Clifford Circuit Initialisation for Variational Quantum Algorithms

M. H. Cheng,1, 2, ∗ K. E. Khosla,1, † C. N. Self,3, 1 M. Lin,1 B. X. Li,1 A. C. Medina,2 and M. S. Kim1

1
QOLS, Blackett Laboratory, Imperial College London, London, SW7 2AZ, UK
2
Fraunhofer Institute for Industrial Mathematics,
Fraunhofer-Platz 1, 67663 Kaiserslautern, Germany
3
Quantinuum, Partnership House, Carlisle Place, London, SW1P 1BX, United Kingdom
(Dated: July 5, 2022)
We present an initialisation method for variational quantum algorithms applicable to interme-
diate scale quantum computers. The method uses simulated annealing of the efficiently simulable
Clifford parameter points as a pre-optimisation to find a low energy initial condition. We numeri-
cally demonstrate the effectiveness of the technique, and how it depends on Hamiltonian structure,
number of qubits and circuit depth. While a range of different problems are considered, we note
that the method is particularly useful for quantum chemistry problems. This presented method
could help achieve a quantum advantage in noisy or fault-tolerant intermediate scale devices, even
arXiv:2207.01539v1 [quant-ph] 4 Jul 2022

though we prove in general that the method is not arbitrarily scalable.

I. INTRODUCTION However, scalability remains a concern as the numbers

of qubits and parameters increase [17], notwithstanding
Intermediate Scale quantum devices are rapidly im- the issue with noise [7]. The main obstacles to scala-
proving and are potential candidates for demonstrating bility are the onset of barren plateaus [18, 19] and the
quantum advantage. While algorithms such as the phase trade-off between expressibility and trainability [20] of a
estimation [1], factorisation [2], or Deutsch-Jozsa [3] are parameterized quantum circuit.
too resource intensive to demonstrate quantum advan- To tackle these problems, we propose a classical ini-
tage on current devices, intermediate scale devices have tialization strategy for the VQE: the Clifford Circuit Pre-
inspired an important class of hybrid classical-quantum Optimisation, which classically searches for a good initial
algorithms based on classical optimisation – Variational condition before running the VQE on device. Clifford cir-
Quantum Algorithms (VQA)[4–6]. Device noise is a cuits are quantum operations that map Pauli strings to
significant barrier to achieving supremacy with VQAs Pauli strings, and when applied to an initial stabilizer
on noisy Intermediate Scale Quantum (noisy ISQ) de- state (i.e. an eigenstate of a Pauli string), the resulting
vices [7]; however, due to their ability to efficiently es- measurement outcomes are known to be efficiently sim-
timate classically intractable cost functions, VQAs re- ulable [21]. This pre-optimisation can help avoid barren
main a possible candidate for demonstrating quantum plateaus and remove the influence of device noise dur-
advantage on small, or intermediate scale error corrected ing the classical optimisation. Given an ansatz circuit
devices. This class of algorithms divides an eigenvalue consisting of parameterized Rz , Rx , and Ry rotations
problem into a classical optimisation task, using a quan- as well as CNOTs, there exists a finite set of parame-
tum computer to estimate the cost function. The cal- ters (Clifford points) such that the final quantum state
culation of the cost function requires only a relatively is a stabilizer state and Pauli measurement outcomes are
short quantum circuit, replacing the requirement for deep classically simulable. Since stabilizer states are uniformly
circuits and ancillary qubits with a run-time overhead. distributed in the Hilbert space, one can envisage a search
The Variational Quantum Eigensolver (VQE), is a spe- algorithm over these discrete points to seek a good ini-
cific variational algorithm that aims to minimize the ex- tialization. In this work, we consider classical annealing
pectation value of a given Hamiltonian. The VQE is with threshold resetting as the search method, however,
sufficiently flexible to solve many optimization problems, in writing the current manuscript, we became aware of
including MAXCUT [8, 9], portfolio optimisation [10], a complimentary work using Bayesian optimisation with
financial transaction settlement [11], and finding the encouraging results [22].
ground state of solid-state Hamiltonians [12] or interact- A good initialization needs not just to minimize the
ing fermions [13]. cost function; ideally we would like an initialization with
The VQE has had some initial success in demon- both low cost, and high gradient. This gradient can be
strating simulations of quantum chemistry energy levels efficiently evaluated using the parameter-shift rule [23],
within the chemical accuracy [14], exotic phases of matter which itself also only requires Clifford circuits for a suit-
such as time crystals [15], and high energy physics [16]. able choice of parameters.
In this paper, we initially study the advantages of
the Clifford pre-optimization by considering the initial
condition with low cost before running the continu-
∗ [email protected] ous parameter variational optimisation. We numerically
† [email protected] show how Clifford pre-optimisation works with differ-
These two authors contributed equally to this work ent Hamiltonian models, and that starting at a low-
 2

