PRXQuantum.4.040341

PRX QUANTUM 4, 040341 (2023)
Statistical Phase Estimation and Error Mitigation on a Superconducting

Quantum Processor
Nick S. Blunt ,1,*,‡ Laura Caune,1,†,‡ Róbert Izsák ,1 Earl T. Campbell,1,2 and Nicole Holzmann1,3
1
Riverlane, Cambridge CB2 3BZ, United Kingdom
2
Department of Physics and Astronomy, University of Sheffield, Sheffield S3 7RH, United Kingdom
3
Astex Pharmaceuticals, 436 Cambridge Science Park, Cambridge, CB4 0QA, United Kingdom
(Received 13 April 2023; revised 14 June 2023; accepted 23 October 2023; published 12 December 2023)
Quantum phase estimation (QPE) is a key quantum algorithm, which has been widely studied as a
method to perform chemistry and solid-state calculations on future fault-tolerant quantum computers.
Recently, several authors have proposed statistical alternatives to QPE that have benefits on early fault-
tolerant devices, including shorter circuits and better suitability for error-mitigation techniques. However,
experimental investigations of the algorithm on real quantum processors are lacking. Here, we implement
statistical phase estimation on Rigetti’s superconducting processors. Specifically, we use a modification of
the Lin and Tong [PRX Quantum 3, 010318 (2022)] algorithm with the improved Fourier approximation of
Wan et al. [Phys. Rev. Lett. 129, 030503 (2022)] and apply a variational-compilation technique to reduce
the circuit depth. We then incorporate error-mitigation strategies including zero-noise extrapolation and
readout-error mitigation with bit-flip averaging. We propose a new method to estimate energies from the
statistical phase estimation data, which is found to improve the accuracy in the final energy estimates
by 1–2 orders of magnitude with respect to prior theoretical bounds, reducing the cost of performing
accurate phase-estimation calculations. We apply these methods to chemistry problems for active spaces
up to four electrons in four orbitals, including the application of a quantum embedding method, and use
them to correctly estimate energies within chemical precision. Our work demonstrates that statistical phase
estimation has a natural resilience to noise, particularly after mitigating coherent errors, and can achieve
far higher accuracy than suggested by previous analysis, demonstrating its potential as a valuable quantum
algorithm for early fault-tolerant devices.
DOI: 10.1103/PRXQuantum.4.040341
I. INTRODUCTION the limit of current classical algorithms, often estimating

the need for tens or hundreds of millions of qubits using
Quantum phase estimation (QPE) [1,2] is one of the
current QPE and QEC schemes [3–6].
most widely studied quantum algorithms, due to its poten-
In the past few years, statistical modifications to the
tial for exponential speed-ups in a wide range of problems.
QPE algorithm have been proposed [7–14] that use lower-
While this potential is promising, the quantum circuits
depth circuits and far fewer auxiliary qubits than tech-
involved in useful applications of QPE have a high depth.
niques based on qubitization [6,15] and so are better
Because of this, QPE is often described as a fault-tolerant
suited to near-term devices. Importantly, statistical QPE
quantum algorithm, which will require large-scale quan-
is also more readily combined with error-mitigation tech-
tum error correction (QEC) for nontrivial applications.
niques, many of which are primarily designed for esti-
Many studies have been performed in recent years to assess
mating expectation values. “Textbook” and many other
the resources required to apply QPE to active spaces at
QPE approaches measure a discrete output, namely, the
bits of the energy estimate, and are therefore not com-
*
[email protected] patible with most error-mitigation techniques. In con-
†
[email protected] trast, the circuits involved in statistical phase estima-
‡
These authors contributed equally to this work. tion methods are typically Hadamard tests, the output
of which is an expectation value, as shown in Fig. 1.
Published by the American Physical Society under the terms of
Multiple such circuits are performed, with the result-
the Creative Commons Attribution 4.0 International license. Fur-
ther distribution of this work must maintain attribution to the ing expectation values used to construct an appropriate
author(s) and the published article’s title, journal citation, and function, from which the desired eigenvalues may be
DOI. estimated.
2691-3399/23/4(4)/040341(23) 040341-1 Published by the American Physical Society

NICK S. BLUNT et al. PRX QUANTUM 4, 040341 (2023)
In this paper, we present the first combination of sta-

tistical phase estimation and error mitigation on a current
quantum device. The methodology presented allows us to
perform ground-state energy calculations to accuracy bet-
ter than 0.1 mHa for the systems considered, which is
FIG. 1. The Hadamard test circuits that are considered in
comparable to, or better than, the accuracy that could be
this work. Setting V = 1 or V = S † allows estimation of previously achieved with the variational quantum eigen-
Re[ ψ| e−iτ Hk |ψ ] and Im[ ψ| e−iτ Hk |ψ ], respectively. solver (VQE) and current hardware [41–43]. Statistical
phase estimation also avoids problematic barren plateaus
Error mitigation has been a major theme in early that hamper the scalability of VQE, although the speed-up
practical applications of quantum computing, including of all QPE algorithms does depend on the quality of the ini-
applications to chemistry and condensed-matter physics tial trial states [44]. For error mitigation, we focus on ZNE,
problems. Current quantum processors are often referred symmetrized readout-error mitigation [17], and RC meth-
to as noisy intermediate-scale quantum (NISQ) devices ods. Although our results are performed on current NISQ
due to their high error rates and low qubit counts. In the devices, the error-mitigation techniques presented can also
absence of QEC, which requires higher qubit counts and be applied in the early fault-tolerant regime. For example,
lower error rates than currently available, error mitiga- we apply RC to mitigate coherent errors for controlled-Z
tion has been widely investigated. This includes techniques (CZ) gates in this work but, as described in Ref. [25], it can
for readout error mitigation [16,17], as well as numer- equally be applied to T gates in an error-corrected scheme.
ous methods for mitigating gate errors, such as zero-noise We focus on a statistical QPE approach based on the
extrapolation (ZNE) [18–20], probabilistic error cancel- cumulative distribution function (CDF) of the spectral
lation (PEC) [18,21,22], Clifford data regression (CDR) measure of the Hamiltonian, which has been introduced by
[23,24], noise tailoring techniques such as randomized Lin and Tong [8], and use the improved Fourier approx-
compiling (RC) [25,26], and symmetry constraints or post- imation derived by Wan et al. [9] (but do not investi-
selection [27–29]. See also Ref. [30] for a recent review of gate the randomized compilation approach for Hamilto-
error-mitigation techniques and Ref. [31] for a recent study nian simulation introduced in the same paper). We also
testing ZNE and PEC on multiple quantum computing implement and test the quantum eigenvalue estimation
platforms. algorithm (QEEA) by Somma [7]. In both approaches,
In addition to NISQ studies, there is good reason to we apply importance sampling as described in Ref. [8].
believe that error mitigation will continue to be important We apply these methods to several molecules, including
in the early fault-tolerant regime, where partial error cor- two examples motivated by pharmaceutical applications,
rection is possible [32–35]. For example, a recent study using a chemical-embedding approach to construct rel-
has investigated the resources required to perform even the evant active spaces [45]. These results are achieved by
simplest ground-state calculation on an H2 molecule using using a variational-circuit-compilation strategy to allow
the surface code [36]; even here, it has been found that the required operations to be performed with a low circuit
thousands of physical qubits are required. The largest over- depth, suitable for current quantum processors. We also
head comes from the need for magic state distillation facto- study a Trotterized example, which avoids the exponential
ries and decomposing rotation gates into T gates. This cost preprocessing step of the variational-compilation method,
could be significantly reduced using error mitigation. For using two-qubit gate depths up to 100.
example, Piveteau et al. [32] have discussed approaches to We show that mitigating coherent errors [25,26] is
perform noisy T gates and apply error mitigation. It should important in statistical phase estimation and that this
be emphasized that error-mitigation techniques in general methodology can be effectively combined with importance
have a cost that scales exponentially with the circuit depth sampling, allowing energies to be extracted with confi-
[37]. However, in the near-term regime of NISQ and early dence, even in the presence of significant QPU errors. We
fault tolerance, they offer valuable methods to improve the further introduce a simple approach to estimate energies
performance of statistical QPE. from the QPU data with significantly improved accuracy,
Examples of prior QPE calculations on quantum hard- compared to the theoretical bounds derived in Ref. [9]. The
ware include an early study of Kitaev’s iterative QPE largest example that we study is a model of a pharmaceu-
algorithm using superconducting qubits [38]; an applica- tically relevant molecule, where the ground-state energy is
tion of textbook QPE on a neutral-atom quantum computer again obtained with an error of less than 0.1 mHa.
[39]; and, very recently, a study of Bayesian QPE on a The structure of the paper is as follows. In Sec. II A, we
trapped-ion quantum computer [40]. All of these studies cover the theory of statistical phase estimation, particularly
consider a minimal H2 model, using a Hamiltonian of the methods of Refs. [8,9], including importance sampling.
only one or two qubits, demonstrating the challenge of In Sec. II B, we discuss a variational-compilation method
performing QPE in practice. to reduce the circuit depth. Section II C introduces the
040341-2
STATISTICAL PHASE ESTIMATION AND ERROR MITIGATION PRX QUANTUM 4, 040341 (2023)
error-mitigation techniques to be applied. Results are then where we define pi = |νi |2 . Therefore, gk will in general
presented from Rigetti’s Aspen devices, first applying the consist of a sum of oscillating signals, the frequencies of
statistical phase estimation method to an example with two which are determined by the energies of H and the ampli-
electrons in two orbitals, followed by a study of larger tudes of which are determined by the components of the
active spaces in combination with error-mitigation strate- corresponding eigenstates in |ψ. The goal of statistical
gies, and a final example using a Trotterized expansion of phase estimation methods is to extract (some of) the phases
the time-evolution operator. λi from the noisy estimates of gk and hence estimate the
energies of H .
II. THEORY Multiple such methods have been introduced in recent
A. Statistical phase estimation years, each suggesting different techniques to construct
eigenvalue estimates from the gk estimates. Roughly
We consider n-qubit Hamiltonians of the form speaking, these methods involve identifying some function
f (H ) that allows eigenvalues to be identified. The function

L
H= cl Pl , (1) f (H ) is then expanded in a Fourier series, which can be
l=1 constructed, after truncation, using the gk estimates. Trun-
cation is needed to put a finite limit on the unitary time
where Pl are n-qubit Pauli operators, and denote the eigen- evolution τ k. Note that once gk has been constructed, the
values and eigenvectors of H by {λi ; |i }. We will be task of estimating the desired λi is a purely classical task
concerned with estimating {λi } using statistical phase- and is related to similar problems in signal processing.
estimation methods. For these techniques, it is necessary to We begin by reviewing the approach of Wan et al. [9].
bound the Hamiltonian in a known range and we therefore This approach is based on a similar method by Lin and
work with a scaled Hamiltonian τ H , where τ > 0. Tong in Ref. [8] but uses an alternative Fourier approxi-
In this paper, we focus on circuits of the form shown in mation that allows the authors to prove a better asymptotic
Fig. 1. This is a Hadamard test circuit, where setting V = complexity. Note that Ref. [9] also introduces a random-
1 or V = S † = |00| − i|11| allows measurement in the ized compiling approach to implement e−iτ Hk with reduced
X or Y bases, respectively. For V = 1, defining a random circuit depth, but this approach is not tested in this paper.
variable X equal to +1 for |0 measurements and −1 for
|1 measurements, it can be shown that 1. CDF-based statistical phase estimation
−iτ Hk
E[X ] = Re[ ψ| e |ψ ]. (2) In the approach of Refs. [8,9], we wish to calculate a
cumulative distribution function (CDF) associated with the
where throughout E[·] denotes an expectation value. Sim- Hamiltonian and state |ψ,
ilarly, for V = S † measurements, defining a random vari-

able Y equal to +1 for |0 measurements and −1 for |1 C(x) = pi . (7)
measurements gives i:τ λi ≤x
E[Y] = Im[ ψ| e−iτ Hk |ψ ]. (3)

