profiles to avoid local minima, so the smoothening
3
regularization is crucial for obtaining continuous pulse profiles which can realistically be implemented experimentally. The
combined cost function is therefore
Cλ ({θ, ϕ}) = (1 − F({θ, ϕ}))

" #
X ϕ(kτ ) − ϕ((k − 1)τ ) 2 θ(kτ ) − θ((k − 1)τ ) 2
+λ + . (S8)
2 2
k
Here, we penalize the derivative of ϕ and θ, which correspond to the detuning of the Rydberg drive and amplitude of the hyperfine
drive.
Next, we discuss a systematic procedure for calculating high-fidelity smooth gates (see Extended Data Fig. 7). To bias the
optimizer towards finding smooth gates, we first minimize the cost function Cλ with λ = 10−3 , and evaluate each profile by
calculating C0 . For sufficiently long evolution times T , we are able to reduce C0 to be about 10−3 , and see the gate is reasonably
smooth. However, the regularization introduces a small amount of bias. As such, we subsequently re-optimize θ, ϕ using C0 as
the cost function, and see the ideal fidelity improves dramatically but the final gates remain smooth (see Extended Data Fig. 7).
Intuitively, the first step where rapid changes in the pulse profile are penalized confines the subsequent second search into a
region of search space where gates are smooth.
Alternating Ansatz
A simpler but less time efficient method to realize the required multi-qubit gates involves applying Ωr (t) and Ωq (t) separately.
This alternating approach is a natural extension of the capabilities demonstrated in Ref. [S57], and further generalizes more
readily to other reconfigurable architectures [S53, S55].
We parameterize a generic symmetric diagonal operation, and a global qubit rotation, as
 
XN
D(ϕ) = exp −i ϕj Ẑj  (S9)
j=0
 
X
Q(θ) = exp −i θα Ŝ α  (S10)
α={x,y,z}
P
where Zj = i1 ,...,ij Zi1 Zi2 ...Zij is a linear combination of all weight-j Pauli-Z strings supported on the N qubits. In
particular, a diagonal gate is characterized by time-evolution under the Rydberg part of the Hamiltonian Ωr (t), such that all
qubit states return to qubit states at the end of the evolution, with a path-dependent phase. Different qubit states acquire different
phases, producing a phase gate. The intuition is that D(ϕ) parameterizes all possible diagonal gates which can be generated
by the all-to-all Rydberg blockade interaction. After removing the global phase ϕ0 , and the part which can be generated by
a single-qubit Hamiltonian ϕ1 , there are N − 1 remaining parametersP which characterize the interaction. The global qubit
2S
rotations Q(θ) generate single-qubit off-diagonal rotations since Ŝ α = a=1 ŝα a , but generate no interactions between qubits.
As such, alternately applying diagonal and off-diagonal operations enables generation of a large class of off-diagonal interaction
operations. To find sequences which realize US and UP , we again use GrAPE-style optimization. In particular, we parameterize
the unitary as
nseq
Y
Ualternating = Q(θk )D(ϕk ) (S11)
k=1
and variationally optimize θk , ϕk to try and match Ualternating to the target unitary. We identify the shortest sequences which are
capable of implementing the target unitary, and present the corresponding sequence lengths (nseq ) in Table I. We further note
nseq N = 2 N = 3 N = 4 N = 5 N = 6 N = 7
US 2 3 6 6 6 6
UP 2 3 7 7 10 10
TABLE I. Optimized sequence lengths nseq for target error below ∼10−3 .
that for US , the sequences only require using the two-qubit phase ϕ2 , likely because the evolution Hamiltonian Ŝ2 is composed
4
of two-qubit operators. We use this fact below to construct shorter two-qubit decompositions for US compared to UP . Similarly,
we find for all sequences (US and UP ), the odd-qubit phases ϕ3 , ϕ5 , ... can be set to zero.
In neutral-atom setups, the global qubit rotation Q(θ) can be realized with very high fidelity [S73]. As such, the gate errors
are dominated by the interaction step D(ϕ), which involves coupling to the Rydberg state. To minimize gate errors, we therefore
use GrAPE to identify Rydberg driving profiles Ωr (t) that can implement a generic symmetric diagonal gate D(ϕ). For a given
system size N , by randomly sampling target phases ϕ, and using GrAPE to optimize the profile independently for each setting,
we identify a threshold evolution time T ∗ (n) above which the GrAPE algorithm always reliably identifies a high-fidelity gate
(Fig. S1). These times, combined with nseq are used to compute the values in Fig. 2b.
Finally, we discuss how gates D(ϕ) can be conveniently promoted to controlled-D(ϕ) gates. A generic control gate can be
parameterized by
 
X n
1 + Z a
CD(ϕ) = exp −i ϕ j Zj  (S12)
2 j=0
1+Za
where 2 = |0⟩⟨0| is a projector onto the |0⟩ state. We see that CUdiag can be decomposed into two symmetric diagonal
gates.
 
n+1
−i X
U1 = exp  ϕj−1 Z̃j  (S13)
2 j=1
 
n
−i X
U2 = exp  (ϕj − ϕj−1 )Zj  (S14)
2 j=1
The first gate applies a symmetric diagonal operation on the ancilla and target qubits, where Z̃j is the linear combination
of weight-j Pauli-Z strings supported on the n + 1 ancilla and target qubits. This realizes the interactions terms with Za
on the ancilla site. However, it also creates additional unwanted terms within the target qubits. Therefore, the second term
simultaneously produces interactions terms with 1 on the ancilla site, and removes the additional terms from the first step, such
that CUdiag = U1 U2 . This diagonal gate can be combined with the optimized alternating ansatz to construct controlled-spin
operations CUS and CUP . Such operations can be used for controlled probe-state preparation in Fig. 2a.
Lastly, we note that the MPO construction for the symmetric spin projectors Ps , presented in the final section, could be turned
into an ancilla-mediated scalable construction for generating spin-gates using two-qubit gates. In particular, the constructed
MPO corresponds to a sequence of controlled-gates from the system qubits onto an ancilla register which effectively measures
the total-spin information. Then, spin-operations can be systematically generated by applying the MPO circuit between the
system and ancilla, applying a diagonal gate on the ancilla, and then undoing the MPO circuit [S11].
Heuristic Simulation Time Estimates
In this section, we provide more details on the heuristics used to estimate coherent simulation times in Fig. 3. The estimates
involve breaking down the error contribution into two parts: simulation error and gate error. Simulation errors arise from
imperfect Hamiltonian engineering, due to mismatch between the engineered Hamiltonian HF and the target Hamiltonian HT .
Gate errors arise from experimental imperfections which accumulate during implementation of few-qubit interactions. Our goal
is to study qualitatively how our Hamiltonian engineering techniques, including both novel Hamiltonian simulation schemes and
hardware-optimized multi-qubit operations, can improve the performance of a quantum simulation.
To characterize simulation errors, we look at the decay of the many-body fidelity between the state evolved under HF and
HT . For symmetrized pulse sequences the decay should exhibit a characteristic t2 and τ 4 dependence.
Z Z
dψ 1 − |⟨ψ|eiHT t e−iHF t |ψ⟩|2 = dψ t2 ⟨ψ|(HT − HF )2 |ψ⟩ − ⟨ψ|(HT − HF )|ψ⟩2