Hilbert space tor H (which is decomposed into a weighted sum of

(a)
Circuit accessible Pauli strings). Selecting the structure and initial param-
Stabilizer eters for this ansatz circuit is an active area of VQE re-
states search [19, 24–26]. For typical ansatz circuits, there is
a finite set of parameter points that implement a Clif-
VQA ford circuit, with the resulting states being the so-called
target stabilizer states, Fig. 1 (a). In the following we make
state use of the Gottesman-Knill theorem [27] to efficiently
classically sample the cost function (over these Clifford
circuit-parameterized stabilizer states) before any quan-
tum hardware is needed. A hardware-efficient Clifford
circuit can be constructed by restricting the parameter
angles in single-direction rotation gates, Fig. 1 (b). Alter-
natively, one could consider a regular circuit of random
Circuit single qubit Clifford gates between entangling C-NOT (or
Z-basis states
inaccessible other two-qubit Clifford) layers, and later decide which
stabilizer states possible angles to relax for a continuous variable circuit.
For some problems the Clifford points of the relevant
(b) ansatz may not be the physically interesting part of the
variational parameter space, e.g. small perturbations us-
ing the Universal Coupled Cluster ansatz [28]. For the
present work, we restrict our analysis to the problem ag-
nostic ansätze, where the physically relevant part of the
variational parameter space is unknown. Before present-
ing numerical results, we first consider in which case this
FIG. 1. (a) Schematic representation of Clifford searching Clifford pre-optimisation search may be useful and iden-
for the VQE, with crosses depicting the stabilizer states in tify potential limitations.
the Hilbert space. A circuit typically parameterizes a subset
(darker crosses) of all stabilizer states — generally including Implementations of continuous parameter VQE are
the logical Z-basis states (blue crosses). A cost function can known to have a trade-off between expressibility and
be efficiently classically simulated at these points, and a pre- trainability [20]. The same problem is expected in the
optimisation method can be used to find a state close to the stabilizer pre-optimisation: (i) the ansatz may not be
VQE target state (green dot). (b) A hardware efficient ansatz sufficiently expressive to include the best stabilizer state
can become a Clifford circuit by restricting (θi )-parameterised (i.e. the stabilizer state that minimises the energy), or
rotations to a finite set, shown here as single direction Pauli (ii) the circuit can express too many stabilizer states and
rotations with angles being multiples of π/2. the pre-optimisation itself becomes difficult. If there are
too many parameters, the finite-size ansätze begin to ap-
proximate a 2-design [29], where barren plateaus and uni-
cost initial condition drastically reduces the gradient de- form cost function conditions apply [18]. Thus, the right
scent times. In section II we overview the Clifford pre- number of parameters and circuit design will be the key
optimisation method, comparing and contrasting it to to a successful Clifford pre-optimisation.
standard continuous parameter VQE. We then present For quantum chemistry tasks, a Fermion-to-qubit
numerical results in section III for quantum chemistry, transform, e.g. Jordan-Wigner [30], Parity [31], or
Transverse Field Ising Model (TFIM), and Binary opti- Bravyi-Kitaev [32], is required to generate a suitable
misation problems using the Quantum Approximate Op- qubit Hamiltonian. This transform has a direct impact
timisation Algorithm (QAOA). We also investigate the on “where” (in Hilbert space) the ground state is located,
relationship between ansatz types, circuit depth and nu- in contrast to parametrically accessible stabilizer states
merical scalability. Finally, in section IV, we use a count- which are fixed in the Hilbert space. Hence different
ing argument to show the method is not arbitrarily scal- Fermion-to-qubit transforms may be more or less suitable
able, but is still relevant for pushing true quantum ad- for pre-optimisation given a particular hardware-efficient
vantage for ISQ circuit sizes of several tens of qubits. ansatz. Unfortunately, it is not possible to a priori know
the best transform, but running the stabilizer search over
each transform is a linear overhead and the best trans-
II. CLIFFORD PRE-OPTIMISATION form for the quantum device can therefore be selected in
a scalable way.
The VQE algorithm requires a parameterised ansatz Finally, we note that a given Pauli string has far more
circuit U (θ) to construct the trial wave function |ψ(θ)i, 0 expectation value stabilizer states than ±1 eigenvalue
combined with a classical optimisation that minimizes stabilizers, and therefore the expectation value of a Pauli
the expectation value hψ(θ)| H |ψ(θ)i for some opera- string with a random stabilizer state is likely zero. These
 3

Trotter ansatz
Depth: 2 Depth: 5
−2
Energy
−3

−4
Jordan-Wigner

Real ansatz
Parity
−2
distance

... H H H H ... Bravyi-Kitaev

Energy

−3 d1

−4

0.6 0.8 1.0 1.2 1.4 0.6 0.8 1.0 1.2 1.4
Distance Distance

FIG. 2. Comparison of different ansatz types (rows), ansatz depths (columns), and Fermion transforms (line colours) for a
linear chain of eight hydrogen atoms (see inset). The energy curve is obtained from ten thousand iterations of Clifford pre-
optimisation, with no quantum VQE. Grey dashed lines show ground and first excited states obtained via exact diagonalization.
The real-valued ansatz of depth two, and Trotter ansatz of depth five perform equally well at approximating the ground state
energy for low bond lengths. The Trotter ansatz at depth two is insufficiently expressive, while the real-valued ansatz at depth
five is over parameterized. The plots highlight how the pre-optimisation is not a panacea to all problems, and the ansatz
structure, depth and Fermion-to-qubit transform each play an interesting role. Further on-device VQE is necessary to find
better energy solutions (e.g. up to chemical accuracy). We note that these curves are found classically and can therefore be
used to select the best ansatz for on-device optimisation.

