Implementation of the Density-functional Theory on Quantum

Computers with Linear Scaling with respect to the Number of Atoms

Taehee Ko,1, ∗ Xiantao Li,1, † and Chunhao Wang2, ‡

1
Department of Mathematics, Pennsylvania State University
2
Department of Computer Science and Engineering, Pennsylvania State University
Density-functional theory (DFT) has revolutionized computer simulations in chemistry and mate-
rial science. A faithful implementation of the theory requires self-consistent calculations. However,
this effort involves repeatedly diagonalizing the Hamiltonian, for which a classical algorithm typically
requires a computational complexity that scales cubically with respect to the number of electrons.
arXiv:2307.07067v1 [quant-ph] 13 Jul 2023

This limits DFT’s applicability to large-scale problems with complex chemical environments and
microstructures. This article presents a quantum algorithm that has a linear scaling with respect to
the number of atoms, which is much smaller than the number of electrons. Our algorithm leverages
the quantum singular value transformation (QSVT) to generate a quantum circuit to encode the
density-matrix, and an estimation method for computing the output electron density. In addition,
we present a randomized block coordinate fixed-point method to accelerate the self-consistent field
calculations by reducing the number of components of the electron density that needs to be esti-
mated. The proposed framework is accompanied by a rigorous error analysis that quantifies the
function approximation error, the statistical fluctuation, and the iteration complexity. In partic-
ular, the analysis of our self-consistent iterations takes into account the measurement noise from
the quantum circuit. These advancements offer a promising avenue for tackling large-scale DFT
problems, enabling simulations of complex systems that were previously computationally infeasible.

I. INTRODUCTION

One of the breakthroughs in computational chemistry is the development of the density-functional

theory (DFT) [40], which led to the Nobel Prize in Chemistry in 1998. The theory is founded on
the observation that the electronic structures are fully determined by the underlying electron den-
sity n(r); r ∈ R3 → [0, +∞) which, thanks to the remarkable work of Kohn and Sham [46], can
be represented through an auxiliary system of non-interacting electrons with an effective Kohn-
Sham Hamiltonian H. The electron-electron interactions are captured by an exchange-correlation
energy functional of n(r), which is part of the Hamiltonian operator. In the implementation, the
electron density is computed self-consistently to meet the self-consistent field (SCF) requirement
[68], which mathematically corresponds to a fixed-point problem [54, 84]. Meanwhile, direct com-
putation amounts to calculating many eigenvalues of a large-dimensional matrix, for which the
computational cost typically scales cubically with the dimension [93]. Such scaling has been the
major limiting factor for large-scale DFT calculations. The most expensive component in comput-
ing the new electron density is the eigenvalue calculations. Due to such cubic scaling, large-scale
DFT calculations are still an outstanding challenge in connecting electron structure to macroscale
material properties, e.g., perovskite materials [38], high-entropy alloys [17], bi-layer two-dimensional
materials with small twist angle [92], and biomaterials [26].

∗ [email protected]
† [email protected].
‡ [email protected]
 2

The purpose of this paper is to demonstrate a quantum speedup with a new quantum algorithm
that leverages many unique capabilities of quantum computing devices. Instead of explicitly com-
puting the eigenvalues and eigenvectors, we apply an eigenvalue transformation and construct a
quantum circuit for the density-matrix in DFT. The electron density is then extracted from the
diagonals of the density-matrix. We will show that the gate and query complexity of the algorithm
only scales linearly with respect to the dimension of the problem, which we will compress down
to be proportional to the number of atoms. In addition, we propose an efficient self-consistent
iteration algorithm, where only some components of the electron density are updated. We provide
theoretical analysis for the convergence of the iteration methods. Our numerical results indicate
that the overall complexity can be well below the theoretical bound.
The paper is organized as follows. In the remainder of this introduction, we provide informal
problem statements and summarize our main results, followed by discussions of related works. In
Section II, we detail the problem setup and highlight the computational aspects, including the
spatial discretization and SCF iterations. Our quantum algorithms will be presented in Section III,
together with error estimates and complexity bounds. We show some numerical results in Sec-
tion IV.

A. Problem Statements and Summary of Results

The self-consistent field (SCF) in DFT asserts that the electron density n(r) that enters the
effective Kohn-Sham Hamiltonian H has to be the same as the electron density determined from
the eigenvalues and eigenvectors of H. This is often achieved by iterations, n → F (n), with
n and F (n) being respectively the input and updated density represented at grid points (see
the precise definition in Eq. (12)). The components of F (n) can be linked to a density-matrix,
expressed as the diagonals of a matrix function, f (H), with f being the Fermi-Dirac function
−1
f (x) = (1 + exp β(x − µ)) at finite temperature (β = O(1)). The first problem addresses the
computation of F (n).
Problem 1 (Updating the electron density). Assume that the effective Hamiltonian H is approx-
imated by a Hermitian matrix H with sparsity s on a set of grid points. Suppose we are given an
oracle to access H and its nonzero elements (See Eqs. (13) and (14)). Determine an estimate for
the updated electron density F̂ (n) ∈ RNI at NI grid points, such that F̂ (n) − F (n) < ϵ.

Theorem 1 (Informal version of Theorem 11). There is a quantum algorithm that outputs an
approximate electron density F̂ (n) such that F̂ (n) − F (n) < ϵ with probability at least 1 − δ.
e sNI .1


Under the assumptions above, the algorithm involves a query complexity O ϵ

The updated density F (n) can be used as the input density at the next iteration, and the
iterations continue until the two densities coincide. In the limit, the electron density converges to
the ground state density n∗ , i.e., n∗ = F (n∗ ).
Problem 2 (Determining the ground state electron density). Given an input electron density n0 (r),
such that ∥n∗ (r) − n0 (r)∥ < γ, with sufficiently small γ. Determine an estimate n̂(r) such that
∥n∗ (r) − n̂(r)∥ < ϵ.

