Hartree-Fock On A Superconducting Qubit Quantum Computer

Hartree-Fock on a superconducting qubit quantum computer
Google AI Quantum and Collaborators∗

(Dated: September 22, 2020)
As the search continues for useful applications of noisy intermediate scale quantum devices, vari-
ational simulations of fermionic systems remain one of the most promising directions. Here, we
perform a series of quantum simulations of chemistry which involve twice the number of qubits and
more than ten times the number of gates as the largest prior experiments. We model the binding
energy of H6 , H8 , H10 and H12 chains as well the isomerization of diazene. We also demonstrate
error-mitigation strategies based on N -representability which dramatically improve the effective fi-
delity of our experiments. Our parameterized ansatz circuits realize the Givens rotation approach
to noninteracting fermion evolution, which we variationally optimize to prepare the Hartree-Fock
wavefunction. This ubiquitous algorithmic primitive corresponds to a rotation of the orbital basis
and is required by many proposals for correlated simulations of molecules and Hubbard models.
arXiv:2004.04174v4 [quant-ph] 18 Sep 2020
Because noninteracting fermion evolutions are classically tractable to simulate, yet still generate
highly entangled states over the computational basis, we use these experiments to benchmark the
performance of our hardware while establishing a foundation for scaling up more complex correlated
quantum simulations of chemistry.
The prediction of molecular properties and chemical We simulated quantum chemistry in a second-
reactions from ab initio quantum mechanics has emerged quantized representation where the state of each of N
as one of the most promising applications of quantum qubits encoded the occupancy of an orbital basis func-
computing [1]. This fact is due both to the commer- tion. We used what are commonly referred to as core
cial value of accurate simulations as well as the relatively orbitals as the initial orbitals (shown for H12 on the left
modest number of qubits required to represent interesting of Fig. 1a), which are the eigenfunctions of the molecu-
instances. However, as the age of quantum supremacy lar Hamiltonian without the electron-electron interaction
dawns [2], so has a more complete appreciation of the term. The goal of this experiment was to use a quan-
challenges required to scale such computations to the tum computer to implement the Hartree-Fock procedure,
classically intractable regime using near-term interme- which is a method for obtaining the best single-particle
diate scale quantum (NISQ) devices. Achieving that ob- orbital functions assuming each electron feels the average
jective will require further algorithmic innovations, hard- potential generated from all the other electrons. This as-
ware with more qubits and low error rates, and more sumption is enforced by constraining the wavefunction
effective error-mitigation strategies. Here, we report a to be a product of one-particle functions which has been
variational quantum eigensolver (VQE) [3] simulation of appropriately antisymmetrized to satisfy the Pauli exclu-
molecular systems with progress in all three directions. sion principle. An initial guess for the Hartree-Fock state,
We used the Google Sycamore quantum processor to from which we can optimize the orbitals, was obtained by
simulate the binding energy of hydrogen chains as large filling the lowest energy η/2 orbitals, each with a spin-up
as H12 , as well as a chemical reaction mechanism (the electron and a spin-down electron, where η is the number
isomerization of diazene). The Sycamore quantum proof electrons. Since we simulated the singlet ground state
cessor consists of a two-dimensional array of 54 transmon for all molecules considered here, there is no spin com-
qubits [2]. Each qubit is tunably coupled to four nearest ponent to the mean-field approximation; thus, we only
neighbors in a rectangular lattice. Our largest simula- needed to explicitly simulate the η/2 spin-up electrons.
tions used a dozen qubits – twice the size as the largest By performing a unitary rotation of the initial (core)
prior quantum simulations of chemistry [4] – and required orbital basis ϕp (r), one can obtain a new valid set of
only nearest-neighbor coupling (depicted in Fig. 1). Prior orbitals ϕep (r) as a linear combination of the initial ones:
simulations of chemistry on superconducting qubit de-
vices and trapped ion systems demonstrated the possibil- N
X
ity of error mitigation through VQE [4–10], albeit on a ϕ
ep (r) = [eκ ]pq ϕq (r) , (1)
small scale. We demonstrated that, to within the model, q=1
achieving chemical accuracy through VQE is possible for
intermediate scale problems when combined with effec- where κ is an N × N anti-Hermitian matrix and [eκ ]pq
tive error mitigation strategies. Furthermore, we argue is the p, q element of the matrix exponential of κ. A
that the circuit ansatz we used for VQE is especially ap- result due to Thouless [11] is that one can express the
pealing as a benchmark for chemistry. unitary that applies this basis rotation to the quantum
state as time-evolution under a non-interacting fermion
Hamiltonian. Specifically, if we take a†p and ap to be
fermionic creation and annihilation operators for the core
∗ Corresponding author (Nicholas Rubin): [email protected]; orbital ϕp (r) then we can parameterize |ψκ i, an anti-
Corresponding author (Ryan Babbush): [email protected] symmetric product state in the new basis ϕ ep (r), as non-
2
a
Initial basis set Basis rotation circuit the generator for Uκ corresponds to a non-interacting
fermion Hamiltonian, its action on a product state in sec-
ond quantization can be classically simulated in O(N 3 )
+Post-selection +Purification
by diagonalizing the one-body operator and in some cases
the Hartree-Fock procedure can be made to converge
with even lower complexity. Despite that fact, we ar-
gue that this procedure is still a compelling experiment
for a quantum computer.
The Hartree-Fock state is usually the initial state for
classical correlated electronic structure calculations such
b c as coupled cluster and configuration interaction methods,
as well as for many quantum algorithms for chemistry.
Thus, often one chooses to work in the molecular orbital
basis, which is defined so that the Hartree-Fock state
is a computational basis state. However, the molecular
orbital basis Hamiltonian has a large number of terms
FIG. 1. Basis rotation circuit and compilation. a) To which can be challenging to simulate and measure with
the left of the circuit diagram are the initial orbitals for the low complexity. Accordingly, the most efficient quan-
H12 chain with atom spacings of 1.3 Å, obtained by diago- tum algorithms for chemistry [12–15] require that one
nalizing the Hamiltonian ignoring electron-electron interac- perform the simulation in more structured bases with
tions. The circuit diagram depicts the basis rotation ansatz asymptotically fewer terms [16–18], necessitating that
for a linear chain of twelve hydrogen atoms. Each grey box Uκ? is applied explicitly at the beginning of the com-
with a rotation angle θ represents a Givens√rotation gate. b)
putation. Even when simulating chemistry in an arbi-
Compilation of the Givens rotation gate to iswap gates and
trary basis, the most efficient strategies are based on a
single-qubit gates that can√be realized directly in hardware.
The H12 circuit involves 72 iswap gates and 108 single-qubit tensor factorization of the Hamiltonian which requires
Z rotation gates with a total of 36 variational parameters. c) many applications of Uκ to simulate [19, 20]. Exploiting
Depiction of a twelve qubit line on a subgrid of the entire this tensor factorization with basis rotations is also key
54-qubit Sycamore device. All circuits only require gates be- to the most efficient strategy for measuring hHi in vari-
tween pairs of qubits which are adjacent in a linear topology. ational algorithms, and requires implementing Uκ prior
to measurement [21].
We used this variational ansatz based on basis rota-
tions to benchmark the Sycamore processor for linear
interacting fermion dynamics from a computational basis
hydrogen chains of length 6, 8, 10, and 12 and two path-
state |ηi = a†η · · · a†1 |0i in the core orbital basis: ways for diazene bond isomerization. We modeled hy-
! drogen chains of length N with N qubits. Our simula-
N
X tions required N qubits to simulate 2N spin-orbitals due
|ψκ i = Uκ |ηi , Uκ = exp κpq a†p aq . (2) to the constraint that the α-spin-orbitals have the same
p,q=1
spatial wavefunction as the β-spin-orbitals. For diazene
Such states are referred to as Slater determinants. we required 10 qubits after pre-processing. The hydro-
To complete the accurate preparation of Hartree-Fock gen chains are a common benchmark in electronic struc-
states, we implemented variational relaxation of the κ ture [22–24] and the diazene bond isomerization provides
parameters to minimize the energy of |ψκ i starting from a system where the required accuracy is more representa-
the optimal κ determined by solving the Hartree-Fock tive of typical electronic structure problems and has been
equations classically. This is an idealized implementa- used as a benchmark for coupled cluster methods [25].
tion of VQE that allowed us to demonstrate error miti- For the diazene isomerization our goal was to resolve the
gation of coherent errors through variational relaxation. energetic difference between the transition states of two
We defined the Hartree-Fock state |ψHF i to be the lowest competing mechanisms, requiring accuracy of about 40
energy Slater determinant for the molecular Hamiltonian milliHartree. This objective differs from prior quantum
H, i.e. simulations of chemistry which have focused on bond dis-
sociation curves [4–7].
|ψHF i = |ψκ? i κ? = argminκ hψκ | H |ψκ i . (3) One motivation for this work was to calibrate and val-
idate the performance of our device in realizing an im-
We applied Uκ to |ηi using our quantum computer and portant algorithmic primitive for quantum chemistry and
then performed the optimization over κ through feed- lattice model simulation. Our experiment was also ap-
back from a classical optimization routine. The energy pealing for benchmarking purposes since the circuits we
decreased because the initial core orbitals were obtained explored generated highly entangled states but with spe-
by ignoring the electron-electron interaction and varia- cial structure that enabled the efficient measurement of
tional relaxation compensates for coherent errors. Since fidelity and the determination of systematic errors. Fur-
3
a b c 1
2.00 6 qubits 10 H-H distance
Raw 0.5
1.3
Energy error [Hartree]

Energy error [Hartree]
1
1 - [Fidelity Witness]
1 10 2.1
2.25 10 +PS
Energy [Hartree]
4 2
10
2.50 2
10 2
Chemical 10
2.75 accuracy +Purification
3
3 10
3.00 10
3
10 +VQE
3.25 4
10
0.5 1.0 1.5 2.0 2.5 3.0 1 2 3 1 2 0 2 4 6 8 10 12 14 16 18
Bond distance [Å] Bond distance [Å] Bond distance [Å] Optimizer Iterations
d e f 3
2.5 8 qubits 3.0 10 qubits 12 qubits
3.5
Energy [Hartree]
Energy [Hartree]
Energy [Hartree]
3.0 4
6 8 10
4.0
3.5 5
0
4.5
Energy error
Energy error
10 0
Energy Error
10
1
10
1 1
10 10
4.0 3
5.0 10
2 6 2
10 10
0 2 4 0 2 4 0 2 4
0.5 1.0 1.5 2.0 2.5 3.0 0.5 1.0 1.5 2.0 2.5 3.0 0.5 1.0 1.5 2.0 2.5 3.0
Bond distance [Å] Bond distance [Å] Bond distance [Å]
FIG. 2. Static and VQE performance on hydrogen chains. Binding curve simulations for H6 , H8 , H10 , and H12 with
various forms of error mitigation. Subfigures (a, d, e, f) compare Sycamore’s raw performance (yellow diamonds) with post-
selection (green squares), purification (blue circles), and error mitigated combined with variational relaxation (red triangles).
For all hydrogen systems the raw data at 0.5 Å bond length is off the top of the plot. The yellow, green, and blue points
were calculated using the optimal basis rotation angles computed from a classical simulation; thus, the variational optimization
shown here is only used to correct systematic errors in the circuit realization. Subfigure (b) contains the absolute error and
infidelity for the H6 system. For all points we calculated a fidelity witness described in Appendix D. The error bars for all points
were computed by estimating the covariance between simultaneously measured sets of 1-RDM elements and resampling those
elements under a multivariate Gaussian model. Energies from each sample were tabulated and the standard deviation is used
as the error bar. The “+PS” means applying post-selection to the raw data, “+Purification” means applying post-selection
and McWeeny purification, and “+VQE” means post-selection, McWeeny purification, and variational relaxation. Subfigure
(c) contains optimization traces for three H6 geometries (bond distances of 0.5 Å, 1.3 Å, and 2.1 Å). All optimization runs used
between 18 and 30 iterations. The lowest energy solution from the optimization trace was reported.
ther motivation was to implement the largest variational main components: ansatz specification in the form of a
quantum simulation of chemistry so that it is possible parameterized quantum circuit (the function), observable
to better quantify the current gap between the capabili- estimation (the functional), and outer-loop optimization
ties of NISQ devices and real applications. Even though (the minimization). Each component is distinctively af-
the Hartree-Fock ansatz is efficient to simulate classi- fected by our choice to simulate a model corresponding
cally, the circuits in our experiment are far more complex to non-interacting fermion wavefunctions. Symmetries
than prior experimental quantum simulations of chem- built into this ansatz allowed for reduction of the num-
istry. Finally, the structure of the Hartree-Fock state ber of qubits required to simulate molecular systems, a
enabled us to sample the energy and gradients of the reduction in the number of measurements needed to es-
variational ansatz with fewer measurements than would timate the energy, and access to the gradient without
typically be required, allowing us to focus on other as- additional measurements beyond those required for en-
pects of quantum simulating chemistry at scale, such ergy estimation. See Appendix A for details on how we
as the effectiveness of various types of error-mitigation. realized Hartree-Fock with VQE.
Thus, our choice to focus on Hartree-Fock for this exper- The unitary in Eq. (2) can be compiled exactly (with-
iment embraces the notion that we should work towards out Trotterization) using a procedure based on Givens
valuable quantum simulations of chemistry by first scal- rotations. This strategy was first suggested for quan-
ing up important components of the exact solution (e.g., tum computing in work on linear optics in [26] and later
error-mitigation strategies and basis rotations) in a fash- in the context of fermionic simulations in [27]. Here,
ion that enables us to completely understand and perfect we implemented these basis rotations using the optimal
those primitives. compilation of [28] that has gate depth N/2 and requires
Variational algorithms are specified in the form of a only η(N − η) two qubit “Givens rotation” gates on a
functional minimization. This minimization has three linearly connected architecture, giving one rotation for
4
108.400 a protocol described in Appendix C. This protocol en-