zero expectation value states give no information to the to-qubit transformations (for quantum chemistry prob-
optimiser resulting in a barren plateau-like flattening of lems) on the pre-optimisation. In each case, simulated
the optimisation landscape. Higher numbers of non- annealing with a threshold resetting method is used to
commuting Pauli strings will help combat this, hence optimize each Clifford circuit. Our annealing method be-
the more Pauli strings, the more likely a given stabi- gins with random Clifford parameters and changes two
lizer state will give you useful information. These ar- Clifford gates each iteration, keeping the new gates if a
guments are expanded in detail in section IV, but for lower energy state is found, and probabilistically keep-
now we simply postulate that Hamiltonians with more ing the new gate if a higher energy state is found (see
(non-Commuting) Pauli strings may be better suited appendix A for details).
to Clifford pre-optimisation. Hence the Clifford pre- First, as a typical benchmark, we consider ground state
optimisation cannot be expected to be useful for all prob- problems in quantum chemistry. We use the qiskit nature
lems at every scale. Since quantum chemistry problems library to compute the molecular orbitals of a second
typically require many non-commuting Pauli strings, we quantized Fermionic Hamiltonian with frozen core elec-
might expect that the stabilizer search is more useful trons (if applicable). The same library is used to gen-
for chemistry problems compared to binary optimisation erate a qubit Hamiltonian via the parity [31], Jordan-
(e.g. QAOA), or quadratic condensed matter interac- Wigner [30], or Bravyi-Kitaev [32] transforms. These
tions. quantum chemistry Hamiltonians are expressed as a
weighted sum of Pauli strings, with many groups of non-
commuting terms.
III. NUMERICAL RESULTS Following on from quantum chemistry problems, we
then analyze Clifford pre-optimization as it applies to
the MAXCUT and Transverse Field Ising Model (TFIM)
Here we consider several example problems to test
Hamiltonians. Both of these Hamiltonians can be ex-
the pre-optimisation method. Different Hamiltonians
pressed as,
are chosen to span local (MAXCUT and Ising model)
and global (quantum chemistry examples) cost functions.
X X
H= wij Zi Zj + (gx Xi + gz Zi ). (1)
Here the distinction between local and global loosely ij i
refers to how many non-identity Pauli operators are in
each of the Pauli strings in the decomposition of the For MAXCUT, wij is the graph adjacency matrix and
Hamiltonian [33]. Furthermore, we investigate the ef- gx = gz = 0, while for the TFIM wij = Jδi,i+1 with cou-
fect of ansatz circuit type, circuit depth, and Fermion- pling constant J (and Kronecker-delta δij ), and gx (gz )
 4

the transverse (longitudinal) field. Other constrained Next, we consider the ten qubits H2 O Hamiltonian
and unconstrained quadratic binary optimisation prob- with varying O-H bond length and H-O-H bond angle,
lems can be written in this Ising-like form [11, 34]. While Fig. 3. Again we note the parity transform outper-
the TFIM Hamiltonian is a local cost function Hamilto- forms the Jordan-Wigner transform (Fig. 3). In both
nian, it is known to have a phase transition where the transformations, there is a noticeable bias in the min-
ground state exhibits volume law entanglement [35], al- imum towards shorter bond lengths, and larger bond
lowing the potential for a non-trivial stabilizer state to angles. That is, Clifford pre-optimisation finds a more
be found via the Clifford search. accurate energy solution near shorter bond lengths and
As well as different types of Hamiltonians and trans- larger bond angles. Since quantum chemistry Hamilto-
forms, we compare the performance of different types nians have far more Pauli strings (typically O(poly(n)))
of ansatz: (i) an ansatz that parameterizes only real- compared to solid-state (O(n)) or binary optimisation
valued stabilizer states, consisting of only controlled Y (O(n2 )), we expect them to benefit the most from circuit-
rotations and CNOT gates, (ii) an ansatz inspired by agnostic Clifford pre-optimisation, due to the higher
the Trotterized unitary for the TFIM Hamiltonian that probability of finding non-zero expectation value stabi-
allows complex-valued stabilizer states, and (iii) for the lizer states.
MAXCUT task we consider the standard QAOA ansatz.
The real-valued and QAOA ansätze are more constrained
than the trotterized ansatz and parameterise fewer sta- B. TFIM
bilizer states for a given depth, but may be more suited
to problems with known real-valued ground state eigen- Here we benchmark the Clifford pre-optimisation with
vectors, such as some quantum chemistry problems and an eight-qubit TFIM Hamiltonian, with varying gx /J
binary optimisation. It turns out however, that this is and gz /J fields (see Eq. (1) with wij = Jδi,i+1 ). Fig. 4
not necessarily the case, as we shall later show. shows the Clifford search finds low energy states for low
fields, but begins struggling for higher fields. We also
plot the solution after two hundred rounds of continuous
A. Quantum chemistry parameter SPSA descent to resolve the effects of pre-
optimisation and circuit expressibility. For gx = gz ≈
Fig. 2 shows the effect of ansatz type, ansatz depth, 0.9, the ground state exhibits long-range entanglement;
and Fermion transform for a chain of eight hydrogen however, since the continuous parameter gradient decent
atoms. The H8 chain is parameterized by the dimeri- (using Simultaneous Perturbation Stochastic Approxi-
sation length (see inset in Fig. 2), and the plots show the mation — SPSA) quickly converges within two hundred
best expectation value found over 104 iterations of simu- iterations, this suggests the ansatz is sufficiently expres-
lated annealing of the Clifford circuits (see appendix A). sive for the required entanglement, but a non-trivial sta-
The real ansatz outperforms the Trotterized ansatz at bilizer solution isn’t found.
depth D = 2, indicating the Trotterized ansatz struggles For gx ≥ gz , the Clifford ansatz finds the Xi stabi-
to parameterize Clifford points close to the ground state lizer states (i.e. the eigenstate of each of the Pauli Xi
despite having the same entanglement structure. In con- operators), independent of the gz field. This can be un-
trast, the Trotterized ansatz performs better at depth derstood by noting that Zi generates Zi Zi+1 terms in
D = 5 as it allows more flexible (complex valued) sta- the stabilizer group (i.e. all Zi Zi+1 are products of the
bilizer states. Interestingly, although the depth D = 5 single Zi operators), thereby giving correlated eigenval-
real-valued ansatz can parameterise a super-set of the ues as they all commute. These correlations preclude a
D = 2 stabilizer states, it performs significantly worse. low energy solution: the low energy (i.e. -1 eigenstates)
We attribute this to ansatz over-parameterization as the states of Zi and Zi+1 , result in a positive energy contri-
circuit expressibility and parameter dimension are both bution for Zi Zi+1 (i.e. is a +1 eigenstate), while no such
significantly increased. Correspondingly the D = 2 Trot- restriction applies for Xi and Xi+1 stabilizers. Likewise,
terized ansatz is underparameterized, as the more com- all low energy (i.e. odd parity) stabilizer states of Zi Zi+1
plex (and more difficult to optimize) D = 5 circuit seems will give identically zero energy for Xi , Xi+1 terms, and
to find a good solution. There is also a noticeable dif- zero energy for Zi + Zi+1 terms[36]. For large gx , the
ference between the Fermion transforms: each transform Clifford pre-optimisation simply finds the ground state
results gives a different set of weights and Pauli strings, of the J = gz = 0 Hamiltonian, and there appears to be
and therefore the ground state has different coefficients in little benefit in running it. This emphasises that Clifford
the logical basis. This places the ground state at different pre-optimisation can be parameter-specific and not just
parts of the Hilbert space, and therefore has different pro- problem-specific.
jections onto the set of circuit-accessible stabilizer states. Following this logic, we could expect that when
While here it looks like the Bravyi-Kitaev slightly un- the Hamiltonian contains many non-commuting equally
derperforms compared to the Parity and Jordan-Wigner weighted Pauli strings, the lowest energy Clifford state
maps, we have no current understanding as to if this is may not give a good ground state estimate. Nonetheless,
expected in general. in some cases (e.g. gz ≈ gx ) it can give a substantially
 5