1 We use O
e to neglect poly-logarithmic factors.
 3

Theorem 2 (Informal version of Theorem 17). Assume that the effective Hamiltonian H is approx-
imated by H on a set of grid points and H has sparsity s. Suppose we are given an oracle to access
H and its nonzero elements (See Eqs. (13) and (14)). There is a hybrid quantum-classical algo-
rithm that outputs an approximate ground state electron density n̂(r) such that ∥n∗ (r) − n̂(r)∥ < ϵ.
e sNI


Neglecting logarithmic factors, and under the assumptions above, the algorithm involves O ϵ
queries to the Hamiltonian matrix H.

Overall, our approach is a quantum-classical hybrid algorithm, where the quantum algorithm
produces the density-matrix f (H) (see also Γ in Eq. (9)), while the classical algorithm employs an
SCF iteration to provide updated values of the electron density to reprogram the quantum algorithm
by updating H at the next step. As a result, the updated density is subject to measurement noise.
This gives rise to a stochastic SCF problem. In addition to a straightforward application of a
mixing scheme [12, 39], we propose a random coordinate method, in which, for each fixed-point
iteration (see Eq. (12)), one only chooses to update some randomly selected components of F̂ (n),
rather than computing all the components. This significantly reduces the number of measurements
needed at each iteration. The theoretical analysis shows that the new method has a comparable
convergence rate as the full coordinate method that computes all components at each iteration
step, and numerical tests suggest that the new method can be significantly more efficient overall.
Namely, the complexity can be sublinear in NI , the number of grid points.
Compared to classical algorithms, the hybrid algorithm has far better scaling in terms of the
number of electrons. This quantum speed-up offers a promising opportunity for treating large-scale
DFT problems, and it has the exciting potential to lead to accelerated discoveries enabled by DFT.

B. Related work

Classical algorithms for computing the updated electron density In classical computing,
the most expensive part of typical DFT implementations to compute the electron density is the
step of solving the Kohn-Sham equations, which is equivalent to finding eigenpairs corresponding to
the Hamiltonian matrix. Many efficient techniques have been proposed over the last two decades,
such as polynomial filtering methods [9, 56, 93], direct energy minimization [87, 88] and spectrum
slicing type methods [47, 77]. For numerical implementations of DFT, the readers are referred to
[49, 50, 58], and a large collection of software packages [28, 35, 56, 61, 78, 80, 91]. The complexity of
these algorithms typically scales cubically with the number of electrons: O(Ne3 ). There are several
classical algorithms for electron structure calculations that exhibit linear (potentially sublinear
scaling) with respect to the number of electrons [22, 29, 30, 34, 81]. Although such complexity is
particularly attractive for large-scale problems in material science and chemistry, these methods do
not strictly satisfy self-consistency. Therefore it is difficult to quantify the error when compared to
the true ground-state electron density.
Quantum algorithms for electron structure calculations. Studying electron structures on
quantum computers has been envisioned to be one of the first few applications of quantum com-
puting. Many algorithms have been proposed to push this effort forward based on the many-body
Schrödinger equation [1, 2, 4, 5, 37, 53, 67, 82, 92]. But the ability of these algorithms to handle
large-scale problems, e.g., those in [17, 26, 38, 92] that are of direct practical interest, has not been
demonstrated. We expect that even for fault-tolerance quantum computers, mean-field models,
such as DFT, will still be an important alternative. Meanwhile, due to its origin in many-body
quantum physics, DFT has been studied in the context of quantum computing [7, 27, 79]. The
 4

work of Baker and Poulin [7] attempts to compute the Kohn-Sham potential with quantum com-
puting; Gaitan and Nori [27] demonstrated that DFT can be formulated for a quantum system
consisting of qubits; Senjean et al. [79] showed how the Kohn-Sham Hamiltonian, constructed
based on an auxiliary non-interacting system, can be mapped to an interacting Hamiltonian on
quantum computers. None of the aforementioned works, however, solves the self-consistent DFT
model directly on quantum computers. More importantly, precise error estimates and the overall
computational complexity were not addressed. To the authors’ knowledge, this paper is the first
attempt to faithfully implement DFT on quantum computers with rigorous complexity estimates
and direct comparison of the complexity with classical algorithms.
Classical algorithms for the self-consistent iterations. The most widely used algorithm
for the SCF iterations is the mixing schemes [12, 39, 54], which can be proved to be at least
linearly convergent [54, 83, 84]. Therefore, the complexity, in terms of the number of iterations
until convergence, involves a log 1ϵ scaling. Our hybrid algorithm, which uses a block encoding [33]
of the density-matrix (Γ in Eq. (9)) and the amplitude amplification [13] to estimate the electron
density, has the same complexity in the fixed-point iterations. Motivated by the remarkable success
of stochastic approximation methods [21, 31, 62, 75, 89] in the optimization of large-scale machine
learning models [11], the authors Ko and Li [45] have recently developed a classical algorithm to
carry out the self-consistent iterations stochastically, so that each iteration has a linear scaling
complexity. The core of the algorithms in both [45] and [6] is a randomized numerical method
called the trace estimator [3, 10, 36, 48, 51, 60, 73]. Such an approach was applied in computational
chemistry [6] for the DFT, albeit without rigorous error bounds. For the computational complexity,
the total number of iterations has a polynomial dependence on ϵ. In addition, each step of the
iteration still requires a diagonalization of a small matrix. These limitations will be removed in the
present framework, by using quantum computing algorithms.
Random coordinate methods in machine-learning and reinforcement-learning. Coordinate-
wise iterative algorithms are the state of the art for many large-scale problems, due to their simplic-
ity, low cost per iteration, and overall efficiency. Many variants have been developed and improved
convergence than the full coordinate counterparts has been demonstrated for optimization prob-
lems, see [18, 43, 55, 63, 64, 66, 76, 90] and the references therein. What is closely related to the
current work is the idea of the coordinate-wise update rule has been used in the context of fixed-
point problems [20, 23, 42, 69, 86] but less explored than in the optimization tasks. Among these
works, the asynchronous coordinate update rules in [69, 86] have a resemblance to the proposed
algorithm, as those rules randomly update a portion of components of fixed-point mappings at
each iteration. But there are fundamental differences in our approach. For example, the focus of
[86] is on the Q-learning in the context of reinforcement learning, which requires more restrictive
assumptions due to its complicated problem setup. In [69], the authors consider situations where
the estimation of a component of a mapping is exact while in our setting, it involves random noise
due to the nature of quantum measurement. Besides, we provide convergence analysis of the block
coordinate case which was not analyzed in [69].