If C(x) could be constructed, then it would allow us to
Performing the circuits of Fig. 1 therefore allows estima- identify the eigenvalues of H through its discontinuities.
tion of In practice, the CDF will be constructed as a sum of terms
eikx for integer values of k and so it is necessary to define
gk = ψ|e−iτ Hk |ψ (4) C(x) to be 2π periodic. We therefore instead define
up to statistical errors, which are controlled by averaging π/2

over multiple repetitions of the circuit, or “shots.” This gk C(x) = p(y)(x − y)dy, (8)
−π/2
is the main quantity of interest that we seek to estimate by
quantum computation.
where (x) is a 2π -periodic Heaviside step function and
In general, |ψ will not be an exact eigenstate of H . We
p(y) is the probability distribution of τ H associated with
denote the expansion of |ψ in the eigenbasis of H by
|ψ,

|ψ = νi |i , (5)
i p(y) = pi δ(y − τ λi ). (9)
i
which gives
It is straightforward
to check that this definition gives the
gk = pi e−iτ λi k , (6) desired C(x) = i:λi ≤x pi for |x| ≤ π/2. For |x| > π/2,
i this is not true and we should therefore choose τ such that
040341-3
τ H ≤ π/2, where . . . denotes the operator norm. The 2. Importance sampling

discontinuities of C(x) in the range |x| ≤ π/2 can then be The summation to be estimated in the CDF-based QPE
used to identify eigenvalues of τ H . approach takes the form given in Eq. (15), where each term
One can proceed by considering a Fourier-series expan- is weighted by a Fourier coefficient, |Fk |. These Fourier
sion of the step function. Reference [9] defines coefficients decay rapidly, such that contributions at large
k may be several orders of magnitude smaller than those at
F(x) = Fk eikx , (10)
low k.
|k|≤N
For this reason, Ref. [8] suggests performing importance
where F(x) is an approximation to (x), which is derived sampling of this summation. Here, terms are randomly
by truncating the Chebyshev polynomial expansion for sampled with probabilities proportional to |Fk |,
an erf(·) function. The authors carefully constructed this
approximation to satisfy various bounds on its error and |Fk |
on the scaling of |k|<N |Fk | with N . The derived Fourier Pk = , (16)
S
coefficients are

F0 = 1/2, where S = Nk=1 |Fk |. We obtain a set of NS values
{k1 , . . . , kNS }, where each ki is sampled with probability
β −β Ij (β) + Ij +1 (β) Pki . The importance-sampled CDF (which we denote by
F2j +1 = −i e , 0≤j ≤d−1
2π 2j + 1 H̃ (x)) can then be constructed as

β −β Id (β)
F2d+1 = −i e , (11)
2d + 1 1 2S
NS
2π
H̃ (x) = + Re[gki ] sin(ki x) + Im[gki ] cos(ki x) ,
where d is related to N above by N = 2d + 1. In addition, 2 NS i=1
we have F−k = −Fk for all k = 0. In (β) is the nth modified (17)
Bessel function of the first kind. The Fourier coefficient Fk
is nonzero for odd k only; even k do not contribute.
The Fourier coefficients, and hence the approximate which is an unbiased estimator for C̃(x). The estimates of
Heaviside function F(x), depend on a parameter β > 0, Re[gki ] and Im[gki ] for each ki are then each obtained by
which represents the sharpness of the erf(·) function. For performing Hadamard tests as in Fig. 1, with the result-
untruncated k, the approximation becomes more accurate ing real and imaginary components denoted by ri and si ,
as β increases. However, the {Fk } decay more slowly with respectively. In our experiments, each ri and si estimate
increasing β and so for a truncated summation, there is a will be averaged over multiple shots in practice, since on
trade-off in the choice of N and β. current cloud-based platforms it is inefficient to perform a
An approximate periodic CDF can then be expressed as circuit for a single shot, due to the overhead in submitting
π/2 a circuit. The CDF can then be constructed by
C̃(x) = p(y)F(x − y)dy (12)
−π/2
1 2S
NS
G̃(x) = + [ri sin(ki x) + si cos(ki x)] , (18)
2 NS i=1
−iτ Hk
= Fk e ψ|e
ikx
|ψ (13)
|k|≤N
which again is an unbiased estimator for C̃(x). Note that
new estimates of ri and si are taken for each sample
= Fk eikx gk . (14)
ki (rather than only obtaining single estimates for each
|k|≤N
unique k). Also note that we use the same set of sam-
Noting that g−k = gk∗ and that Fk = −i|Fk | for k > 0 and ples {k1 , . . . , kNS }, {r1 , . . . , rNS } and {s1 , . . . , sNS } for every
Fk = i|Fk | for k < 0, this can be simplified to the following value of x when constructing G̃(x) (rather than performing
final expression: a fresh sample for each x).
We emphasize that there are two separate lev-
1 N
els of sampling here. We refer to |C̃(x) − H̃ (x)| as
C̃(x) = +2 |Fk | [Re[gk ] sin(kx) + Im[gk ] cos(kx)] .
2 “importance-sampling error,” whereas |C̃(x) − G̃(x)| con-
k=1
(15) tains importance-sampling error and also “shot noise.”
To aid with discussion later, it will be helpful to write
In practice, one only needs to estimate gk for k ≥ 1 and Eq. (18) in an alternative form. Let nk denote the number
only for odd values of k. of times that k is sampled during importance sampling. We
040341-4
1.0 (a) 350 (b)

300
0.8
CDF derivative
250
0.6 200
CDF
150
0.4
100
0.2
Exact energies 50
Without sampling errors
0.0 With importance-sampling errors 0
−0.5 −0.4 −0.3 −0.2 −0.1 0.0 0.1 −0.368 −0.367 −0.366 −0.365 −0.364 −0.363 −0.362
x x
FIG. 2. (a) The CDF for H4 STO-3G, taking the Hartree-Fock wave function as the initial state, calculated with and without sampling
errors. In this example, the results are simulated. The parameters are β = 106 and d = 5000. The “Without sampling errors” CDF is
obtained via Eq. (15), using exact values of gk . The CDF “with importance-sampling errors” is obtained from Eq. (17) with NS = 5000
but again using exact gk values. The dashed vertical lines are exact energies. (b) The derivative of the CDF, enlarged in the region
around the ground-state energy. The maximum of the CDF derivative provides an accurate energy estimate.
can then write the CDF-QPE method. This rapid decay is the reason for
the large efficiency gain in using importance sampling.
1 2S
N
G̃(x) = + nk [r̃k sin(kx) + s̃k cos(kx)] , (19)
2 NS k=1
where 3. Estimating energies in the CDF-QPE method

Equation (15) provides a formula to construct the
1 1
r̃k = ri , s̃k = si (20) approximate CDF, which in the limit of large β and N can
nk i:k =k nk i:k =k be used to obtain the exact energies τ λi of τ H through
i i
its jump discontinuities. In practice, this function is only
are estimates of the real and imaginary parts of gk , aver- constructed to a finite precision and a method is needed to
aged over all repeated samples of k. In order to mitigate estimate λi from the approximate C̃(x).
coherent errors when running on a QPU, we will perform Reference [9] proves that F(x) can be constructed with a
a separate Pauli twirl for each sample of k. Therefore, r̃k guaranteed level of accuracy, provided that sufficient β and
(s̃k ) will denote the estimate of Re[gk ] (Im[gk ]) averaged d are chosen. In particular, it is proven that for any > 0
over nk Pauli twirls of the appropriate Hadamard test (with
each single-Pauli-twirl estimate, ri or si also averaged over
multiple shots in practice).
As an example, Fig. 2(a) presents simulated results for 10−1
H4 in a STO-3G basis, with a square geometry of side

10−3
length 1.28 Å. The state |ψ is taken to be the Hartree-
Fock state. The CDF estimates used here are C̃(x) (red) and 10−5
H̃ (x) (blue) with NS = 5000 (thus shot noise is not present

|Fk |
10−7
in these examples). Because this problem is multirefer-
ence, the CDF has multiple “jumps,” each corresponding 10−9
to an energy eigenvalue of τ H . While importance sam- 10−11 QEEA

pling introduces noise into the CDF estimate, the first few CDF QPE
eigenvalues can still be clearly identified. 0 500 1000 1500 2000 2500 3000 3500 4000
k
Figure 3 plots the values |Fk | against k for the CDF-
QPE method and compares them to those from Somma’s FIG. 3. A comparison of the Fourier coefficients from the two
QEEA, which is discussed in Appendix B. In the QEEA, statistical phase estimation methods considered (rescaled so that
we take the half bin width as = 3 × 10−3 . In the CDF- |F1 | = 1), for similar target accuracies. In the QEEA [7], we
QPE method, we set β = 105 , which has been chosen to choose a bin size of 3 × 10−3 . In the CDF-QPE method, we set
target an equivalent accuracy of around 3 × 10−3 in esti- β = 105 , which gives at least a similar level of accuracy with
mates of λi . The Fourier coefficients Fk decay rapidly in high probability. The Fourier coefficients decay rapidly in both
both methods, although the decay is much more rapid in methods, which allows efficient importance sampling.
040341-5
and δ ∈ (0, π/2), the condition restricting the jump to a region of width approximately
2δ. However, from the Fourier definition of F(x), it can
|(x) − F(x)| ≤ ∀ x ∈ [−π + δ, −δ] ∪ [δ, π − δ] be seen that the maximum of the derivative of F(x) lies at
(21) exactly x = 0, even for small β and d. To see this, consider
the derivative of F(x),
is satisfied, provided that
F (x) = Sk eikx , (25)
1 3 |k|≤N
β = max W ,1 (22)
4 sin δ
2 π 2
where we have denoted Sk = ikFk . The coefficientsSk obey
and with a sufficient d, which can be chosen as d = Sk = S−k ≥ 0 for all k. Therefore, F (0) = 2 Nk=1 Sk .
O(δ −1 log( −1 )), and W(·) denotes the principal branch of Since Sk ≥ 0, this is the maximum value of F (x), as
the Lambert W function. Throughout this paper, we use Eq. claimed. Now consider the simple case when p0 = 1, so
(22) to choose β for a given target accuracy of δ, after that the initial state |ψ is an exact eigenstate of τ H with
loosely setting = 0.1. Note that it is always possible to energy τ λ0 . In this case, p(y) = δ(y − τ λ0 ). Since C̃(x)
construct F(x) via Eq. (10) to check the accuracy of F(x) is a convolution between F(x) and p(y), the maximum of
for a given β and d; we use this approach to choose d after its derivative will then lie at exactly τ λ0 , even for small
first choosing β from Eq. (22). Given the accuracy guar- β. In the more general case where multiple pi values are
antees above, the authors of Ref. [9] estimate τ λi with a nonzero, the CDF derivative will be a sum of such contri-
procedure similar to a binary search, suggested in Ref. [8]. butions that will overlap and this argument no longer holds
This approach allows a careful proof of the scaling of the exactly. However, if the gap between eigenvalues τ λi is
algorithm for a given target accuracy δ. much greater than δ, then it is expected to remain a signif-
In this paper, we take a different practical approach to icantly better approximation than the bound provided by
estimating τ λi from C̃(x), which we find can give more Eq. (21).
accurate estimates than the target accuracy δ by 1–2 orders These arguments will be affected by the presence
of magnitude or more. This is performed by maximizing of noise, including shot-noise and importance-sampling
the derivative of C̃(x) in the region of each jump. From Eq. errors. Also note that the additional factors of k in G̃ (x)
(15), the derivative can be calculated (up to an unimportant will increase noise from high-k contributions, potentially
constant) as making the derivative more susceptible to errors. We will
show in our results, however, that this approach is often

N robust in practice.
C̃ (x) = |Fk | k [Re[gk ] cos(kx) − Im[gk ] sin(kx)] . Figure 2(b) plots the CDF derivative for the H4 example
k=1 described above, enlarged in the region of the ground-state
(23) energy. It can be seen that the maximum is an accurate
estimate of τ λ0 , even after applying importance sampling.
This can be viewed as an objective function and the loca- By numerically maximizing this function, the estimate of
tions of its local maxima (in the regions of jumps in τ λ0 is in error by only 2.5 × 10−7 Ha without importance
C̃(x)) can be used to estimate each τ λi . When performing sampling [i.e., using Eq. (15)], which increases to 9.2 ×
importance sampling and averaging over shots, we instead 10−6 Ha with importance sampling [using Eq. (17)]. This
take can be compared to the width of the jump region, which is
approximately 10−3 Ha.