(S15)
Z h i
(2)
= dψ t2 ⟨ψ|(∆2 HF )|ψ⟩ (S16)
≈ (c2 τ 2 t)2 (S17)
where we have only kept leading contributions in t and τ , averaged over valid initial states |ψ⟩, and introduced a coefficient
c2 which captures the magnitude of the leading order error term in HF − HT . This coefficient depends on the details of the
5
10−1
ΩT ∗ (N )/2π
4
10−3 2
Gate Infidelity
10−5 0
2 3 4 5 6 7
Group Size N
10−7
10−9
10−11
10−13
0 2 4 6 8
Evolution Time ΩT /2π
Extended Data – FIG. S1. We determine the maximum evolution time T ∗ (n) required to implement arbitrary symmetric diagonal gates D(ϕ)
using the blockade interaction for clusters of N = 2, ..., 7 qubits. To do this, for each evolution time ΩT we randomly select a target rotation
angles ϕ, and use GrAPE to find phase profiles implementing D(ϕ). Above the critical times T ∗ (vertical dotted lines), the infidelity achieved
was always below 10−6 . Identified values for T ∗ (N ) are also presented in the inset.
Hamiltonian simulation protocol, and can be understood as arising from the third order term in the Magnus expansion
K k2
k1 X
τ2 X X

(2) δk1 ,k2 + δk2 ,k3
HF = 1− ([H(k1 ), [H(k2 ), H(k3 )]] + [[H(k1 ), H(k2 )] , H(k3 )]) . (S18)
K 2
k1 =1 k2 =1 k3 =1
In what follows, we characterize simulation errors via c2 , and focus on the effect of theKcycle length K on the value of c2 .
At a first glance, we observe that in (S18) there are K

3 terms with k 1 ̸
= k2 ̸
= k 3 , and 2 terms with k1 = k2 or k2 = k3 ,
K

and 1 terms where k1 = k2 = k3 . In general, the nested commutators contain a lot of additional structure which will affect
c2 , including large cancellations between individual terms. For example, the terms where k1 = k2 = k3 automatically vanish.
Further, for symmetrized pulse sequences, H(k)P = H(K −k +1), and many pairs of triple commutators alsoP vanish. Finally, the
fact that each H(k) can be decomposed as n H (n) e−iΘk n , paired with choose pulse sequences such that k e−iΘk n , implies
many of the phases will interfere destructively. A more detailed analysis of the Floquet engineering would take these structures
into account explicitly, but is beyond the scope of this work. Nevertheless, the structure of the third-order term suggests that
c2 scales as most as O(K 2 ), but also may have a large O(K) contribution. As such, we will utilize numerical simulations to
determine a heuristic relationship between K and c2 .
To characterize the contribution from incoherent gate errors, we assume the error is accumulated linearly in time. If we let g be
the typical gate error per cycle, then the total incoherent error accumulated during simulation to time T is approximately gT /τ .
We neglect coherent gate errors, which will grow as T 2 , for simplicity. Our estimates are based on using hardware-optimized
operations for the Rydberg platform, so we estimate g based on the dual-driving gate time from Fig. 2b, and assume an error rate
g0 = 10−3 per physical Rabi time ΩT = 2π.
When combined with the Floquet error, sourced by c2 , the total error of a simulation can be written as
gT
ε = (c2 τ 2 )2 T 2 + (S19)
τ
There is a tradeoff between these two error sources, which can be tuned by adjusted the step-size τ . Minimizing the total error ε
for a fixed evolution time T , we see the optimal step-size is
1/5
g
τopt = (S20)
4T c22
which results in a minimum error of
1/5
c2 g 4 T 6

εopt = 5 (S21)
24
6
(a) Spin-1 Spin-5/2 (b) Kagome Heisenberg
Short-time error (1 − F (t))/t2
Short-time error (1 − F (t))/t2

10−1 N = 6, 3Q gates
−3 N = 6, 2Q gates
10
N = 12, 3Q gates
10−3 N = 12, 2Q gates
10−5
−5
10
S = 1, Proj.
S = 1, Trot. 10−7
10−7 S = 5/2, Proj.
S = 5/2, Trot.
10−9
10−1 0.6 0.7 0.8 0.9 1.0 1.1 1.2
Step-step τ Step-size τ (×10−2 )
Extended Data – FIG. S2. Visualization of procedure used to fit c2 from numerical simulations, for two interacting spins (a) and the
Kagome Heisenberg model (b). In (a), we compare a Trotterization approach to one using dynamical projection and multi-qubit gates.
For an ensemble of randomly chosen initial states in the symmetric subspace, we time evolve to a target t and compute the average fidelity
†
F (t) = |⟨ψ|Uideal UF (t)|ψ⟩|2 . We further vary the stepsize τ , and observe the characteristic τ 4 scaling associated with symmetrization, which
2
eliminates the τ term. The overall coefficient c2 is also extracted from the same fits.
or equivalently, in a maximum evolution time given a particular error tolerance
22/3 ϵ5/6
Tmax = (S22)
55/6 (cg 2 )1/3
Notice that both c2 and g are extensive quantities, growing linearly in the number of qubits N . As such, we will typically
work with intensive quantities, c̃2 = c/N and g̃ = g/N , and absorb the N dependence into the target error tolerance
22/3 ϵ5/6 /L
Tmax = . (S23)
55/6 (c̃2 g̃ 2 )1/3
This corresponds to selecting an error which grows (almost) extensively as well, i.e. a (nearly) constant error density.
Two interacting spin-S’s
We next describe how c2 and g are estimated heuristically for the fourPmodels studied in Fig. 3. We start by simulating the two
spin-S model with generic two-body interactions described by H2 = αβ J αβ Ŝ1α Ŝ2β . In particular, we consider a mixture of
Heisenberg and DM interactions, J αβ = Jδ αβ + D γ ϵαβγ . The c2 coefficient is fitted numerically from noise-free dynamics,
P
for both a two-qubit Trotter and multi-qubit dynamical projection approach. In particular, we simulate ideal and simulated
Hamiltonian dynamics from random product states |ψ⟩ in the symmetric subspace, and compute the error at short times t. Then,
we extract the coefficient c2 τ 2 by averaging over a small time window, and fit c2 by varying τ (see Fig. S2).
In the Trotter approach, the Hamiltonian simulation protocol involves applying a sequence of 2S Hamiltonians, each of which
act pairwise between the two clusters.
2S
αβ α β
J¯ij
XX
(H2,Trotter )i = ŝi,a ŝj,a+i mod(2S) (S24)
αβ a=1
This decomposition ensures every pair of constituent qubits between the two clusters interacts exactly once during the 2S step-
sequence. Note that we say this sequence has K = 2S, even though the symmetrized sequence we simulate actually has period
2K.
7
In the dynamical projection approach, we select H2,I = (H2,Trotter )0 to be the first pairing of the cluster of qubits. It turns
(±1)
out, for this problem, a K = 2 sequence is sufficient to symmetrize H2,I . The key observation, is to notice that H2,I = 0, due
to the specific structure of the Hamiltonian. Decomposing into all terms we get,
2S
(0) β (0)
J¯αβ ŝα
X
H2,I (k) = 1,a ŝ2,a
a=1
(+1) β (0) iΘ (0) β (+1) iΘ
+ ŝα
1,a ŝ2,a e k
+ ŝα
1,a ŝ2,a ek
(−1) β (0) −iΘ (0) β (−1) −iΘ