II. PROBLEM SETUP

Notations. We use bold fonts for vectors, e.g., r, and the entries will be labeled in parenthesis,
e.g., r(j) being the jth entry of r. D ⊂ R3 will be used to denote the physical domain. n(r) : D →
[0, +∞) is a function representing the electron density. Here is a summary of the notations that
will be used in this section. Ne and Na are respectively the number of electrons and the number of
atoms. In the numerical discretization, Dδ is a set of grid points in D, with Ng being the number
 5

of grid points, which is often comparable to Ne . We choose Ng = 2m to map functions defined at

the grid points to quantum states in a m-qubit system. H ∈ RNg ×Ng refers to the Hamiltonian
represented at the grid points. Meanwhile, D∆ is a set of coarse grid points in D and NI = |D∆ |,
which is comparable to Na .
The DFT formulations. DFT is formulated to find the ground state energy of a system by
solving the following eigenvalue problem [46],

H[n] |ψj ⟩ = Ej |ψj ⟩ ,

∇2
H[n] = − + V [n](r),
2 (1)
V [n](r) := VH [n](r) + Vxc [n](r) + Vext ,
Z
ψi (r)∗ ψj (r)dr = δij ,

where Ej ’s are the Kohn-Sham eigenvalues and ψj ’s are the Kohn-Sham wavefunctions. The no-
tation [·] indicates a dependence on the function n(r) of a functional. The first term in H is the
one-electron kinetic energy. VH [n](r) is the Hartree potential, which is a functional of n. More
precisely, this potential can be obtained by solving the Poisson equation [32],
1 2
− ∇ VH (r) = n(r) + b. (2)
4π
In the above equation, b comes from the pseudocharges from the nuclei and other possible charge
corrections. The second term Vxc [n] embodies the electron-electron interactions. This function
is universal and it has been parameterized in function forms that are easily implementable, e.g.,
[57, 71]. Finally, the external potential energy Vext accounts for the interaction between electrons
and nuclei.
From the eigenvalue problem in Eq. (1), the finite temperature density matrix operator, which is
known as the first-order density matrix [68], is defined as

ρ1 (r, r ′ ) := f (H), (3)

where f is the Fermi-Dirac function with the inverse temperature β and the chemical potential µ,
1
f (x) = . (4)
1 + exp(β(x − µ))
The computation of ρ1 requires a given chemical potential µ. In the case where the number of
electrons Ne is fixed, µ is chosen such that
X
f (Ej ) = Ne , (5)
j

where f (Ej ) is referred to as occupation numbers and Ne is the number of electrons. Here we
neglect the spin orbitals for simplicity.
The eigenvalue problem in Eq. (1) provides an implicit representation of the electron density. In
particular, the electron density n(r) defined by
X 2
n(r) := f (Ej )|ψj (r)| , (6)
j
 6

which in turn determines VH and Vxc in Eq. (1) and therefore the Hamiltonian H[n].
Using the matrix function notation, we can express the problem of determining the electron
density as the following fixed-point problem,


n(r) = ⟨r| f H[n] |r⟩ , (7)

where H[n] has been defined in Eq. (1).

Remark 3. It is important to point out that although it has been customary in the DFT literature
[58, 68] to refer to f (H) as a density-matrix, it may not have trace one, which is usually required
for a density-matrix in quantum information.

A. Real-space Discretization

To solve Eq. (1) in a computation, we assume that the Hamiltonian operator is properly dis-
cretized in a three-dimensional domain D by a finite-difference method [8] with grid size δ. We
denote H ∈ CM ×M to be the Hamiltonian matrix; M = 2m , so that it can be directly mapped to
the Hilbert space associated with a quantum circuit with m qubits.
Within the discretization, the electron density at the grid points is expressed as a vector n.
Following the Hamiltonian operator in Eq. (1), we can express the matrix H as follows,
1
H(n) = − ∇2δ + Vδ (n), (8)
2
where ∇2δ is a finite-difference approximation of the kinetic energy operator. Vδ , which enters the
Hamiltonian through the diagonals, is the potential evaluated at the grid points and it collects all
the potential terms in the Hamiltonian operator.
In terms of mthe mmatrix H from the finite-difference approximation, we can define the density-
matrix Γ ∈ C2 ×2 that is similar to Eq. (3),