N
G̃ (x) = nk k [r̃k cos(kx) − s̃k sin(kx)] (24) We note that a comparable approach has recently been
k=1
suggested by Wang et al. [13]. In this, the method of Lin
and Tong is used to find an approximate region where the
as an objective function, which provides an unbiased esti- ground-state energy is located. A more accurate estimate
mator of C̃ (x), up to an unimportant overall constant. is then obtained by finding the maximum of (fσ ∗ p)(x)
Here, nk , r̃k , and s̃k are as defined in Sec. II A 2. Note that in this region, where fσ (x) is a Gaussian filter kernel. We
although C(x) is discontinuous in the exact case, for any expect that our approach is comparable from a practical
finite d and β, the function C̃ (x) will be well defined and point of view, although it avoids working with a separate
so this does not lead to practical issues. Gaussian kernel. Instead, we work with the derivative of
To motivate why the above provides accurate esti- F(x) in place of fσ (x).
mates of τ λi , recall that C̃(x) is defined as a convolution
between the approximate Heaviside function, F(x), and B. Variational-circuit compilation
the probability density function, p(y), as in Eq. (8). F(x) Applying statistical phase estimation requires estimat-
is constructed to meet the accuracy condition in Eq. (21), ing gk = ψ|e−iτ Hk |ψ using the Hadamard test of Fig. 1.
040341-6
This requires implementation of the controlled-e−iτ Hk uni- methodology is not scalable to systems beyond classi-
tary for a range of k values. One commonly consid- cal computation but is nonetheless valuable for near-term
ered approach to approximately implement this unitary NISQ studies. Alternative loss functions based on the
is through Trotter product formulas. On current NISQ reduced density matrices can be used instead [46]; this
devices, the circuit depth required to achieve this is far compilation strategy could then be applied to subcircuits
too high for nontrivial problems, particularly for ab initio on fewer qubits than the total circuit, allowing the approach
Hamiltonians, where there are many terms in H , and for to scale to large numbers of total qubits. Alternatively,
the high values of k required for good precision in QPE. approaches based on tensor networks could be used [49].
Instead, in this work we primarily use a variational- We do not consider these alternative approaches here and
compilation technique that allows the action of each instead work with Eq. (27). While the use of constant-
controlled-e−iτ Hk operation to be compiled to a constant depth circuits simplifies some aspects of the error mitiga-
circuit depth, up to a negligible error. This technique has tion, there are many aspects of the error-mitigation task
recently been used in other studies applying closely related that remain challenging and important to treat carefully, as
circuits on superconducting processors [46,47]. More gen- we shall see. To investigate the additional challenges intro-
erally, the task of variationally optimizing the action of duced by using Trotterization, we also study a Trotterized
time-evolution operators on a given input state has been example on a QPU at the end of Sec. IV. Additionally,
considered in Ref. [48] and generalized to approximate the a Trotterized example is studied in Appendix C in the
whole unitary using tensor-network methods in Ref. [49]. presence of both unitary and depolarizing errors.
The results in these papers suggest that variationally opti- The circuit optimization is implemented using the JAX
mized circuits can be made significantly shorter and more library [51], which enables automatic differentiation of
accurate than Trotter product circuits. A related approach PYTHON functions. We use the BFGS algorithm imple-
has also been applied to run quantum signal processing on mented in the JAX library to perform the minimization of
a trapped-ion device, using parametrized quantum circuits L(p). We find BFGS to be far more robust than optimiza-
that are again optimized variationally [50]. tions using stochastic gradient-descent methods for this
We consider a circuit ansatz as shown in Fig. 4, con- task.
sisting of alternating one- and two-qubit layers. The one-
qubit layers consist of U3 gates, each of which allows
an arbitrary one-qubit rotation. The two-qubit layers are C. Error mitigation
constructed using CZ gates, entangling alternating pairs of 1. Zero-noise extrapolation
qubits in each layer with a “brickwork” pattern. The U3
In this paper, we apply ZNE [18–20,52] to mitigate
gates are parametrized by Euler angles (θ, φ, and λ) and
errors in the expectation values of the Hadamard tests.
implemented in native gates for Rigetti’s processors as
ZNE is one of the most commonly studied error-mitigation
methods in the literature. The core idea of ZNE is to exe-
U3 (θ, φ, λ) = RZ (φ) RX (−π/2) RZ (θ ) RX (π/2) RZ (λ), cute the target circuit at varying error rates, denoted by
(26) λ, and extrapolate the results to obtain an estimate at a
reduced error rate. Expectation values are estimated for
where rotation gates are defined by RZ (θ ) = e−iθ Z/2 and
RX (θ) = e−iθX /2 . The parameters in the U3 gates can be
variationally optimized such that the circuit ansatz closely
matches the action of the desired unitary on a given input
state. Specifically, following Refs. [47,48], we define the
loss function
L(p) = U| − Ũ(p)|, (27)
where the L2 norm is used. Here, U is the target uni-

tary, Ũ(p) is that of the ansatz circuit with parameters p,
and | = |+ ⊗ |ψ is the input state to U in the cir-
cuit. We apply this variational-compilation procedure to FIG. 4. The circuit ansatz used to compile controlled-e−iτ Hk
the controlled-e−iτ Hk operation. Other components in the operations. The one-qubit layers consist of U3 gates. Each U3
circuit are constructed directly as native gates. gate is specified by three parameters that are optimized to approx-
We perform the minimization of L(p) classically. Con- imately match the action of the desired unitary. Each U3 gate is
structing and optimizing this loss function requires con- applied as five native gates on the quantum processor (see Eq.
structing the action of U on |. As such, the current (26)). The two-qubit layers are formed from CZ gates.
040341-7
the original circuit, defined by λ = 1, in addition to cir- shots, as we do in this work (for details, see Sec. III B).
cuits at increased error rates λ > 1. A function is fitted to After inserting the Pauli layers, each circuit contains sub-
these expectation values and used to extrapolate to error sequent layers of Pauli and U3 gates, which can be merged
rate λ = 0, which gives the error-mitigated estimate. into a single U3 gate layer. This procedure is known to con-
There are various possible methods to increase the error vert an arbitrary error channel into a Pauli error channel,
rate λ. Examples in the literature include parameter noise thus eliminating coherent errors.
scaling and pulse stretching [20,52]. In our implementation Our implementation of randomized compiling differs
of ZNE, we increase λ using identity insertion (sometimes from previous descriptions due to the use of importance
referred to as “unitary folding”) [53]. Identity insertion n > sampling. Remember that our goal is to estimate the CDF,
0 times replaces a unitary operation U according to C̃(x), defined in Eq. (15). As discussed in Sec. II A 2,
we importance sample this summation, as contributions at
U → U(U† U)n . (28) high k will typically be orders of magnitude smaller than
at small k. We incorporate the twirling procedure into the
The error rate λ is then defined as importance sampling of C̃(x). If nk denotes the number
of times that k is sampled during importance sampling,
λ = 1 + 2n. (29) then the estimates for Re[gk ] and Im[gk ] are each averaged
over nk independent twirls of the corresponding circuits.
According to this definition, λ = 1 corresponds to n = 0, These estimates are denoted by r̃k and s̃k , as defined in Eq.
meaning that no identity insertion is performed. (20). Therefore, estimates for k = 1 will typically be aver-
The circuits in this study are performed in layers of one- aged over a large number of twirls, while many circuits for
or two-qubit gates that are executed in parallel (see Sec. large values of k will be performed for just a single twirl
II B and Fig. 4). In our ZNE implementation, we fold full (i.e., without averaging). This will lead to poorer results
two-qubit gate layers, which typically have higher error at large k (both larger coherent errors and larger statisti-
rates than one-qubit gates on superconducting devices. cal errors), which will also impact on the performance of
Folding layers of gates rather than individual gates helps to ZNE. However, because these terms are weighted by nk ,
ensure a consistent error profile with folding, e.g., ensuring their contribution will be small and thus it is to be expected
that the crosstalk will be consistent in each layer. that the corresponding errors will not significantly impact
In our implementation of ZNE, we execute circuits at C̃(x). The converse is also true; using this approach, errors
error rates λ = 1, 3, and 5. To obtain circuits at error rates at low k will be much smaller, which is beneficial for
λ = 3 and λ = 5, we apply identity insertion as in Eq. (28), the final estimate of C̃(x) due to their high weight in the
with n = 1 and n = 2, respectively. A possible drawback summation.
of using such large error rates as 3 or 5 is that for low The mitigation of coherent errors is known to improve
gate fidelities, the final error can be too large to perform a the performance of ZNE, allowing a more reliable fitting
reliable extrapolation. However, in the examples studied of the expectation values with λ. This has been demon-
in this paper, this is not found to be a significant prob- strated in, e.g., Ref. [58], which has provided theoreti-
lem. Furthermore, the use of only odd-integer λ values cal arguments to justify this finding, and will be further
means that every two-qubit gate layer is folded, avoiding demonstrated in our results. Important improvements have
complications around having to pick a subset of layers to also been demonstrated in the VQE algorithm [59]. Specif-
fold. ically, we find an exponential fit to be accurate in most
cases. It should be pointed out, however, that the theoret-
2. Mitigating coherent errors ical arguments for well-behaved exponential ZNE extrap-
As discussed in Sec. II C 1, we fold the two-qubit lay- olations in Ref. [58] only hold for depolarizing channels
ers, which consist of CZ gates in this work. Such CZ gates and, indeed, the authors show that this result does not hold
typically suffer from significant coherent errors on cur- in general for Pauli channels with nonequal Pauli weights.
rent superconducting devices. To mitigate these coherent Despite this, we will see that ZNE performs well after ran-
errors, we apply a form of RC [25,26]. This is achieved domized compiling. We mention that the noiseless output
by applying random Pauli gates, uniformly sampled from extrapolation (NOX) method [60] has been proposed to
{I , Z, X , Y}, to each qubit before a CZ layer. Because CZ overcome the above potential shortcomings, although we
gates are Clifford operators, it is always possible to then do not consider this method here.
apply corresponding cancelling Paulis after the CZ layer Given the above, our strategy for ZNE is to attempt an
[54,55]. This essentially has the effect of Pauli twirling exponential fit for Re[gk ] and Im[gk ] at every value of k
the CZ gate layer [56,57]. This twirling process is usu- sampled. However, in some cases (particularly at large k
ally averaged over multiple instances of the same circuit, where nk is small), this fit may be unstable or of poor
with a new set of random Paulis applied in each. One can quality. Since we know that all Re[gk ] and Im[gk ] values
additionally average each twirl of the circuit over multiple must lie in the range [−1, +1], we loosely check that the
040341-8
extrapolated estimate is less 1.2 in magnitude. If this is III. METHODS

not the case, then we instead switch to a quadratic fit for
A. Software
that data point. This strategy avoids extremely large gk
estimates due to unstable exponential fits. To generate phase-estimation circuits, including
variational-circuit compilation, we use software devel-
oped by Riverlane. These circuits are then converted to
3. Readout-error mitigation pyQuil format and compiled to executables that are run on
We apply readout-error mitigation to our results using Rigetti’s Aspen quantum processors, using their Quantum
the symmetrized approach described in Ref. [17]. Other Cloud Services (QCS) platform [62] and associated soft-
readout-mitigation strategies include methods based on ware [63]. We also use pyQuil extensively to perform prior
assuming readout noise to be local or based on continuous- testing on quantum virtual machines (QVMs).
time Markov processes [16,61]. We use the ORCA program package [64] to perform
Suppose that we are mitigating the readout of n qubits. embedding calculations and PySCF [65,66] to perform
Define a calibration matrix A as Hartree-Fock to generate the fermionic Hamiltonian for
other systems. The Gaussian 16 program package [67] is
Aij = P(measure |i | prepared |j ), (30) used to perform geometry optimization for one structure
detailed in Appendix A. The OpenFermion library [68] is
used to perform mapping from fermionic to qubit operators
where |i and |j are computational basis states. Also using the Bravyi-Kitaev mapping [69], discussed further
define a vector C such that the ith element is equal to the below. Variational-circuit compilation is performed using
number of times the n-qubit state is measured to state |i. the JAX library [51].
Then let Cideal be such a vector C under perfect readout.
Applying matrix A to Cideal gives an estimate of the results
under noisy readout, B. Implementation details
Here, we discuss some specifics related to implementa-
E[Cnoisy ] = ACideal , (31) tion of circuits for Rigetti’s software stack and QPUs.
As discussed in Sec. II C 2, we perform a separate
where E[·] denotes an expectation value. Therefore, an Pauli twirl of the circuit for each value ki obtained dur-
estimate of Cideal can be obtained by inverting A, ing importance sampling. For example, if k = 1 is sam-
pled n1 = 50 times during the importance-sampling step,
then the estimates for r̃1 and s̃1 will each be averaged
Cideal ≈ A−1 Cnoisy . (32) over 50 independent twirls of the corresponding k = 1
circuits.
To estimate the matrix A, we repeatedly prepare and mea- Ideally, we would like to perform as many Pauli twirls
sure each of the 2n computational basis states. We use the as possible and so would seek to perform an independent
measurement outcomes to estimate the probabilities as in twirl for each shot. In practice, each separate twirl must be
Eq. (30) and thus the matrix A. submitted to the QPU as a separate circuit. Because circuit
In practice, the readout error when measuring state |1 loading takes significantly longer than circuit running, it is
is often higher than the readout error when measuring state inefficient to perform only one shot per twirl. Instead, we
|0, so that the calibration matrix A will not be symmetric. perform 100 shots of each, with 50 each for the original and
Moreover, readout errors often drift quite rapidly, so that bit-flipped versions of the circuit in order to symmetrize
a given estimate of A may not be accurate throughout an readout errors. Therefore, when constructing the estimates
experiment. We can nonetheless symmetrize the calibra- r̃k and s̃k as in Eq. (20), it should be understood that the
tion matrix by bit-flip averaging [17], in which an X gate estimates ri and si are each averaged over 100 shots in
is applied directly before measurement for half of the shots this manner, before averaging over nk independent twirls
performed (for every circuit involved in estimating each to construct r̃k and s̃k . These are the final estimates plotted
gk ). This symmetrizes any errors that remain after apply- for Re[gk ] and Im[gk ] in subsequent sections and also used
ing readout mitigation, resulting in a better-behaved final in constructing the final CDF estimates via Eq. (19).
CDF estimate. Nonsymmetric readout errors will generally To improve the effectiveness of ZNE, it is important
result in additive errors in the gk estimates, which can show for each gate layer to be as consistent as possible, thus
up as a spurious signal at λi = 0 in the Fourier transform. ensuring a similar noise profile for each. To help achieve
Symmetrizing readout errors in this way can therefore sig- this, we add FENCE statements around all two-qubit layers
nificantly improve the quality of the results. This simple in the Quil circuit, which ensures identical pulse tim-
approach of applying X before measurement in 50% of ings. Additionally, we decompose all one-qubit gates with
shots is found to be very effective in practice. the structure RZ (φ) RX (−π/2) RZ (θ) RX (π/2) RZ (λ). This
040341-9
1.0 1.0
(a) (c)
0.5 0.5
Re(gk )
Re(gk )
0.0 0.0
−0.5 −0.5
−1.0 −1.0
10 20 30 40 50 60 10 20 30 40 50 60
k k
1.0 (b) 1.0 (d)