+ ŝα
1,a ŝ2,a e k
+ ŝα
1,a ŝ2,a e k
(1) β (1) 2iΘk (−1) β (−1) −2iΘk

+ ŝα
1,a ŝ2,a e + ŝα
1,a ŝ2,a e . (S25)
(0)
However, notice that ŝβj,a = 2S 1
Sj , and a ŝα (±n)
P
1,a = 0 for any n ̸= 0 due to symmetry. As such, the terms which change
HP by ±1 vanish, leaving only the ±2 terms. This means the cancellation condition can be satisfied using the two-step sequence
Θk = (0, π/2).
The fitted values of c2 from numerical simulations are presented in Table. II. In particular, we present an intensive version
c˜2 which is normalized by the total number of qubits in the system. Furthermore, note that the value of c2 depends on the
characteristic energy scale of the Hamiltonian. To handle this systematically, we scale the simulation time by a simple measure
of the local Hamiltonian strength |H|local . This is computed by simply adding the number of Pauli operators which intersect with
any given qubit i weighted by the magnitude of their coefficients, and dividing by the weight of each operator. For example, for
H2 , there are 2S × 3 two-body Pauli operators with coefficient J, and 2S × 6 two-body operators with coefficient D. Therefore,
|H|local = 6S(J + 2D)/2 = 9SJ for D = J.
c̃2 S = 1 S = 3/2 S = 2 S = 5/2

Trotter 0.135 0.215 0.285 0.397
Dynamical Projection 0.12 0.17 0.17 0.18
TABLE II. Numerically estimated values for the normalized, intensive coefficients c̃2 , for two interacting large spin S’s. The Trotter error is
consistent with roughly linear growth, arising the growth of the sequence length K = 2S. In contrast, dynamical projection appears consistent
with a constant c̃2 , matching the constant sequence length K = 2.
The computation of g involves counting the total number of gate per step. In the Trotter approach, there are 2S two-qubit
interactions. Because each interaction involves a symmetric and anti-symmetric part, it needs to be further decomposed, e.g.
into a sequence of symmetric gates and single-qubit rotations. In particular, a three step sequence of symmetric interactions
h1 = J(sx1 sx2 + sy1 sy2 + sz1 sz2 ), h2 = D(sx1 sy2 + sy1 sz2 + sz1 sx2 ), and h3 = D(sx1 sz2 + sy1 sx2 + sz1 sy2 ) can realize the target
Hamiltonian on average. Importantly, h2 and h3 can be brought into the same form as h1 by applying a single-qubit rotation
on the second site around the sx + sy + sz axis. Note that, while this introduces additional simulation error at higher-order,
variational gate optimization as in the alternating ansatz could be used to identify higher-fidelity local sequences. As such, each
cycle in the Trotter approach utilizes 2S × 3 two-qubit symmetric gates. In the dynamical projection approach, there are still
2S × 3 two-qubit gates per cycle, but also two additional 2S-qubit gates to apply HP . However, in practice some pairs of gates
applied during separate steps can be merged, such as when the interaction frame Θk = Θk+1 is the same. In symmetrized
sequences, this happens during the 1st and K-th gates in the 2K cycle. As such, we multiply g by K−1 K to account for this
reduction. The intensive quantity g̃ is presented in Table. III.
Gate error g̃ S = 1 S = 3/2 S = 2 S = 5/2

Trotter 3.1 4.1 4.7 5.0
TABLE III. Numerically estimated values for the gate error per cycle (×10−3 ), per qubit, g̃. Note that despite the overhead, the projection
approach becomes more efficient than Trotter due to the ability to combine gates during the 1st and K-th cycle. Further, the projection overhead
decreases with 2S since this quantity is intensive.
Since K is constant for dynamical projection, and grows with 2S for Trotterization, the dynamical projection approach
becomes more favorable as the spin-size grows. Combining the estimated values for c2 and g produces the simulation times in
Table IV.
8
Sim. Times (×|H|local ) S = 1 S = 3/2 S = 2 S = 5/2

Trotter 5.6 3.9 3.3 2.8
TABLE IV. Numerically estimated simulation times, for an (approximately) per-qubit error tolerance of ϵ = 0.1, in units of |H|local .
Kagome Heisenberg
Next, we study the Kagome Heisenberg model. Since this model has no large-spins, there is no need for the projection
approach. However we can still utilize multi-qubit gates to improve the simulation. To illustrate this, we consider two cases, one
with two-qubit gates and another with three-qubit gates. In the two-qubit case, we consider a K = 6 protocol, where two-qubit
ŝi ·ŝj gates are applied during each step. In the three-qubit case, we develop a K = 2 protocol, applying three-qubit Ŝ2 evolution
along upwards-facing triangles and downwards-facing triangles in alternating fashion.
We follow the same fitting procedure for c2 with N = 6 and N = 12 size systems with periodic boundary conditions, and
again estimate g using the gate-times from Fig. 2b. The N = 12 values are used for the estimates in Fig. 3.
N c̃2 g̃ Tmax
N = 6, two-qubit gates 2.6 1.7 3.1
N = 6, three-qubit gates 0.25 1.7 6.9
N = 12, two-qubit gates 3.5 1.7 2.8
N = 12, three-qubit gates 0.36 1.7 6.1
TABLE V. Parameters used to estimate effective simulation time for the Kagome Heisenberg
We notice that the c̃2 coefficients are signficantly higher for the K = 6 protocol compared to the similar K = 5 spin-5/2
value. This discrepancy could be partly due to the fact that the local Hamiltonian norm is overestimated in the two spin-5/2
case, which has more frustrated (non-commuting) interaction terms. In particular, under a time rescale t → λt, the c2 coefficient
varies by λ3 , making its specific value quite sensitive to the choice of units. Nevertheless, the overall trend is clear, shortening
the sequence appears to accumulate simulation errors more slowly.
Spin-2 models
Finally, we use this approach to present heuristic estimates of the simulation time for a representative complex spin-model.
In particular, we consider a system of spin-2’s on a square lattice, with pairwise higher-order terms (Si · Sj )k , making the
interactions quite complicated.
n X
X Jn
Hsq,n = (Si · Sj )k (S26)
(2S)n−1
k=1 ⟨ij⟩
One of the key limitations towards realizing this model with the conventional two-qubit Trotter approach stems from the fact