10Q energy
TS2 abled us to optimally parallelize measurement of all N 2
108.425 12Q energy
1-RDM elements with N + 1 distinct circuits. For each
Energy [Hartree]
Out of plane
108.450
In plane
distinct circuit we made 250,000 measurements.
trans We performed two types of error mitigation on our
108.475 cis
measured data: post-selection on particle number (con-
108.500 served in basis rotations) and pure-state projection. To
108.525
TS1 apply post-selection we modified our circuits by first ro-
tating into a basis that diagonalizes a†p aq + a†q ap for N
108.550 different pairs of p and q so that these elements could be
0 5 10 15 20 25 30 35 40 sampled at the same time as the total particle-number
Reaction Coordinate operator. Following the strategy in Appendix C, this
measurement√ was accomplished at the cost of two T gates
FIG. 3. VQE performance on distinguishing the mech- and one iswap gate per pair of qubits. We then post-
anism of diazene isomerization. Hartree-Fock curves for selected to discard measurements where the total number
diazene isomerization between cis and trans configurations. of excitations changed from η/2.
T S1 and T S2 are the transition states for the in-plane and For pure-state purification, we leveraged the fact that
out-of-plane rotation of the hydrogen, respectively. The yel- the 1-RDM for any single-Slater determinant wavefunc-
low arrows on T S1 and T S2 indicate the corresponding reaction |ψκ i has eigenvalues restricted to be 0 and 1 [30].
tion coordinate. The solid curve is the energy obtained from We performed projection back to the pure-set of 1-
optimizing a 10-qubit problem generated by freezing the core
RDMs using a technique known as McWeeny purifi-
orbitals generated from two self-consistent-field cycles. The
transparent lines of the same color are the full 12 qubit system
cation [29]. Details on the procedure and sampling
indicating that freezing the lowest two levels does not change bounds for guaranteeing the procedure has a fixed-point
the characteristics of the model chemistry. Nine points along 1-RDM corresponding to a Slater determinant can be
the reaction paths are simulated on Sycamore using VQE. found in Appendix E. Although McWeeny purification
We allowed the optimizer 30 iterations for all points except only works for Slater determinant wavefunctions, pure-
for fifth and sixth point from the left of the in-plane rota- state N -representability conditions are known for more
tion curve which we allowed 60 steps. The error bars for all general systems [31] and we expect that a computational
points were computed by estimating the covariance between procedure similar to enforcing ensemble constraints could
simultaneously measured sets of 1-RDM elements and resam- be employed [32, 33].
pling those elements under a multivariate Gaussian model.
Energies from each sample were tabulated and the standard
A variety of circuit optimization techniques based on
deviation is used as the error bar. No purification was applied gradient- and gradient-free methods have been proposed
for the computation of the error bar. If purification is applied in the context of NISQ algorithms. Here, we developed
the error bars become smaller than√the markers. Each ba- an optimization technique that exploits local gradient
sis rotation for diazene contains 50 iswap gates and 80 Rz and Hessian information in a fashion which is distinctive
gates. to the Hartree-Fock model. It is based on a proposal
for iterative construction of a wavefunction to satisfy
the Brillouin condition for a single-particle model [34].
each element in the unitary basis change. These Givens Our optimization protocol used the property that at
rotation
√ gates were implemented by decomposition into a local optima the commutator of the Hamiltonian H
two iswap gates and three Rz gates. In Fig. 1, we with respect to any generator of rotation G is zero
depict the basis change circuit for the H12 chain, which (i.e. hψ| [H, G] |ψi = 0) and the fact that sequential basis
has a diamond shaped structure. We further review the change circuits can be concatenated into a single basis
compilation of these circuits in Appendix B. change circuit (i.e.
The average energy of any molecular system can be P Ua Ub = Uab ). Using these relations
and taking G = pq κpq a†p aq , as in our experiment, the
evaluated with knowledge of the one-particle reduced double commutator hψ|[[H, G], G]|ψi determined an aug-
density matrix (1-RDM), ha†p aq i, and the two-particle re- mented Hessian (matrix of derivatives) which we could
duced density matrix (2-RDM), ha†p a†q ar as i. In general, use to iteratively update the wavefunction such that the
it is not possible to exactly reconstruct the 2-RDM from first order condition was approximately satisfied. Regu-
knowledge of just the 1-RDM. However, for single-Slater larization was added by limiting the size of update pa-
determinants (as in our Hartree-Fock experiment), the rameters [35]. For details, see Appendix H.
2-RDM is completely determined by the 1-RDM [29]: As a benchmark, we studied symmetrically stretched
ha†p a†q ar as i = ha†p as i ha†q ar i − ha†q as i ha†p ar i . (4) hydrogen chains of length 6, 8, 10, and 12 atoms, Fig. 2.
The initial parameters were set to the parameters ob-
Thus, in our experiment we only needed to sample the 1- tained by solving the Hartree-Fock equations on a classi-
RDM to estimate the energy. As the 2-RDM has quadrat- cal computer. The data from the quantum computer is
ically more elements than the 1-RDM, this approach is a plotted along with classical Hartree-Fock results, showing
significant simplification. We measured the 1-RDM using better and better agreement as we added post-selection,
5
system estimate raw +ps +pure +VQE respect to the simulated model. Correctly identifying
H6 0.571 0.674(2) 0.906(2) 0.9969(1) 0.99910(9) this pathway requires resolving the energy gap of 40 milli-
H8 0.412 0.464(2) 0.827(2) 0.9879(3) 0.99911(8) Hartree between the two transition states. The pathways
H10 0.277 0.316(2) 0.784(3) 0.9704(5) 0.9834(4) correspond to the motion of the hydrogen in the process
H12 0.174 0.010(2) 0.654(3) 0.9424(9) 0.9913(3)
of converting cis-diazene to trans-diazene. One mecha-
TABLE I. Average fidelity lower bounds for hydro- nism is in-plane rotation of a hydrogen and the other is
gen chain calculations. We report values of the fi- an out-of-plane rotation corresponding to rotation of the
delity witness from [36], averaged across H-H separations of HNNH dihedral angle. Fig. 3 contains VQE optimized
{0.5, 0.9, 1.3, 1.7, 2.1, 2.5} Å, starting from circuits with the data simulating nine points along the reaction coordi-
theoretically optimal variational parameters (κ). “estimate” nates for in-plane and out-of-plane rotation of hydrogen.
corresponds to an estimate of the fidelity derived by multiply- For all points along the reaction coordinate the initial
ing gate errors assuming 0.5 percent single-qubit gate error, parameter setting was the solution to the Hartree-Fock
1 percent two-qubit gate error and 3 percent readout error. equations. VQE produced 1-RDMs with average fidelity
“Raw” corresponds to fidelities from constructing the 1-RDM greater than 0.98 after error-mitigation. Once again, we
without any error mitigation. “+ps” corresponds to fidelities
see that our full error mitigation procedure significantly
from constructing the 1-RDM with post-selection on particle
number. “+pure” corresponds to fidelities from constructing
improves the accuracy of our calculation.
the 1-RDM with post-selection and applying purification as Our VQE calculations on diazene predicted the correct
post-processing. Finally, “+VQE” corresponds to fidelities ordering of the transition states to within the chemical
from using all previously mentioned error mitigation tech- model with an energy gap of 41 ± 6 milliHartree and
niques in conjunction with variational relaxation. Note that the true gap is 40.2 milliHartree. We provide a more
for small values (such as the “raw” value for H12 ) we expect detailed analysis of the error mitigation performance on
the fidelity lower-bound is more likely to be loose. the diazene circuits in Appendix F considering that the
√
iswap gates we used had a residual cphase(π/24) and
Rz gates had stochastic control angles. This simulation
post-selection and purification, and then error mitigated reinforced VQEs √ ability to mitigate systematic errors at
variational relaxation. The 6- and 8-qubit data achieved the scale of 50 iswap gates and over 80 Rz gates.
chemical accuracy after VQE, and even the 12-qubit data In this work we took a step towards answering the
followed the expected energy closely. The error data in question of whether NISQ computers can offer quantum
Fig. 2b and the other inserts are remarkable as they show advantage for chemical simulation by studying VQE per-
a large and consistent decrease, about a factor of 100, formance on basis rotation circuits that are widely used in
when using these protocols. Fig. 2c details the signifi- quantum algorithms for fermionic simulation. The con-
cant decrease in error using a modest number of VQE sidered ansatz afforded ways to minimize the resource
iterations. requirements for VQE and study device performance for
A fidelity witness can be efficiently computed from the circuits that are similar to those needed for full Hamilto-
experimental data [36]; see Appendix D 2. This value is nian simulation. These basis rotation circuits also made
a lower bound to the true fidelity, and thus potentially an attractive benchmark due to their prevalence, opti-
loose when fidelity is small. However, Fig. 2b demon- mal known compilation, the ability to extract fidelity
strates that this fidelity generally tracks the measured and fidelity witness values and the fact that they pa-
errors. Table I shows how fidelity increased as we added rameterize a continuous family of analytically solvable
various forms of error mitigation, starting on the left col- circuits demonstrating a high degree of entanglement.
umn where the optimal angles were computed classically. The circuits also serve as a natural progression towards
Uncertainties in the last digit, indicated in the paren- more correlated ansatze such as a generalized swap net-
thesis, are calculated by the procedure described in Ap- work [28] or a non-particle conserving circuit ansatz fol-
pendix C 5. The first column of Table I is an estimate lowed by particle number projection.
of the fidelity based on multiplying the fidelity for all We demonstrated the performance of two error mitiga-
the gates and readout assuming 99.5% fidelity for single tion techniques on basis rotation circuit fidelity. The first
qubit gates, 99% fidelity for two-qubit gates, and 97% fi- is post-selection on total occupation number when mea-
delity for readout. We see that this estimate qualitatively suring all elements of the 1-RDM. This step was accom-
follows the “raw” fidelity witness estimates except when plished by permuting the basis rotation circuit such that
the witness value is very small. For all hydrogen systems all measurements involved estimating nearest-neighbor
studied, we observed drastic fidelity improvements with observables and measuring each pair of observables such
combined error mitigation. that the total occupation number is preserved. The sec-
Diazene isomerization. We simulated two isomer- ond is the application of McWeeny purification as a post-
ization pathways for diazene, marking the first time that processing step. The energy improvements from project-
a chemical reaction mechanism has been modelled using ing back to the pure-state N -representable manifold was
a quantum computer. It is known that Hartree-Fock the- evidence that generalized pure-state N -representability
ory reverses the order of the transition states; however, conditions would be instrumental in making NISQ chem-
here we focused on the accuracy of the computation with istry computations feasible. This fact underscores the
6
importance of developing procedures for applying pure- However, we find the accuracy of these experiments and
state N -representability conditions in a more general con- the effectiveness of these error-mitigation procedures to
text. The post-selection and RDM measurement tech- be an encouraging signal of progress in that direction.
niques can be generalized to measuring all 1-RDM and 2-
RDM elements when considering a less restrictive circuit
ansatz by permuting the labels of the fermionic modes. ACKNOWLEDGEMENTS
For ansatz such as the generalized swap network [28] the
circuit structure would not change, only the rotation an- D.B. is a CIFAR Associate Fellow in the Quantum In-
gles. Thus, the measurement schemes presented here are formation Science Program. Funding: This work was
applicable in the more general case. Furthermore, it is supported by Google. Competing Interests: The au-
important to understand the performance of these error thors declare no competing interests. Author Contri-
mitigation techniques when combined with alternatives butions: N.C.R. designed the experiment. C.N. assisted
such as noise extrapolation [37]. with data collection. Z.J., V.S., and N.W. assisted with
Finally, we were able to show further evidence that analytical calculations and gate synthesis. N.C.R. and
variational relaxation effectively mitigates coherent er- R.B. wrote the paper. Experiments were performed us-
rors arising in implementation of physical gates. The ing a quantum processor that was recently developed and
performance of our problem specific optimization strat- fabricated by a large effort involving the entire Google
egy motivates the study of iterative wavefunction con- Quantum team. Data and materials availability:
structions [38] in a more general setting. The combi- The code used for this experiment and a tutorial for run-
nation of these error mitigation techniques with VQE ning it can be found in the open source library Recirq,
unambiguously resolved a chemical mechanism to within located at https://github.com/quantumlib/ReCirq/
the model chemistry using a quantum computation. It tree/master/recirq. All data needed to evaluate the
is still an open question whether NISQ devices will be conclusions in the paper are present in the paper or the
able to simulate challenging quantum chemistry systems Supplementary Materials. Data presented in the figures
and it is likely that major innovations would be required. can be found in the Dryad repository located at [39]
Google AI Quantum and Collaborators