0.8 0.8
0.6 0.6
C̃(x)
C̃(x)
0.4 0.4 Exact energies
without QPU errors
0.2 0.2 Error rate = 1
Error rate = 3
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
x x
FIG. 5. Results from the CDF-QPE method performed on the Aspen-11 QPU at three different error rates (ZNE-extrapolated results
are not presented here), for methanethiol (2e, 2o) with a stretched SH bond. (a),(b) Results performed without mitigation of coherent
errors. (c),(d) Results with mitigation of coherent errors by RC. There are three significant eigenstates present, the exact eigenvalues
of which are marked by dashed lines in the CDF. The mitigation of coherent errors leads to significantly better-behaved estimates
of gk with an increasing error rate, which leads to improved estimates of C̃(x). The locations of jumps in the CDF match the exact
eigenvalues closely, even at high error rates.
includes trivial Pauli gates such as I and Z. Our aim here is The qubit Hamiltonian is generated using the Bravyi-
to again ensure that gate layers are as consistent as possi- Kitaev qubit mapping [69] for all systems. Specifically, we
ble. It should be noted that RZ gates are performed virtually use the approach of Ref. [71], which allows two qubits to
on Rigetti’s quantum processors [70]. be tapered due to spin- and particle-number symmetries,
and is implemented in OpenFermion [68]. Thus, for an
C. Chemical systems active space of M spatial orbitals, the corresponding qubit
Hamiltonian requires 2M − 2 qubits. For the Trotterized
We study five separate systems in active spaces from H2 example, we taper a further qubit, which is possible due
two to four spatial orbitals. Here, the active space refers to reflection symmetry. This allows the Hamiltonian for H2
to a particular set of orbitals and a number of electrons STO-3G to be represented by a single qubit.
used to occupy those orbitals. An active space with n elec- The Hartree-Fock wave function is taken as the initial
trons in m spatial orbitals (2m spin orbitals) is denoted wave function, |ψ, for all systems. The PySCF input files
(ne, mo). Two of our example systems are motivated by used to generate fermionic integrals for H2 , H+ −
3 , H3 , and
pharmaceutical applications and use a recently developed H4 , including the molecular geometries, are included in the
embedding method to target a chemically relevant region additional data [72].
of the molecule with a small active space [45]. The first of
these systems is methanethiol, using a (2e, 2o) active space
for a minimal model of hydrogen abstraction. The second
is a structure that we refer to as “clusterTS,” taking a (4e, IV. RESULTS
4o) active space. Both of these systems are described in
Appendix A. A. Methanethiol (2e, 2o)
We also study H+ −
3 and H3 in the STO-3G basis as As a first example, we consider application to
example three-orbital systems. The geometry is an equi- methanethiol in a (2e, 2o) active space, using orbitals cen-
lateral triangle in both cases, with a bond distance of tered on the sulfur-hydrogen (SH) bond, as described in
0.9 Å in H+ −
3 and 1.75 Å in H3 . Lastly, we study H2 to Appendix A. To model the dissociation limit, we take
investigate a minimal example of Trotterization. Here, a the SH distance to be 4 Å. This is a minimal model but
STO-3G basis is used once again, with a stretched bond it results in a multireference problem. We focus first on
length of 2.0 Å. demonstrating how the CDF is constructed from the QPU
040341-10
−0.4 (a) 1.2 (b)

−0.5 1.0
−0.6 0.8
Re(g3 )
C̃(x)
0.6
−0.7
0.4
−0.8
Exact value 0.2 Exact energies
−0.9 Exponential fit without QPU errors
Without RC 0.0 Without RC
With RC With RC
−1.0
−0.2
0 1 2 3 4 5 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
λ x
FIG. 6. The results for methanethiol (2e, 2o) with a stretched SH bond, using gk estimates obtained from Aspen-11. (a) Estimates of
Re[gk ] for k = 3 as an example, performed at error rates λ = 1, 3, and 5. An exponential fit is accurate after performing randomized
compiling, leading to an improved ZNE estimate at λ = 0. Such a fit cannot be reliably performed for data obtained without mitigation
of coherent errors. (b) CDFs constructed using ZNE-extrapolated gk estimates.
calculations and, in particular, on the effect of incorporat- λ = 1 can be used to correctly estimate the energies of
ing RC into the importance-sampling procedure. In Sec. τ H through its jumps but this becomes challenging for the
IV B, we then apply this approach to more challenging excited states at higher error rates and the general qual-
Hamiltonians with four and six qubits. ity of the CDF is poor. This is improved significantly by
The qubit Hamiltonian for this system can be con- applying RC. The energies are clearly identifiable at all
structed using two qubits, so that three qubits are required error rates and the shape of the CDF is largely correct.
for each Hadamard test. We variationally compile each By inspecting the values of Re[gk ], it can be seen that the
controlled-e−iτ Hk operation to a brickwork circuit ansatz. expectation values decay with increasing error rate in a
Here, the variational compilation can be performed with more systematic manner when RC is applied than with-
negligible errors using only three layers of CZ gates; the out. Figure 6(a) emphasizes this behavior by showing an
value of L(p) is typically smaller than 10−6 after optimiza- example ZNE extrapolation for the real component of gk
tion. The total number of CZ gates is three, nine, and 15 for at k = 3. An exponential fit is seen to be accurate with RC
error rate λ = 1, 3, and 5 circuits, respectively. applied, which leads to an improved estimate. This expo-
The calculations are performed on Aspen-11 using nential decay is not observed when RC is not incorporated,
qubits 11, 26, and 27, with qubit 11 taken as the ancilla. making ZNE less effective.
CDF Fourier parameters are taken as β = 105 and d = Figure 6(b) shows the CDFs obtained from ZNE-
2 × 103 . From Eq. (22) together with = 0.1, this corre- extrapolated estimates of gk . The same general behavior
sponds to an accuracy of δ ∼ 0.003. NS = 2 × 103 samples is observed; with RC applied, the ZNE-corrected CDF
are taken for the importance sampling and 100 shots are has roughly the correct amplitude compared with the
performed per sample, thus the total number of shots is exact result. In contrast, the ZNE-corrected CDF obtained
2 × 105 . Due to the very rapid decay of Fk , most sam- without RC is of relatively poor quality and has large fluc-
ples are performed at small k. For example, 530 samples tuations (even beyond the expected importance-sampling
are performed at k = 1 and 153 samples are performed at noise) away from the jump regions. These results demon-
k = 3, whereas only 28 samples are taken in total for all strate the importance of mitigating coherent errors in
values k > 1000. statistical phase estimation experiments.
In all of the results, we compare with the CDF labeled
“Without QPU errors,” which is calculated numerically
with importance sampling using exactly the same sam- B. Four- and six-qubit Hamiltonians
ples {ki } but in the absence of QPU errors or shot noise; We next apply these methods to larger systems, first
this corresponds to H̃ (x) in Eq. (17). ZNE aims to cor- to H− 3 , which requires five-qubit circuits, and then to the
rect gate errors by improving estimates of gk but cannot clusterTS system defined in Appendix A, which requires
correct importance-sampling noise, and so this is the fair seven-qubit circuits. Given the improvements observed by
comparison to make. applying RC in Sec. IV A, it is applied for all results in this
Figure 5 presents results for the real components of section.
gk (up to k = 79) and the corresponding CDF estimates, Figure 7 shows the results for H− 3 . Here, controlled-
for error rates λ = 1, 3, and 5, performed both with and e−iτ Hk
unitaries are compiled with seven CZ layers for
without randomized compiling. Without RC, the CDF at the λ = 1 circuits, with two CZ gates per layer. The CDF
040341-11
1.0
(a) 1.0 (c)
0.8
0.5
0.6
Re(gk )
G̃(x)
0.0
0.4
0.2
−0.5
0.0
−1.0
10 20 30 40 50 60 70 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5
k x
500
0.8
(b) (d) Exact energies
0.6 without QPU errors
400
0.4 Error rate = 1
Error rate = 3
0.2 300
Im(gk )
G̃ (x)
Error rate = 5
Extrapolated
0.0
200
−0.2
−0.4 100
−0.6
0
10 20 30 40 50 60 70 −0.2040 −0.2035 −0.2030 −0.2025 −0.2020 −0.2015 −0.2010 −0.2005 −0.2000
k x
FIG. 7. Results from the CDF-QPE method for H− 3 in a STO-3G basis, performed on Aspen-M-3. Here, the ground-state wave
function is multireference, leading to three jump regions in the CDF, each indicating an energy eigenvalue. (a),(b) Estimates of the real
and imaginary parts of gk for error rates 1, 3, and 5, and the subsequent ZNE-extrapolated estimates. Note that we set d = 5000 but
only present results up to k = 79 for clarity. (c) The CDF itself and (d) its derivative, enlarged in the region of the ground-state energy.
The derivative of the CDF can be used to obtain an extremely accurate estimate of each energy. Extrapolation improves the amplitude
of the CDF, although the location of the jump is not affected.
Fourier parameters are β = 106 and d = 5 × 103 . NS = by averaging over a large number of circuits; when
4 × 103 samples are taken for importance sampling. The considering both X and Y measurement bases, each ZNE
circuits are performed on Aspen-M-3 using qubits 30 and error rate, the large number samples NS (each correspond-
34–37, with 34 taken as the ancilla. The corresponding CZ ing to a separate Pauli twirl), and bit-flip averaging to
fidelities, as estimated by randomized benchmarking, are symmetrize readout errors, 57 600 separate circuits are per-
between 97.5% and 99.3%. formed to construct the CDFs, with 50 shots of each.
It can be seen that ZNE does a good job at correcting The circuits are again performed on Aspen-M-3, using
gk estimates, particularly at low k, where more samples are qubits 30–32 and 34–37, with qubit 34 again taken as the
taken. This system is multireference, with three eigenstates ancilla. The corresponding CZ fidelities, as estimated by
having a significant overlap with the initial Hartree–Fock randomized benchmarking, are between 96.7% and 99.2%.
state, leading to three jump regions visible in the CDF. Figures 8(a) and 8(b) show the real and imaginary com-
The ground-state energy is clearly identifiable at all error ponents of gk at each error rate and the ZNE-extrapolated
rates, although there is no clear signal from excited states at results. It is again observed that gk estimates decay sensi-
error rate λ = 5. Also shown is the CDF derivative, G̃ (x), bly with λ within statistical errors and that the extrapolated
enlarged in the region of the ground-state energy; the max- estimates are a significant improvement for most k values.
imum of this objective function provides an extremely The final CDF estimates are shown in Fig. 8(c). Here, the
accurate estimate of the true energy at all error rates. ground-state wave function is single reference, leading to
Figure 8 presents equivalent results for the clusterTS a single jump in the CDF, which is distinguishable at each
system, which has four orbitals in the active space. Here, error rate. The maximum of the CDF derivative in Fig. 8(d)
we take CDF Fourier parameters β = 106 and d = 5 × again allows an accurate estimate of λ0 to be obtained.
103 , and NS = 4.8 × 103 for importance sampling. We are As one method of quantifying the improvement made
able to obtain a good representation of each circuit using to the CDF estimates by applying ZNE, we consider the
nine CZ layers for each controlled-e−iτ Hk operation. Each distance metric
layer contains three CZ gates. Thus, for λ = 1, each circuit α
contains 27 CZ gates. For the highest error rate, λ = 5, each
W= |C̃approx (x) − C̃exact (x)| dx, (33)
circuit contains 135 CZ gates. The final CDF is constructed −α
040341-12
1.0 (a) 1.0 (c)