Γ = f (H). (9)

Similarly, we generalize the continuous fixed-point problem in Eq. (7) to a discrete one,

n = f (H(n)) . (10)

Real-space discretizations usually lead to sparse Hamiltonian matrix H. Therefore, we assume

that in the following quantum algorithms, the Hamiltonian H is s-sparse, in the sense that there
are at most s nonzero entries in each row/column. s depends on the choice of the finite difference
methods. For example, if the second-order central difference method is used, then s = 7, and a
fourth-order method would give the sparsity s = 13 (see [32]).
To further make the algorithms more practical, we discuss how the problem size M can be reduced
by means of interpolation techniques. As observed in Fig. 1, the electron density typically forms
a smooth function peaked around Na atoms. From this observation, we can take NI interpolation
points, with NI comparable to Na , so that the original electron density on the entire M grid
points can be effectively represented without compromising the accuracy. The electron density will
be used to compute the potential Vδ at the interpolation points and interpolated to reconstruct
the Hamiltonian. More precisely, assuming that we have the potential Vδ (r) obtained only at NI
interpolation points (NI ≪ M ), we construct its interpolation such that for any k ∈ [M ],
 7

X
Vδ (rk ) ≈ Vδ (r)N (rk − r), (11)
r∈D∆

where D∆ is the set of the interpolation points (namely |D∆ | = NI ) and the function N are the
shape functions. In classical algorithms for DFT, this interpolation is part of the multigrid scheme to
compute VH in the Poisson equation (Eq. (2)) in DFT [59]. Besides, Vxc can be simply determined
from the interpolated electron density with explicit functional evaluation (e.g. the local density
approximation [72]). Therefore, we can efficiently construct the Hamiltonian matrix H ∈ CM ×M
from the NI -dimensional interpolated electron density. As we will discuss in Section III A, this in-
terpolation leads to an appreciable reduction of quantum random-access memory (QRAM) storage,
thereby yielding an Õ(NI ) scaling implementation of the Hamiltonian on quantum hardware.
With this interpolation, we reduce the fixed-point problem in Eq. (10) to one that is defined on
D∆ ,

n(j) = f (H(n)) (j, j) =: F (n)(j), j ∈ [NI ]. (12)

Namely, each diagonal must match with the input electron density at the grid point rj ∈ D∆ . With
a slight abuse of the notations, we will still denote this reduced fixed-point problem as n = F (n).

FIG. 1: A quick illustration of the ground state electron density of a one-dimensional

Lithium-Hydrogen chain from [44]. Also shown are the atoms (6 Li atoms in the middle and H
atoms on the two sides).

B. Self-consistent iterations

Like many mean-field theories in quantum chemistry, Eqs. (1) and (6) have to be solved self-
consistently. At the level of numerical discretization, this is manifested as the fixed-point problems
in Eq. (12).
In terms of implementation, the nonlinear mapping F is implicitly determined by the procedure
of obtaining the output from input n within the SCF iteration as shown in Fig. 2. A simple
procedure to obtain a fixed point is to apply iterations nk+1 = F (nk ) repeatedly until convergence.
 8

To guarantee and speed up convergence, mixing schemes are typically applied in practice, such as
simple mixing and Pulay mixing. This will be explained in Section III D.

FIG. 2: An illustration of the SCF procedure in DFT.

III. QUANTUM ALGORITHMS

A. Preliminaries

As pointed out in the previous section, the matrix H from the real-space discretization is usually
sparse. The sparsity implies that the matrix is efficiently row/column computable. To access H, we
assume we have access to a procedure OS that can perform the following mapping:
OS : |i⟩ |k⟩ 7→ |i⟩ |rik ⟩ , (13)
where rik is the k-th nonzero entry of the i-th row of A. In addition, OH can also perform the
following mapping:
OH : |i⟩ |j⟩ |0⟩ 7→ |i⟩ |j⟩ |H(i, j)⟩ . (14)
One key ingredient of our quantum algorithm is block encoding. We say that UA is an (α, a, ϵ)-
block-encoding of A if UA is a (m + a)-qubit unitary, and
A − α( ⟨0⊗a | ⊗ I)UA ( |0⊗a ⟩ ⊗ I) 2
≤ ϵ. (15)
Intuitively, the block encoding constructs a unitary with the upper-left block being proportional to
H,
 
A ·
UA = .
· ·
 9

To implement OS and OH efficiently, we use the interpolation in Eq. (11) to generate the electron
density in Eq. (7) approximately. For this, we need to store O(NI ) parameters in QRAM in order
to update the diagonals of H input oracle. The gate complexity for implementing such QRAM is
O(NI ). Moreover, the circuit depth of QRAM is O(log NI ) [65]. The input oracles OS and OH for
H can be implemented as a procedure that reads data in the QRAM.
In the next three sections, we will present our quantum algorithm. We first outline a high-level
description of the algorithm in Fig. 3, which consists of a quantum singular value transformation
(QSVT) to construct a quantum circuit for the density-matrix, an amplitude amplification (AA) to
estimate the updated electron density and a classical fixed-point iteration to provide the electron
density (and chemical potential if Ne is given ) for the next iteration.

Initial Effective
density potential

Block-encoding QSVT

Input Hamiltonian

AA
Output diagonal estimates
Update the Electron density
& the Fermi energy

FIG. 3: An illustration of the hybrid algorithm.