that each term (Si · Sj )k decomposes into a sum of 2k-qubit interactions. Further, decomposing higher-order interactions into
two-qubit gates comes with large overhead, as depicted in Fig. 2c. The second is the the rapid growth of the number of terms
2
with the order of the interaction. For example, since each S = 2 site has four qubits, there are 42 = 36 groups of four-qubit
interactions coupling adjacent sites i and j in (Si · Sj )2 . Both of these problems are mitigated in the multi-qubit approach.
In particular, the hardware-level optimization introduces native multi-qubit gates to directly realize higher-qubit operations.
Further, dynamical projection realizes the multitude of terms using simpler operations.
We first consider the H2 model with Heisenberg k = 1 and bi-quadratic k = 2 interactions. The Trotter version involves
cycling through the 144 four-qubit interactions, between each spin and its four neighbors. Since each interaction acts on two-
qubits per cluster, this can be parallelized into a cycle of period K = 72. In contrast, the dynamical projection approach
alternates between vertical interactions, horizontal interactions, and an intra-cluster interaction (HP ). Since nmax = 2, this
needs to be cycled three times, leading to K = 6. In order to get a heuristic estimate for c2 , we extrapolate a linear-scaling, and
simply approximate c̃2 ∼ K, which is within an order-of-magnitude of the estimates from simulation above.
Decomposition of the four-qubit spin interaction into two-qubit operations requires roughly 100 gates, resulting in a gate-error
per qubit per step of 100g0 /4 = 25g0 . In contrast, the large-angle four-qubit simultaneous driving gate takes time 3.74g0 . Since
there are two interaction gates for every intra-cluster gate, this results in a gate-error per qubit per step of 3.74 ∗ 1.5g0 /4 =
1.395g0 . Note for simplicity, we neglect the additional K−1K factor arising from combining gates.
9
Putting this together, our estimated simulation time is Tmax = 0.17 for the conventional, two-qubit Trotterization approach,
and Tmax = 2.7 for the multi-qubit dynamical projection approach. We can also isolate the relative contribution from multi-
qubit gates and dynamical projection. The multi-qubit Trotter approach, (using c2 = 77 and g = 1.395g0 ) is estimated to be
Tmax = 1.2. The two-qubit, dynamical projection approach, (using c2 = 77 and g = 1.395g0 ) is estimated to have Tmax = 0.4.
Finally, we consider a model with up to bi-quartic k = 4 interactions. Here, the k = 4 terms act on all eight qubits between
two clusters, so no dynamical projection is needed. Instead, all of the advantage comes from the efficiency of the multi-qubit
approach. The trotterization period is K = 4, and the two-qubit decomposition estimates suggests over 104 gates are needed
to realize the k = 4 interaction, so we plug in g ∼ 104 g0 /8 ∼ 103 g0 . In contrast, the eight-qubit gate times is ΩT /2π = 5.0,
producing g = 5.0g0 /8. Again using c2 = K heuristically, this results in estimated evolution times of 0.03 and 5.3 respectively.
MANY-BODY SPECTROSCOPY
The algorithmic framework we develop utilizes classical co-processing to extract detailed spectral information from time-
evolution and snapshot measurements in a resource-efficient way. In the following sections, we discuss the following key aspects.
First, we describe how measurements from quantum simulation (Fig. 4a) enable parallel measurement of exponentially many
two-time correlation functions of the form CO,R (t). Next, we explain how from M snapshots, and for simple reference states
|S⟩, this information is in principle sufficient to efficiently reconstruct any linear functional of the time-evolved probe states
U (t)|R⟩. We show how this enables efficient calculation of the density-of-states, and how the convergence of the estimator can
be system size independent for appropriately chosen ensembles. Then, we discuss the estimation of thermal expectation values,
and finally present a poly(N ) MPO construction for computing spin-projected DOS, enabling efficient classical evaluation of
DPS (ω).
Estimator (31) for CO,R (t) from snapshots
The core quantities which enable access to spectral properties of the Hamiltonian are correlation functions of the form
CO,R (t) = ⟨S|O(t)R(0)|S⟩, For a given evolution time t and perturbation R, these can be efficiently obtained using the
circuit in Fig. 4a. Further assume the system qubits are measured in a fixed basis, such as the X-basis, and O is diagonal in this
basis. Then, CO,R can be efficiently estimated in parallel from the resulting snapshot measurements. During the i-th run of the
experiment, let µ(i) = {x, y} be the randomly selected ancilla measurement basis each with probability 1/2, and a(i) = {0, 1},
and |b(i) ⟩ be the ancilla and system measurement outcomes respectively. Then, as presented in (31), the estimator can be written
as
M
1 X
CO,R (t) = 2σ(µ(i) , a(i) )⟨b(i) |O|b(i) ⟩. (S27)
M i=1
To study the performance of this estimator, first we show that it is unbiased, in the sense that it’s expectation value matches
the target correlation function. Then, we compute it’s variance, which determines how quickly the estimator converges to it’s
expectation value.
To show the estimator is unbiased, we need to compute the probability that during any given run, the output of the simulation
is µ, a, b. The conditional measurement probabilities P (a, b|µ) can be straightforwardly computed using the form of of |ψf ⟩.
1 1
P (0, b|x) = ⟨1 ⊗ Πb ⟩ψf + ⟨X ⊗ Πb ⟩ψf (S28)
2 2
1 1
P (1, b|x) = ⟨1 ⊗ Πb ⟩ψf − ⟨X ⊗ Πb ⟩ψf (S29)
2 2
1 1
P (0, b|y) = ⟨1 ⊗ Πb ⟩ψf + ⟨Y ⊗ Πb ⟩ψf (S30)
2 2
1 1
P (1, b|y) = ⟨1 ⊗ Πb ⟩ψf + ⟨Y ⊗ Πb ⟩ψf (S31)
2 2
where Πb = |b⟩⟨b| is the projector onto the measured state. For observables O which are diagonal in the X-basis, it can be
uniquely decomposed in the operator basis spanned by the projectors Πb ,
X
O= Πb ⟨b|O|b⟩. (S32)
b
10
Therefore, we see that the expectation value of the estimator is equal to