Parity Jordan Wigner Exact

2.5

-74.8

Bond angle (rad)

Energy (H)
2.0

-74.9
1.5

-75.0
1.0
0.8 1.0 1.2 1.4 0.8 1.0 1.2 1.4 0.8 1.0 1.2 1.4
o
O-H Bond length (A)

FIG. 3. Ground state energy surface of H2 O. (left) Clifford pre-optimisation with Parity mapping (mid) Clifford pre-optimisation
with Jordan-Wigner mapping, and (right) exact diagonalization. There is a noticeable bias towards shorter distances/large
bond angles where the pre-optimisation gets closer to the ground state. The patches in the Jordan-Wigner plot suggest the
optimisation landscape under this transform is more difficult for the simulated annealing. The numerical results show that the
energy obtained from Clifford Pre-optimisation can find a good initialization for the ground state energy.

CliffordAnnealing Post SPSA Exact

-10
0.1

-14
gx

Energy
-18

1.5 -22

0.1 1.5 0.1 1.5 0.1 1.5

gz gz gz

FIG. 4. Eight-qubit transverse field Ising model [Eq. (1)] energy as a function of dimensionless transverse (gx ) and longitudinal
(gz ) fields with J = 1. Plots show the lowest energy seen over the first 104 annealed stabilizer states (left), lowest energy
after two hundred further rounds of SPSA starting from the best initial point (middle), and solution from exact diagonalization
(right). The Clifford results (left) show that as gx becomes larger than gz , the ansatz seems to find the same energy (independent
of gz ). This suggests that the stabilizer state is preferentially finding Xi eigenstates over Zi , or Zi Zi+1 eigenstates (a similar
effect is seen for large gz and gx ≤ 0.5).

better initial condition, thereby allowing a continuous nature of the circuit, the stabilizer search is not a useful
parameter VQE optimisation to convergence faster. initialisation strategy. For binary optimisation problems
the target state is a logical basis state, and while there
are guarantees that the optimal circuit gets close to the
C. MAXCUT Problem target basis states [8], the circuit cannot exactly param-
eterize a good stabilizer state. We note that since the
target state is a logical basis state, the optimal Clifford
Finally, we consider Clifford pre-optimisation for the ansatz is constructed only from a single layer of single-
MAXCUT problem. The MAXCUT task is qualitatively qubit X-gates — in which case the Clifford search triv-
different from quantum chemistry and TFIM tasks, as ially becomes identical to an implementation of classical
all of the O(n2 ) Pauli Strings commute with each other. simulated annealing.
Furthermore, it is an example of a fundamentally classical
problem of unconstrained quadratic binary optimisation.
When the ansatz circuit is not expressive enough to
include the ideal stabilizer state (i.e. the solution), one IV. SCALING
might expect the Clifford search to be unsuccessful in
finding a low expectation value initialisation. Inter- Here we analytically argue that a Clifford based ini-
estingly, this is also the case for the problem-specific tialisation is not arbitrarily scalable, but due to practical
(QAOA) ansatz circuits. The QAOA ansatz can be read- considerations, may still be useful for VQE tasks involv-
ily tuned to a Clifford circuit (by setting all parameters ing several tens of qubits. In agreement with the numeri-
to be multiples of π/2). However, due to the constrained cal results above, we show that the method is particularly
 6

0
where M is the number of terms in the Pauli decomposi-
− 20 tion of H. If M scales as poly(n), then the Hamiltonian
cannot escape the fate of having an exponentially shrink-
− 40 ing size of stabilizer states, giving an effective barren
Energy