B. Preparing the density-matrix using quantum singular value transformation

Since H is Hermitian, one can use the spectral map and approximate the density-matrix in Eq. (9)
by polynomial approximations of the Fermi-Dirac function (Eq. (4)). For the error analysis, we use
a result from function approximation [85, Theorem 8.2], which is restated as follows,
Lemma 4. For any analytic function f such that can be analytically extended to an Berstein ellipse
Br with some r > 1, there exists a polynomial pℓ of degree ℓ such that
r−ℓ
max |f (x) − pℓ (x)| ≤ 2 sup |f (z)| · . (16)
x∈[−1,1] z∈Br r−1
To apply this technique to the density-matrix in Eq. (9), we rescale the Hamiltonian matrix as
follows
  
λ+ + λ−
  −1
f (H) = 1 + exp β −µ exp β̂ H̃ , (17)
2
 10

where
 
λ+ − λ− 2 λ+ + λ−
β̂ := β, H̃ := H− I . (18)
2 λ+ − λ− 2
Here λ− and λ+ are some lower and upper bounds of the eigenvalues of H. The scaling is simply
to map the eigenvalues of H to the interval [−1, 1]. One way to roughly estimate an upper bound
is to apply Gershgorin’s Circle Theorem. A tighter upper bound can be efficiently obtained by
running only a few steps of the Lanczos algorithm, as pointed out in [93], which will take O(sM )
operations. Once we get an estimate of an upper bound of λ+ , we can also obtain a lower bound
of λ− in a similar manner after shifting H properly.
Noticing that σ(H̃) ⊂ [−1, 1], we can apply Lemma 4 for the polynomial approximation of the
density matrix. The following lemma, as in [45, Remark 4.8], shows that the quality of approxima-
tion depends on a given temperature. With a slight abuse of notations, we will continue to use H
as the scaled Hamiltonian.
Lemma 5. For a given inverse temperature β, the degree of the Chebyshev expansion to approximate
the f (H), up to a precision ϵ, requires at least,
 
1
ℓ = Θ logr . (19)
ϵ
√
c(β)+ c(β)2 +4
Here the constant r satisfies that r ∈ (1, 2 ) with c(β) = λ+4π 1
−λ− β .

2π
Remark 6. We observe that at the low temperature where β̂ ≫ 1, we have r ≈ 1 + β̂
, and
!
β̂
ℓ=O . (20)
ϵ
Therefore, the QSVT approach is more efficient in the finite temperature regime.

Polynomial approximations of the density-matrix in Eq. (9) are not new. In fact, it has been
used in [25]. But in this classical algorithm, the matrix multiplications will introduce significant
computational overhead. In contrast, the quantum singular value transformation (QSVT) [33]
can efficiently prepare the density-matrix with a complexity that does not depend on the matrix
dimension explicitly.
m m
Lemma 7 ([33, Lemma 48]). Let H ∈ C2 ×2 be an m-qubit operator with at most s nonzero
entries in each row and column. Suppose H is specified by the following sparse-access oracles OS
and OH defined in Eqs. (13) and (14). Suppose |Hi,j | ≤ 1 for i ∈ [m] and j ∈ [m]. Then for all
ϵ ∈ (0, 1), an (s, m + 3, ϵ)-block-encoding of H can be implemented using O(1) queries to OH and
OS , along with O(m + polylog(1/ϵ)) 1- and 2-qubit gates.
Remark 8. According to Appendix VI A, the condition that |Hi,j | ≤ 1 for i ∈ [m] and j ∈ [m] is
automatically satisfied due to the scaling in Eq. (18).
The QSVT builds a block-encoding of the following matrix function,
   i−1
pℓ (H) ·
h
Upℓ (H) = , pℓ (x) ≈ 1 + exp β̂x on [−1, 1], (21)
· ·

where β̂ is defined in Eq. (18). This is summarized as follows,

 11

Lemma 9. [33, Theorem 56] Let UH be a block encoding of H. Then there is a quantum circuit
†
Upℓ (H) which is a block encoding of f (H). The circuit involves ℓ application of UH and UH , one
application of controlled-UH gate, and O(ℓ) other one- and two-qubit gates.

In light of Lemma 5, at finite temperature, the complexity of the block encoding only has a
logarithmic dependence on ϵ.

C. Estimating the electron density

Recall that the electron density at different locations corresponds to the diagonals of f (H):


F (n)(j) = tr ρj f (H) . (22)

where,

ρj = |rj ⟩⟨rj | , j ∈ [NI ]. (23)

with rj being an interpolation point in D∆ .

The QSVT uses the polynomial approximation f (H) ≈ pℓ (H), and it provides an approximate
block encoding of pℓ (H). Therefore, we use the following estimator for the electron density,


F̂ (n)(j) = tr ρj pℓ (H) . (24)

In Eq. (24), we have treated pℓ (H) as observables. To estimate the expectation in Eq. (24),
we consider the techniques in Rall [74]. Rall’s approach involves the purification of the density
operators, the block encodings of the observables, and amplitude amplification [13]. Fortunately,
the density operators in Eq. (23) are pure states, and the observable f (H) is already block-encoded.

Lemma 10 ([74]). If a Hermitian matrix A with ∥A∥2 ≤ α can be block-encoded by Q elementary

gates and a density operator ρ can be purified as

ρ = trCk (|ρ⟩⟨ρ|), |ρ⟩ = U |0⟩|0⟩k (25)

with an unitary U implementable by R elementary gates, then for every ϵ, δ > 0 there exists an
algorithm that produces an estimate ξ of tr(ρA) such that

|ξ − tr(ρA)| ≤ ϵ, (26)

with probability 1 − δ. The gate complexity of the algorithm is O (R + Q) αϵ log 1

δ .