Frank Arute1 , Kunal Arya1 , Ryan Babbush1 , Dave Bacon1 , Joseph C. Bardin1, 2 , Rami Barends1 , Sergio Boixo1 ,
Michael Broughton1 , Bob B. Buckley1 , David A. Buell1 , Brian Burkett1 , Nicholas Bushnell1 , Yu Chen1 , Zijun
Chen1 , Benjamin Chiaro1, 3 , Roberto Collins1 , William Courtney1 , Sean Demura1 , Andrew Dunsworth1 , Daniel
Eppens1 , Edward Farhi1 , Austin Fowler1 , Brooks Foxen1 , Craig Gidney1 , Marissa Giustina1 , Rob Graff1 , Steve
Habegger1 , Matthew P. Harrigan1 , Alan Ho1 , Sabrina Hong1 , Trent Huang1 , William J. Huggins1, 4 , Lev Ioffe1 ,
Sergei V. Isakov1 , Evan Jeffrey1 , Zhang Jiang1 , Cody Jones1 , Dvir Kafri1 , Kostyantyn Kechedzhi1 , Julian Kelly1 ,
Seon Kim1 , Paul V. Klimov1 , Alexander Korotkov1, 5 , Fedor Kostritsa1 , David Landhuis1 , Pavel Laptev1 , Mike
Lindmark1 , Erik Lucero1 , Orion Martin1 , John M. Martinis1, 3 , Jarrod R. McClean1 , Matt McEwen1, 3 , Anthony
Megrant1 , Xiao Mi1 , Masoud Mohseni1 , Wojciech Mruczkiewicz1 , Josh Mutus1 , Ofer Naaman1 , Matthew Neeley1 ,
Charles Neill1 , Hartmut Neven1 , Murphy Yuezhen Niu1 , Thomas E. O’Brien1 , Eric Ostby1 , Andre Petukhov1 ,
Harald Putterman1 , Chris Quintana1 , Pedram Roushan1 , Nicholas C. Rubin1 , Daniel Sank1 , Kevin J. Satzinger1 ,
Vadim Smelyanskiy1 , Doug Strain1 , Kevin J. Sung1, 6 , Marco Szalay1 , Tyler Y. Takeshita7 , Amit Vainsencher1 ,
Theodore White1 , Nathan Wiebe1, 8, 9 , Z. Jamie Yao1 , Ping Yeh1 , Adam Zalcman1
1
Google Research
2
Department of Electrical and Computer Engineering, University of Massachusetts, Amherst, MA
3
Department of Physics, University of California, Santa Barbara, CA
4
Department of Chemistry, University of California, Berkeley, CA
5
Department of Electrical and Computer Engineering, University of California, Riverside, CA
6
Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor, MI
7
Mercedes-Benz Research and Development, North America, Sunnyvale, CA
8
Department of Physics, University of Washington, Seattle, WA
9
Pacific Northwest National Laboratory, Richland, WA
[1] Alan Aspuru-Guzik, Anthony D Dutoi, Peter J Love, tation of Molecular Energies,” Science 309, 1704 (2005).
and Martin Head-Gordon, “Simulated Quantum Compu-
7
[2] Frank Arute, Kunal Arya, Ryan Babbush, Dave Bacon, insulator transition,” Physical Review A 100, 022517
Joseph C. Bardin, Rami Barends, Rupak Biswas, Sergio (2019).
Boixo, Fernando G. S. L. Brandao, David A. Buell, Brian [10] R. Sagastizabal, X. Bonet-Monroig, M. Singh, M. A.
Burkett, Yu Chen, Zijun Chen, Ben Chiaro, Roberto Rol, C. C. Bultink, X. Fu, C. H. Price, V. P. Os-
Collins, William Courtney, Andrew Dunsworth, Ed- troukh, N. Muthusubramanian, A. Bruno, M. Beekman,
ward Farhi, Brooks Foxen, Austin Fowler, Craig Gidney, N. Haider, T. E. O’Brien, and L. DiCarlo, “Experimen-
Marissa Giustina, Rob Graff, Keith Guerin, Steve Habeg- tal error mitigation via symmetry verification in a vari-
ger, Matthew P. Harrigan, Michael J. Hartmann, Alan ational quantum eigensolver,” Physical Review A 100,
Ho, Markus Hoffmann, Trent Huang, Travis S. Humble, 010302 (2019).
Sergei V. Isakov, Evan Jeffrey, Zhang Jiang, Dvir Kafri, [11] David J Thouless, “Stability conditions and nuclear ro-
Kostyantyn Kechedzhi, Julian Kelly, Paul V. Klimov, tations in the Hartree-Fock theory,” Nuclear Physics 21,
Sergey Knysh, Alexander Korotkov, Fedor Kostritsa, 225–232 (1960).
David Landhuis, Mike Lindmark, Erik Lucero, Dmitry [12] Guang Hao Low and Nathan Wiebe, “Hamiltonian Sim-
Lyakh, Salvatore Mandrà, Jarrod R. McClean, Matthew ulation in the Interaction Picture,” arXiv:1805.00675
McEwen, Anthony Megrant, Xiao Mi, Kristel Michielsen, (2018).
Masoud Mohseni, Josh Mutus, Ofer Naaman, Matthew [13] Ryan Babbush, Dominic W. Berry, Jarrod R. McClean,
Neeley, Charles Neill, Murphy Yuezhen Niu, Eric Os- and Hartmut Neven, “Quantum Simulation of Chemistry
tby, Andre Petukhov, John C. Platt, Chris Quintana, with Sublinear Scaling in Basis Size,” npj Quantum In-
Eleanor G. Rieffel, Pedram Roushan, Nicholas C. Ru- formation 5, 92 (2019).
bin, Daniel Sank, Kevin J. Satzinger, Vadim Smelyan- [14] Andrew M Childs, Yuan Su, Minh C Tran, Nathan
skiy, Kevin J. Sung, Matthew D. Trevithick, Amit Wiebe, and Shuchen Zhu, “A theory of trotter error,”
Vainsencher, Benjamin Villalonga, Theodore White, arXiv:1912.08854 (2019).
Z. Jamie Yao, Ping Yeh, Adam Zalcman, Hartmut Neven, [15] Ryan Babbush, Craig Gidney, Dominic Berry, Nathan
and John M. Martinis, “Quantum supremacy using a Wiebe, Jarrod McClean, Alexandru Paler, Austin
programmable superconducting processor,” Nature 574, Fowler, and Hartmut Neven, “Encoding Electronic Spec-
505–510 (2019). tra in Quantum Circuits with Linear T Complexity,”
[3] Alberto Peruzzo, Jarrod McClean, Peter Shadbolt, Man- Physical Review X 8, 041015 (2018).
Hong Yung, Xiao-Qi Zhou, Peter J Love, Alan Aspuru- [16] Ryan Babbush, Nathan Wiebe, Jarrod McClean, James
Guzik, and Jeremy L O’Brien, “A Variational Eigenvalue McClain, Hartmut Neven, and Garnet Kin-Lic Chan,
Solver on a Photonic Quantum Processor,” Nature Com- “Low-Depth Quantum Simulation of Materials,” Physi-
munications 5, 1–7 (2014). cal Review X 8, 011044 (2018).
[4] Abhinav Kandala, Antonio Mezzacapo, Kristan Temme, [17] Steven R White, “Hybrid grid/basis set discretizations
Maika Takita, Markus Brink, Jerry M Chow, and of the Schrödinger equation,” The Journal of Chemical
Jay M Gambetta, “Hardware-efficient variational quan- Physics 147, 244102 (2017).
tum eigensolver for small molecules and quantum mag- [18] Jarrod R McClean, Fabian M Faulstich, Qinyi Zhu,
nets,” Nature 549, 242–246 (2017). Bryan O’Gorman, Yiheng Qiu, Steven R White,
[5] P J J O’Malley, R Babbush, I D Kivlichan, J Romero, Ryan Babbush, and Lin Lin, “Discontinuous galerkin
J R McClean, R Barends, J Kelly, P Roushan, A Tran- discretization for quantum simulation of chemistry,”
ter, N Ding, B Campbell, Y Chen, Z Chen, B Chiaro, arXiv:1909.00028 (2019).
A Dunsworth, A G Fowler, E Jeffrey, A Megrant, J Y [19] Mario Motta, Erika Ye, Jarrod R. McClean, Zhendong Li,
Mutus, C Neill, C Quintana, D Sank, A Vainsencher, Austin J. Minnich, Ryan Babbush, and Garnet Kin-Lic
J Wenner, T C White, P V Coveney, P J Love, H Neven, Chan, “Low Rank Representations for Quantum Simula-
A Aspuru-Guzik, and J M Martinis, “Scalable Quantum tion of Electronic Structure,” arXiv:1808.02625 (2018).
Simulation of Molecular Energies,” Physical Review X 6, [20] Dominic Berry, Craig Gidney, Mario Motta, Jarrod Mc-
31007 (2016). Clean, and Ryan Babbush, “Qubitization of Arbitrary
[6] Cornelius Hempel, Christine Maier, Jonathan Romero, Basis Quantum Chemistry Leveraging Sparsity and Low
Jarrod McClean, Thomas Monz, Heng Shen, Petar Ju- Rank Factorization,” Quantum 3, 208 (2019).
rcevic, Ben P. Lanyon, Peter Love, Ryan Babbush, Alán [21] William J Huggins, Jarrod McClean, Nicholas Rubin,
Aspuru-Guzik, Rainer Blatt, and Christian F. Roos, Zhang Jiang, Nathan Wiebe, K Birgitta Whaley, and
“Quantum chemistry calculations on a trapped-ion quan- Ryan Babbush, “Efficient and noise resilient measure-
tum simulator,” Physical Review X 8, 031022 (2018). ments for quantum chemistry on near-term quantum
[7] Abhinav Kandala, Kristan Temme, Antonio D Córcoles, computers,” arXiv:1907.13117 (2019).
Antonio Mezzacapo, Jerry M Chow, and Jay M Gam- [22] Mario Motta, David M. Ceperley, Garnet Kin-Lic Chan,
betta, “Error mitigation extends the computational reach John A. Gomez, Emanuel Gull, Sheng Guo, Carlos A.
of a noisy quantum processor,” Nature 567, 491–495 Jiménez-Hoyos, Tran Nguyen Lan, Jia Li, Fengjie Ma,
(2019). Andrew J. Millis, Nikolay V. Prokof’ev, Ushnish Ray,
[8] James I Colless, Vinay V Ramasesh, Dar Dahlen, Gustavo E. Scuseria, Sandro Sorella, Edwin M. Stouden-
Machiel S Blok, Jarrod R McClean, Jonathan Carter, mire, Qiming Sun, Igor S. Tupitsyn, Steven R. White,
Wibe A de Jong, and Irfan Siddiqi, “Robust Determi- Dominika Zgid, and Shiwei Zhang (Simons Collaboration
nation of Molecular Spectra on a Quantum Processor,” on the Many-Electron Problem), “Towards the solution
Physical Review X 8, 011021 (2018). of the many-electron problem in real materials: Equation
[9] Scott E. Smart and David A. Mazziotti, “Quantum- of state of the hydrogen chain with state-of-the-art many-
classical hybrid algorithm using an error-mitigating n- body methods,” Physical Review X 7, 031059 (2017).
representability condition to compute the mott metal- [23] Peter A Limacher, Paul W Ayers, Paul A Johnson, Stijn
8
De Baerdemacker, Dimitri Van Neck, and Patrick Bult- mura, Andrew Dunsworth, Edward Farhi, Austin Fowler,
inck, “A new mean-field method suitable for strongly cor- Brooks Foxen, Craig Gidney, Marissa Giustina, Rob
related electrons: Computationally facile antisymmetric Graff, Steve Habegger, Matthew Harrigan, Alan Ho,
products of nonorthogonal geminals,” Journal of chemi- Sabrina Hong, Trent Huang, William Huggins, Lev
cal theory and computation 9, 1394–1401 (2013). Ioffe, Sergei Isakov, Evan Jeffrey, Zhang Jiang, Cody
[24] Johannes Hachmann, Wim Cardoen, and Garnet Kin- Jones, Dvir Kafri, Kostyantyn Kechedzhi, Julian Kelly,
Lic Chan, “Multireference correlation in long molecules Seon Kim, Paul Klimov, Alexander Korotkov, Fedor
with the quadratic scaling density matrix renormaliza- Kostritsa, David Landhuis, Pavel Laptev, Mike Lind-
tion group,” The Journal of chemical physics 125, 144101 mark, Erik Lucero, Orion Martin, John Martinis, Jarrod
(2006). McClean, Matt McEwen, Anthony Megrant, Xiao Mi,
[25] Rajat K Chaudhuri, Karl F Freed, Sudip Chattopadhyay, Masoud Mohseni, Wojciech Mruczkiewicz, Josh Mutus,
and Uttam Sinha Mahapatra, “Potential energy curve Ofer Naaman, Matthew Neeley, Charles Neill, Hartmut
for isomerization of N2 H2 and C2 H4 using the improved Neven, Murphy Yuezhen Niu, Thomas O’Brien, Eric Os-
virtual orbital multireference moller-plesset perturbation tby, Andre Petukhov, Harald Putterman, Chris Quin-
theory,” The Journal of Chemical Physics 128, 144304 tana, Pedram Roushan, Nicholas Rubin, Daniel Sank,
(2008). Kevin Satzinger, Vadim Smelyanskiy, Doug Strain, Kevin
[26] Michael Reck, Anton Zeilinger, Herbert J. Bernstein, Sung, Marco Szalay, Tyler Takeshita, Amit Vainsencher,
and Philip Bertani, “Experimental realization of any dis- Theodore White, Nathan Wiebe, Z. Jamie Yao, Ping
crete unitary operator,” Physical Review Letters 73, 58– Yeh, and Adam Zalcman, “Hartree-fock on a supercon-
61 (1994). ducting qubit quantum computer,” (2020).
[27] Dave Wecker, Matthew B Hastings, Nathan Wiebe, [40] Attila Szabo and Neil S Ostlund, Modern quantum chem-
Bryan K Clark, Chetan Nayak, and Matthias Troyer, istry: introduction to advanced electronic structure theory
“Solving strongly correlated electron models on a quan- (Courier Corporation, 2012).
tum computer,” Physical Review A 92, 62318 (2015). [41] Zhang Jiang, Kevin J. Sung, Kostyantyn Kechedzhi,
[28] Ian D Kivlichan, Jarrod McClean, Nathan Wiebe, Craig Vadim N. Smelyanskiy, and Sergio Boixo, “Quantum
Gidney, Alan Aspuru-Guzik, Garnet Kin-Lic Chan, algorithms to simulate many-body physics of correlated
and Ryan Babbush, “Quantum Simulation of Electronic fermions,” Physical Review Applied 9, 044036 (2018).
Structure with Linear Depth and Connectivity,” Physical [42] Anmer Daskin and Sabre Kais, “Decomposition of uni-
Review Letters 120, 110501 (2018). tary matrices for finding quantum circuits: application
[29] R. McWeeny, “Some recent advances in density matrix to molecular hamiltonians,” The Journal of chemical
theory,” Reviews of Modern Physics 32, 335–369 (1960). physics 134, 144112 (2011).
[30] A. J. Coleman, “Structure of fermion density matrices,” [43] Tyler Takeshita, Nicholas C. Rubin, Zhang Jiang, Eun-
Reviews of Modern Physics 35, 668–686 (1963). seok Lee, Ryan Babbush, and Jarrod R. McClean, “In-
[31] David A. Mazziotti, “Pure-n-representability conditions creasing the representation accuracy of quantum simu-
of two-fermion reduced density matrices,” Phys. Rev. A lations of chemistry without extra quantum resources,”
94, 032516 (2016). Physical Review X 10, 011004 (2020).
[32] Nicholas C. Rubin, Ryan Babbush, and Jarrod McClean, [44] Alexander J McCaskey, Zachary P Parks, Jacek
“Application of fermionic marginal constraints to hybrid Jakowski, Shirley V Moore, Titus D Morris, Travis S
quantum algorithms,” New Journal of Physics 20, 053020 Humble, and Raphael C Pooser, “Quantum chemistry
(2018). as a benchmark for near-term quantum computers,” npj
[33] Alexander A Klyachko, “Quantum marginal problem and Quantum Information 5, 1–8 (2019).
n-representability,” in Journal of Physics: Conference [45] Roger A Horn and Charles R Johnson, Matrix analysis
Series, Vol. 36 (IOP Publishing, 2006) p. 72. (Cambridge university press, 2012).
[34] Werner Kutzelnigg, “Generalized k-particle brillouin con- [46] RM Wilcox, “Exponential operators and parameter dif-
ditions and their use for the construction of correlated ferentiation in quantum physics,” Journal of Mathemat-
electronic wavefunctions,” Chemical Physics Letters 64, ical Physics 8, 962–982 (1967).
383–387 (1979). [47] Trygve Helgaker, Poul Jorgensen, and Jeppe Olsen,
[35] Qiming Sun, “Co-iterative augmented hessian method for Molecular Electronic Structure Theory (Wiley, 2002).
orbital optimization,” arXiv:1610.08423 (2016). [48] P. V. Klimov, J. Kelly, Z. Chen, M. Neeley, A. Megrant,
[36] M. Gluza, M. Kliesch, J. Eisert, and L. Aolita, “Fidelity B. Burkett, R. Barends, K. Arya, B. Chiaro, Yu Chen,
witnesses for fermionic quantum simulations,” Physical A. Dunsworth, A. Fowler, B. Foxen, C. Gidney,
Review Letters 120, 190501 (2018). M. Giustina, R. Graff, T. Huang, E. Jeffrey, Erik
[37] Kristan Temme, Sergey Bravyi, and Jay M. Gam- Lucero, J. Y. Mutus, O. Naaman, C. Neill, C. Quintana,
betta, “Error mitigation for short-depth quantum cir- P. Roushan, Daniel Sank, A. Vainsencher, J. Wenner,
cuits,” Phys. Rev. Lett. 119, 180509 (2017). T. C. White, S. Boixo, R. Babbush, V. N. Smelyan-
[38] Harper R Grimsley, Sophia E Economou, Edwin Barnes, skiy, H. Neven, and John M. Martinis, “Fluctuations
and Nicholas J Mayhall, “An adaptive variational algo- of energy-relaxation times in superconducting qubits,”
rithm for exact molecular simulations on a quantum com- Phys. Rev. Lett. 121, 090502 (2018).
puter,” Nature communications 10, 1–9 (2019). [49] Jarrod R McClean, Kevin J Sung, Ian D Kivlichan,
[39] Frank Arute, Kunal Arya, Ryan Babbush, Dave Bacon, Yudong Cao, Chengyu Dai, E Schuyler Fried, Craig Gid-
Joseph Bardin, Rami Barends, Sergio Boixo, Michael ney, Brendan Gimby, Thomas Häner, Tarini Hardikar,
Broughton, Bob B. Buckley, David Buell, Brian Bur- Vojtch Havlı́ček, Oscar Higgott, Cupjin Huang, Josh
kett, Nicholas Bushnell, Yu Chen, Zijun Chen, Ben Izaac, Zhang Jiang, Xinle Liu, Sam McArdle, Matthew
Chiaro, Roberto Collins, William Courtney, Sean De- Neeley, Thomas O’Brien, Bryan O’Gorman, Isil Ozfidan,
9
Maxwell D Radin, Jhonathan Romero, Nicholas Rubin, drew M James, Harley R McAlexander, Ashutosh Ku-
Nicolas P. D. Sawaya, Kanav Setia, Sukin Sim, Damian S mar, Masaaki Saitow, Xiao Wang, Benjamin P Pritchard,
Steiger, Mark Steudtner, Qiming Sun, Wei Sun, Daochen Prakash Verma, Henry F Schaefer, Konrad Patkowski,
Wang, Fang Zhang, and Ryan Babbush, “OpenFermion: Rollin A King, Edward F Valeev, Francesco A Evange-
The Electronic Structure Package for Quantum Comput- lista, Justin M Turney, T Daniel Crawford, and C David
ers,” arXiv:1710.07629 (2017). Sherrill, “Psi4 1.1: An Open-Source Electronic Structure
[50] Robert M Parrish, Lori A Burns, Daniel G A Smith, An- Program Emphasizing Automation, Advanced Libraries,
drew C Simmonett, A Eugene DePrince, Edward G Ho- and Interoperability,” Journal of Chemical Theory and
henstein, Uur Bozkaya, Alexander Yu. Sokolov, Roberto Computation 13, 3185–3197 (2017).
Di Remigio, Ryan M Richard, Jrme F Gonthier, An-
10
Appendix A: Hartree-Fock Theory via Canonical Transformations
In this section we derive Hartree-Fock theory from the perspective of canonical transformations. This derivation
follows an original work by David Thouless [11] and is reproduced here due to its foundational importance to the
formulation of this experiment. In Hartree-Fock theory one attempts to solve the time-independent Schrödinger
equation using a state ansatz that is an antisymmetrized product of one-particle functions. Starting from an arbitrary
orthogonal basis {φi } the goal is to variationally optimize the wavefunction
|ψ(r1 , ..., rn )i = (n!)−1/2 An (χ1 (r1 )...χn (rn )) (A1)
where An is the antisymmetrizer and χi (r) = j cji φj (r) in terms of the coefficients for χ. This antisymmetrized
P
product of one-particle functions is commonly expressed in a more compact form as a determinant of a matrix whose
elements are the functions χi (rj ) with i indexing the column and j indexing the row of this matrix. This representation
of the antisymmetrized product through a determinant is why this wavefunction ansatz is commonly referred to as a
Slater determinant.
The variational principle for the Schrödinger equation can be stated as
hδψ|H|ψi = 0 (A2)
which is a statement that the energy is stationary with respect to first order changes in the wavefunction. In second
quantization a single antisymmetrized product of orbitals corresponds to a product of ladder operators acting on the
vacuum to “create” a representation of the antisymmetrized wavefunction
 