XX
E [2σ(µ, a)⟨b|O|b⟩] = (P (a, b|x)P (x)2σ(x, a) + P (a, b|y)P (y)2σ(y, a)) ⟨b|O|b⟩ (S33)
a b

1 1 1 1
= ⟨1 ⊗ Πb ⟩ψf + ⟨X ⊗ O⟩ψf − ⟨1 ⊗ Πb ⟩ψf − ⟨X ⊗ O⟩ψf (S34)
2 2 2 2

1 1 1 1
+i ⟨1 ⊗ Πb ⟩ψf + ⟨Y ⊗ O⟩ψf − i ⟨1 ⊗ Πb ⟩ψf − ⟨Y ⊗ O⟩ψf (S35)
2 2 2 2
= ⟨(X + iY ) ⊗ O⟩ψf = CO,R (S36)
as desired. Note that the factor of two in the coefficient 2σ(µ, a) accounts for the fact that each basis is only sampled half the
time, cancelling the contribution from P (µ).
The variance of the estimator can be computed by performing similar calculations. Formally, the second moment of the
estimator is given
E |2σ(µ, a)⟨b|O|b⟩|2 = E 4|⟨b|O|b⟩|2

(S37)
X
=4 ⟨Πb ⟩ψf |⟨b|O|b⟩|2 (S38)
b
X
=2 (⟨R(t)|b⟩⟨b|R(t)⟩ + ⟨S(t)|b⟩ ⟨b|S(t)⟩)|⟨b|O|b⟩|2 (S39)
b
Here, we used the fact that |σ(µ, a)|2 = 1 for all choices of µ and a; thus, the properties of the ancilla drop out of the expression.
Finally, subtracting the expected value gives an expression for the variance.
Beyond diagonal operators
When the operator O is not necessarily diagonal in the measurement basis, estimating CO,R (t) will require varying the
measurement basis of the system qubits as well. State-of-the-art shadow tomography methods could be utilized for this pur-
pose [S84, S121]. We discuss here the simplest case, random Pauli measurements, where the joint ancilla and system measure-
ment basis is µν = µν1 ...νn . As before, µ determines whether the ancilla is measured in X or Y , while ν1 , ..., νn determines
whether each systemP qubit is measured in X, Y , or Z. An operator O can be decomposed into a linear combination of Pauli
operators, O = P O(P )P , where P = P1 ...Pn is a product of Pauli operators, and O(P ) = tr(OP ). Furthermore, define
|b(i) ⟩ as the system state measured during the i-th run, which depends on the choice of basis setting ν (i) and measured bitstring.
Then, the estimator CO,R (t) can be constructed as
M
1 X X
CO,R (t) ≃ 2σ(µ(i) , a(i) ) 3|P | O(P )⟨b(i) |P |b(i) ⟩ (S40)
M i=1
P
The additional factor of 3|P | comes from inverting the probability that the measurement basis is aligned with the Pauli operator P ,
where |P | is the number of sites non-identity sites [S84]. Indeed, if the measurement basis is not aligned, then ⟨b(i) |P |b(i) ⟩ = 0
instead of ±1, and so no information about the expectation value of P is learned.
Interpretation as shadow tomography of |R(t)⟩ from single-basis measurements
Snapshot measurements enable construction of a classical representation, i.e., a classical shadow [S84], of |R(t)⟩ = U (t)R|S⟩,
from which linear functionals (such as wavefunction amplitudes ⟨ϕ|R(t)⟩) can be efficiently estimated. The amplitude used for
the DOS calculation ⟨R|R(t)⟩ is just one such example.
To illustrate this formalism, let us still again consider the case where |S⟩ = |0⟩⊗N , and snapshot measurements of the system
⊗N
are performed in the X-basis, where all Pauli strings are enumerated as Xs . Then, {Xs |0⟩ , s ∈ {0, ..., 2N − 1}} forms a
complete, orthonormal basis for the Hilbert space and we can again utilize the resolution of the identity in terms of Xs acting on
a classical state,
X
1= Xs |0N ⟩⟨0N |Xs , (S41)
s
11
to rewrite |R(t)⟩ as a linear combination correlation functions CXs ,R (t), which can be measured simultaneously for each s,
X
|R(t)⟩ = Xs |0N ⟩⟨0N |Xs |R(t)⟩
s
X
= Xs |0N ⟩CXs ,R (t)e−iE0 t , (S42)
s
⊗N
where we additionally use the fact that |0⟩ is an eigenstate of the Hamiltonian with energy E0 and does not evolve in time.
We can use the estimator of CXs ,R (t) from Eq. (31) to form an estimator for linear functionals of |R(t)⟩, which makes Eq. (S42)
analagous to a classical shadow. The classical shadow is an object in Ref. [S84], composed of snapshot measurements, which
replicates the properties of a density matrix ρ as the number of measurements M → ∞ tends to infinity. The verison here, tends
to the wavefunction |R(t)⟩, and is hence estimators for wavefunction amplitudes ⟨ϕ|R(t)⟩ can be written as
M
e−iE0 t X
⟨ϕ|R(t)⟩ = 2σ(µ(i) , a(i) )2N/2 ⟨ϕ|b(i) ⟩ (S43)
M i=1
where |b(i) ⟩ is the i-th measured X-basis bitstring. To be able to efficiently evaluate this expression, it is important that |ϕ⟩ has
an efficient classical representation, so that the coefficient ⟨ϕ|b(i) ⟩ are computable classically. Note that, while any wavefunction
amplitude can be efficiently computed.
Unlike the original versions of shadow tomography [S84], here we utilize partial knowledge of the measured state (namely
the form of |S⟩) in order to extract all of the unknown information (functionals of |R(t)⟩) from a single-basis measurement.
Further, our procedure is sensitive to global phase information in |R(t)⟩, since it works by performing interferometry relative to
a reference state |S⟩. This phase information is crucial for estimation of spectral information via the Fourier-transformed state
|R(ω)⟩. In Ref. [S47], snapshot measurements were combined with time-series data, to extract a different Hamiltonian spectrum
that has peaks corresponding to energy differences ϵi − ϵj between eigenstates. Our method, utilizing the ancilla and reference
state, is capable of detecting individual energies ϵi and filtering their corresponding eigenstates [S81].
The operator-resolved DOS estimator
A
The key quantity required to evaluate the DOS estimator (9) is the calculation of DR defined in (37). One way of arriving at
A
the expression for the estimator, is by noticing that DR = ⟨R|A|R(t)⟩, which is equivalent to selecting |ϕ⟩ = A|R⟩ in (S42).
This estimator is also equivalent to computing CO,R for a non-hermitian operator O. Focusing for simplicity on the polarized
reference state |S⟩ = |0⟩⊗N = |0N ⟩, the observable is
X
A
O := OR = ⟨R|AXs |0N ⟩X̂s , (S44)
s
A
and can be written as a linear combination of Pauli-X operators. We can verify that CORA ,R reproduces DR up to the global
time-dependent phase shift,
X
CORA ,R (t) = ⟨R|AXs |0N ⟩⟨0N |X̂s (t)|R⟩ = eiE0 t DR
A
(t), (S45)
s
N
where we assumed that |0 ⟩ is an eigenstate of H with energy ES .
Despite the exponentially large sum over {s}, the density-of-states estimator can often be evaluated efficiently from snapshot
data, as long as the states |R⟩ and |S⟩ are chosen correctly. For the polarized reference state the diagonal matrix elements take a
simple form,
X
A
⟨b|OR |b⟩ = ⟨R|AXs |0N ⟩⟨b|X̂s |b⟩
s
X
= ⟨R|A (−1)b·s Xs |0N ⟩
s
= 2N/2 ⟨R|A|b⟩, (S46)

which is straightforward to evaluate for many choices of the probe state |R⟩ and operators A. A large class of states for which
this can be efficiently evaluated classically, is when |R⟩, |S⟩ are matrix product states (MPS) with poly(N ) bond-dimension,
and A is an matrix product operator (MPO) with poly(N ) bond-dimension. For example, the projectors onto different total
spin-sectors Ps we leverage in Fig. 4 and Fig. 5 can be written efficiently using the MPO formalism, as we discuss further below.
12
Convergence of the DOS estimator
To compute how quickly the DOS (A = 1) calculation will converge for polarized reference |S⟩ = |0⟩⊗N and Os = Xs , we
bound the variance of the estimator, and discuss the dependence on the ensemble of probe states. For simplicity, let us focus on
t = 0, where COR ,R (0) = 1 by construction. The second moment is equal to
X X
E|COR ,R |2 = 2N +1 (|⟨R|b⟩|2 + |⟨S|b⟩|2 )|⟨R|b⟩|2 = 2 + 2N +1 |⟨R|b⟩|4 , (S47)
b b
2 −N
where we used |⟨S|b⟩| = 2 . We evaluate this expression Q in a few relevant limits. First, we consider states |R⟩ which come
from applying random single-qubit rotations of the form i e−iηi Xi , where ηi encodes the random rotation angles. For this
choice, it is straightforward to show that ⟨R|b⟩⟨b|R⟩ = 2−N for every X-basis eigenstate |b⟩. Intuitively, each qubit of |R⟩ lives
in the Y Z-plane, and hence is orthogonal to the X-axis. Then, the second moment evaluates to
X
E|COR ,R |2 = 2 + 2N +1 |⟨R|b⟩|4 → 2 + 2N +1 × 2N × 2−2N = 4, (S48)
b
and the variance is Var(COR ,R ) = E[|COR ,R | ] − |E[COR ,R ]|2 = 3. More generally, states |R⟩ generated by evolving under a
2
Hamiltonian diagonal in the X basis will satisfy these results.

Next, we consider another limiting case, whereby |R⟩ is an N -qubit Haar random state. The Haar-averaged second moment
can be written as,
XZ
2 N +1
ER∈Haar E|COR ,R | = 2 + 2 dR |⟨R|(|b⟩ ⟨b|)|R⟩|2 , (S49)
b
which is expressed via a random-state average of the |b⟩ ⟨b| operator. We use the formula for Haar-random averages (Eq. (1) in
Ref. [S120]) with M = |b⟩ ⟨b|,
1
Z
dR |⟨R|M |R⟩|2 = N N Tr(M M † ) + |Tr(M )|2 ,