In light of Lemmas 5 and 10, we immediately have,

Theorem 11. For each j ∈ [NI ], there is a quantum algorithm that outputs an estimate F̂ (n)(j)
of F (n)(j), with ϵ accuracy, i.e.,

F̂ (n)(j) − F (n)(j) < ϵ,

s
log sϵ log 1


with probability 1 − δ. The algorithm uses O ϵ δ queries to OH .
 12

D. Hybrid algorithms and overall complexity

By far, we have built a procedure for estimating the electron density using QSVT and amplitude
amplification on quantum computers. To perform the self-consistent calculation of the DFT, we will
use the estimate of the electron density to interface with fixed-point iteration methods on classical
computers. Overall, this constitutes a hybrid algorithm for implementing the SCF iteration in the
DFT. An iteration on a classical computer produces a new electron density at the interpolation
points in D∆ . One then evaluates VH and Vxc and then interpolates them onto the fine grid in Dδ ,
as illustrated in Fig. 3. We make the following assumption,
Assumption 12. Given the electron density at NI interpolation points, the potential V in the
Hamiltonian matrix (esp. VH and Vxc ) can be evaluated with precision ϵ with cost O(NI ), excluding
logarithmic factors.
Let us elaborate on this assumption. First, a simple implementation of the interpolation proce-
dure is the multi-grid approach, which has been used in [59] to accelerate the DFT calculations.
In this case, the interpolation points correspond to a coarse grid. Second, the calculation of the
exchange-correlation potentials [70, 72] at the interpolation points is quite straightforward. Third,
the Poisson equation that leads to the Hartree potential can be solved with classical algorithms,
e.g., via Fast Fourier transform, which has complexity O(NI log NI ) [14]. It is also possible to solve
2/3
Poisson’s equation with quantum algorithms [19, 24, 52], in which case the complexity is O(sNI ).
Finally, as we will show in the next section, even without the interpolation step, i.e., NI = Ng , our
algorithm still has a cubic speedup over classical algorithms in terms of the number of electrons.
Therefore, the computational gain from the interpolation is only moderate, and it is meant for a
further reduction of the complexity.
Notice that since the major computational cost in classical algorithms comes from the compu-
tation of roughly Ne eigenvalues and eigenvectors, such an interpolation procedure will not signifi-
cantly improve the complexity there. In contrast, in the quantum algorithm, the complexity can be
mostly attributed to the computation of the expectations, in which case the interpolation provides
an important means to reduce the complexity.
To quantify convergence, we make a stability assumption.
∂F
Assumption 13. The Jacobian ∂n (n∗ ) has eigenvalues with real parts less than 1.
In this section, we consider two fixed-point methods and show the runtime analysis by establishing
the convergence theorems of those methods. The first method is known as the standard fixed-point
iteration with simple mixing, which we will call the full coordinate fixed-point method (FCFP), in the
sense that the method updates all components of F̂ (n) in Eq. (22). We will show that the iterations
converge linearly under suitable conditions. However, the cost for estimating all components of the
electron density scales linearly with respect to NI . As an alternative, we propose a method that
requires only some components of F̂ (n) to be updated at each iteration. We will call this method
the randomized block coordinate fixed-point method (RBCFP), which will be made more precise
later.
The convergence of fixed-point iterations usually requires a contraction property of the fixed-
point function. For generality, this contraction property is expressed in terms of a weighted vector
norm,
sX
2
∥x∥w = w(j)|x(j)| , (27)
j
 13

where w ̸= 0 is a vector with nonnegative entries that will be regarded as weights.

Definition 14. We say that a mapping F (n) is locally-contractive if there exist some weighted
vector norm ∥·∥w and some w-dependent c ∈ (0, 1) such that
∥F (n) − F (n′ )∥w ≤ c∥n − n′ ∥w , (28)
′
for all n, n ∈ Bγ (n∗ ) which denotes the ball centered at the fixed point n∗ with radius γ. Here the
norm ∥ · ∥w does not have to be the standard Euclidean norm.
For the DFT calculations, such property is connected to the structural stability of the underlying
physical system [16, 54]. Here we give a mathematical condition based on Assumption 13 that
ensures a contraction.
Lemma 15 (Theorem 3.3 [45]). Under Assumption 13, then there exists a γ > 0, a damping
parameter a ∈ [0, 1], an (a, ∂F
∂n (n∗ ))-dependent weighted norm ∥·∥w and a w-dependent c ∈ (0, 1)
such that the mapping F = (1 − a)n + aF (n) is contractive in the neighboring Bγ (n∗ ), that is,
e

Fe(n1 ) − Fe(n2 ) < c∥n1 − n2 ∥w , ∀n1 , n2 ∈ Bγ (n∗ ). (29)

w
In fact, the weighted norm ∥ · ∥w is induced by an inner product [45, Theorem 3.3]. This fact
implies that the weighted norm can be used in place of the standard Euclidean norm in convergence
analysis, as shown in Appendix VI, due to the equivalence property of norms in a finite-dimensional
Banach space [15]. In other words, as long as the contraction holds for one vector norm, the
convergence property is guaranteed in any other norm.