n χ1 (r1 ) ... χ1 (rn )
1
ai |0i = √ Det  ... .. ..  .
Y †
hr|ψi =hr| (A3)

. . 
i=1
n!
χn (r1 ) ... χn (rn )
Assuming we are working in a fixed particle manifold and given the aforementioned complete set of one-particle
functions is used as a basis we can index the functions used in the product wavefunction by i and those not used are
labeled by a then any change in the wavefunction is generated by
hδψ| = hψ|a†i aa ζ (A4)
where ζ is the first order change to an orbital χi . This fact is because any unitary generator that has only indices
{a} or {i} merely changes the phase on the state and thus is not observable [40]. Evaluating Eq. (A2) one arrives at
an expression for the stationarity of the state
hψ|a†i aa H|ψi = 0. (A5)
All the quantities in Eq. (A5) can be evaluated using Wick’s theorem given the initial state ψ is a product state
and ar |0i = 0. This variational condition naturally leads to the self-consistent-field Hamiltonian commonly derived
through a Lagrangian technique for the Hartree-Fock equations. In order to design a VQE style approach to solving the
Hartree-Fock equations we take a different approach that leverages the fact that we can determine any basis rotation
through a linear-depth quantum circuit. Thouless demonstrated [11] that any non-orthogonal product wavefunction
can be obtained from a product wavefunction by a unitary generated by one-body fermionic operators of the form a†p aq .
The underlying reason for why this fact is true is that the one-body fermionic generators form a closed Lie-algebra.
Given,
†
ap aq , a†r as = δq,r a†p as − δp,s a†r as

(A6)
the adjoint representation of any element of the algebra κ where

X
κ= κp,q a†p aq , (A7)
p,q
and its commutator with any other element can be efficiently represented as matrix that is m × m where m is the
number of fermionic modes.
κ, a†p = a†q κp,q , [κ, ap ] = aq κ∗p,q

(A8)
11
Using the BCH expansion, we can express the similarity transformed ladder operators as
X X
eK a†p e−K = a†q uq,p , eK ap e−K = aq u∗q,p (A9)
q q
where u is the matrix given by the exponentiation of the coefficient matrix for the generator operator κ
u = eκ (A10)
which is the proof for Eq. (1). Any rotation of the underlying basis can now be represented as a similarity transfor-
mation of each fermionic mode
|φ(κ)i = eK a†1 e−K ...eK a†n e−K |0i = eK |ψi. (A11)
Thus any non-orthogonal state can be generated by implementing eK as a circuit acting on an initial product state.
Given the Hartree-Fock wavefunction ansatz the energy is given by
E(κ) = hφ(κ)|H|φ(κ)i = hψ|eK He−K |ψi. (A12)
With the energy expressed in the form of Eq. (A12) it is not immediately clear that it can be evaluated without
knowledge of the 2-RDM. To see this fact, we used the BCH expansion and notice that all nested commutator terms
involve a†p aq -like terms and the original Hamiltonian. The commutator of a two-mode number conserving fermionic
operator with a four-mode number conserving fermionic operator produces a linear combination of four four-mode
number conserving fermionic operators. Therefore, all terms in the expansion can be evaluated with knowledge of only
the 2-RDM. If we start with a product state defined from an orthogonal set of states the 2-RDM can be constructed
directly from the 1-RDM [29]
1
Dij =hφ|a†j ai |φi
pq
2
Dij =hφ|a†p a†q aj ai |φi = 1 Dip 1 Djq − 1 Diq 1 Djp . (A13)
This expression also demonstrates that we only need to measure the 1-RDM to evaluate the energy. It is important to
note that the reconstruction of the 2-RDM from the 1-RDM described in Eq. (4) is only exact for Slater determinant
wavefunctions. The energy is evaluated as a function of the 1- and 2-RDM by
hij hφ(κ)|a†i aj |φ(κ)i + Vijkl hφ(κ)|a†i a†j ak al |φ(κ)i =
X X X X ij
E(κ) = hij 1 Dji + Vijkl 2 Dlk (A14)
ij ijkl ij ijkl
where hij and Vijkl

!
ZA
Z X
hi,j = drχ∗i (r) 2
−∇ (r) + χj (r) (A15)
|r − RA |
A
1
Z Z
i,j
dr1 dr2 χ∗i (r1 )χ∗j (r2 ) |r1 − r2 |−1 χk (r2 )χl (r1 )

Vl,k = (A16)
2
are the molecular integrals in the original basis. These orbitals are determined by diagonalizing the matrix of one-body
integrals hij = [h]ij described in the STO-3G atomic basis. In summary, to measure the energy of our system given
basis rotation circuit ansatz we need the following steps:
1 Measure the entire 1-RDM.
2 Compute the 2-RDM by evaluating. Eq. (A13)
3 Compute the energy by evaluating. Eq. (A14)
1. Classical simulation of non-interacting fermion circuits
Given a particular set of parameters {κp,q } the 1-RDM resulting from a wavefunction ψ = U (κ)φ, where φ is an
initial product state, is
up,i a†i up,i u∗q,j hφ|a†i aj |φi.
X X X
D̃q =hφ|e−K a†p eK e−K ap eK |φi = hφ|
1 p
u∗q,j aj |φi = (A17)
j q,j ij
With this 1-RDM one can evaluate the energy and gradients with respect to κp,q . This expression requires two matrix
multiplications to evaluate along with the 1-RDM of the starting state.
12
Appendix B: Implementing the Basis Change Circuit and Circuit Concatenation
In order to implement the basis rotation circuits we leverage a number of recent works that provide asymptotically
optimal circuit compilations. We review a circuit construction that is analogous to a QR decomposition as motivation
before highlighting the salient features of the optimal circuit compilation. The basis rotation circuit is first expressed
in fermionic modes which we then provide a compilation to the gate set used in this work. Our goal is to implement
a unitary corresponding to
κi,j a†i aj
X
U (eκ ) = eK K= κ† = −κ. (B1)
i,j
Not all terms in K commute and thus naively one would expect an approximate method such as Trotterization to
be required. In Reference [28] the connection of the QR decomposition of eκ via Givens rotation to the sequence of
untiaries Rpq (u)
† †
R(u)pq = elog[u]pq (ap aq −aq ap ) (B2)
was established allowing for the exact evolution of the one-body component of the Hamiltonian without Trotter error
and a circuit to implement any basis rotation–i.e. any Slater determinant. A distinctive feature of one-body rotations
is that the map U (eκ ) is a homomorphism under matrix multiplication
0 0
U (eκ ) · U (eκ ) = U (eκ · eκ ) (B3)
We use this homomorphism through the observation that
Rpq (θ)U (u) = U (rpq (θ)u) (B4)
where
 
1 ... 0 ...... 0 0
 .. .. .. ..
.. 
.
 . . .
.
0 ... cos(θ) ... − sin(θ) ... 0
=  ... .. .. .. 
 
r(θ)p,q (B5)

 . . . 
0 ... sin(θ) ... cos(θ) ... 0
 .. .. .. . . .. 
 
. . . . .
0 ... 0 ... 0 ... 1
which given an appropriate selection of a sequence of rp,q (θ) brings u into diagonal form
Y X † X
Rk (θk )U (u) = e−iφp ap ap = e−iφp |pihp| (B6)
k p p
The sequence of Rk (θk ) can be determined by a QR decomposition of the matrix u. This fact was first recognized by
Reck [26] and used in a variety of quantum optics experiments to implement universal unitary operations–limited to
unitaries associated with one-body fermionic Hamiltonians. Jiang et. al and Kivlichan et. al [28, 41] point out that
in a fixed particle manifold the circuit depth can be further minimized. This fact is clearly shown by considering the
state in the basis that is being prepared through the Givens rotation network and back transforming to the original
basis
η η η X
ã†i |vaci = e−K ã†i eK |vaci =
Y Y Y
|ψ(κ)i = [eκ ]i,p a†p |vaci (B7)
i=1 i=1 i=1 p
we only need the first η-columns of the matrix [eκ ]. Therefore, we can focus on Givens network elimination on these
columns. Jiang et al. provide a further circuit minimization by noting that any rotation amongst the occupied orbitals
merely shifts the observable by a global phase. Given a unitary V
η X
η η
Vi,j ã†i |vaci = det [V ] ã†i |vaci
Y Y
(B8)
i=1 j=1 i=1
13
where the det [V ] is a phase and thus not observable. The V can be chosen such that the lower left triangle or
eκ are zeroed out by Givens rotations. In chemistry parlance, this transformation is called an occupied-occupied
orbital rotations and is known to be a redundant rotation. For restricted Hartree-Fock the number of non-redundant
parameters in κ is equal to the number of occupied spatial orbitals times the number of virtual orbitals. We also note
that this decomposition is exact and likely asymptotically optimal. While the authors of [28] argue that in terms of
gate count their Givens rotation circuits are likely optimal, we note that approximate unitary constructions such as
those in [42] may provide a route to approximating the compilation of similar circuits with even fewer gates.
X 1.0534π 1.0π
X 1.5π 1.0π -0.053π 1.5π 1.0π
X 1.3280π 1.0π -0.5π 0.9916π 1.0π -0.5π 0.9403π 1.0π
X 1.5π 1.0π) -0.328π 1.5π 1.0π 0.0083π 1.5π 1.0π 0.0596π 1.5π 1.0π
-0.5π 0.7960π 1.0π -0.5π 1.1698π 1.0π -0.5π 0.9257π 1.0π -0.5π
0.2039π 1.5π 1.0π -0.169π 1.5π 1.0π 0.0742π
-0.5π 0.6931π 1.0π -0.5π
0.3068π
FIG. 4. Givens rotation