− 60
plateau for Clifford points. This exponentially growing
Depth
5 zero-eigenvalue stabilizer state will become the main ob-
− 80 10 Exa
20
c t (e stacle to scaling up the Clifford pre-optimisation, and is
xtra
30 pol equivalent to the barren plateau in the continuous param-
− 100 ate
50 d)
eter case. Nevertheless, stabilizer state pre-optimisation
10 20 30 40 50 60 70 80 90 is useful for device bench-marking, and possible (non-
Number of qubits scalable) small-scale demonstration of quantum advan-
tage, and selecting useful ansatz circuits.
FIG. 5. Minimum pre-optimised energy for the TFIM prob- For less-than-linear depth circuits, the number of ac-
lem (gz = gx = 0.85J), as the number of qubits increases, but cessible stabilizer states is far smaller than the (expo-
keeping the Clifford annealing time fixed to 104 iterations. nential in n2 ) number of all stabilizer states, NT (this
The dashed line shows the (linearly extrapolated) ground is also the case for any depth circuits with a restricted
state. The relationship between the depth (different colours) number of parameterized gates). This smaller set of
and number of qubits suggest the method is not arbitrarily
circuit-constrained set of states could allow a higher-
salable for the TFIM, but for shallower circuits and several
tens of qubits, Clifford pre-optimisation can find a good initial than-random chance of a given state being stabilized by
condition. Pauli strings in the Hamiltonian. This has the poten-
tial for limited scalability of classical preoptimisatoin,
perhaps into the quantum advantage regime, see Fig. 5.
attractive for quantum chemistry problems, compared to Moreover, classically finding a good initial state, or suit-
binary optimisation or solid-state problems. able ansatz circuit will allow increased throughput of
The ansatz circuit evaluated at the Clifford parameter quantum devices.
points results in a stabilizer state. The utility of the pre- Section III showed the quantum chemistry prob-
optimisation depends on how many stabilizer states give lems (with their many groups of non-commuting Pauli
a non-zero expectation value of the cost function operator strings), benefited more from the Clifford search, when
H. For an n qubit cost function, this can be compared to the TFIM (with only two groups of non-
Pestimated
given the Pauli string decomposition, H = i ci Pi , for commuting Pauli strings), or MAXCUT (where all Pauli
Pauli strings Pi and coefficients ci . Stabilizer states (|si) strings commute). Scaling arguments notwithstanding,
either return zero expectation value hs| P |si = 0 for a the more non-commuting Pauli strings in the Hamilto-
Pauli string P , or are eigenstates P |si = ± |si, in which nian, the more useful a stabilizer state search. We reiter-
case P is said to stabilize |si. ate that quantum chemistry problems fit this paradigm,
Following Ref. [21], the number NP of stabilizer states suggesting that quantum chemistry problems, in contrast
which can be stabilized by a single n qubit Pauli string to binary optimisation and solid-state models, are partic-
is NP = S(n − 1), with ularly well suited to stabilizer pre-optimisation in the ISQ
regime.
n
Y
S(n) = 2n (2k + 1) (2)
k=1
V. CONCLUSION
while the total number of stabilizer states is given by
NT = S(n). NP is exponentially smaller than NT . In this paper, we presented Clifford pre-optimisation;
Hence, for a Hamiltonian which is decomposed into a classical tool that reduces the load on the quantum
poly(n) Pauli strings, the probability of getting a non- part for the VQE. This algorithm optimizes for stabi-
zero expectation value becomes vanishing small asymp- lizer states accessible from a given quantum circuit (with
totically. Therefore in the most general case, to keep the aid of Clifford gates) to minimize a cost function.
up the scaling of the stabilizer state space, the number Many stabilizer states are accessible by a parameterized
of Pauli Strings in the Hamiltonian must also grow ex- circuit, and these uniformly distributed, problem inde-
ponentially. Furthermore, anti-commuting Pauli strings pendent states, make them ideal for comparing the cost
do not share any stabilizers, and two commuting strings functions before a quantum device is used for full contin-
share a set of NC = S(n − 2) common stabilizers. Hence, uous parameter optimization. The total number of stabi-
finding a single stabilizer state that is stabilized by two lizers scales faster-than-exponentially with the number of
commuting Pauli strings also becomes vanishing small qubits (exponential with n2 ), but exponentially with the
asymptotically. number of ansatz parameters, thereby highlighting the
Combing the above counting arguments, the number of importance of the ansatz circuit. Furthermore, the num-
stabilizer states that give a non-zero expectation value to ber of stabilizer states that are Pauli string eigenvalues
the Hamiltonian (NH ) has an upper bound, NH < NP M , also scales exponentially in n2 (although it is exponen-
 7

tially small compared to the total numbers of stabilizer bench marking, and likely (albeit not provably) useful
states), but a given restricted depth ansatz circuit does for ISQ quantum advantage, in particular for quantum
not uniformly sample all states, potentially allowing more chemistry problems, Fig. 2 and 3. However, we noted
useful states to be found. that for many qubits and deep circuits, the counting ar-
guments of Sec. IV, show the eventual onset of a barren
We presented numerical studies using Clifford cir-
plateau, and limit the arbitrary scalability of the method,
cuit annealing as the pre-optimisation method for quan-
even when individual Clifford circuits themselves are ef-
tum chemistry, TFIM, and QAOA problems. We un-
ficiently evaluated.
raveled the effects of circuit choice, Hamiltonian struc-
ture (global, local, and the number of commuting/non-
commuting Pauli terms), and Fermion transform on the VI. ACKNOWLEDGEMENTS
pre-optimisations performance. We identified that the
Clifford search finds a single stabilizer of a (commuting)
We thank Tobias Hung for his helpful insights during
set of Pauli strings in the cost function, and this is not
this work, and David Amaro for constructive comments
necessarily the maximally commuting set, Fig. 4.
on the manuscript. KEK and MSK, acknowledge funding
Our results show that Clifford pre-optimisation pro- from the Samsung GRC grant, KIST grant and EPSRC
vides a powerful new method to lower the cost of re- Quantum computing and simulation hub grant. MHC
quired quantum processing time but pushing the classical and ACM acknowledge funding from the Fraunhofer In-
limit of the VQE. The method is useful for both device stitute. Source code available on request.