1. The full coordinate fixed-point method

In this section, we establish the convergence rate of the FCFP method in conjunction with the
simple mixing scheme [16, 54]. Algorithm 1 outlined the implementation of the FCFP method. In
addition, we present the overall query complexity of the hybrid algorithm equipped with the FCFP
method.
Recall that we denote F̂ (n) as the vector in RNI , whose component is defined by Eq. (22).
Theorem 16. Assume that there exists a > 0 and c ∈ (0, 1) under the assumption in Lemma 15.
For a given initial guess n0 ∈ Bγ (n∗ ), the FCFP iteration obtained from the simple mixing scheme,
nk+1 = (1 − a)nk + aF̂ (nk ), (30)
∥n0 −n∗ ∥2w
converges to the fixed-point linearly with probability at least 1 − γ2 ,

E[∥nk − n∗ ∥2w ] ≤ c2(k−1) ∥n0 − n∗ ∥2w . (31)

The proof of Theorem 16 can be found in Appendix VI B. Similar results regarding linear con-
vergence have been obtained in [16, 54, 83].
Theorem 17. The hybrid algorithm (Algorithm 1) can be implemented to obtain ∥nk − n∗ ∥w < ϵ
∥n −n ∥2
with probability at least 1 − δ − 0 γ 2 ∗ w with
 
sNI 1 1
O log log , (32)
ϵ ϵ δ
queries to OH .
 14

Algorithm 1: Full coordinate fixed-point iteration

Data: initial guess n0 , damping parameter a ∈ (0, 1)
Result: n∗
for k = 0 : T do
Estimate F̂ (nk ) using the QSVT and the amplitude amplification
nk+1 = (1 − a)nk + aF̂ (nk );
Update the Hamiltonian;
end

2. The randomized coordinate fixed-point method

In this section, we introduce an alternative to the FCFP method. Rather than updating all
components of F (n), we only update the components selectively. The key idea is similar to the
randomized coordinate iterative algorithms [63, 69, 86]. The new method will be termed the ran-
domized coordinate fixed-point method (RCFP). The basic steps are outlined in Algorithm 2.
Formally, we define the RCFP method
Definition 18. Given a fixed-point mapping F̂ (n), a randomized block coordinate fixed-point map-
ping (RBCFP) is defined as
X X
F̂R,m (n) = (uk , F̂ (n))uk + (uk , n)uk , (33)
k∈{kRj }m
j=1 k̸∈{kRj }m
j=1

where {kRj }m
j=1 is the set of m indices randomly sampled from the index set [NI ], uniformly without
replacement, and the parenthesis ( , ) refers to the standard inner product between vectors.
We remark that the method in Theorem 16 corresponds to the special case m = NI . The
following theorem shows that despite the partial update of the density, the method still has linear
convergence.
Theorem 19. Assume that there exist a > 0 and c ∈ (0, 1) as in Lemma 15. Let m ∈ {2, .., NI } be
given. For a given initial guess n0 ∈ Bγ (n∗ ), the RBCFP iteration obtained from the simple mixing
scheme,

nk+1 = (1 − a)nk + aF̂R,m (nk ) (34)

∥n0 −n∗ ∥2w
converges to the fixed-point linearly with probability at least 1 − γ2 ,
k
m(1 − c2 )

E[∥nk − n∗ ∥2w ] ≤ 1− ∥n0 − n∗ ∥2w . (35)
NI
The proof of Theorem 19 can be found in Appendix VI D.
Remark 20. It is worthwhile to highlight the differences between the FCFP and RBCFP methods.
First, the admissible range of the damping parameter in Theorem 19 can be different from that
of the damping parameter in Theorem 16. This is because the Jacobian of the mapping in the
RBCFP method (Eq. (33)) is different from that of the FCFP method (Eq. (30)). More precisely,
the RBCFP method can perform with a larger damping parameter without the loss of stability.
 15

Second, the convergence rate in Theorem 19 is proven for the worst-case scenario with the same
choice of the damping parameter in Theorem 16. In practice, we expect that the RBCFP has the
potential for faster convergence. This has been observed in our numerical results in Section IV.
Theorem 21. The hybrid algorithm (Algorithm 2) can be implemented ∥nk − n∗ ∥w < ϵ with
∥n −n ∥2
probability at least 1 − δ − 0 γ 2 ∗ w with
 
sNI 1 1
O log log , (36)
ϵ ϵ δ
queries to OH .

Algorithm 2: Randomized block coordinate fixed-point iteration

Data: initial guess n0 , damping parameter a ∈ (0, 1), index parameter m ∈ [NI ]
Result: n∗
for k = 0 : T do
Sample m different indices {kR,j }m
j=1 ⊂ [NI ] uniformly
Estimate F̂R,m [nk ] using the QSVT and the Amplitude Amplification.
nk+1 = (1 − a)nk + aF̂R,m [nk ];
Update the Hamiltonian;
end

3. Estimating the Chemical potential

Within the hybrid algorithms in Algorithms 1 and 2, we have so far focused on the case with
given chemical potential µ. If Ne is given instead, we can incorporate the constraint in Eq. (5) to
determine µ. At the continuous level, this implies that
Z
n(r) dr = Ne , n(r) = ⟨r|f (H − µI)|r⟩. (37)

The first equation can be cast into a nonlinear equation,

Z
G(n, µ) = 0, G(n, µ) := n(r) dr − Ne . (38)

Given n(r), G is a monotone function of µ.

To incorporate the constraint in Eq. (5) in our quantum algorithm, we update the chemical po-
tential on classical computers together with the update of the electron density, e.g., in Algorithm 2.
This extended algorithm consists of the following steps,
nk+1 = (1 − a)nk + aF̂R,m (nk ),
(39)
µk+1 = µk − η Ĝ(nk+1 ),
where
X
Ĝ(n) := |n(j)|δ 3 − Ne ,
j
 16

is a discretization of Eq. (38).

Here δ 3 is the infinitesimal volume from the finite-difference approach with grid size δ; η ∈ (0, 1)
is the damping parameter for updating the Fermi energy. This solver for µ is motivated by the
stochastic approximation method by Robbins and Monro for solving nonlinear equations [75].