√ circuit for H8 simulating a random basis transformation in the half filling sector. Each Givens rotation
is compiled into iswap (hexagon two-qubit gates) and Rz gates (square gates with an angle depicted).
An example of an eight qubit half-filling circuit is given in Fig. S4. When we are away from half filling the nice
symmetry of the circuit is lost. For example, Fig. S5 is Diazene which has 8-electrons in 12 orbitals.
X 1.2489π 1.0π
X 0.7830π 1.0π -0.248π 1.0213π 1.0π
X 0.9853π 1.0π 0.2169π 0.8735π 1.0π -0.021π 1.1037π 1.0π
X 1.4994π 1.0π 0.0146π 1.2810π 1.0π 0.1264π 0.9626π 1.0π -0.103π 0.5177π 1.0π
X 1.4525π 1.0π -0.499π 0.5004π 1.0π -0.281π 0.9549π 1.0π 0.0373π 1.1683π 1.0π 0.4822π
X 0.5803π 1.0π -0.452π 1.4300π 1.0π 0.4995π 1.4951π 1.0π 0.0450π 1.0048π 1.0π -0.168π
X 0.6504π 1.0π 0.4196π 1.4417π 1.0π -0.430π 1.4899π 1.0π -0.495π 0.9922π 1.0π -0.004π
X 0.7146π 1.0π 0.3495π 1.2264π 1.0π -0.441π 0.5934π 1.0π -0.489π 0.5000π 1.0π 0.0077π
0.2853π 0.5532π 1.0π -0.226π 0.5449π 1.0π 0.4065π 0.5000π 1.0π 0.4999π
0.4467π 0.6895π 1.0π 0.4550π 0.5000π 1.0π 0.4999π
0.3104π 1.4999π 1.0π 0.4999π
-0.499π
FIG. 5. Givens rotation circuit for diazene prior to freezing the two lowest energy orbitals. Away from half filling the basis
rotations have a parallelogram structure.
Appendix C: Optimal Measurement of the 1-RDM
In this section we present a methodology that allows us to measure the 1-RDM in N + 1 measurement settings and
no additional quantum resources. We will also discuss a method√that allows us to perform post selection on all the
Monte Carlo averaged terms at the cost of an additional row of iswap gates at the end of the circuit. The 1-RDM
is an N × N hermitian positive semidefinite matrix with elements equal to the expectation values ha†i aj i where {i, j}
index the row and column of the matrix. The matrix of expectation values is depicted in Fig. S6. As a motivator for
our measurement protocol we start by describing circuits required to measure the diagonal elements of the 1-RDM of
a six qubit system at half filling–i.e. ha†i ai i.
1. Diagonal terms
Given a circuit U implementing the basis rotation eκ the diagonal elements of the 1-RDM are obtained by measuring
the Z expectation value on each qubit. The correspondence between a†i ai , measurement result Mi from qubit i, qubit
operators is derived using the Jordan-Wigner transform
I − hZi i
ha†i ai i = = hMi i (C1)
2
where Zi is the Z-qubit operator on qubit labeled i. The expectation value ha†i ai i is equivalent to the probability of
measuring a 1 bit on qubit i–i.e hMi i. Because we are measuring in the computational basis we can post-select on
the three excitations in the measurement result. This process is depicted in Fig. S6.
14
Post-select half filling
FIG. 6. Measurement circuit associated with estimating all diagonal elements of the 1-RDM simultaneously. The elements that
are acquired with this circuit are highlighted in red.
2. One-off-diagonal terms
The hermiticity of the 1-RDM demands that ha†i ai+1 i = ha†i+1 ai i∗ . The 1-RDM has no imaginary component
because we use an initial basis built from real valued orbitals and the basis rotation circuit implements an element
of SO(N )–i.e. the basis rotation circuit involves a unitary matrix with real values. Therefore, we only measure the
real part of all one-off-diagonal terms a†i ai+1 + a†i+1 ai which corresponds to 2<ha†i ai+1 i. Using the Jordan-Wigner
transform to map fermionic ladder operators to qubits
1
ha†i ai+1 + a†i+1 ai i =
(hXi Xi+1 i + hYi Yi+1 i) = 2<ha†i ai+1 i (C2)
2
we see that we must measure XX on all pairs and Y Y on all pairs. This measurement can be accomplished with two
circuits depicted in Fig. S7.
FIG. 7. The two circuits allowing for the measurement of all one-off-diagonal elements of the 1-RDM simultaneously. The teal
circuit involves performing an Ry rotation (to measure in the X basis) at the end of the circuit and the purple circuit contains
an Rx rotation (to measure in the Y basis). The 1-RDM elements that are acquired with these circuits are highlighted in
red. We label which pairs contribute to which expectation values with grey dashed lines. The thinner dashes are for the even
1-RDM pairs and the thicker dashes are for the odd 1-RDM pairs. Because Ry and Rx operations do not preserve particle
number we cannot post-select on total particle number with these measurement circuits.
3. General off-diagonal terms and virtual swapping
The label of each fermionic mode is an arbitrary choice, so we are free to reorder the labels such that measuring
nearest-neighbor pairs of qubits corresponds to measuring different off-diagonal 1-RDM elements. Every relabeling
of the qubits requires us to recompile the Givens rotation circuit. The structure of the circuit stays the same but
the rotation angles are different. In this section we describe how to recompute the Givens rotation angles based on a
new label ordering. Using the label sets {1, 3, 0, 5, 2, 4} and {3, 5, 1, 4, 0, 2} we are able to use the two measurement
circuits in Fig. S7 to measure the remaining off-diagonal 1-RDM elements.
Formally, we build the new qubit labels by virtually swapping fermionic modes at the end of the original circuit
implement eκ . We note that performing nearest-neighbor fermionic swaps between adjacent pairs twice (even swaps
and odd swaps) we obtain a new ordering of qubits. For example, consider six fermionic modes {0, 1, 2, 3, 4, 5}.
Performing a set of fermionic swaps on modes labeled {(0, 1), (2, 3), (4, 5)} followed by swaps on {(1, 2), (3, 4)} leaves
our mode ordering as {1, 3, 0, 5, 2, 4}. We can then perform X-Pauli and Y -Pauli measurements on each qubit to
recover expectation values associated with
{<(a†1 a3 + a†3 a1 ), <(a†3 a0 + a†0 a3 ), <(a†0 a5 + a†5 a0 ), <(a†5 a2 + a†2 a5 ), <(a†2 a4 + a†4 a2 )}. (C3)
15
T √
iswap
T†
FIG. 8. Two-mode fermionic fast Fourier transform that diagonalizes the XX + Y Y Hamiltonian.
This procedure can be repeated once more to measure all the required two-body fermionic correlators to construct
the 1-RDM. Though it appears that each new label set incurs additional circuit by requiring fermionic swaps between
neighboring modes we can exploit the fact that one-body fermionic swaps generated by exp(−iπfswap/2) where
fswap is
fswap = a†p aq + a†q ap − a†p ap − a†q aq . (C4)
This one-body permutation can be viewed as a basis rotation which can be concatenated with the original circuit at
no extra cost due to Eq. (B3). The swapping unitary simply shuffles the columns of eκ that is used to generate the
Givens rotation network. The same effect could have been achieved by relabeling the fermionic modes which would
have been equivalent to permuting the rows and columns of eκ . This relabeling technique can be applied beyond basis
rotation circuits. For example, one can relabel the fermionic modes of a generalized swap network such that different
sets of RDM elements can be measured as nearest-neighbor pairs. The same logic can be applied to k-RDM elements.
In conclusion we need N/2 circuits, where each of the N/2 circuits gets measured in two or three different ways, for
an N -qubit system to measure the 1-RDM. This is a quadratic improvement over the naive measurement scheme which
would require O(N 2 ) different measurement settings. To make this savings concrete we consider the number of Pauli
terms one would need to measure for a 12-qubit system. If no grouping is applied then there are 276 measurement
circuits. With greedy grouping considering locally commuting Pauli terms then there are 149 different measurement
circuits. With the measurement strategy outlined above we require 13 different circuits.
4. Off-diagonal terms with post-selection
The circuits depicted in Fig. S7 did not allow for post-selection because the rotations to measure in the X-basis
and Y -basis do not commute with the total number operator. In this section we design a basis rotation circuit
that commutes with the total number operator and diagonalizes the 12 (XX + Y Y ) Hamiltonian. The diagonal form
means that after performing the basis rotation we can measure in the computational basis to obtain expectation
values 12 hXX + Y Y i.
The circuit that diagonalizes 12 (XX + Y Y ) is described in Fig. S8 and is denoted UM below. Its commutation
with the total number operator can √ be easy seen by recognizing that the T -gate (Rz(π/4)) commutes with the total
number operator and so does the iswap. Applying UM to the 21 (XX + Y Y ) Hamiltonian
0 0 0 0 0 0 0 0
   
0 0 1 0 † 0 1 0 0
UM  U = (C5)
0 1 0 0 M 0 0 −1 0
0 0 0 0 0 0 0 0
transforms the operator into a diagonal representation. Given an ordered pair of qubits {a, a + 1} the last matrix
in (C5) is 12 (Za − Za+1 ) in qubit representation. Finally, we can relate the Z expectation values, the transformed
XX + Y Y expectation values, fermionic ladder operators, and binary measurements {Ma , Ma+1 } via
1 1 1
hUm a†a aa+1 + a†a+1 aa Um
† †
i = hUm (Xa Xa+1 + Ya Ya+1 ) Um i = hZa − Za+1 i = (Ma+1 − Ma ) . (C6)
2 2 2
The measurement circuit can only be applied to non-overlapping pairs and thus we can obtain estimates of Xa Xa+1 +
Ya Ya+1 for a values corresponding to even integers or a corresponding to odd integers. More concretely, we describe
this process in Fig. S9 for a six qubit problem. All experiments involved circuits that allowed for post-selection based
on total Hamming weight. The “raw” data indicates analysis of the resulting bitstrings without post-selection.
16
Post-select half filling
FIG. 9. Two circuit measuring the one-off-diagonal of the 1-RDM such that the total particle number can be measured
simultaneously. This circuit allows us to post select on the correct number of excitations in the measured bitstring. The top
circuit measures the even pairs and the bottom circuit measures the odd pairs. Local Z expectation values are measured on
all the qubits and used to construct the expecation value for ha†i ai+1 i.
5. Computing error bars for elements of the 1-RDM
We use two methods to estimate error bars for all quantities in our experiments. The procedures differ in how the
covariance between 1-RDM terms is estimated. In the first procedure, error bars are generated by estimating the
covariance between terms in the 1-RDM at the same time as the mean estimation. Mean values of off-diagonal 1-RDM
terms involve estimating the expectation values for (Za − Zb )/2. Therefore, the covariance between two off-diagonal
elements of the 1-RDM is

1 1 1
Cov (Za − Zb ) , (Zp − Zq ) = (Cov [Za , Zp ] − Cov [Za , Zq ] − Cov [Zb , Zp ] + Cov [Zb , Zq ]) (C7)
2 2 4
for all pair sets {(a, b), (p, q)} measured simultaneously. All quantities can be estimated from the simultaneous
measurement of all qubits. Therefore, for each circuit permutation we obtain two covariance matrix of size N/2 × N/2
and N/2 − 1 × N/2 − 1. For the circuit with no label permutation we also obtain the covariances for all a†i ai terms.
In the second procedure for estimating covariance matrices we assume we are sampling from a pure Gaussian state.
This assumption is applicable when the fidelity is high enough as any change to the covariance matrix would be a second
order effect. For these states the 2-RDM is exactly described by the 1-RDM and therefore all covariances between
the 1-RDM elements are perfectly defined by a non-linear function of the 1-RDM elements. For any wavefunction
ψ corresponding to the output of a basis rotation circuit the covariance of 1-RDM elements computed from such a
wavefunction are as follows:
h i
Cov a†i aj + a†j ai , a†p aq + a†q ap = Dqi δpj − Dqi Djp + Dpi δqj − Dpi Djq + Dqj δpi − Dqj Dip + Dpj δqi − Dpj Diq . (C8)
ψ
With the estimates of the covariances we are able to re-sample the 1-RDM assuming central-limit theorem statistics.
We use a multinomial distribution where the mean values are ha†σ(i) aσ(i+1) i and the covariance matrix of the multino-
mial distribution is obtained by dividing the estimates of the covariance matrix above by α× 250,000. α is a number
less than 1 reflecting the probability that a bitstring is rejected. α is estimated from prior N -qubit experiments. Once
the new 1-RDM is obtained it can be purified, used to estimate a fidelity witness, and compute the energy. For all
error bars we re-sample the 1-RDM 1000 times and compute a mean value and standard deviation from this set. All
quantities estimated are sensitive to the N -representability of the resampled 1-RDM. We use the fixed trace positive
projection described in [32] to ensure that each resampled 1-RDM is positive semidefinite and has the correct trace.
The correction procedure is only applied when the resampled 1-RDM has eigenvalues below zero.
17
Appendix D: Computing the fidelity and a fidelity witness from the 1-RDM
1. Fidelity Witness
The class of quantum circuits simulating non-interacting fermion dynamics have the special property that an efficient
fidelity witness can be derived. The formal derivation for general non-interacting fermion wavefunctions is described
in Ref. [36]. Here we adapt this result to the special case of particle conserving dynamics generated by one-body
fermionic generators. A fidelity witness is an observable that provides a strict lower bound to the fidelity for all input
states. The fidelity witness is efficient in the sense that for an L-qubit system only L2 expectation values are required
to evaluate the fidelity witness. Given that U is a unitary corresponding to a basis transformation circuit and |ωi is
the initial computational basis state corresponding to ω = (ω1 , ..., ωL ) any L-bit string which satisfies nj |ωi = ωj |ωi
for j = 1, ..., L allows us to define a basis state annihilator operator
L
X L
X L
X
n(ω) = [(1 − ωj )nj + ωj (I − nj )] = [nj − ωj nj + ωj I − ωj nj ] = [nj + ωj I − 2ωj nj ] (D1)
j=1 j=1 j=1
which satisfies n(ω) |ωi = 0. The computational basis state |ωi is the zero energy eigenstate of nω and any other
computational basis state an excitation from this state. The excitation energy is exactly the number of bits that are
different from ω for each Fock basis state which can be computed by summing the resulting bit string from the XOR
operation between the two Fock basis states being considered. The fidelty witness
W = U (I − nω ) U † (D2)
can be evaluated with knowledge of the measured 1-RDM. To relate the fidelity witness to the 1-RDM it is important
to note the following
h i
Tr U ρp U † a†i aj = uDu† i,j

(D3)
where D is the matrix of expectation values hρp , a†i aj i and u = eκ because any one-body rotation on the state ρp can
be equated to a similarity transform of the generating matrix for that one-body transformation. This logic is similar
to logic used in [43] which moved one-body basis rotations at the end of the circuit into the Hamiltonian as an error
mitigation technique. Using this relationship we can evaluate the fidelity witness with the following expression
L
X †
u Du j,j + ωj − 2ωj u† Du j,j