[1] A. Y. Kitaev, Quantum measurements and the abelian [12] H. Ma, M. Govoni, and G. Galli, Quantum simulations
stabilizer problem (1995). of materials on near-term quantum computers, npj Com-
[2] P. Shor, Algorithms for quantum computation: discrete putational Materials 6, 85 (2020).
logarithms and factoring, in Proceedings 35th Annual [13] M. Consiglio, W. J. Chetcuti, C. Bravo-Prieto, S. Ramos-
Symposium on Foundations of Computer Science (1994) Calderer, A. Minguzzi, J. I. Latorre, L. Amico, and
pp. 124–134. T. J. G. Apollaro, Variational quantum eigensolver for
[3] D. Deutsch and R. Jozsa, Rapid solution of problems by su(n) fermions (2021).
quantum computation, Proceedings of the Royal Society [14] F. Arute, K. Arya, R. Babbush, D. Bacon, J. C.
A 439, 553–558 (1992). Bardin, R. Barends, S. Boixo, M. Broughton, B. B.
[4] A. Peruzzo, J. McClean, P. Shadbolt, M. H. Yung, X. Q. Buckley, D. A. Buell, B. Burkett, N. Bushnell, Y. Chen,
Zhou, and P. J. Love, A variational eigenvalue solver on Z. Chen, B. Chiaro, R. Collins, W. Courtney, S. De-
a photonic quantum processor, Nature Communications mura, A. Dunsworth, E. Farhi, A. Fowler, B. Foxen,
5, 4213 (2014). C. Gidney, M. Giustina, R. Graff, S. Habegger, M. P.
[5] J. R. McClean, J. Romero, R. Babbush, and A. Aspuru- Harrigan, A. Ho, S. Hong, T. Huang, W. J. Huggins,
Guzik, The theory of variational hybrid quantum- L. Ioffe, S. V. Isakov, E. Jeffrey, Z. Jiang, C. Jones,
classical algorithms, New Journal of Physics 18, 023023 D. Kafri, K. Kechedzhi, J. Kelly, S. Kim, P. V. Klimov,
(2016). A. Korotkov, F. Kostritsa, D. Landhuis, P. Laptev,
[6] J. Romero, R. Babbush, J. R. McClean, C. Hempel, P. J. M. Lindmark, E. Lucero, O. Martin, J. M. Martinis, J. R.
Love, and A. Aspuru-Guzik, Strategies for quantum com- McClean, M. McEwen, A. Megrant, X. Mi, M. Mohseni,
puting molecular energies using the unitary coupled clus- W. Mruczkiewicz, J. Mutus, O. Naaman, M. Neeley,
ter ansatz, Quantum Science and Technology 4, 014008 C. Neill, H. Neven, M. Y. Niu, T. E. O’Brien, E. Ostby,
(2018). A. Petukhov, H. Putterman, C. Quintana, P. Roushan,
[7] D. S. França and R. Garcı́a-Patrón, Limitations of op- N. C. Rubin, D. Sank, K. J. Satzinger, V. Smelyanskiy,
timization algorithms on noisy quantum devices, Nature D. Strain, K. J. Sung, M. Szalay, T. Y. Takeshita,
Physics 17, 1221 (2021). A. Vainsencher, T. White, N. Wiebe, Z. J. Yao, P. Yeh,
[8] E. Farhi, J. Goldstone, and S. Gutmann, A quantum ap- and A. Zalcman, Hartree-fock on a superconducting
proximate optimization algorithm (2014). qubit quantum computer, Science 369, 1084 (2020),
[9] D. Amaro, C. Modica, M. Rosenkranz, M. Fioren- https://www.science.org/doi/pdf/10.1126/science.abb9811.
tini, M. Benedetti, and M. Lubasch, Filtering vari- [15] X. Mi, M. Ippoliti, C. Quintana, A. Greene, Z. Chen,
ational quantum algorithms for combinatorial opti- J. Gross, F. Arute, K. Arya, J. Atalaya, R. Bab-
mization, Quantum Sci. Technol. 7, 015021 (2022), bush, J. C. Bardin, J. Basso, A. Bengtsson, A. Bilmes,
arXiv:2106.10055 [quant-ph]. A. Bourassa, L. Brill, M. Broughton, B. B. Buck-
[10] R. Orús, S. Mugel, and E. Lizaso, Quantum computing ley, D. A. Buell, B. Burkett, N. Bushnell, B. Chiaro,
for finance: Overview and prospects, Reviews in Physics R. Collins, W. Courtney, D. Debroy, S. Demura, A. R.
4, 100028 (2019). Derk, A. Dunsworth, D. Eppens, C. Erickson, E. Farhi,
[11] L. Braine, D. J. Egger, J. Glick, and S. Woerner, Quan- A. G. Fowler, B. Foxen, C. Gidney, M. Giustina, M. P.
tum algorithms for mixed binary optimization applied to Harrigan, S. D. Harrington, J. Hilton, A. Ho, S. Hong,
transaction settlement, IEEE Transactions on Quantum T. Huang, A. Huff, W. J. Huggins, L. B. Ioffe, S. V.
Engineering 2, 1 (2021). Isakov, J. Iveland, E. Jeffrey, Z. Jiang, C. Jones,
 8