(S50)
2 (2 + 1)
which results in the variance,
1+1 4
Var(COR ,R )Haar = 2 + 2N +1 × 2N × −1=1+ , (S51)
2N (2N + 1) 1 + 2−N
which is of order 1 for large systems. Moreover, the above formula has a product structure, which allows us to also analyze the
case where the perturbation R is a product of locally Haar-random operations. Concretely, we can divide N qubits into n groups
of k = N/n qubits and assume that |R⟩ is a 2-design within each group (e.g. is a Clifford circuit). Then, we get
n n
Y
k
X
2 2
Var(COR ,R )local Haar = 2 + 2 2 |⟨Ri |bi ⟩⟨bi |Ri ⟩| − 1 = 1 + 2 , (S52)
i=1
1 + 2−k
bi
which reproduces the previous result for n = 1. Now, if we consider a fixed subsystem size k that is independent of N , the
variance will grow exponentially with N . However, if the subsystem size increases with N (fixed number of partitions α), such
that k = αN , then the expression will saturate to a constant 2α . Thus, the variance is larger than for k = N , but still independent
of the system size. Since variance is directly related to sample complexity, it is important to avoid exponentially growing
variances in practical settings. We note that maintaining low variance while considering optimized state-preparation strategies,
such as MPS state preparation, is an interesting problem we leave for future work.
More general reference states
Spectroscopy from more general reference states |S⟩ can also be performed, by considering different measurement strategies.
One approach is to apply the inverse state-preparation unitary S † and then measure in the X-basis. This is enables parallel
access to the 2N operators {SXs S † }, which by construction form an orthonormal basis when applied to the reference state |S⟩.
In particular, we can use this ensemble to construct an alternate resolution of the identity
!
X X
† †
1= SXs S |S⟩⟨S|SXs S = S Xs |0 ⟩⟨0 |Xs S † = SS † ,
N N
(S53)
s s
13
A
which can be used as before to compute DR (t),
X
A
DR (t) = ⟨R|ASXs S † |S⟩⟨S|SXs S † |R(t)⟩
s
X
= ⟨R|ASXs |0N ⟩CSXs S † ,R (t)e−iE0 t . (S54)
s
As with (S46), this sum can be efficiently evaluated classically as long as the state-preparation unitary S has an efficient MPO
description.
Randomized measurements also provide an alternate route for extracting the density-of-states, where the measurement basis
can be independent of the specific reference state |S⟩. Here, we consider random Pauli measurements. The set of all Pauli
operators can be used to construct a resolution of the identity from the over-complete basis {Pr |S⟩}
1 X
Pr |S⟩⟨S|Pr = 1. (S55)
2N r
A
For this choice, the expression for DR (t) becomes
1 X
⟨S|AU (t)R|S⟩ = ⟨S|APr U (t)|S⟩⟨S|U † (t)Pr U (t)R|S⟩ (S56)
2N r
X
= e−iES t ⟨S|APr |S⟩CPr ,R (t) (S57)
r
where ES is the energy of the eigenstate |S⟩. This quantity can be efficiently evaluated classically as well. However, note the
variance of the estimator for this quantity is generically large, due to the exponential “shadow norm” for high-weight opera-
tors [S84], see also (S40). Improving sample complexity for the randomized-measurement version of many-body spectroscopy
is an exciting direction for continued research.
Beyond MPS reference states with doubled Hilbert spaces
Here, we have proposed a hybrid classical-quantum approach for evaluating DA (ω), which leverages an efficient classical
representation for |R⟩ and |S⟩, that is suitable for near-term applications. In some cases, this requirement can be replaced by a
fully quantum measurement, while preserving the sample-efficiency of the procedure, by coherently simulating two systems in
parallel and performing entangled measurements between them [S39].
We illustrate such a scheme for estimation of the bare-DOS D1 (ω). Then, we observe that the sum over Pauli operators in
(S56) can be re-written as a two-copy measurement of the form
1 1 X
DR (t) = ⟨S|U (t)R|S⟩ = N ⟨R|Pr |S⟩⟨S|Pr U (t)R|S⟩ (S58)
2 r
1 X
= N ⟨R|⟨S|(Pr ⊗ Pr )(1 ⊗ U (t))|S⟩|R⟩ (S59)
2 r
This resulting correlation function has the same form as the two-time correlators measured by Fig. 4a. Further, all the observables
commute Pr ⊗ Pr , indicating that they can be written as diagonal operators in the correct measurement basis. This basis
corresponds to performing pairwise Bell measurements between the two systems. To see this, notice that a single pairwise Bell
measurement simultaneously measures 1, XX, Y Y, ZZ. Therefore, all 4N Pauli-pairs Pr ⊗ Pr can be measured in parallel
by factorizing into individual pairs between the two systems. These entangled measurements can be readily implemented in
reconfigurable architectures, as demonstrated in Ref. [S56].
Thermal expectation values
Once we have DA (ω), a simple way to compute thermal expectation values is by reweighting the density of states, and
computing the following ratio
R −βω A
e D (ω)dω
⟨A⟩β = R −βω 1 (S60)
e D (ω)dω
14
as discussed also in Ref. [S81].

We further note that this ratio is theoretically robust to uniform spectral broadening, e.g. arising from a finite maximum
evolution time. To illustrate this, we can broaden the delta function in DA (ω) to be
X
DTA (ω) = ⟨n|A|n⟩δT (ω − ϵn ) (S61)
n
where δT is a broadened delta function with width ∼ 1/T . When we add the Boltzmann weight e−βω and integrate over
frequencies, the numerator can be rewritten as
Z Z X
dωDTA (ω)e−βω = dω ⟨n|A|n⟩δT (ω − ϵn )e−βω (S62)
n
X Z
= ⟨n|A|n⟩ dωδT (ω − ϵn )e−βω (S63)
n
X Z
= ⟨n|A|n⟩e−βϵn dωδT (ω)e−βω (S64)
n | {z }
CT
X
= CT ⟨n|A|n⟩e−βϵn . (S65)
n
Therefore, we see this is equal to the the numerator when T → ∞, rescaled by a T -dependent constant CT . The denominator is
similarly rescaled by the same constant,
Z Z
1
X
−βω
dωDT (ω)e = dω δT (ω − ϵn )e−βω (S66)
n
X Z
−βωn
= e dωδT (ω)e−βω (S67)
n
X
= CT e−βϵn . (S68)
n
As such, the constant CT cancels out in the ratio, producing

⟨n|A|n⟩e−βϵn
P
⟨A⟩β,T = nP −βϵn = ⟨A⟩β,T →∞ (S69)
ne
which is equal to the ideal value.
Magnetic Susceptibility
The magnetic susceptibility for the Heisenberg models studied in Fig. 4 and Fig. 5 can be computed straightforwardly by
selecting A = (S z )2 . To understand why this is the case, consider coupling the system to an external magnetic field B [S85].
P
Define global spin operators Ŝ = i Ŝi . Then, the Hamiltonian in the presence of a field becomes
H → H + B · Ŝ (S70)
The density matrix at finite temperature is further given by
1 −βH
ρβ = e , Z = Tr[e−βH ]. (S71)
Z
The magnetization is a vector quantity measured by
1
M µ (B, β) = Tr[S µ e−βH ] (S72)
Z
and the susceptibily is a response matrix
∂
χµν = M µ (B, β) (S73)
∂B ν
15
0.225
0.200
Signal-to-noise ratio
0.175
0.150
0.125
0.100
0.075
0.050
0.025
0.000 0.005 0.010 0.015 0.020