FW (ρp ) = 1 − (D4)
j=1
where D is the 1-RDM that is measured, u = eκ is the unitary rotation representing the new Slater determinant.
2. Fidelity
Given an idempotent 1-RDM and the basis rotation unitary u = eκ , the fidelity can be determined by the following
procedure:
1. Perform an eigen decomposition on the purified 1-RDM and use the eigenvectors associated with eigenvalues
equal to 1 as the columns of a unitary matrix v corresponding to the measured basis rotation.
2. Use the expression for the overlap between two basis rotation unitaries |hψu |ψv i|2 = |det v † u |2 to compute

the fidelity. The function det is the determinant of a matrix. This is the inner product between two Grassmann
representatives and is independent of choice of orbitals.
Appendix E: Error mitigation through purification
A distinctive feature of the Slater determinant wavefunction ansatz is that their 1-RDMs are idempotent matrices.
The manifold of states with idempotent 1-particle density matrices is significantly smaller than the space of possi-
ble wavefunctions. Thus our error mitigation strategy will rely on projecting the measured 1-RDM to the closest
idempotent 1-RDM. This projection procedure can be represented by the following mathematical program
min ||D − D̃|| (E1)
T r[D]=η,D0,D 2 =D
18
that seeks to determine a 1-RDM D that is close to the measured 1-RDM D̃ with has fixed trace, is positive semidef-
inite, and is a projector. A practical implementation of the the program in Eq. (E1) is challenging due to the
idempotency constraint. Instead of solving Eq. (E1) directly we rely on an iterative procedure that under mild
conditions projects a measured 1-RDM D̃ towards the set of idempotent matrices. This procedure is the McWeeny
purification commonly used in linear scaling electronic structure techniques [29] and is defined by the iteration
Dn+1 = 3Dn2 − 2Dn3 . (E2)
After each iteration the eigenvalues are closer to {0, 1}. Prior work [44] proposed to use McWeeny purification on the
2-RDM, but it is not clear what that accomplishes. This is because, in general, 2-RDMs are not idempotent matrices
and applying pure-state purification requires more general pure-state N -representability conditions [33]. Due to the
fact that McWeeny iteration has no effect on the eigenvectors, it merely pushes the eigenvalues of D towards {0, 1},
we could have achieved this projection by diagonalizing D and rounding the eigenvalues to 0 or 1. We performed the
purification iteration because we are able to analyze the convergence when D is obtained by sampling.
Here we will estimate the number of samples needed to ensure that the 1-RDMs can be faithfully reconstructed
within arbitrarily small error using our protocol. This analysis assumes we sample from a perfect state and thus our
goal is to provide evidence that McWeeny purification is convergent under sampling noise. Consider the purification
process in Eq. (E2). Now let us assume that the principal eigenvalue of D is Pk . In absentia of numerical error we
would have that Pk = 1 for Hartree-Fock theory. However, sampling error incurs an error in this eigenvalue such that
Pk = 1 + ∆, (E3)
2 k 2
where ∆ is a random variable with mean 0 and variance σ . Further, let µk = E(∆ ), where µ2 = σ for example.
Now given these quantities we wish to evaluate
E(Pk+1 ) = E(3Pk2 − 2Pk3 ) = 1 − 3σ 2 − 2µ3 . (E4)
Similarly we have that
2
E(Pk+1 ) = 1 − 6σ 2 − 4µ3 + 9µ4 + 12µ5 + 4µ6 . (E5)
This implies that the variance is
2
V(Pk+1 ) = E(Pk+1 ) − E(Pk+1 )2 = 9(µ4 − σ 4 ) + 12(µ5 − σ 2 µ3 ) + 4(µ6 − µ23 ). (E6)
Further, let us assume that µj ≤ αj σ j , for all j .
V(Pk+1 ) ≤ 9σ 4 (α4 − 1) + 12|σ|5 (α5 + α3 ) + 4|σ|6 α6 . (E7)
Assuming that σ ≤ 1 we have that
V(Pk+1 ) ≤ 9σ 4 (α4 − 1) + 4|σ|5 (α6 + 3α5 + 3α3 ). (E8)
It is clear from this recurrence relation that the variance for this method converges quadratically (assuming σ is
sufficiently small). Specifically, we have that V(PK ) ≤ for K ∈ O (log log(1/)) if appropriate convergence criteria
are met. A criterion for convergence is that 9σ 4 (α4 − 1) + 4|σ|5 (α6 + 3α5 + 3α3 ) ≤ σ 2 . This is guaranteed if,
 p 
1 β −β + β 2 + 36α − 36
4
σ2 ≤ 1 − , (E9)
9(α4 − 1) 18(α4 − 1)
where β = 4(α6 + 3α5 + 3α3 ).

The precise values of αj depend on the nature of the underlying distribution. However, if we assume that it is
Gaussian then we have that α2j+1 = 0 ∀ j, α4 = 3, α6 = 15. Furthermore, we have under these Gaussian assumptions
(for any σ > 0) that
V(Pk+1 ) ≤ 18σ 4 + 12σ 6 . (E10)
In this case, we find that the McWeeny iteration will converge if V(Pk+1 ) ≤ σ 2 which is implied by
√
2 −3 564
σ ≤ + ≈ 0.048. (E11)
20 120
This relatively broad distribution implies that even if the uncertainty in the principal eigenvalue of the reconstructed
RDM is large then the algorithm will with high probability converge to a pure state after a small number of iterations
(if the underlying distribution is Gaussian). If the distribution is non-Gaussian then Eq. (E9) can be used to show
convergence given that the moments of the distribution are appropriately small.
19
1. Errors in Eigenvalues
The errors in the eigenvalues of the RDM are easy to compute from known results. We have from Corollary 6.3.4
from [45] that if ρ is the true density operator and ρ̃ = ρ + sE for some matrix E of errors and some scalar s ∈ [0, 1]
then the error in a particular eigenvalue is at most
|λ(ρ) − λ(ρ + sE)| ≤ skEk, (E12)
where kEk is the spectral norm of the error matrix. We are of course most interested in the case where s = 1, however
below we will need the above formula for general values of s and so we give it for generality.
Now let E be a matrix consisting of M elements, each of which is independently distributed with zero mean and
2
variances at most σM . We then have that
 
X
E (λ(ρ) − λ(ρ + E))2 ≤ E  [E 2 ]i,j  ≤ M σM 2

. (E13)
i,j
Thus
2
V(λ(ρ + E)) ≤ M σM . (E14)
2 2
Hence σ ≤ M σM ,which allows the upper bounds in Eq. (E10) to be easily computed (under assumptions of
Gaussianity). In particular, we then have convergence under the Gaussianity assumption if
√ !
2 1 −3 564
σM ≤ + . (E15)
M 20 120
Recall that the 1-RDM constists of N (N + 1)/2 independent matrix elements, which implies that M = N (N + 1)/2
in our case.
2. Errors in Eigenvectors
Although the above criteria give conditions for the convergence of McWeeny purification starting from a sampled
1-RDM, there remains the question of whether the pure state that it converges to is -close to the true value. This is
relevant because if the errors are large enough that an eigenvalue crossing occurs, then the purification process can
fail to yield the desired state. Our aim here is to bound the distance between the eigenvectors.
First, rather than arguing about the difference in eigenvectors for ρ and ρ + E we will instead consider R time slices
and will be interested in the eigenvectors of ρ(j) := ρ + (j/R)E. Let the principal eigenvector of ρ be |λi and more
generally at step j let us denote the eigenvector to be |λ(j)i and the correspeonding eigenvalue to be λ(j). We then
have from first order perturbation theory, assuming that there is an eigenvalue gap that for any state |ν([j − 1])i that
is orthogonal to |λ([j − 1])i,
1 hν(j − 1)| E |λ(j − 1)i
hν(j − 1)|λ(j)i = + O(1/R2 ) (E16)
R ν(j − 1) − λ(j − 1)
Thus if we define γ(j) to be the minimum eigenvalue gap between |λ(j)i and the remainder of the spectrum of ρ(j)
we have that
X X 1 | hν(j − 1)| E |λ(j − 1)i |2
|hν(j − 1)|λ(j)i|2 ≤ + O(1/R3 )
R2 (λ(j − 1) − ν(j − 1))2
ν6=λ ν6=λ
X 1
≤ | hν(j − 1)| E |λ(j − 1)i |2 + O(1/R3 )
γ 2 (j − 1)R2
ν6=λ
X 1
= hλ(j − 1)| E |ν(j − 1)ihν(j − 1)| E |λ(j − 1)i |2 + O(1/R3 )
γ 2 (j − 1)R2
ν6=λ
1
hλ(j − 1)| E 2 |λ(j − 1)i − (hλ(j − 1)| E |λ(j − 1)i)2 + O(1/R3 )

=
γ 2 (j − 1)R2
kE 2 k
≤ 2 + O(1/R3 ) (E17)
γ (j − 1)R2
20
It then follows from Eq. (E17) that
kE 2 k
|hλ(j − 1)|λ(j)i − 1|2 ≤ + O(1/R3 ) (E18)
γ 2 (j − 1)R2
This gives us that, for the Euclidean distance between two vectors,
p
2kE 2 k
||λ(j)i − |λ(j − 1)i| ≤ + O(1/R2 ). (E19)
γ(j − 1)R
Next we have from the triangle inequality that for any integer R,
R R p
X X 2kE 2 k
||λ(R)i − |λ(0)i| ≤ ||λ(j)i − |λ(j − 1)i| ≤ + O(1/R) (E20)
j=1 j=1
γ(j − 1)R
In particular, this holds as we take R → ∞, which yields

R p p √
X 2kE 2 k 2kE 2 k 2kEk
lim + O(1/R) ≤ = . (E21)
R→∞
j=1
γ(j − 1)R γmin γmin
Unfortunately, we do not know what γmin is apriori, however we can bound it modulo some weak assumptions. Let
kEk ≤ 1/4, it is then straight forward to verify from Eq. (E12) that
γmin ≥ 1 − 2kEk. (E22)
Under the exact same assumptions we then have from a series expansion of the denominator that
√ √
||λ(ρ)i − |λ(ρ + E)i| ≤ 2kEk (1 + 4kEk) ≤ 2 2kEk (E23)
As E is a sum of M elements each with zero mean and variance at most σM we then have under the above assumptions
(and the additive property of variance) that
2
V ||λ(ρ)i − |λ(ρ + E)i| ≤ 8M σM . (E24)
Therefore if we demand that the variance is atmost 2 it suffices to pick
2 2
σM = , (E25)
8M
which sets a sufficient condition on the number of samples of Nsamp ≥ 2√2M . The remaining caveat is that in the
above analysis we needed to assume that kEk ≤ 1/4. If each of the entries of the matrix E are Gaussian random
variables, for example, it then follows that regardless of the value of σ there will always be a tail probability that this
eigenvalue condition is not met. We can bound the tail probability using Chebyshev’s inequality. Using the exact
same reasoning as in Eq. (E14) we have that
2
V(kEk) ≤ M σM . (E26)
Thus the probability that kEk ≥ 1/4 is

2
P ≤ 16M σM ∼ 22 . (E27)
Thus even under the pessimistic assumptions of Chebyshev’s inequality, we have that the probability of failure is
asymptotically negligible if σM is chosen in accordance with Eq. (E25). Note that the number of samples needed
taken in this case is in Θ(/N ) as there are M ∈ Θ(N 2 ) independent matrix elements in the 1-RDM.
Appendix F: Effect of CPHASE and Givens Rotation Error
In this section we consider two known gate errors that occur in the Givens rotation√circuits and attempt to
analytically and numerically benchmark the effect of these errors. When implementing the iswap operation there is
21
a known |11ih11| phase

√ error of approximately π/24. We model this phase error as a cphase(π/24) gate that occurs
directly after the iswap gate (Eq. (F1)). We find that the always on cphase(π/24) has negligible effect on the
outcome of the experiment and the stochastic Rz(θ) errors coherently corrupt the output of the circuit.
1 0 0 0
 
√ 1 i √
 0 √2 √2 0 
iswap ≈ 0 √i √1
 = CPHASE(π/24) iswap (F1)
2 2
0 
0 0 0 eiπ/24
To benchmark the effect of the parasitic cphase we simulate the diazene experiment with this interaction turned on
and evaluate the results with error mitigation. We can counteract the cphase(π/24) by performing local Rz gates.
108.40
Out of plane
In plane
108.42 Out of plane-Corrected
Out of plane-VQE
108.44 In plane-Corrected
Energy [Hartree]
In plane-VQE
108.46
108.48
108.50
108.52
108.54
108.56
0 5 10 15 20 25 30 35
Reaction Coordinate
FIG. 10. Analysis of the diazene isomerization curve where the Givens rotations are corrupted by a parasitic cphase(π/24). All
points are after purification. Without purification all curves are significantly higher in energy. The light dots are the circuits
executed without optimization, the darker dots are with an angle adjustment to counteract the known parasitic cphase, and
the plus markers are VQE optimization of the cphase circuits.
Consider the imperfect gate

√
U1 = diag(1, 1, 1, e−iφ ) iswap (F2)
we can use a different imperfect gate which differs only by single qubit phases,
√
U2 = diag(1, eiφ/2 , eiφ/2 , 1) iswap. (F3)
The error associated with U1 can be approximated by considering the Pauli expansion of the cphase part of U1 ,

iφ iφ iφ
diag(1, 1, 1, e−iφ ) ≈ e−iφ × II + IZ + ZI − ZZ , (F4)
4 4 4
and thus the error is approximately
2
φ 3 2
Err1 ≈ 3 = φ . (F5)
4 16
Similarly for U2

iφ/2 iφ/2 iφ/4 iφ
diag(1, e ,e , 1) ≈ e II − ZZ (F6)
4
with an associated Pauli error of
1 2
Err2 ≈ φ . (F7)
16
22
√
Very crudely, since for iswap φ = π/24, we expect Err1 to be approximately 0.32% and Err2 to be approximately
0.11%. This improvement is shown to be most beneficial for simulating the in plane rotations of diazene in Fig. S10.
The light dots are with the original cphase gate whereas the solid dots are with this local Rz correction. We also
include a VQE optimization to numerically determine the noise floor for this experiment. This suggests that VQE +
error mitigation can mitigate not only control error but more fundamental gate physics issues. For in-plane rotation
circuits the dynamics during the circuit execution are apparently more sensitive to these types of coherent errors near
transition states, although the exact reason for increased sensitivity is unclear.
To determine the error budget on the Rz rotation angles we can determine the degree of corruption from Gaussian
noise on the control angle. Consider the Rz rotation
Rz(θ, δα) = e−iZθ(1+δα)/2 (F8)
where θ is the desired rotation angle and δα is a stochastic variable. We can build a simplified model of control angle
error as Givens rotation error
† †
G(θ, δα) = eθ(1+δα)(ai aj −aj ai ) (F9)
which can be expressed as

√ † √
G(θ) = iswapi,j e−iθ(1+δα)Zi /2 eiθ(1+δα)Zj /2 iswapi,j . (F10)
For numerical simplicity we consider the effect on elements of the 1-RDM

† †
ai cos(θ(1 + δα)) + aj sin(θ(1 + δα)) if r = i

G(−θ, δα, i, j)a†r G(θ, δα, i, j) = a†j cos(θ(1 + δα)) − a†i sin(θ(1 + δα)) if r = j (F11)
a†