D. Kafri, T. Khattar, S. Kim, A. Kitaev, P. V. Klimov, [30] E. Fradkin, Jordan-wigner transformation for quantum-
A. N. Korotkov, F. Kostritsa, D. Landhuis, P. Laptev, spin systems in two dimensions and fractional statistics,
J. Lee, K. Lee, A. Locharla, E. Lucero, O. Martin, Phys. Rev. Lett. 63, 322 (1989).
J. R. McClean, T. McCourt, M. McEwen, K. C. Miao, [31] J. T. Seeley, M. J. Richard, and P. J. Love, The bravyi-
M. Mohseni, S. Montazeri, W. Mruczkiewicz, O. Naa- kitaev transformation for quantum computation of elec-
man, M. Neeley, C. Neill, M. Newman, M. Y. Niu, T. E. tronic structure, The Journal of Chemical Physics 137,
O’Brien, A. Opremcak, E. Ostby, B. Pato, A. Petukhov, 224109 (2012).
N. C. Rubin, D. Sank, K. J. Satzinger, V. Shvarts, [32] S. B. Bravyi and A. Y. Kitaev, Fermionic quantum com-
Y. Su, D. Strain, M. Szalay, M. D. Trevithick, B. Villa- putation, Annals of Physics 298, 210 (2002).
longa, T. White, Z. J. Yao, P. Yeh, J. Yoo, A. Zalcman, [33] J. Kempe, A. Kitaev, and O. Regev, The
H. Neven, S. Boixo, V. Smelyanskiy, A. Megrant, J. Kelly, complexity of the local hamiltonian problem
Y. Chen, S. L. Sondhi, R. Moessner, K. Kechedzhi, 10.48550/ARXIV.QUANT-PH/0406180 (2004).
V. Khemani, and P. Roushan, Time-crystalline eigenstate [34] G. G. Guerreschi, Solving quadratic unconstrained bi-
order on a quantum processor, Nature 601, 531 (2021). nary optimization with divide-and-conquer and quantum
[16] C. W. Bauer, B. Nachman, and M. Freytsis, Sim- algorithms (2021).
ulating collider physics on quantum computers using [35] A. Dutta, G. Aeppli, B. K. Chakrabarti, U. Divakaran,
effective field theories, Physical Review Letters 127, T. F. Rosenbaum, and D. Sen, Quantum phase tran-
10.1103/physrevlett.127.212001 (2021). sitions in transverse field spin models: from statistical
[17] D. Fedorov, B. Peng, and N. e. a. Govind, Vqe method: a physics to quantum information (Cambridge University
short survey and recent developments, Materials Theory Press, 2015).
6, 10.1186/s41313-021-00032-6 (2022). [36] A low energy hZi Zj i stabilizer state is either an odd
[18] A. Peruzzo, J. McClean, P. Shadbolt, M. H. Yung, X. Q. parity logical state (in i, j) in which case hZi i = ±1 =
Zhou, and P. J. Love, Barren plateaus in quantum neural − hZj i, or a Bell-like state (in i, j) in which case hZi i =
network training landscapes, Nature Communications 5, hZj i = 0.
4812 (2018). [37] Here local minima is referring to the θ vector locality, not
[19] J. Kim and Y. Oz, Entanglement diagnostics for efficient Hilbert-space locality as change in a gate angle can have
quantum computation (2021). a large change of the parameterised state in the Hilbert
[20] Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, space.
Connecting ansatz expressibility to gradient magnitudes
and barren plateaus, PRX Quantum 3, 10.1103/prxquan-
tum.3.010313 (2022).
[21] S. Aaronson and D. Gottesman, Improved simulation of
stabilizer circuits, Physical Review A 70, 10.1103/phys-
reva.70.052328 (2004).
[22] G. S. Ravi, P. Gokhale, Y. Ding, W. M. Kirby, K. N.
Smith, J. M. Baker, P. J. Love, H. Hoffmann, K. R.
Brown, and F. T. Chong, Cafqa: Clifford ansatz for quan-
tum accuracy (2022).
[23] K. Mitarai, M. Negoro, M. Kitagawa, and K. Fu-
jii, Quantum circuit learning, Physical Review A 98,
10.1103/physreva.98.032309 (2018).
[24] E. Grant, L. Wossnig, M. Ostaszewski, and M. Benedetti,
An initialization strategy for addressing barren plateaus
in parametrized quantum circuits, Quantum 3, 214
(2019).
[25] J. Dborin, F. Barratt, V. Wimalaweera, L. Wright, and
A. G. Green, Matrix product state pre-training for quan-
tum machine learning (2021).
[26] A. Skolik, J. R. McClean, M. Mohseni, P. van der Smagt,
and M. Leib, Layerwise learning for quantum neural net-
works, Quantum Machine Intelligence 3, 10.1007/s42484-
020-00036-4 (2021).
[27] D. Gottesman, The heisenberg representation of quan-
tum computers 10.48550/ARXIV.QUANT-PH/9807006
(1998).
[28] A. Anand, P. Schleich, S. Alperin-Lea, P. W. K. Jensen,
S. Sim, M. Dı́ az-Tinoco, J. S. Kottmann, M. Degroote,
A. F. Izmaylov, and A. Aspuru-Guzik, A quantum com-
puting view on unitary coupled cluster theory, Chemical
Society Reviews 51, 1659 (2022).
[29] A. W. Harrow and R. A. Low, Random quantum circuits
are approximate 2-designs, Communications in Mathe-
matical Physics 291, 257 (2009).
 9