0.5 0.8
0.6
Re(gk )
G̃(x)
0.0
0.4
−0.5 0.2
−1.0 0.0
10 20 30 40 50 60 70 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

k x
800
1.0 (b) (d) Exact energy
without QPU errors
Error rate = 1
0.5 600
Error rate = 3
Error rate = 5
Im(gk)
G̃ (x)
0.0 400 Extrapolated
−0.5 200
−1.0
0
10 20 30 40 50 60 70 −0.3895 −0.3890 −0.3885 −0.3880 −0.3875 −0.3870 −0.3865 −0.3860
k x
FIG. 8. The results from the CDF-QPE method for the six-qubit “clusterTS” system, performed on Aspen-M-3. Here, the ground-
state wave function is single reference, leading to a single jump region in the CDF. (a),(b) Estimates of the real and imaginary parts of
gk for error rates 1, 3, and 5, and the subsequent ZNE-extrapolated estimates. (c) The CDF itself and (d) the CDF derivative, enlarged
in the region of the ground-state energy.
where [−α, α] is the range in which the CDF is calculated, two important metrics are the final energy estimates and
with α = 1.5 here, and we calculate the integral numeri- the ability to distinguish these jumps from the sampling
cally. An exact estimate of the CDF corresponds to W = 0. noise. Here, we find no improvement upon performing
Table I gives the percentage reduction in W after perform- ZNE. Table I gives errors in ground-state energy estimates
ing ZNE (λ = 0) compared to the unmitigated estimates for each system, calculated as λ0 = λestimate
0 − λexact
0 . No
(λ = 1). The CDF is improved by around 77–80% for the systematic improvement is observed by performing ZNE
larger systems studied. and in many cases the error is slightly increased (although
This improvement demonstrates the potential of ZNE statistical errors may account for this). Similarly, while
to mitigate errors in expectation values, and agrees with ZNE increases the amplitude of the CDF, the importance-
previous ZNE studies. However, we find that the bene- sampling noise is also inevitably amplified. This latter
fit of performing ZNE is somewhat limited in statistical result may be expected; remember that ZNE only aims
phase estimation. Ultimately, the energies of the Hamil- to address QPU errors, and not statistical sampling errors.
tonian are estimated through the jumps in the CDF; thus, Moreover, ZNE comes with a significant sampling over-
head in order to estimate each gk with sufficient precision
at high λ, as required for a reliable extrapolation.
TABLE I. Metrics assessing the effect of ZNE. All results use The lack of improvement by performing ZNE seems to
RC and readout-error mitigation. The application of ZNE leads be associated with the natural resilience of statistical phase
to a significant improvement in the distance metric W for all sys- estimation to noise, particularly after mitigating coherent
tems studied. The error in the ground-state energy estimate, λ0 , errors and symmetrizing readout errors. Indeed, even in
is extremely small both with and without ZNE applied; how- the presence of significant QPU errors, the ground-state
ever, there is no improvement made within statistical errors by energy errors are found to be extremely small, typically
applying ZNE and the estimate is even worsened in some cases.
smaller than 0.1 mHa. Again, note that for all of the exam-
Improvement ples studied here, β is chosen according to Eq. (22) for
in W after λ0 before λ0 after a target accuracy of δ ∼ 1 mHa. Thus the final accuracy
System ZNE (%) ZNE (mHa) ZNE (mHa) is often found to significantly exceed this target, even in
Methanethiol 63.1 0.022 0.009 the presence of noise. Therefore, while applying ZNE is
H+
3 79.8 0.050 0.071 found to give little practical benefit, we find that miti-
H−
3 77.6 −0.002 0.019 gating coherent errors by RC is very beneficial and can
ClusterTS 77.3 0.089 0.149 lead to an algorithm with natural noise tolerance. Since
040341-13
statistical phase estimation requires averaging over a large where c1 = 0.121256 and c2 = 0.259138 and we have dis-
number of circuits, the required twirls can be incorpo- carded a constant shift of −0.662537. The two eigenstates
rated at minimal cost compared to the bare method, unlike of H correspond to the lowest bonding and antibond-
ZNE, which requires additional circuits to be performed, ing states of H2 . We choose τ = 1.5/(c1 + c2 ) = 3.943
the signal of which decays exponentially with increasing and work with the scaled Hamiltonian τ H . We then per-
λ. Furthermore, incorporating the twirls directly into the form first-order Trotterization with a single Trotter step,
importance-sampling procedure is found to be practically i.e., taking e−iτ Hk ∼ (e−iτ c2 X e−iτ c1 Z )k . The exact energies
effective. of τ H before and after Trotterization are ±1.128189 Ha
and ±1.089119 Ha, respectively. Thus there is a Trot-
C. Trotterization ter error of 39 mHa, or 10 mHa after rescaling by τ −1 .
The previous sections have investigated constructing the We are not concerned with Trotter error here but, rather,
CDF for circuits the depth of which is independent of with the performance of the statistical phase estimation
k. This requires mitigation of various errors that reduce method and error mitigation and thus we only compare to
the performance of the algorithm. However, for a scalable the Trotterized energies from now on. Note that a similar
approach, we require circuits the length of which grows at H2 Hamiltonian has been used in a study of textbook QPE
least linearly with k. As a final example, we investigate a on a neutral-atom quantum computer, performed to three
minimal model of H2 using Trotterization and investigate bits of precision [39].
the performance of the same error-mitigation techniques. The circuit for a single Trotter step is given in Appendix
We consider H2 in a STO-3G basis set, with a stretched C and has a CZ depth of 4. Therefore, circuits to estimate
internuclear distance of 2.0 Å. Using the Bravyi-Kitaev gk have a CZ depth of 4k. The results are performed on
transformation with qubit tapering, this Hamiltonian can Aspen-M-2 using qubits 121 and 122. The estimated CZ-
be represented by a single data qubit [71]. In particular, gate fidelity from randomized benchmarking is 99.22 ±
qubits are tapered due to particle number, spin, and reflec- 0.1662%. The ancilla is taken as qubit 121, with an
tion symmetries. This allows the qubit Hamiltonian to be estimated readout fidelity of 98.4%.
written as The simulation parameters are chosen as β = 50, d =
30, and NS = 500. These are considerably lower than those
H = c1 Z + c2 X , (34) used in previous sections due to the need to limit the
1.0 1.0
(a) (b)
0.5 0.5
Re(gk )
Re(gk )
0.0 0.0
−0.5 −0.5
−1.0 −1.0
5 10 15 20 25 5 10 15 20 25
k k
3.5 3.5
3.0
(c) 3.0
(d) Exact energy
without QPU errors
GS ES GS ES Error rate = 1
Error rate = 5
2.0 2.0 Extrapolated
G̃ (x)
G̃ (x)
1.5 1.5
1.0 1.0
0.5 0.5
0.0 0.0
−1.4 −1.3 −1.2 −1.1 −1.0 −0.9 0.8 1.0 1.2 1.4 −1.4 −1.3 −1.2 −1.1 −1.0 −0.9 0.8 1.0 1.2 1.4
x x
FIG. 9. The results for a stretched H2 molecule, using first-order Trotterization to construct e−iτ Hk , performed on the Aspen-M-2
QPU. (a),(b) Estimates of the real part of gk , estimated (a) without and (b) with randomized compiling. The dashed lines are added
between estimates for clarity. (c),(d) The derivative of the CDF constructed from gk , (c) without and (d) with randomized compiling.
The subplots labeled “GS” and “ES” are enlarged in the region of the ground state and excited state, respectively. The results are
performed at error rates 1, 3, and 5 and extrapolated. RC leads to better final energy estimates and a more accurate relative signal
between the two states. ZNE improves the signal but does not improve the energy estimates.
040341-14
circuit depth. Taking = 0.1, this corresponds to δ ≈ 0.13 chemical-embedding methods, allowing accurate energy
in Eq. (22); thus we expect a low resolution in the CDF. estimation for several small chemical problems. In addi-
The highest value of k sampled is k = 25. The highest CZ tion, a variational-compilation technique has been used
depth is therefore 100 with λ = 1 or 500 with λ = 5. to reduce the circuit depth. We find this combination
The results are presented in Fig. 9, performed with (left) of techniques to be robust in practice, allowing accurate
and without (right) RC applied: they display behavior sim- estimation of the ground-state energy with high confi-
ilar to that observed in Sec. IV A. In particular, the decay dence, even in the presence of significant QPU errors.
of gk with increasing λ is better behaved for the twirled The variational-compilation technique is also found to be
results, leading to more accurate extrapolations for ZNE. robust and should be seen as a valuable tool for near-term
A better energy estimate is also obtained (as determined NISQ studies. In the longer term, there is an interesting
by maximizing the CDF derivative) with RC applied than possibility of using this technique to optimize repeated
without. For the ground-state energy, the error in the ZNE- sub-blocks within larger circuits.
extrapolated estimate is reduced from −22 mHa to −8 We have demonstrated that the CDF-based approach of
mHa after applying RC. These errors remain large in both Lin and Tong [8], using the optimized Fourier approxima-
cases but this is expected due to the low precision used. As tion of Wan et al. [9], can be used to give significantly
observed in Sec. IV B, we again find that ZNE improves better energy estimates in practice than suggested by pre-
the amplitude of the CDF toward the exact value but does vious bounds. In particular, the derivative of the estimated
not lead to improvements in the energy estimates. CDF can be viewed as an objective function, the max-
We note that a related result has been observed in simu- imum of which gives accurate energy estimates. This
lations performed in Ref. [73], where the authors consider estimate can be orders of magnitude more accurate than
a control-free variant of phase estimation. They show that the rigorous bound derived in Ref. [9], even in the pres-
coherent noise causes errors in the phases of the unitary ence of both QPU and importance-sampling errors. This
and prove that these errors are removed to first order with improved accuracy will permit the use of shorter circuits,
RC, verified with simulations. Such a result can be sim- allowing useful applications of phase estimation to be per-
ilarly
motivated in statistical phase estimation. For gk = formed sooner. In the future, it would be interesting to
−iτ λi k
p
i i e , the Fourier transform gives a sum of delta derive a theoretical explanation of our observed improved
functions, shifted by the energies, λi . In the ideal case accuracy.
of a depolarizing noise model, we expect gk to decay We have shown that error mitigation and noise tailoring
exponentially
with k relative to the exact result, i.e., gk ∼ are important for improving the quality of the estimated
e−γ |k| i pi e−iτ λi k , for some decay rate γ . The energies CDF. In particular, there is a significant benefit in mitigat-
λi can still be extracted exactly by taking a Fourier trans- ing coherent errors. The twirling procedure in randomized
form of this gk ; the corresponding poles will be broadened compiling is incorporated into the importance sampling of
but the maxima of the poles, and therefore the energy the CDF, thus performing a much larger number of twirls
estimates, are unchanged. This also motivates why ZNE at small k than at large k. Indeed, many circuits at large
should not be expected to improve energy estimates fur- k are performed for only a single twirl in this approach.
ther. In Appendix C, simulated results are performed for Despite this, we find the approach to be robust, leading to
the same system but with higher precision, applying depo- noise resilience in the statistical phase estimation results.
larizing noise in one example and unitary errors in another. We do not find further improvements in the energy esti-
The results are found to confirm these ideas; increasing mates after applying ZNE; although ZNE does improve
the depolarizing error rate broadens each peak in the CDF the signal of the CDF in each case, the energy estimates
derivative but does not affect the energy estimates, beyond themselves and the signal-to-noise ratio are not systemat-
statistical noise. Under unitary errors, there is a signifi- ically improved. Given the very high additional sampling
cant error in the ground-state energy estimate, which is cost associated with performing ZNE, overall we find it
largely removed by incorporating RC into the importance- preferable to not use ZNE for this application. Lastly, we
sampling procedure, at the cost of broadened signals in the mention that an alternative error-mitigation approach has
CDF. Lastly, we note that Ref. [74] has also analyzed a been suggested recently, which involves postselecting the
statistical version of phase estimation and given theoreti- data qubits of the Hadamard test circuits to be measured in
cal arguments to justify the approach having tolerance to the starting state [29]; we have not tested this approach but
noise under certain models. believe that it could be combined with the methodology
developed here to improve the results further.
The results presented suggest the possibility of perform-
V. CONCLUSIONS
ing large-scale phase-estimation experiments in a manner
In this study, we have implemented statistical that allows noise resilience and also that the required cir-
phase estimation techniques on Rigetti’s quantum cuit depth for a given accuracy can often be made much
processors, in combination with error-mitigation and lower than in traditional QPE approaches. The use of
040341-15
shorter circuits and accurate results in the presence of noise methanethiol was carried out with the Gaussian 16 pro-
will be crucial for successful applications of phase estima- gram package [67] at the MP2 level of theory [77] and
tion on early fault-tolerant devices. Taken together with using the aug-cc-pVQZ orbital basis set [78–80] level of
other recent results [13,14], we believe that this shows theory. The SH distance has then been fixed to 4 Å. The
significant promise for statistical phase estimation tech- transition-state structure for the drug-protein cut-out has
niques, emphasizing their potential as valuable techniques been taken from one of our previous studies [81].
for chemistry and materials problems and motivating fur-
ther development work in this direction. As we enter the
regime of early fault-tolerant quantum computing, it will APPENDIX B: SOMMA’S QUANTUM
be important to explore whether our observed results hold. EIGENVALUE ESTIMATION ALGORITHM
This suggests further work to repeat our experiments on a (QEEA)
real or simulated early fault-tolerant quantum computer. Some alternative statistical phase estimation methods
primarily differ in their classical analysis. Therefore, once
ACKNOWLEDGMENTS a set of values {gk } have been estimated, it is relatively easy
to test alternative statistical phase estimation approaches to
We are grateful to Andy Patterson and Marco Paini for estimate the eigenvalues {λi }.
helpful discussions during this work, to Andrew Arrasmith In addition to the CDF-QPE approach discussed in
for discussions on error mitigation, and to Bram Evert for the main text, we have also implemented the quantum
technical assistance with Rigetti’s devices. This work was eigenvalue estimation algorithm (QEEA) of Somma [7]
performed as part of Astex’s Sustaining Innovation Post- and tested it with estimates of {gk } from Rigetti’s QPUs.
doctoral Program. This work was supported by Innovate Although the underlying approach to estimate {λi } is quite
UK via the Quantum Commercialisation program of the different, it will be seen that the final objective function
Industrial Strategy Challenge Fund (ISCF) [Project No. takes an identical form, primarily differing in the Fourier
10001505]. coefficients of the target function. We also extend the
QEEA method by implementing importance sampling and
APPENDIX A: DETAILS OF CHEMICAL briefly demonstrate its performance here.
SYSTEMS
Here, we provide details of the two systems studied 1. Theory
in this paper that have been motivated by pharmaceutical In the QEEA, a range [−α, α] is considered in which
applications. the Hamiltonian is known to be bounded and where α ≤ π .
Methanethiol is a model system for the amino acid α = 1/2 is chosen in the original presentation. The range
cysteine, a naturally occurring amino acid in proteins. [−α, α] is divided into M bins, each with width 2 for a
Hydrogen abstraction from the thiol group of cysteine small > 0. These bins are constructed to overlap with
is an essential step in many enzymatic and drug-binding each other, such that the center of the j th bin is given
processes [75]. Methanethiol is the smallest neutral self- by −α + j . The use of bins that overlap helps to ensure
contained system that features a carbon-connected thiol appropriate normalization of results; we will return to this
group. The orbitals in the (2e, 2o) active space selected point shortly.
are located on the SH bond, allowing a minimal model of Consider a state |ψ, which will be chosen to have a
hydrogen abstraction. large overlap with a state (or set of states) the energy
As a larger example, we have studied a cut-out of ibruti- of which we seek to estimate. For example, for a chem-
nib in its binding pocket, a drug approved for treatment of ical system where we wish to estimate the ground-state
non-Hodgkin lymphoma [76] that binds covalently to the energy, |ψ might be taken as the Hartree-Fock state. For
cysteine of Bruton’s tyrosine kinase (BTK). Correspond- each bin, the QEEA aims to assess whether corresponding
ing to the hydrogen abstraction in the model system, we eigenstates (the eigenvalues of which lie within the bin)
have considered a transition-state structure for the hydro- have a significant overlap with |ψ. By making suffi-
gen transfer from thiol to an adjacent water molecule, ciently small, we can then obtain accurate estimates of the
which we refer to as “clusterTS.” There are the two elec- desired λi .
trons of the thiol bond and the forming bond between More precisely, for each bin, we define a function fj that
the transferred hydrogen and water, which need to be acts as a window function for that bin. This means that
included in a (4e, 4o) active space. This pharmaceutically fj (λi ) will be nonzero only if λi is within the j th bin. For
relevant system has been used in a previous study of quan- an eigenstate |i of τ H with eigenvalue τ λi ,
tum algorithm resource estimation by some of the current
authors [5] and we refer to a more detailed description of
this system in that work. The geometry optimization of fj (τ H )|i = fj (τ λi )|i . (B1)
040341-16
The goal of the QEEA is then to construct the vector As stated in Ref. [7, Appendix A], the Fourier coeffi-
cients can be derived as
pj = ψ|fj (τ H )|ψ. (B2)
sin(k /2)
Fj (k) = H (k /2)e−iλj k , (B8)
k
This is referred to as the probability vector. Only bins con-
taining eigenstates supported by |ψ will have nonzero ≡ Fk e−iλj k , (B9)
values pj and the magnitudes pj will be related to the cor-
responding overlaps, i |ψ. Thus, by assessing the bins where λj is the center of the j th bin and H (k) is the Fourier
with a large pj , one can estimate the desired eigenvalues transform of h(x), which we calculate numerically through
of τ H . a fast Fourier transform after discretization. The coeffi-
We can estimate pj from a set of gk estimates by cients Fj (k) only depend on the bin index j through a phase
expanding each fj as a Fourier series, factor, which allows us to write the second line and thus
define Fk independent of j . We can therefore rewrite Eq.
∞ (B4) as
1
fj (τ H ) = √ Fj (k)eiτ Hk . (B3)
2π k=−∞ 1
p̃j = √ Fk e−iλj k gk∗ . (B10)
2π |k|<N
Inserting this expansion into Eq. (B2) and truncating such
that |k| < N , for some N ∈ N, gives Note that this expression has an almost identical form to
that in Eq. (14) in the CDF-QPE method. The main differ-
1 ence is in the Fourier coefficients used. These are plotted
p̃j = √ Fj (k)gk∗ . (B4) for = 3 × 10−3 in Fig. 3 and compared to those in the
2π |k|<N method of Wan et al., aiming for a similar final accuracy by
using Eq. (22) to select β. The Fourier coefficients decay
Here, p̃j denotes the approximate probability vector, with less rapidly in the QEEA. However, as pointed out by
an error introduced due to truncation. Somma [7], there are advantages to this binning approach;
It only remains to choose the precise form of fj (x). in particular, if the gaps between eigenvalues are very
There are many choices that could be made but because the small (e.g., in a solid-state system with bands of energy
Fourier series is truncated, it is important that the Fourier values), then solving the QEEA is much less ambitious
coefficients Fj (k) decay rapidly. Somma chooses than distinguishing individual eigenvalues to very high
precision and so there may be advantages in such cases.
∞ Lastly, noting that g−k = gk∗ and F−k = Fk allows us to
fj (x) = h (y − x)1j (y) dy, (B5) write
−∞
N −1
2
where 1j is the indicator function for the j th bin, and h is p̃j = + Fk Re[gk ] cos(kλj )
a rescaled bump function. Specifically, 2π π k=1