6 i & r 6= j
if r =
r

ai cos(θ(1 + δα)) + aj sin(θ(1 + δα))
 if s = i
G(−θ, δα, i, j)as G(θ, δα, i, j) = aj cos(θ(1 + δα)) − ai sin(θ(1 + δα)) if s = j (F12)

a
s if s 6= i & s 6= j
We can determine the expected 1-RDM with respect to a Gaussian distribution of noise by integrating with respect
to the perturbation
1 (δα)2
ρ(δα) = √ e− 2σ2 (F13)
σ 2π
 2 2 2 2
† −θ σ † −θ σ
Z ∞ ai cos(θ)e 2 + aj sin(θ)e 2
 if r = i
2 2 2 2
ρ(δα, σ)G(−θ, δα, i, j)a†r G(θ, δα, i, j)dδα = a†j cos(θ)e− θ 2σ − a†i sin(θ)e− θ 2σ if r = j (F14)
−∞ 
 †
ar 6 i & r 6= j
if r =
 2 2 2 2
−θ σ −θ σ
Z ∞ ai cos(θ)e 2 + aj sin(θ)e 2
 if s = i
2 2 2 2
ρ(δα, σ)G(−θ, δα, i, j)as G(θ, δα, i, j)dδα = aj cos(θ)e− θ 2σ − ai sin(θ)e− θ 2σ if s = j (F15)
−∞ 
as 6 i & s 6= j
if s =

Therefore, propagating the 1-RDM with stochastic Rz errors corresponds to evaluating the map in Eq. (F15). This
calculation assumes that the stochasticity has a time-scale that is much faster than a single energy evaluation. We
find that with σ > 0.22 purification projects to the wrong 1-RDM.
Appendix G: Gradient for the Basis Rotation Ansatz
Another benefit of restricting our ansatz to Slater determinants is the fact that the gradient with respect to the
parameters is accessible via the elements of the 1-RDM. The gradient of the energy with respect to the parameters
23
= 0.22 Out of plane

108.42 = 0.22 In plane
=0
108.44
Energy [Hartree]
108.46
108.48
108.50
108.52
108.54
108.56
0 5 10 15 20 25 30 35
Reaction Coordinate
FIG. 11. Stochastic Rz simulation of diazene with σ = 0.22 radian fluctuation on the Givens rotation gates. The plotted points
are after applying purification to the result 1-RDM

a†b ai − a†i ab is
P
of a one-body generator Z = i<b cb,i
dE de−Z deZ
= hφ0 | HeZ |φ0 i + hφ0 |e−Z H |φ0 i. (G1)
dcb,i dcb,i dcb,i
Due to the structure of this operator we expect the gradient to involve the commutator of the Hamiltonian with
respect to the anithermitian operator that becomes the prefactor to the right gradient. We call this prefactor ∇f (Z)
to indicate that it is a different operator from just the rotation generator associated with cb,i .
dE
= hφ0 |e−Z H, ∇f (Z)cb,i eZ |φ0 i

(G2)
dcb,i
All quantities in the commutator above can be evaluated with knowledge of the 1-RDM when φ0 is a computational
basis state. In this work we utilized this gradient for a classical implementation and provide it here as justification
for the ansatz and for future studies. The formal derivation of ∇f (Z) can be found in [46] and [47].
As a sketch for the form of ∇f (Z) consider the unitary performed in the Hartree-Fock experiment
−
cb,i (a†b ai −a†i ab )
P P
cb,i Eb,i
U (cb,i ) = e b,i =e b,i . (G3)
We now want to consider the energy derivative with respect to cb,i . Using the formulas in [46] we obtain
Z 1 P
dU (c) P − P −
− −x b,i cb,i Eb,i −
= dxex b,i cb,i Eb,i Eb,i e e b,i cb,i Eb,i . (G4)
dcb,i 0
In order to evaluate this integral we need to have an analytical form for the similarity transform of the integrand.
The integrand can be expressed in series form with the Baker-Campbell-Hausdorff identity where each term involves
−
nested commutators. Each nested commutator can be expressed more succinctly as the adjoint action of Z on Eb,i
 n
X
ad  cb0 ,i0 E −0 0  E − . b ,i b,i (G5)
b0 ,i0
A general strategy for evaluating sums of adjoint actions is to represent the operator b0 ,i0 cb0 ,i0 Eb−0 ,i0 in its eigenbasis
P
and directly evaluate the commutator as a matrix power. In our case this would involve diagonalizing a large 2n × 2n
matrix. Fortunately, due to the connection between one-particle-basis rotations and rotations by one-body operators
on the full Hilbert space we can find a n × n unitary that can diagonalize the matrix of cb,i coefficeints and represent
−
the operator Eb,i in this one-particle basis. Following this step of the derivation in [47] we form the C matrix of
coefficients cb,i which is antihermitian and diagonalize. Therefore, C is represented in its eigenbasis as
X
iC = i λr ã†r ãr (G6)
r
24
where λ are purely imaginary and we have used the fact that
X X
ãp = u∗p,q aq , ã†p = up,q a†q . (G7)
q q
−
We represent Eb,i term in the basis that diagonalizes iC
Ykl ã†k ãl

X
Y= (G8)
k,l

−
Ykl = U † Eb,i U (G9)
k,l
−
here Eb,i is an antisymmetric matrix with 1 at the (b, i) position and −1 at (i, b) position which is a representation of
−
the operator Eb,i . Therefore,
 
X h i
iλr Ykl ã†r ãr , ã†k ãl
X
ad  cb0 ,i0 Eb−0 ,i0  Eb,i
−
= (G10)
b0 ,i0 rkl

iλr Ykl ã†r ãl δkr − ã†k ãr δrl
X
= (G11)
rkl
(λk − λl ) Ykl ã†k ãl .

X
=i (G12)
kl
Furthermore, powers of the adjoint action are

 n

(λk − λl ) Ykl ã†k ãl .
X X n
ad  cb0 ,i0 Eb−0 ,i0  Eb,i
−
= in (G13)
b0 ,i0 kl
Armed with the adjoint power we can now evaluate the integrand of Eq. (G4) via fundemental theorem of calculus
and arrive at an expression for the gradient
" #
dU (c) X ei(λk −λl ) − 1 † P −
= Ykl ãk ãl e b,i cb,i Eb,i (G14)
dcb,i i (λk − λl )
kl
 
−
U M U † kl a†k al  e b,i cb,i Eb,i
X P
= (G15)
k,l
i(λk −λl )
where Mkl = Ykl e i(λk −λl−1
) . The expression in the parenthesis is a new one-body operator that we previous denoted
∇f (Z).
Appendix H: Optimization Technique
The optimizer we use in the experiment is based on Kutzelnigg’s approach to iteratively constructing a wavefunction
that satisfies the Brillouin condition [34]. In the following section we include the derivation and modifications of this
procedure from Reference [34] for completeness. This approach starts from the Lie-algebraic perspective on the
variational principle. The generators for variations in a norm conserved wavefunction are elements of a complex Lie
algebra. The variational principle which states
δ Ẽ = δhψ̃|H|ψ̃i = 0 (H1)
can be cast as stationarity with respect to a unitary group
U = eR R = −R† (H2)
25
where R is an element of the Lie algebra L supporting H. Formulation of the variations in Ẽ with respect to R is
formulated using the BCH expansion
1
Ẽ → Ẽ 0 = Ẽ + hψ̃| [H, R] |ψ̃i + hψ̃| [[H, R] , R] |ψ̃i + ... (H3)
2
and thus stationarity with respect to infinitesimal variations in R implies
hψ̃| [H, R] |ψ̃i = 0 ∀R = −R† (H4)
1. Iteratively constructing wavefunctions
Given an R that does not satisfy the first order stationarity condition Eq. (H4) we can propose a new wavefunction
that is approximately stationary with respect to R.
AR = hφ| [H, R] |φi =

6 0 (H5)
We want to determine an update of the generator R such that the first order condition holds. We consider the update
to the wavefunction
ψ = e−fR R φ (H6)
where fr is a real number. Considering how the energy changes as a function of fR
fR2
EfR = hφ|efR R He−fR R |φi ≈ hφ|H + fR [H, R] + [[H, R] , R] |φi (H7)
2
In a similar fashion to deriving a Newton-Raphson update in optimization we can differentiate to find an fR that
approximately satisfies Eq. (H4).
dEfR fR
= hφ| [H, R] + [[H, R] , R] |φi = 0 fR = −AR /BR,R BR,R = hφ| [[H, R] , R] |φi (H8)
dfR 2
Alternatively, one can determine the change in the stationary condition with respect to fR
fR2
0 = hφ|efR R [H, R] e−fR R |φi = hφ| [H, R] + fR [[H, R] , R] + [[[H, R] , R] , R] + ...|φi (H9)
2
and enforce the stationarity approximately by truncating at first order and solving for fR
fR = −AR /BR,R (H10)
which provides the same type of update. The error in the residual for R, AR , is now of the magnitude O(fR2 ) at leading
order. This update inspires a possible iterative procedure for improving the wavefunction that will quadratically
converge to the correct state if we are in a convex region away from the exact solution [34].
One can use the above procedure where R is not an element of the operator basis {Xk } of the Lie algebra L
X
R= ck X , X ⊂ L (H11)
k
and to determine a set of ck which approximately satisfy Eq. (H4).

X
0 ≈ hφ| [H, Xk ] + [[H, Xk ] , Xl ] cl |φi (H12)
l
Again, approximating the expansion in Eq. (H12) to first order we get a system of equations to solve for ck that
ensures the Brillouin condition is satisfied up to leading error of O(c2k ).
In the context of a NISQ machine one needs to consider the family of generators {R} that is tractable and the cost
of the measurements associated with measuring AR and Bk,l . In this work we use
Ap,q = hψ| H, a†p aq |ψi Bp,q;r,s = hψ| H, a†p aq , a†r as |ψi.

(H13)
26
Both the gradient and the Hessian term can be evaluated with knowledge of the 1-RDM under the assumption that
ψ corresponds to a Slater determinant. The update parameters to κ, fp,q are computed by solving the augmented
Hessian eigenvalue problem

0 A 1 1
= (H14)
A† B fp,q fp,q
which provides an optimal level shift to Newton’s method
A + (B − )fp,q = 0. (H15)
As described in [35] we add regularization by limiting the size of the update fp,q by rescaling under the condition that
the max update is above a parameter γ
(
fp,q max(fp,q ) < γ
γ (H16)
max(fp,q ) fp,q max(fp,q ) ≥ γ
The algorithm then dictates that the wavefunction is updated through Eq. (H6) which is yet another non-interacting
fermion wavefunction. We concatenate this basis rotation with the original using Eq. (B3) so the circuit depth remains
constant. The optimization procedure is iterated for a fixed number of steps or the commutator h[H, Xk ]i falls below
a predefined threshold.
Appendix I: Additional Performance Analysis
1. Post-selection performance
In this section we examine the percentage of measurements rejected by post-selection as a function of system size
and fidelity metrics across the systems studied in the hydrogen chain and diazene experiments. In Table II we plot
the ratio of the total number of circuit repetitions that result in the correct excitation number. As expected this
ratio decreases with system size, almost perfectly tracking a joint readout fidelity of 95%. We believe the discrepancy
between the two 10-qubit experiments (H10 and diazene experiments) stems from the fact that the diazene circuits
have more idle circuit moments where the qubits are free to decay.
Molecule Post-selection Shot Ratio

H6 0.764(7)
H8 0.66(1)
H10 0.56(1)
H12 0.46(2)
diazene 0.44(1)
TABLE II. The average fraction of the 250,000 circuit repetitions used to measure observables for each circuit. The average
is collected across all hydrogen geometries and diazene geometries for every circuit required to estimate the 1-RDM for these
systems.
Plotted another way, we can examine the distribution of local qubit expectation values hMi i where Mi is the
measurement result of qubit i. In Fig. S12 we plot the integrated histogram of Mi –i.e. the probability of a 1
bit being measured from qubit i–(denoted P1) on all the qubits for all circuits in all hydrogen chain experiments.
This is compared to the theoretical value obtained by the perfect 1-RDM simulation described in Appendix A 1. The
significant improvement in readout scatter from post-selection is a fundamental driver in the success of this experiment
due to the sensitivity of quantum chemistry energies to electron number.
27
1.0 6 qubits
8 qubits
Integrated Histogram, ecdf

0.8 10 qubits
12 qubits
0.6
0.4
0.2
Without Postselection
0.0
0.100 0.075 0.050 0.025 0.000 0.025 0.050 0.075 0.100
P1(ideal) - P1(measured)
FIG. 12. Integrated histogram of readout performance with and without post-selection on photon number. Grey lines are the
histograms of circuit measurements without post-selection.
2. Natural Occupation Numbers
In this section we tabulate the natural occupation numbers for the “raw” and the “post-selection” data sets.
TABLE III. H6 raw natural orbitals

bond distance λ1 λ2 λ3 λ4 λ5 λ6
0.5 0.0024 0.0181 0.0338 0.8738 0.9109 0.9402
0.9 0.0000 0.0176 0.0295 0.8888 0.9142 0.9405
1.3 0.0083 0.0182 0.0321 0.8808 0.9157 0.9417
1.7 0.0080 0.0209 0.0400 0.8780 0.8999 0.9518
2.1 0.0103 0.0130 0.0378 0.8884 0.9074 0.9438
2.5 0.0098 0.0128 0.0403 0.8868 0.9126 0.9427
TABLE IV. H6 +post-selection natural orbitals

bond distance λ1 λ2 λ3 λ4 λ5 λ6
0.5 -0.0006 0.0113 0.0271 0.9768 0.9861 0.9993
0.9 -0.0037 0.0146 0.0204 0.9815 0.9883 0.9989
1.3 0.0042 0.0139 0.0266 0.9726 0.9818 1.0009
1.7 0.0019 0.0164 0.0322 0.9656 0.9732 1.0106
2.1 0.0075 0.0124 0.0332 0.9711 0.9737 1.0021
2.5 0.0085 0.0094 0.0343 0.9689 0.9801 0.9988
TABLE V. H8 raw natural orbitals

bond distance λ1 λ2 λ3 λ4 λ5 λ6 λ7 λ8
0.5 -0.0033 0.0083 0.0197 0.0365 0.8909 0.8975 0.9046 0.9283
0.9 -0.0047 0.0126 0.0290 0.0435 0.8742 0.8800 0.9181 0.9331
1.3 -0.0060 0.0148 0.0275 0.0469 0.8622 0.8881 0.9129 0.9238
1.7 -0.0089 0.0210 0.0426 0.0538 0.8594 0.8736 0.9118 0.9179
2.1 0.0026 0.0247 0.0458 0.0532 0.8454 0.8634 0.9139 0.9295
2.5 0.0085 0.0259 0.0572 0.0610 0.8379 0.8801 0.9010 0.9168
28
TABLE VI. H8 +post-selection natural orbitals