Appendix A: Clifford circuit annealing

Here we provide a detailed description of the Clifford pre-optimisation algorithm. The ansatz circuits we consider
are (i) Complex valued ansatz that has the same structure as a Trotterized dynamics of an Ising model, and (ii) Real
valued ansatz consisting of CX and parameterised Ry only, see Fig. 6. Note that for a given depth D and qubit count
n the Trotter ansatz as D(3n − 1) parameters while the Real ansatz has Dn. Each parameter is an Rx , Rz or CZ
(controlled Z) rotation. The algorithm considers the classically simulable sector of an ansatz quantum circuit, i.e.
restricting parameterised rotations in the parameter vector θ = (θ1 , θ2 , ...θN ) to each be multiples of π/2.
To minimize the state over this discrete set we consider classical annealing in the Clifford angles. This method
randomly changes exactly two components (chosen heuristically) of the parameter vector θ — keeping the changed
parameters in {0, π/2, π, 3π/2} — to give a new θ0 . If the new cost hψ(θ0 )| H |ψ(θ0 )i is lower than the previous
cost hψ(θ)| H |ψ(θ)i , the updated θ0 vector is accepted with probability one, otherwise the new θ0 is accepted with
probability exp(−β∆E) with ∆E = hψ(θ0 )| H |ψ(θ0 )i − hψ(θ)| H |ψ(θ)i. Here β is the effective inverse temperature
and is set to be on the order of the (estimated) characteristic energy gap between low energy states. If this gap is
unknown then beta may be chosen to be arbitrarily large corresponding to zero temperature annealing.
To stop the annealing algorithm getting stuck in a local minima [37] we randomly reset the full θ vector if the
annealing hasn’t found a better θ0 point over the last k iterations. For all plots we heuristically choose k in the
range of 400-1000. We note other annealing methods may be applicable, e.g. parallel tempering, or different (perhaps
iteration-dependent) rules for selecting which individual θi parameters to update. Or other discrete optimisation
algorithms, such as Bayesian optimisation [22].
The annealing trajectory is local in the parameter space (vector space of θ), but can be non-local in terms of Hilbert
space jumps of the resulting stabilizer state. There trajectory can therefore be dependent on the initial choice of the
Clifford point. That can lead to initial condition-dependent local minima in the classical optimisation process. We
note that our resetting after k sequential non-improving evaluations helps prevent this issue. An example annealing
iteration series in Fig. 7 shows the utility of threshold-dependent resetting in finding a good initial solution.

Trotter ansatz Real ansatz

H RX(p3) RZ(p6) RY(p1)

CZ(p1)

H RX(p4) RZ(p7) RY(p2)

CZ(p2)

H RX(p5) RZ(p8) RY(p3)

× depth × depth

FIG. 6. Ansatz circuits used for TFIM and quantum chemistry problems. Trotterised inspired ansatz (with independent
parameters for each control-Z and single qubit rotations in each layer) allows complex valued wave functions. The real-valued
ansatz has only Y-rotations and control not gates and ensures the final wave function is real valued.
 10

Increasing
Increasing electrons
circuit
H H H H H H H H H H H H depth
r r r r r r

0
Circuit
1 Depth
1
2

3
Ansatz
Type
Energy (Hartree)

0 Trotter
Circuit
Real
1 Depth
3
2

Dimer
Distance
o
0 (A)
Circuit
1 Depth
5 r=0.8
2 r=1.0
3
0 5000 0 5000 0 5000
Iterations

FIG. 7. Plots shows ansatz circuit selection trajectories for three different quantum chemistry problems: (left) H2 (middle) H4
and (right) H6 , corresponding to four, six and eight qubits respectively (not including symmetry reduction). Rows correspond
to different depth ansatz circuits, (top) depth one, (middle) depth three and (bottom) depth five, with the colors corresponding
to the Trotter ansatz (blue) and real ansatz (orange). Line styles correspond to different dimer distances with a fixed inter
dymer distance. Each curve is efficiently classically calculated and the set can be used to select an appropriate circuit for the
on-device problem. Note the depth one real ansatz seems to find as good a solution as the depth five Trotter ansatz, but due
to the few parameter, tends to find a good solution faster, and more consistently.

Appendix B: Clifford solution vs true ground states

Other than observing the exact ground state energy, we can compare the stabilizer state and the ground state
solution through looking at their respective contribution from individual Pauli strings in the Hamiltonian. The set
of expectation values also tells us about the amount of magic in the exact ground state solution, since a state with
high magic will have expectation values that deviate from 1, 0, −1. If we observe a low magic ground state, it means
that there is a stabilizer state that is close to the exact solution, which is good for the gradient descent. In the case
of a high magic ground states (see Fig. 8), we can still find a stabilizer state with relatively low expectation value
compared to random parameterisation. However, since the ground state solution is far from a stabilzier state, the
stabilizer state found is not guaranteed to be close to ground state, and still may be difficult to optimise..
 11

0.0

Expectation value
stabiliser solution
exact diagonalisation

1.0

5 1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9

9
3
4

6
2

8
4,

6,

8,
2,
3,

5,
1,

7,
ZZ-terms Z-terms X-terms

P
FIG. 8. Expectation values of each individual Pauli string in a nine qubit TFIM Hamiltonian H = i (Zi Zi+1 ) + 2Zi +
2Xi , comparing the exact expectation values (orange circles) and stabilizer solution (blue crosses). The stabilizer solution
(constrained to be ±1 or zero) gets the general pattern of low/high correct, but with an emphasis on the Xi stabilizer states
as discussed in the main text.