− Im[gk ] sin(kλj ) . (B11)
−1/(1−x2 )
ae , if |x| < 1
h(x) = (B6)
0, if |x| ≥ 1 2. Importance sampling
We have additionally implemented importance sampling
1 in the QEEA, which was not considered in the original
with a such that −1 h(x)dx = 1 and then
presentation of the method. As pointed out in Ref. [8],
this can be used to reduce the complexity of the QEEA,
2 bringing it closer to that of other statistical phase estima-
h (x) = h(2x/ ). (B7)
tion approaches, some of which have proven Heisenberg-
limited scaling [8,11]. We do not perform a study of scaling
The key properties for this choice are that the Fourier here. However, importance sampling allows the use of a
coefficients Fj (k) decay superpolynomially and that much smaller bin width (and therefore higher precision)
M
j =1 fj (x) = 1 for x ∈ [−α, α]. This second point is for a given number of circuits to perform.
important to ensure that contributions are appropriately The probability vectors in the QEEA can be constructed
normalized; this is why bins are chosen to overlap, pre- using importance sampling in an identical manner to that
venting potential issues if an eigenvalue lies near the edge described in Sec. II A 2, starting from Eq. (B11). The main
of a bin. difference is that the sign of Fk can vary. Therefore, these
040341-17
0.5
(a) Exact energy
without QPU errors
0.4 Extrapolated
Error rate = 1
0.3 Error rate = 3
Error rate = 5
p̃j 0.2
0.1
0.0
−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

λj
0.25
0.5
(b) (c)
0.20
0.4
0.3 0.15
0.10
p̃j
0.2
0.1 0.05
0.0 0.00
−1.075 −1.070 −1.065 −1.060 −1.055 0.935 0.940 0.945 0.950 0.955
λj λj
FIG. 10. Results from Aspen-11, performing the QEEA on methanethiol (2e, 2o) with a stretched SH bond. (a) The full probability
vector from −1.5 to +1.5. (b),(c) Plots enlarged in the region of the ground-state and first-excited state, respectively. The results
presented are for error rates λ = 1, 3, and 5 and the ZNE-extrapolated (λ = 0) result. The half bin width is set to = 3 × 10−3 and
we set N = 4001. For importance sampling, NS = 2000 samples are taken. The correct energy is estimated for every error rate, to the
precision considered.
signs must also be absorbed into the importance-sampled set to = 3 × 10−3 and the corresponding Fourier coeffi-
summation but the approach is otherwise unchanged. As cients have been importance sampled with NS = 2000. A
described in the main text, we have also performed RC in separate Pauli twirl has been performed for each sample.
the following results, incorporating it into the importance- The Fourier summation has been truncated at N = 4001.
sampling procedure by performing one twirl for each The results are presented in Fig. 10 and can be com-
sample. pared to equivalent results from the CDF-QPE method in
Fig. 5. As for the CDF-QPE method, we find the QEEA
to be robust and that each eigenvalue can be clearly and
3. Results correctly identified from the probability vector, within the
We have performed the QEEA on Rigetti’s Aspen-11 accuracy determined by roughly half the bin width.
for the same methanethiol system studied in Sec. IV A, As found in Sec. IV B, there is no improvement to the
which requires three qubits in total. Each controlled-e−iτ Hk energy estimates after performing ZNE. In this case, the
operation has been compiled to a circuit ansatz with three estimates are already correct within the resolution deter-
CZ layers, as in Sec. IV A. The half bin width has been mined by and so there is no improvement to be made.
FIG. 11. The circuit for a single Trotter step for H2 in a STO-3G basis, where the Hamiltonian has the form H = c1 Z + c2 X . The
circuit on the right is reduced so that the only two-qubit gates are CZ gates, using standard identities. The one-qubit gates are each
implemented in native Rigetti gates via the structure RZ (φ) RX (−π/2) RZ (θ) RX (π/2) RZ (λ), which ensures consistent gate layers.
040341-18
0.7
0.6 (a) 1.0 (c)