Ensemble temperature β
Extended Data – FIG. S3. Importance sampling to improve SNR. We consider the signal-to-noise ratio associated with calculating ground-
state (T = 0) properties of the S2H-1b model of the OEC. We consider perturbed states |R⟩ = R|0N ⟩ generated by applying random
single-qubit rotations (i.e. random product states), and sample from an ensemble R′ with probabilities proportional to e−βensemble E(R) . We
find that there is an optimal temperature βensemble which approximately doubles the signal-to-noise ratio compared to a fully random ensemble
(βensemble = 0). Further increasing βensemble actually causes the signal-to-noise ratio to become worse, since the ground state cannot be
expressed as a product state. Initial states supporting some entanglement (e.g. MPS) should further improve the SNR.
which depends on both the temperature β and external field B. Using the fact that S commutes with H in the Heisenberg models,
the derivative can be simplified to
χα α β
β (B, β) = βTr[Ŝ Ŝ ρβ ]. (S74)
In the main text, we focus on χ at zero-field B = 0, where the rotational symmetry implies the matrix χα α
β is proportional to δβ .
Thus, we focus on one component, and plot
χ(T ) = χzz (B = 0, β) = βTr[(Ŝ z )2 ρβ ]. (S75)
Sampling schemes for low temperatures
As discussed above, it is important for sample efficiency that the ensemble of perturbations R prepare probe states that have
significant overlap with target eigenstates. Classical optimization can be used to prepare variational states with larger overlap
on, e.g the ground state.
However, to accurately compute finite-temperature properties, while preferentially targetting the low-energy sector, it is cru-
cial that we accurately compute the density of states DA (ω), without introducing bias by restricting to a smaller subspace.
Importance sampling is a method to simulate uniform sampling over all of Hilbert space, while still preferentially sampling
low-energy states. The key intuition is that we can estimate the low-energy part of DA (ω) with lower-noise, at the cost of
making the high-energy part of DA (ω) more noisy. Then, since the contribution to low-temperature properties is dominated by
the low-energy part of DA (ω), this produces an overall less noisy estimate.
Specifically, we consider an importance sampling scheme, where the probability of sampling a state |R⟩ from a uniform
distribution R is weighted by the Boltzmann factor PR′ (R) ∝ e−βE(R) where E(R) = ⟨R|H|R⟩ is a clasically computed
energy. Then, we can replace the average over R with a weighted average over R′ .
2N
ER∼R 2N → ER∼R′ (S76)
PR′ (R)
Substitutions of this form can reduce the total sample complexity, as we demonstrate for the random product state case in
Extended data Fig. S3.
16
Efficient evaluation of spin-sector projectors
Our efficient estimator (S46) of the operator-resolved density of states involves evaluating expectation values of the form
⟨R|A|b⟩ where b is the measured (X-basis) product state, |R⟩ is the initially prepared probe state, and A is the operator to
measure. As interesting choices for A, we utilize the projector onto a total S z sector, and the projector onto total-spin sectors
Ps .
To evaluate these overlaps efficiently, we utilize the formalism of 1D tensor networks. In particular, write the projector onto
an S z sector with m spin-flips as a matrix-product operator
N
N Y
P (N = N, S z = πi0 1a + πi1 σa+ |0⟩a

− m) = ⟨m|a (S77)
2 i=1
where in the term πi1 σa+ , the first index acts on the physical site i, and the second term acts on the ancillary mode, which starts
at |0⟩ on the right and at the end is projected onto ⟨m|. Intuitively, the ancilla mode counts the number of qubits in the state |1⟩,
and the ket ⟨m| only picks out strings with exactly m down spins. Thus, this forms the projector onto all states with exactly m
down spin, or S z = N2 − m. This MPO has bond dimension Θ(N ). Therefore, overlaps of the form ⟨R|PS z |b⟩, where |R⟩ and
|b⟩ are both finite-bond dimension MPS, can be efficiently evaluated in O(N 2 ) time via tensor contraction.
Next, we generalize the projector construction to sectors with well-defined total-spin and z-component, S and S z respectively.
The key insight underlying the construction is to utilize three ancilla modes, which track the total S z of the bra and ket, as well
as their Hamming distance.
N N
P (N = N, S = − s, S z = − m) = ⟨N, s, m|a ×
2 2
N
Y
|0⟩i ⟨0|i (111)a + |0⟩i ⟨1|i (1σ + σ + )a + |1⟩i ⟨0|i (σ + 1σ + )a + |1⟩i ⟨1|i (σ + σ + 1)a

i=1
× |000⟩a (S78)
3
The bond-dimension of this MPO scales as O(NP ) in general. To understand this expression, notice that we can always write
the matrix elements of the projector as P = s,s′ Ps,s′ |s⟩⟨s′ | in the z-basis, where the coefficient Ps,s′ = ⟨s|P |s′ ⟩. When the
MPO for P is contracted with the two bistrings ⟨s| · |s′ ⟩, the expression reduces to
Ps,s′ = ⟨N, s, m|w(s)w(s′ )w(s − s′ )⟩a (S79)
where w(s) measures the number of down-spins (weight) of each configuration s. As in the S z case, for this expression to be
non-zero the number of down-spins in both s and s′ must equal m, which imposes the condition w(s) = w(s′ ) = m. However,
in general, the projector also contains off-diagonal terms, where the Hamming distance w(s − s′ ) between the bra and ket state
can be non-zero.
A simple example is the projector onto the triplet state,
1
P (N = 2, S = 1, S z = 0) = (|01⟩⟨01| + |01⟩⟨10| + |10⟩⟨01| + |10⟩⟨10|) (S80)
2
For this example, we see the termination vector should take the form
1
|2, 0, 1⟩ = (|110⟩ + |112⟩) (S81)
2
Note that this object is a virtual state, not a physical state, and hence does not need to be normalized. To simplify notation, we
will in general write each termination vector as
m
X
N
|N, s, m⟩a = Ps,m,d |(m)(m)(2d)⟩a (S82)
d=0
N
where Ps,m,d encodes all of the relevant weights. We note that, in this expression, we use the fact that the Hamming distance
between two configurations with the same S z is always an even number 2d, and is can be no greater than twice the number of
spin-flips m. As such, we can interpret d as the number of exchanges needed to connect the two-strings s and s′ .
N
We present an efficient (poly(N ) time) recursive construction for the coefficients Ps,m,d
17
To illustrate the idea, let us start with the simplest projector, onto S = N2 and S z = N2 , or s = 0, m = 0. The only state which
N
contributes is the all-up state, so P0,0,0 = 1. The projectors onto all other symmetric subspaces, with S = N2 , S z = N2 − m can
be computed using a ladder-like construction. For one spin-flip (m = 1), the projector can be written as
! !
N z N 1 X + N z N X
−
P (N = N, S = , S = − 1) = σi P (S = , S = − 1) σi . (S83)
2 2 N i
2 2 i
By plugging in representative configurations with a single down-spin, we can compute the relevant coefficients to be
N N
P0,1,0 = 1/N P0,1,1 = 1/N (S84)
To compute the normalization factor N , we use the fact that P 2 =P . In particular, we can again plug in representative configu-
rations to match
⟨s|P |s′ ⟩ = ⟨s|P 2 |s′ ⟩ (S85)