bond distance λ1 λ2 λ3 λ4 λ5 λ6 λ7 λ8
0.5 -0.0087 0.0042 0.0146 0.0216 0.9779 0.9889 0.9944 1.0072
0.9 -0.0136 0.0080 0.0207 0.0350 0.9637 0.9731 0.9992 1.0138
1.3 -0.0160 0.0115 0.0237 0.0395 0.9535 0.9748 0.9998 1.0132
1.7 -0.0209 0.0181 0.0357 0.0512 0.9500 0.9634 0.9926 1.0099
2.1 -0.0120 0.0188 0.0411 0.0517 0.9458 0.9502 0.9947 1.0097
2.5 -0.0075 0.0115 0.0513 0.0568 0.9362 0.9660 0.9814 1.0043
TABLE VII. H10 raw natural orbitals

bond distance λ1 λ2 λ3 λ4 λ5 λ6 λ7 λ8 λ9 λ10
0.5 -0.0503 -0.0067 0.0087 0.0493 0.0638 0.8663 0.8747 0.8961 0.9109 0.9222
0.9 -0.0176 0.0109 0.0169 0.0514 0.0829 0.8359 0.8490 0.8779 0.9165 0.9182
1.3 -0.0050 0.0020 0.0215 0.0279 0.0460 0.8302 0.8868 0.8892 0.9142 0.9266
1.7 -0.0289 0.0107 0.0212 0.0343 0.0535 0.8529 0.8747 0.8795 0.9106 0.9372
2.1 -0.0092 0.0048 0.0145 0.0299 0.0596 0.8537 0.8651 0.8985 0.9217 0.9352
2.5 -0.0010 0.0118 0.0216 0.0300 0.0626 0.8470 0.8739 0.8842 0.9107 0.9258
TABLE VIII. H10 +post-selection natural orbitals

bond distance λ1 λ2 λ3 λ4 λ5 λ6 λ7 λ8 λ9 λ10
0.5 -0.0624 -0.0122 0.0021 0.0492 0.0553 0.9488 0.9826 0.9984 1.0130 1.0251
0.9 -0.0224 0.0056 0.0139 0.0408 0.0736 0.9306 0.9585 0.9830 0.9985 1.0179
1.3 -0.0126 -0.0110 0.0185 0.0276 0.0497 0.9262 0.9791 0.9967 0.9978 1.0278
1.7 -0.0397 -0.0005 0.0170 0.0290 0.0561 0.9470 0.9731 0.9825 1.0077 1.0276
2.1 -0.0215 0.0026 0.0057 0.0224 0.0529 0.9559 0.9583 0.9851 1.0069 1.0317
2.5 -0.0184 0.0057 0.0232 0.0248 0.0597 0.9441 0.9672 0.9817 0.9997 1.0122
TABLE IX. H12 raw natural orbitals

bond distance λ1 λ2 λ3 λ4 λ5 λ6 λ7 λ8 λ9 λ10 λ11 λ12
0.5 -0.0037 0.0010 0.0077 0.0182 0.0500 0.0589 0.8361 0.8515 0.8828 0.8910 0.8995 0.9098
0.9 -0.0195 0.0021 0.0118 0.0247 0.0413 0.0713 0.7948 0.8316 0.8816 0.8919 0.9019 0.9442
1.3 -0.0160 0.0066 0.0192 0.0346 0.0521 0.0823 0.8035 0.8179 0.8853 0.8911 0.9099 0.9198
1.7 -0.0016 0.0087 0.0276 0.0288 0.0458 0.0737 0.7967 0.8480 0.8614 0.8906 0.8991 0.9099
2.1 -0.0153 -0.0011 0.0235 0.0325 0.0572 0.0777 0.8198 0.8331 0.8556 0.8646 0.9132 0.9260
2.5 -0.0143 0.0029 0.0207 0.0564 0.0650 0.0821 0.8143 0.8351 0.8719 0.8750 0.8966 0.9131
TABLE X. H12 +post-selection natural orbitals

bond distance λ1 λ2 λ3 λ4 λ5 λ6 λ7 λ8 λ9 λ10 λ11 λ12
0.5 -0.0120 -0.0029 0.0030 0.0137 0.0474 0.0536 0.9349 0.9540 0.9878 0.9968 1.0042 1.0194
0.9 -0.0310 -0.0039 0.0037 0.0222 0.0337 0.0735 0.8947 0.9397 0.9889 0.9970 1.0166 1.0649
1.3 -0.0260 -0.0009 0.0113 0.0216 0.0454 0.0817 0.8979 0.9265 0.9924 0.9976 1.0231 1.0292
1.7 -0.0074 0.0002 0.0147 0.0231 0.0459 0.0719 0.8964 0.9596 0.9665 0.9995 1.0082 1.0215
2.1 -0.0296 -0.0082 0.0172 0.0223 0.0593 0.0735 0.9170 0.9403 0.9715 0.9803 1.0134 1.0430
2.5 -0.0252 -0.0156 0.0096 0.0397 0.0618 0.0878 0.9139 0.9410 0.9782 0.9837 1.0029 1.0222
3. Energy and Fidelity
In Fig. S13 we plot the log-log scatter of absolute error and fidelity witness for all systems studied. The correlation
in the fidelity and absolute energy error suggests that fidelity can be used as an optimization target for this system.
This is a useful property when considering basis rotation states as targets for benchmarks and tune-up protocols.
29
H6 H12
1 - [Purified Fidelity Witness]

H8 Out of plane
2 H10 In plane
10
3
10
4 3 2
10 10 10
| E E * | [Hartree]
FIG. 13. Absolute error versus fidelity witness for VQE optimized with error mitigation for all experiments.
Slow two-level-system (TLS) diffusion on the surface of superconducting processors can alter the performance of
qubits over time periods of hours or days. It is likely that the H10 data set was collected during a time where the best
performing qubits had worse coupling to an itinerant TLS[48] than when we collected the H12 dataset. Thus, there
was some variance in performance across the different days when the chip was used to collect data. We believe that
by showing all of these results without cherry picking and rerunning less performant curves, we give a more accurate
representation of the average performance of the device.
To better describe the consistent quality of VQE optimized 10 qubit calculations we tabulate the perceived fidelity
calculated from purified 1-RDMs in all 10 qubit experiments: six H10 experiments and eighteen diazene points. On
all but one experiment variational relaxation combined with other error mitigation techniques allows us to achieve
> 98.0% average fidelity.
6 H10
Out of plane
5 In plane
4
Counts
00.975 0.980 0.985 0.990 0.995

Fidelity
FIG. 14. Fidelity of 10 qubit experiments: A histogram of fidelity witness values associated with the VQE optimized 10
qubit systems.
Appendix J: Molecular geometries
For the hydrogen chains OpenFermion [49] and Psi4 [50] were used to generate the integrals. All hydrogen chains
were computed at atom-atom separations of 0.5, 0.9, 1.3, 1.7, 2.1, and 2.5 Å. For the diazene curves we used Psi4
to map out the reaction coordinate for each isomerization mechanism by optimizing the geometries of the molecule
simultaneously constraining either the dihedral angle or NNH angle to a fixed value. Table XI and Table XII, below,
contain the geometries we considered for out-of-plane rotation and in-plane rotation of the hydrogen atom. To reduce
30
diazene to a 10 qubit problem we perform two cycles of canonical Hartree-Fock self-consistent field and then integrate
out the bottom two energy levels.
TABLE XI. Out-of-plane rotation geometries TABLE XII. In-plane rotation geometries
Internal coord. Atom Cartesian coordinates Internal Coord. Atom Cartesian coordinates
3.157 H -0.00183 0.61231 -1.23326 108.736 H 0.00000 0.61228 -1.23237
N -0.00183 0.61231 -0.16961 N 0.00000 0.61228 -0.16925
N -0.00183 -0.56366 0.29317 N 0.00000 -0.56613 0.29515
H 0.05269 -1.28820 -0.48362 H 0.00001 -1.25344 -0.51686
26.315 H -0.01473 0.61213 -1.23797 127.473 H 0.00000 0.61339 -1.24223

N -0.01473 0.61213 -0.17381 N 0.00000 0.61339 -0.17528
N -0.01473 -0.56586 0.29104 N 0.00000 -0.55235 0.28364
H 0.42406 -1.25504 -0.39080 H 0.00001 -1.46143 -0.26340
49.473 H -0.02522 0.61175 -1.25087 146.210 H 0.00000 0.61423 -1.26614

N -0.02522 0.61175 -0.18596 N 0.00000 0.61423 -0.18644
N -0.02522 -0.57104 0.28761 N 0.00000 -0.54592 0.27365
H 0.72616 -1.17742 -0.16152 H 0.00001 -1.56334 0.05447
72.631 H -0.03098 0.60150 -1.29530 164.947 H 0.00000 0.61854 -1.26711

N -0.03098 0.60150 -0.23059 N 0.00000 0.61854 -0.18132
N -0.03098 -0.56623 0.30717 N 0.00000 -0.55047 0.24812
H 0.89199 -1.09153 0.23125 H 0.00000 -1.56434 0.33892
95.641 H -0.03338 0.62184 -1.24521 182.0 H 0.00000 0.62468 -1.24838

N -0.03338 0.62184 -0.18592 N 0.00000 0.62468 -0.16646
N -0.03338 -0.60178 0.24302 N 0.00000 -0.56138 0.21599
H 0.96106 -0.90055 0.45184 H 0.00007 -1.50420 0.56008
117.611 H -0.03017 0.62034 -1.22808 200.526 H -0.00002 0.63051 -1.22163

N -0.03017 0.62034 -0.16858 N -0.00002 0.63051 -0.14939
N -0.03017 -0.60975 0.20354 N -0.00002 -0.57522 0.18140
H 0.86843 -0.76736 0.74227 H 0.00060 -1.39873 0.77683
139.581 H -0.02217 0.61500 -1.22656 219.052 H -0.00004 0.63081 -1.20416

N -0.02217 0.61500 -0.16715 N -0.00004 0.63081 -0.14048
N -0.02217 -0.61052 0.18284 N -0.00004 -0.58876 0.15662
H 0.63829 -0.67733 1.00847 H 0.00116 -1.21501 0.97997
161.551 H -0.01085 0.61243 -1.22466 237.578 H -0.00005 0.62175 -1.20948

N -0.01085 0.61243 -0.16601 N -0.00005 0.62175 -0.14932
N -0.01085 -0.61150 0.16941 N -0.00005 -0.59932 0.15347
H 0.31225 -0.62540 1.17746 H 0.00155 -0.93328 1.15189
183.522 H 0.00211 0.61159 -1.22471 256.105 H -0.00004 0.61032 -1.22743

N 0.00211 0.61159 -0.16627 N -0.00004 0.61032 -0.16648
N 0.00211 -0.61155 0.16640 N -0.00004 -0.61200 0.16650
H -0.06064 -0.61209 1.22297 H 0.00111 -0.58711 1.22715

Hartree-Fock On A Superconducting Qubit Quantum Computer

Uploaded by

Document Information

Original Description:

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

Hartree-Fock On A Superconducting Qubit Quantum Computer

Uploaded by

Copyright:

Available Formats

Hartree-Fock on a superconducting qubit quantum computer

Google AI Quantum and Collaborators∗

Energy error [Hartree]

108.400 a protocol described in Appendix C. This protocol en-

Google AI Quantum and Collaborators

Appendix A: Hartree-Fock Theory via Canonical Transformations

|ψ(r1 , ..., rn )i = (n!)−1/2 An (χ1 (r1 )...χn (rn )) (A1)

hδψ| = hψ|a†i aa ζ (A4)

hψ|a†i aa H|ψi = 0. (A5)

the adjoint representation of any element of the algebra κ where

κ, a†p = a†q κp,q , [κ, ap ] = aq κ∗p,q

where hij and Vijkl

1. Classical simulation of non-interacting fermion circuits

Appendix B: Implementing the Basis Change Circuit and Circuit Concatenation

We use this homomorphism through the observation that

Rpq (θ)U (u) = U (rpq (θ)u) (B4)

X 1.5π 1.0π -0.053π 1.5π 1.0π

X 1.3280π 1.0π -0.5π 0.9916π 1.0π -0.5π 0.9403π 1.0π

0.2039π 1.5π 1.0π -0.169π 1.5π 1.0π 0.0742π

-0.5π 0.6931π 1.0π -0.5π

FIG. 4. Givens rotation

X 0.7830π 1.0π -0.248π 1.0213π 1.0π

X 0.9853π 1.0π 0.2169π 0.8735π 1.0π -0.021π 1.1037π 1.0π

0.4467π 0.6895π 1.0π 0.4550π 0.5000π 1.0π 0.4999π

0.3104π 1.4999π 1.0π 0.4999π

Appendix C: Optimal Measurement of the 1-RDM

Post-select half filling

3. General off-diagonal terms and virtual swapping

fswap = a†p aq + a†q ap − a†p ap − a†q aq . (C4)

4. Off-diagonal terms with post-selection

Post-select half filling

5. Computing error bars for elements of the 1-RDM

Appendix E: Error mitigation through purification

where β = 4(α6 + 3α5 + 3α3 ).

It then follows from Eq. (E17) that

In particular, this holds as we take R → ∞, which yields

γmin ≥ 1 − 2kEk. (E22)

Therefore if we demand that the variance is atmost 2 it suffices to pick

Thus the probability that kEk ≥ 1/4 is

Appendix F: Effect of CPHASE and Givens Rotation Error

a known |11ih11| phase

Consider the imperfect gate

Rz(θ, δα) = e−iZθ(1+δα)/2 (F8)

which can be expressed as

For numerical simplicity we consider the effect on elements of the 1-RDM

Appendix G: Gradient for the Basis Rotation Ansatz

= 0.22 Out of plane

Ykl ã†k ãl

(λk − λl ) Ykl ã†k ãl .

Furthermore, powers of the adjoint action are

Appendix H: Optimization Technique

can be cast as stationarity with respect to a unitary group

hψ̃| [H, R] |ψ̃i = 0 ∀R = −R† (H4)

1. Iteratively constructing wavefunctions

AR = hφ| [H, R] |φi =

where fr is a real number. Considering how the energy changes as a function of fR

fR = −AR /BR,R (H10)

and to determine a set of ck which approximately satisfy Eq. (H4).

Ap,q = hψ| H, a†p aq |ψi Bp,q;r,s = hψ| H, a†p aq , a†r as |ψi.

which provides an optimal level shift to Newton’s method

Appendix I: Additional Performance Analysis

Molecule Post-selection Shot Ratio

Therefore if we demand that the variance is atmost 2 it suffices to pick