0.5
0.9
0.4
G̃(x)
G̃(x)
0.3 0.8
0.2
0.7
0.1
0.0 0.6
−1.25 −1.20 −1.15 −1.10 −1.05 −1.00 −0.95 0.95 1.00 1.05 1.10 1.15 1.20 1.25
x x
70 40
(b) (d) Exact energy
60 p = 5 × 10−4
30 p = 1 × 10−3
50
p = 2 × 10−3
40 p = 4 × 10−3
G̃ (x)
G̃ (x)
20
30
20 10
10
0
0
−1.110 −1.105 −1.100 −1.095 −1.090 −1.085 −1.080 −1.075 −1.070 1.070 1.075 1.080 1.085 1.090 1.095 1.100 1.105 1.110
x x
FIG. 12. CDFs and their derivatives for stretched H2 STO-3G, performed with Trotterization using the Trotter step in Fig. 11. The
simulations are performed on a pyQuil quantum virtual machine (QVM). Depolarizing noise is applied to CZ gates with four different
error rates, p. The subplots are enlarged in the region of (a),(b) the ground state and (c),(d) the excited state, respectively. Increasing
the error rate broadens the “jump” region of the CDF and the corresponding peak of the derivative. The peak of the CDF remains
roughly correct regardless, although statistical noise at higher p can lead to errors in the final energy estimate. Note that different x-axis
scales are used between subplots.
ZNE boosts the signal from the probability vector but over- allows us to investigate higher precision and the effect of
shoots considerably for the ground-state bin. One reason varying error rates. The same H2 example is considered
for this is that because the coefficients |Fk | decay much with an identical Trotter step. However, CDF-QPE param-
more slowly in the QEEA, very few samples are performed eters of β = 5 × 104 , d = 511, and NS = 1000 are taken.
for any particular k, even at small k. This makes mitigation This choice of β corresponds to δ ≈ 0.004 in Eq. (22),
of coherent errors less successful and also increases the after choosing = 0.1. As for the results in the main text,
statistical error bars on each gk estimate, thus lowering the we perform 100 shots for each ki value obtained during
quality of each extrapolation and therefore also the ZNE importance sampling.
estimate of the probability vector. First, we consider applying depolarizing noise to each
CZ gate, before next considering the effect of coherent
APPENDIX C: SIMULATED TROTTERIZATION errors. The depolarizing channel is defined as
RESULTS WITH COHERENT AND INCOHERENT
ERRORS p
(ρ) = (1 − p)ρ + 1, (C3)
2n
Figure 11 presents the circuit used for each Trotter step
in Sec. IV C. Since the Hamiltonian has the form H =
where n is the number of qubits, equal to 2 when applied to
c1 Z + c2 X , a single first-order Trotter step is taken as
a CZ gate, and p is the depolarizing error parameter. In the
following results, we vary p from 5 × 10−4 to 4 × 10−3 .
UTrotter = e−ic2 τ X e−ic1 τ Z (C1) All other gates and measurements are applied without
= RX (2 c2 τ ) RZ (2 c1 τ ), (C2) error. Figure 12 presents the CDF, G̃(x), and its derivative,
G̃ (x). The subplots are enlarged in the region of the ground
which leads to the circuit on the left. The circuit on the state (GS) and excited state (ES). As might be expected,
right is then expressed with CZ as the only two-qubit gate, the jumps in the CDF become broader as the error rate is
which can be obtained through standard circuit identities. increased. Despite this, the energy estimate obtained by
In addition to the results in the main text, we have maximizing the CDF derivative remains accurate in each
performed simulated results using the pyQuil QVM. This case.
040341-19
0.8
(a) 1.0 (c)
0.6
0.9
0.4
G̃(x)
G̃(x)
0.8
0.2
0.7
0.0
0.6
−1.25 −1.20 −1.15 −1.10 −1.05 −1.00 −0.95 0.95 1.00 1.05 1.10 1.15 1.20 1.25
x x
50
120 Exact energy
(b) (d)
40 Without RC
100
With RC
80 30
G̃ (x)
G̃ (x)
60 20
40 10
20
0
0
−10
−20
−1.200 −1.175 −1.150 −1.125 −1.100 −1.075 −1.050 −1.025 −1.000 1.000 1.025 1.050 1.075 1.100 1.125 1.150 1.175 1.200
x x
FIG. 13. CDFs and their derivatives for stretched H2 STO-3G, performed with Trotterization using the Trotter step in Fig. 11. The
simulations are performed on a pyQuil QVM. A unitary error of e−i(θ/2)Z⊗Z with θ = 0.1 is applied after every CZ gate. We then
consider how this affects the CDF, both before and after applying randomized compiling. (a),(b) Without RC, there is a large error in
the ground-state energy. This error is effectively removed by RC, at the cost of reduced signal. Interestingly, the error in the excited-
state energy is much smaller but still slightly improved by RC. (c),(d) The excited state is harder to distinguish in the CDF derivative
due to statistical noise. Note that different x-axis scales are used between subplots.
It is straightforward to see that depolarizing noise does We choose θ = 0.1, which is a very large error in prac-
not prevent us from obtaining accurate energy estimates. tice. All other gates and measurements are applied without
Under depolarizing noise, the expectation values will error.
decay as e−γ |k| , for some decay rate γ . We then expect The results are presented in Fig. 13. All simulation
parameters are the same as for the results in Fig. 12,
gk ∼ e−γ |k| pi e−iτ λi k . (C4) including β, d, and NS . Application of the CZ unitary
i error leads to a large error in the ground-state energy
This decay factor does not affect the frequencies present of −7.6 mHa (after rescaling by τ −1 ). This is reduced to
in gk but it should be expected that it becomes more chal- +0.3 mHa after applying RC. Interestingly, the error in
lenging to reliably estimate each λi with increasing γ . To the excited-state energy is less than 1 mHa in both cases,
be precise, the Fourier transform of e−γ |k| is a Lorentzian although there is a slight improvement with RC applied.
function centered about 0, and the width of this function These results demonstrate that coherent errors can cause
grows as γ increases. The results in Fig. 12 are as expected, significant errors in energy estimates from statistical phase
given these arguments. estimation but that RC is a promising approach to help
A separate type of error consists of coherent (or unitary) mitigate these errors in practice.
errors, which preserve the purity of the input state. Refer-
ence [73] considers a type of control-free phase estimation
and demonstrates that unitary errors cause errors in the
[1] A. Y. Kitaev, Quantum measurements and the Abelian
final phase estimates, which can be largely removed with stabilizer problem, ArXiv:quant-ph/9511026.
RC. Here, we demonstrate a similar result with the method- [2] R. Cleve, A. Ekert, C. Macchiavello, and M. Mosca, Quan-
ology developed in this paper, with the CDF-based method tum algorithms revisited, Proc. R. Soc. A 454, 339 (1998).
of Ref. [9] and integrating RC with importance sampling. [3] M. Reiher, N. Wiebe, K. M. Svore, D. Wecker, and M.
Following Ref. [73], we apply each CZ gate with a unitary Troyer, Elucidating reaction mechanisms on quantum com-
error, so that UCZ = UCZ , with puters, Proc. Natl. Acad. Sci. 114, 7555 (2017).
[4] J. Lee, D. W. Berry, C. Gidney, W. J. Huggins, J.
= e−i(θ/2)Z⊗Z . (C5) R. McClean, N. Wiebe, and R. Babbush, Even more
040341-20
efficient quantum computations of chemistry through tensor [22] S. Zhang, Y. Lu, K. Zhang, W. Chen, Y. Li, J.-N. Zhang, and
hypercontraction, PRX Quantum 2, 030305 (2021). K. Kim, Error-mitigated quantum gates exceeding physical
[5] N. S. Blunt, J. Camps, O. Crawford, R. Izsák, S. fidelities in a trapped-ion system, Nat. Commun. 11, 587
Leontica, A. Mirani, A. E. Moylett, S. A. Scivier, C. (2020).
Sünderhauf, P. Schopf, J. M. Taylor, and N. Holzmann, [23] P. Czarnik, A. Arrasmith, P. J. Coles, and L. Cincio, Error
Perspective on the current state-of-the-art of quantum com- mitigation with Clifford quantum-circuit data, Quantum 5,
puting for drug discovery applications, J. Chem. Theory 592 (2021).
Comput. 18, 7001 (2022). [24] P. Czarnik, M. McKerns, A. T. Sornborger, and L. Cincio,
[6] A. V. Ivanov, C. Sünderhauf, N. Holzmann, T. Ellaby, R. N. Improving the efficiency of learning-based error mitigation,
Kerber, G. Jones, and J. Camps, Quantum computation for ArXiv:2204.07109 (2022).
periodic solids in second quantization, ArXiv:2210.02403 [25] J. J. Wallman and J. Emerson, Noise tailoring for scalable
(2022). quantum computation via randomized compiling, Phys.
[7] R. D. Somma, Quantum eigenvalue estimation via time Rev. A 94, 052325 (2016).
series analysis, New J. Phys. 21, 123025 (2019). [26] A. Hashim, R. K. Naik, A. Morvan, J.-L. Ville, B. Mitchell,
[8] L. Lin and Y. Tong, Heisenberg-limited ground-state J. M. Kreikebaum, M. Davis, E. Smith, C. Iancu, K. P.
energy estimation for early fault-tolerant quantum comput- O’Brien, I. Hincks, J. J. Wallman, J. Emerson, and I. Sid-
ers, PRX Quantum 3, 010318 (2022). diqi, Randomized compiling for scalable quantum comput-
[9] K. Wan, M. Berta, and E. T. Campbell, Randomized quan- ing on a noisy superconducting quantum processor, Phys.
tum algorithm for statistical phase estimation, Phys. Rev. Rev. X 11, 041039 (2021).
Lett. 129, 030503 (2022). [27] X. Bonet-Monroig, R. Sagastizabal, M. Singh, and T. E.
[10] T. E. O’Brien, B. Tarasinski, and B. M. Terhal, Quantum O’Brien, Low-cost error mitigation by symmetry verifica-
phase estimation of multiple eigenvalues for small-scale tion, Phys. Rev. A 98, 062339 (2018).
(noisy) experiments, New J. Phys. 21, 023022 (2019). [28] S. McArdle, X. Yuan, and S. Benjamin, Error-mitigated
[11] A. Dutkiewicz, B. M. Terhal, and T. E. O’Brien, digital quantum simulation, Phys. Rev. Lett. 122, 180501
Heisenberg-limited quantum phase estimation of multi- (2019).
ple eigenvalues with few control qubits, Quantum 6, 830 [29] T. E. O’Brien, S. Polla, N. C. Rubin, W. J. Huggins, S.
(2022). McArdle, S. Boixo, J. R. McClean, and R. Babbush, Error
[12] L. Clinton, J. Bausch, J. Klassen, and T. Cubitt, Phase esti- mitigation via verified phase estimation, PRX Quantum 2,
mation of local hamiltonians on NISQ hardware, New J. 020317 (2021).
Phys. 25, 033027 (2023). [30] Z. Cai, R. Babbush, S. C. Benjamin, S. Endo, W. J. Hug-
[13] G. Wang, D. Stilck-França, R. Zhang, S. Zhu, and P. D. gins, Y. Li, J. R. McClean, and T. E. O’Brien, Quantum
Johnson, Quantum algorithm for ground state energy esti- error mitigation, ArXiv:2210.00921 (2022).
mation using circuit depth with exponentially improved [31] V. Russo, A. Mari, N. Shammah, R. LaRose, and W. J.
dependence on precision, ArXiv:2209.06811 (2022). Zeng, Testing platform-independent quantum error miti-
[14] Z. Ding and L. Lin, Even shorter quantum circuit for gation on noisy quantum computers, ArXiv:2210.07194
phase estimation on early fault-tolerant quantum comput- (2022).
ers with applications to ground-state energy estimation, [32] C. Piveteau, D. Sutter, S. Bravyi, J. M. Gambetta, and K.
ArXiv:2211.11973 (2022). Temme, Error mitigation for universal gates on encoded
[15] R. Babbush, C. Gidney, D. W. Berry, N. Wiebe, J. McClean, qubits, Phys. Rev. Lett. 127, 200505 (2021).
A. Paler, A. Fowler, and H. Neven, Encoding electronic [33] Y. Suzuki, S. Endo, K. Fujii, and Y. Tokunaga, Quantum
spectra in quantum circuits with linear T complexity, Phys. error mitigation as a universal error reduction technique:
Rev. X 8, 041015 (2018). Applications from the NISQ to the fault-tolerant quantum
[16] S. Bravyi, S. Sheldon, A. Kandala, D. C. Mckay, and J. computing eras, PRX Quantum 3, 010345 (2022).
M. Gambetta, Mitigating measurement errors in multiqubit [34] M. Lostaglio and A. Ciani, Error mitigation and quantum-
experiments, Phys. Rev. A 103, 042605 (2021). assisted simulation in the error corrected regime, Phys. Rev.
[17] A. W. R. Smith, K. E. Khosla, C. N. Self, and M. S. Kim, Lett. 127, 200506 (2021).
Qubit readout error mitigation with bit-flip averaging, Sci. [35] Y. Xiong, D. Chandra, S. X. Ng, and L. Hanzo, Sampling
Adv. 7, eabi8009 (2021). overhead analysis of quantum error mitigation: Uncoded vs.
[18] K. Temme, S. Bravyi, and J. M. Gambetta, Error mitiga- coded systems, IEEE Access 8, 228967 (2020).
tion for short-depth quantum circuits, Phys. Rev. Lett. 119, [36] N. S. Blunt, G. P. Gehér, and A. E. Moylett, Compilation of
180509 (2017). a simple chemistry application to quantum error correction
[19] Y. Li and S. C. Benjamin, Efficient variational quantum primitives, ArXiv:2307.03233 (2023).
simulator incorporating active error minimization, Phys. [37] R. Takagi, S. Endo, S. Minagawa, and M. Gu, Fundamental
Rev. X 7, 021050 (2017). limits of quantum error mitigation, npj Quantum Inf. 8, 114
[20] T. Giurgica-Tiron, Y. Hindy, R. LaRose, A. Mari, and W. J. (2022).
Zeng, in 2020 IEEE International Conference on Quantum [38] P. J. J. O’Malley, et al., Scalable quantum simula-
Computing and Engineering (QCE) (2020), p. 306. tion of molecular energies, Phys. Rev. X 6, 031007
[21] E. van den Berg, Z. K. Minev, A. Kandala, and K. Temme, (2016).
Probabilistic error cancellation with sparse Pauli-Lindblad [39] T. M. Graham, et al., Multi-qubit entanglement and algo-
models on noisy quantum processors, ArXiv:2201.09866 rithms on a neutral-atom quantum computer, Nature 604,
(2022). 457 (2022).
040341-21
[40] K. Yamamoto, S. Duffield, Y. Kikuchi, and D. M. Ramo, entanglement and faithful teleportation via noisy channels,
Demonstrating Bayesian quantum phase estimation with Phys. Rev. Lett. 76, 722 (1996).
quantum error detection, ArXiv:2306.16608 (2023). [57] M. R. Geller and Z. Zhou, Efficient error models for fault-
[41] A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. tolerant architectures and the Pauli twirling approximation,
Brink, J. M. Chow, and J. M. Gambetta, Hardware-efficient Phys. Rev. A 88, 012314 (2013).
variational quantum eigensolver for small molecules and [58] Y. Kim, C. J. Wood, T. J. Yoder, S. T. Merkel, J. M.
quantum magnets, Nature 549, 242 (2017). Gambetta, K. Temme, and A. Kandala, Scalable error mit-
[42] Y. Nam, et al., Ground-state energy estimation of the water igation for noisy quantum circuits produces competitive
molecule on a trapped-ion quantum computer, npj Quantum expectation values, ArXiv:2108.09197 (2021).
Inf. 6, 33 (2020). [59] T. Kurita, H. Qassim, M. Ishii, H. Oshima, S. Sato,
[43] L. Zhao, J. Goings, K. Shin, W. Kyoung, J. I. Fuks, J.-K. and J. Emerson, Synergetic quantum error mitigation by
Kevin Rhee, Y. M. Rhee, K. Wright, J. Nguyen, J. Kim, and randomized compiling and zero-noise extrapolation for
S. Johri, Orbital-optimized pair-correlated electron simula- the variational quantum eigensolver, ArXiv:2212.11198
tions on trapped-ion quantum computers, npj Quantum Inf. (2022).
9, 60 (2023). [60] S. Ferracin, A. Hashim, J.-L. Ville, R. Naik, A. Carignan-
[44] S. Lee, J. Lee, H. Zhai, Y. Tong, A. M. Dalzell, A. Kumar, P. Dugas, H. Qassim, A. Morvan, D. I. Santiago, I. Sid-
Helms, J. Gray, Z.-H. Cui, W. Liu, et al., Evaluating the evi- diqi, and J. J. Wallman, Efficiently improving the perfor-
dence for exponential quantum advantage in ground-state mance of noisy quantum computers, ArXiv:2201.10672
quantum chemistry, Nat. Commun. 14, 1952 (2023). (2022).
[45] R. Izsák, C. Riplinger, N. S. Blunt, B. de Souza, N. Holz- [61] B. Yang, R. Raymond, and S. Uno, Efficient quantum
mann, O. Crawford, J. Camps, F. Neese, and P. Schopf, readout-error mitigation for sparse measurement outcomes
Quantum computing in pharma: A multilayer embedding of near-term quantum devices, Phys. Rev. A 106, 012423
approach for near future applications, J. Comput. Chem. 44, (2022).
406 (2023). [62] P. J. Karalekas, N. A. Tezak, E. C. Peterson, C. A. Ryan,
[46] S.-N. Sun, M. Motta, R. N. Tazhigulov, A. T. K. Tan, G. K.- M. P. da Silva, and R. S. Smith, A quantum-classical
L. Chan, and A. J. Minnich, Quantum computation of finite- cloud platform optimized for variational hybrid algorithms,
temperature static and dynamical properties of spin systems Quantum Sci. Technol. 5, 024003 (2020).
using quantum imaginary time evolution, PRX Quantum 2, [63] R. S. Smith, M. J. Curtis, and W. J. Zeng, A practi-
010317 (2021). cal quantum instruction set architecture, ArXiv:1608.03355
[47] R. N. Tazhigulov, S.-N. Sun, R. Haghshenas, H. Zhai, (2016).
A. T. K. Tan, N. C. Rubin, R. Babbush, A. J. Minnich, [64] F. Neese, Software update: The ORCA program system, ver-
and G. K.-L. Chan, Simulating models of challenging sion 5.0, Wiley Interdiscip. Rev. Comput. Mol. Sci. 12,
correlated molecules and materials on the Sycamore quan- e1606 (2022).
tum processor, PRX Quantum 3, 040318 (2022). [65] Q. Sun, T. C. Berkelbach, N. S. Blunt, G. H. Booth, S. Guo,
[48] M. Benedetti, M. Fiorentini, and M. Lubasch, Hardware- Z. Li, J. Liu, J. D. McClain, E. R. Sayfutyarova, S. Sharma,
efficient variational quantum algorithms for time evolution, S. Wouters, and G. K.-L. Chan, PySCF: The PYTHON-based
Phys. Rev. Res. 3, 033083 (2021). simulations of chemistry framework, WIREs Comput. Mol.
[49] C. M. Keever and M. Lubasch, Classically optimized Sci. 8, e1340 (2018).
Hamiltonian simulation, ArXiv:2205.11427 (2022). [66] Q. Sun, et al., Recent developments in the PySCF program
[50] Y. Kikuchi, C. M. Keever, L. Coopmans, M. Lubasch, and package, J. Chem. Phys. 153, 024109 (2020).
M. Benedetti, Realization of quantum signal processing on [67] M. J. Frisch, G. W. Trucks, H. B. Schlegel, G. E. Scuse-
a noisy quantum computer, ArXiv:2303.05533 (2023). ria, M. A. Robb, J. R. Cheeseman, G. Scalmani, V.
[51] J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Barone, G. A. Petersson, H. Nakatsuji, et al., Gaussian
Leary, D. Maclaurin, G. Necula, A. Paszke, J. Vander- 16 Revision C.01, Gaussian Inc. Wallingford CT, 2016,
Plas, S. Wanderman-Milne, and Q. Zhang, JAX: Com- https://gaussian.com/citation.
posable transformations of PYTHON+NumPy programs, [68] J. R. McClean, et al., OpenFermion: The electronic struc-
http://github.com/google/jax (2018). ture package for quantum computers, Quantum Sci. Tech-
[52] A. Kandala, K. Temme, A. D. Córcoles, A. Mezzacapo, J. nol. 5, 034014 (2020).
M. Chow, and J. M. Gambetta, Error mitigation extends the [69] S. B. Bravyi and A. Y. Kitaev, Fermionic quantum compu-
computational reach of a noisy quantum processor, Nature tation, Ann. Phys. (NY) 298, 210 (2002).
567, 491 (2019). [70] D. C. McKay, C. J. Wood, S. Sheldon, J. M. Chow, and
[53] A. He, B. Nachman, W. A. de Jong, and C. W. Bauer, Zero- J. M. Gambetta, Efficient Z gates for quantum computing,
noise extrapolation for quantum-gate error mitigation with Phys. Rev. A 96, 022330 (2017).
identity insertions, Phys. Rev. A 102, 012426 (2020). [71] S. Bravyi, J. M. Gambetta, A. Mezzacapo, and K. Temme,
[54] E. Knill, Fault-tolerant postselected quantum computation: Tapering off qubits to simulate fermionic Hamiltonians,
Threshold analysis, ArXiv:quant-ph/0404104 (2004). ArXiv:1701.08213 (2017).
[55] E. Knill, Quantum computing with realistically noisy [72] http://doi.org/10.5281/zenodo.10201197.
devices, Nat. Commun. 434, 39 (2005). [73] Y. Gu, Y. Ma, N. Forcellini, and D. E. Liu, Noise-
[56] C. H. Bennett, G. Brassard, S. Popescu, B. Schumacher, resilient phase estimation with randomized compiling,
J. A. Smolin, and W. K. Wootters, Purification of noisy ArXiv:2208.04100 (2022).
040341-22
[74] R. Kshirsagar, A. Katabarwa, and P. D. Johnson, On [77] C. Møller and M. S. Plesset, Note on an approximation
proving the robustness of algorithms for early fault- treatment for many-electron systems, Phys. Rev. 46, 618
tolerant quantum computers, ArXiv:2209.11322 (2022). (1934).
[75] L. E. S. Netto, M. A. de Oliveira, G. Monteiro, A. P. D. [78] T. H. Dunning Jr, Gaussian basis sets for use in correlated
Demasi, J. R. R. Cussiol, K. F. Discola, M. Demasi, G. M. molecular calculations. I. The atoms boron through neon
Silva, S. V. Alves, V. G. Faria, and B. B. Horta, Reactive and hydrogen, J. Chem. Phys. 90, 1007 (1989).
cysteine in proteins: Protein folding, antioxidant defense, [79] R. A. Kendall, T. H. Dunning Jr, and R. J. Harrison, Elec-
redox signaling and more, Comp. Biochem. Physiol. Part tron affinities of the first-row atoms revisited. Systematic
C: Toxicol. Pharmacol. 146, 180 (2007), Fourth Special basis sets and wave functions, J. Chem. Phys. 96, 6796
Issue of CBP dedicated to “The Face of Latin American (1992).
Comparative Biochemistry and Physiology”. [80] D. E. Woon and T. H. Dunning Jr, Gaussian basis sets
[76] U.S. Food and Drug Administration, FDA expands for use in correlated molecular calculations. III. The atoms
approved use of Imbruvica for rare form of non-Hodgkin aluminum through argon, J. Chem. Phys. 98, 1358 (1993).
lymphoma, https://wayback.archive-it.org/7993/20170112 [81] N. Holzmann and P. Schopf, New insight into the cova-
222810/http://www.fda.gov/NewsEvents/Newsroom/Press lent binding mechanism of ibrutinib to Bruton’s tyrosine
Announcements/ucm432123.htm. kinase—a QM cluster study (to be published).
040341-23