X
= ⟨s|P |s′′ ⟩⟨s′′ |P |s′ ⟩ (S86)
s′′
Taking |s⟩ = |s′ ⟩ = |10N −1 ⟩ as the representative states, we could explicitly sum over the N relevant intermediate states.
However, to simplify the calculation, we note the intermediate states naturally split into two groups. The first is distance-0 from
s and s′ , i.e. |s′′ ⟩ = |s⟩, which contributes (P0,1,0
N
)2 . The second group is distance-2 from s and s′ , such as |s′′ ⟩ = |010N −2 ⟩.
N 2
This contributes (P0,1,1 ) , and there are N − 1 such contributions. Therefore, we get the equality
N N
P0,1,0 = (P0,1,0 )2 + (N − 1)(P0,1,1,
N
)2 (S87)
2
1/N = N/N (S88)
N = N. (S89)
We could similarly select |s⟩ and |s′ ⟩ so the left-hand side is the term P0,1,1
N
corresponding to Hamming distance 2. There, a
similar calculation over intermediate paths leads to the constraint
N N N N
P0,1,1 = 2(P0,1,0 P0,1,1 ) + (N − 2)(P0,1,1 )2 (S90)
N =N (S91)
resulting in the same normalization factor. Indeed, the condition P = P 2 implies the computed value for N must be the same
for all non-zero coefficients appearing in the left-hand side. The key intuition underlying these calculations, is that the valid
paths can be efficiently partitioned into groups, and the coefficients can be efficiently computed via combinatorics.
The same intuition can be generalized to a recursive algorithm for computing Ps,m,d from Ps,m−1,d′ , which allows us to
compute the full ladder of states with m = 0, ..., N from the root cases. First, we compute the unnormalized coefficients using
the formula
′
Ps,m,d = d2 Ps,m−1,d−1 + (m − d)(2d + 1)Ps,m−1,d + (m − d)(m − d − 1)Ps,m−1,d+1 (S92)
Where we assume Ps,m,d = 0 when d > m. To understand the coefficient associated with each term, we consider a pair of
representative states |s⟩ = |1m−d 1d 0d 0N −m−d ⟩ and |s′ ⟩ = |1m−d 0d 1d 0N −m−d ⟩, which have m down-spins and Hamming
distance 2d. For convenience, we label the four groups of size m − d, d, d, and N − m − d as A,B,C, and D respectively. The
formula comes from the following case-by-case breakdown:
1. The term |s⟩⟨s′ | can come from a pair with smaller Hamming distance 2(d − 1) by flipping one spin from B in s and from
C in s′ . There are d2 such choices.
2. The term |s⟩⟨s′ | can come from a pair with the same Hamming distance 2d, by flipping one spin from A in s and one spin
from C in s′ , vice versa, or by flipping the same spin from A for s and s′ . There are 2(m−d)d+(m−d) = (m−d)(2d+1)
such choices.
3. The term |s⟩⟨s′ | can come from a pair with larger Hamming distance 2(d + 1) by flipping two different spins from A in s
and s′ . There are (m − d)(m − d − 1) of these pairs.
18
Then, we need to compute the normalization factor. In order to do this, consider the same pair of representative states, and
enumerate over all intermediate states s′′ . In particular, we can efficiently label intermediate states by counting the number of
down-spins in each of the four groups, mA , mB , mC , mD , where the total number of down-spins satisfies m = mA + mB +
mC + mD . For each intermediate configuration, the distance from s to s′′ and s′ to s′′ is given by
d(s, s′′ ) = (m − d − mA ) + (d − mB ) + mC + mD (S93)

′ ′′
d(s , s ) = (m − d − mA ) + mB + (d − mC ) + mD (S94)
The number of intermediate states s′′ associated with each label is

m−d N −m−d

d d
Z(mA , mB , mC , mD ) = . (S95)
mA mB mC mD
Taken together, by enumerating over valid assignments for the number of down-spins in each group, we can compute the
normalization factor
′ ′
P
mA +mB +mC +mD =m Z(mA , mB , mC , mD )Ps,m,d(s,s′′ ) Ps,m,d(s′ ,s′′ )
N = ′ (S96)
Ps,m,d(s,s ′)
and the normalized projector coefficients Ps,m,d = P ′ /N .

Next, we discuss projectors onto states where the total-spin is smaller then N/2. To illustrate this, we start by computing
P (S = N2 − 1, S z = N2 − 1). To do this, we see that there are only two total-spin sectors which contribute to S z = N2 − 1,
s = 0 and s = 1. Therefore, we know that
N N N N N
P (N = N, S = − 1, S z = − 1) = P (N = N, S z = − 1) − P (N = N, S = , S z = − 1) (S97)
2 2 2 2 2
The termination vector associated with S z = N2 − m sector is |(m)(m)(0)⟩. Therefore, the coefficients of the projector can be
N N 2
written as P1,1,d = δd,0 − P0,1,d . As an example, consider the N = 2 singlet state, for which this formula implies P1,1,0 = 1/2
2
and P1,1,2 = −1/2. This leads to the familiar expression
1
P (N = 1, S = 0, S z = 0) = (|01⟩⟨01| − |10⟩⟨01| − |01⟩⟨10| + |10⟩⟨10|) . (S98)
2
More generally, to construct each of the terms in the root projector with s = m, we orthogonalize w.r.t all higher-spin sectors.
X
Pm,m,d = δd,0 − Ps,m,d (S99)
0≤s<m
Here, each term with s < m can be computed from root terms where s = m using the recursive ladder approach described
above.

Programmable Simulations of Molecules and Materials With Reconfigurable Quantum Processors

Uploaded by

Copyright:

Available Formats

Programmable Simulations of Molecules and Materials With Reconfigurable Quantum Processors

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Programmable Simulations of Molecules and Materials With Reconfigurable Quantum Processors

Uploaded by

Copyright:

Available Formats

Programmable Simulations of Molecules and Materials with Reconfigurable Quantum Processors

(a) (b) 3 Active Electrons (c)

Strongly Correlated Programmable Quantum Simulation

Two Qubit Gates

composed of projectors P [(Ŝi )] onto the symmetric spin

(a) (b) (c)

properties of eigenstates can be computed by choosing A ap-

ensemble of random single-spin rotations. This corresponds (a) (b)

effective spectral resolution in the presence of noise as it spar- 5 11/2 11/2

sifies the signal. Finite-temperature response functions [85], 9/2

(a) tries [97, 98].

models, and expand the class of accessible chemistry prob-

[101] T. V. Zache, D. González-Cuadra, and P. Zoller, Fermion-qudit [cond-mat, physics:quant-ph].

and |b(i) ⟩ is the measured projected state.

evaluated at energies ωn of the eigenstates, which can be

The square lattice Heisenberg calculation was performed on

We measure the Green’s function G(r, t) = ⟨S|Xr (t)X0 |S⟩

which reduces to the Fourier transform G(k, ω) when the en-

(a) (d) 4.0

Phenomenological Noise Rate γ/|J|

(a) (b) 4.6

Variational spin-gate optimization.

Θ = (0, 2π/3, 4π/3, 4π/3, 2π/3, 0) (S1)

H2,extended = J1 (Ŝ1 · Ŝ2 ) + J2 (Ŝ1 · Ŝ2 )2 + J3 (Ŝ1 · Ŝ2 )3 . (S3)

Procedures for Multi-Qubit Gate Optimization

Low-dimensional construction of blockade cluster basis

Simultaneous drive – Smooth gates with GrAPE

Cλ ({θ, ϕ}) = (1 − F({θ, ϕ}))

Heuristic Simulation Time Estimates

≈ (c2 τ 2 t)2 (S17)

(a) Spin-1 Spin-5/2 (b) Kagome Heisenberg

Short-time error (1 − F (t))/t2

Short-time error (1 − F (t))/t2

or equivalently, in a maximum evolution time given a particular error tolerance

Two interacting spin-S’s

(−1) β (0) −iΘ (0) β (−1) −iΘ

(1) β (1) 2iΘk (−1) β (−1) −2iΘk

c̃2 S = 1 S = 3/2 S = 2 S = 5/2

Gate error g̃ S = 1 S = 3/2 S = 2 S = 5/2

Sim. Times (×|H|local ) S = 1 S = 3/2 S = 2 S = 5/2

Estimator (31) for CO,R (t) from snapshots

Therefore, we see that the expectation value of the estimator is equal to

E |2σ(µ, a)⟨b|O|b⟩|2 = E 4|⟨b|O|b⟩|2

Beyond diagonal operators

Interpretation as shadow tomography of |R(t)⟩ from single-basis measurements

The operator-resolved DOS estimator

= 2N/2 ⟨R|A|b⟩, (S46)

Convergence of the DOS estimator

Hamiltonian diagonal in the X basis will satisfy these results.

More general reference states

Beyond MPS reference states with doubled Hilbert spaces

Thermal expectation values

as discussed also in Ref. [S81].

As such, the constant CT cancels out in the ratio, producing

which is equal to the ideal value.

0.000 0.005 0.010 0.015 0.020

χ(T ) = χzz (B = 0, β) = βTr[(Ŝ z )2 ρβ ]. (S75)

Sampling schemes for low temperatures

Efficient evaluation of spin-sector projectors