PRXQuantum.4.040341

Uploaded by

Copyright:

Available Formats

PRXQuantum.4.040341

Uploaded by

Document Information

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

PRXQuantum.4.040341

Uploaded by

Copyright:

Available Formats

PRX QUANTUM 4, 040341 (2023)

Statistical Phase Estimation and Error Mitigation on a Superconducting

I. INTRODUCTION the limit of current classical algorithms, often estimating

2691-3399/23/4(4)/040341(23) 040341-1 Published by the American Physical Society

In this paper, we present the ﬁrst combination of sta-

E[Y] = Im[ ψ| e−iτ Hk |ψ ]. (3)

up to statistical errors, which are controlled by averaging π/2

τ H  ≤ π/2, where . . .  denotes the operator norm. The 2. Importance sampling

1.0 (a) 350 (b)

where 3. Estimating energies in the CDF-QPE method

H4 in a STO-3G basis, with a square geometry of side

H̃ (x) (blue) with NS = 5000 (thus shot noise is not present

to an energy eigenvalue of τ H . While importance sam- 10−11 QEEA

L(p) = U| − Ũ(p)|, (27)

where the L2 norm is used. Here, U is the target uni-

extrapolated estimate is less 1.2 in magnitude. If this is III. METHODS

1.0 (b) 1.0 (d)

−0.4 (a) 1.2 (b)

1.0 (a) 1.0 (c)

10 20 30 40 50 60 70 −1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

0.6 (a) 1.0 (c)

You might also like

τ H ≤ π/2, where . . . denotes the operator norm. The 2. Importance sampling

L(p) = U| − Ũ(p)|, (27)