IV. NUMERICAL RESULTS

A. Experiment details

To mimic our hybrid quantum algorithm on a classical computer, we conducted numerical tests
for the approximation of the density-matrix in Eq. (24) within the MATLAB platform M-SPARC,
a real-space density functional electronic structure code [32]. We chose Barium titanate (BaTiO3)
and a water molecule H2 O-sheet as our test models from the set of examples in M-SPARC 1 . In the
models, temperatures are set to T = 300K for BaTiO3 and T = 2320 for H2 O, respectively. The
BaTiO3 system is set up in a supercell in a cubic domain with periodic boundary conditions. The
H2 O system is treated with periodic boundary conditions in the x−y plane where the three atoms are
positioned and a Dirichlet boundary condition in the z direction. The local density approximation
(LDA) is used for exchange and correlation. We should point out that the M-SPARC code uses a
pseudopotential, which we did not consider in our quantum algorithm. Our emphasis, however, is to
use the corresponding Hamiltonian H to test the polynomial approximation of the density-matrix,
and more importantly, the convergence of the SCF iterations.
The initial electron density n0 in M-SPARC is given as a sum of isolated atom densities. We
perform the calculation of the ground state electron density with either a given chemical potential
µ or by fixing a number of electrons Ne . In all tests, the ground truth, i.e., n∗ is the converged
electron density obtained from the simple mixing scheme of SCF iteration, based on the Fermi-
Dirac smearing and direct eigenvalue computation in M-SPARC. In monitoring the convergence
of the SCF iterations, we measure the error between the true density n∗ and one obtained from
FCFP or RBCFP, i.e., nk together with the Chebyshev approximation method (see Algorithm 1
and Algorithm 2).

B. The efficiency of the RBCFP method

To first fully focus on the performance of the FCFP and RBCFP methods, we computed the
density matrix in Eq. (9) exactly as shown in Fig. 4. For each of the two physical systems, we run
the RBCFP with three different block sizes and then compare the convergence to that of the FCFP
method. The error is shown on a logarithmic scale in the figure. To compare the performance on
an equal footing, we rescaled the x axis to indicate the number of coordinate evaluations. The SCF
iterations were terminated when the relative error between the electron density and the true one is
below 10−6 as default in M-SPARC. There are several interesting aspects to note from the results
in Fig. 4. First, while it is well-known that the simple mixing scheme of the direct SCF calculations
leads to linear convergence [16, 54], the RBCFP method also exhibits linear convergence, which
supports our theoretical results Theorem 16 and Theorem 19. Second, for the convergence of both

1 https://github.com/SPARC-X/M-SPARC/tree/master/tests
 17

methods, it is important to select proper damping parameters. For example, in Table I, we checked
different damping parameters for the two systems and found the best damping parameters for the
FCFP method in terms of the number of iterations until convergence, where the optimal values
are found to be around 0.4 for both test cases. However, as shown in Fig. 4, it turns out that the
RBCFP method can perform well with much larger damping parameters that are very close to 1.
A similar observation has been made in the context of coordinate descent optimization methods in
machine learning [63, 66]. In addition, Fig. 4 shows that the RBCFP method can converge faster
than the FCFP method by an order of 2 (BaTiO3) and 1.5 (H2 O), which supports the different
convergence rates proven in Theorems 16 and 19. From the efficiency of the RBCFP method shown
in Fig. 4, we highlight that the practice of updating only a few coordinates randomly selected at
each iteration step can result in the estimation of only a few diagonal elements from quantum
computation in our hybrid algorithm, which amounts to a reduction of the overall complexity.

Damping parameter 0.3 0.33 0.35 0.37 0.38 Damping parameter 0.3 0.4 0.51 0.55 0.58
SCF iterations 39 40 76 482 diverge SCF iterations 39 28 30 112 diverge
(a) system BaTiO3 (b) system H2 O

TABLE I: The role of the damping parameter in the convergence of direct SCF iterations. The
table shows the number of SCF iterations for relative error 10−6 with the simple mixing scheme
applied to the exact SCF formulation in Eq. (10) for the given damping parameters for two
systems BaTiO3 (Left) and H2 O (Right).

FIG. 4: Comparison of the FCFP and RBCFP methods from Algorithms 1 and 2. In both panels,
the x-axis labels the number of coordinate evaluations. The y-axis labels the error of the electron
density on a logarithmic scale. Left: BaTiO3 system with Ne = 40 electrons fixed; Right H2 O
with Ne = 8 electrons fixed. In the left panel, the FCFP method runs with damping parameter
a = 0.3, but the RBCFP method with a = 0.95. In the right panel, the FCFP runs with a = 0.4,
but the RBCFP with a = 0.99.

Our next numerical experiment incorporates the Chebyshev polynomial approximation of the
density-matrix, which mimics the QSVT implementation of the density-matrix on a classical com-
puter. For the system BaTiO3, we applied the Chebyshev approximation method with degree
ℓ = 500 as Eq. (24) for implementing the RBCFP method. We used a fixed chemical potential
µ = 0.3403(eV) that is associated with the ground truth n∗ used in Fig. 4. In Fig. 5, we observe
 18

that the RBCFP methods still converge faster than the FCFP method in terms of coordinate eval-
uations to a given precision. Similarly in Fig. 6, we applied the polynomial approximation method
for system H2 O-sheet. One difference is that we used the variable chemical potential in Eq. (39) to
satisfy the constraint on Ne during the iteration. Still, we can clearly see that the RBCFP methods
converge to a given precision faster than the FCFP method in terms of the electron density. Fur-
thermore, it is observed that the chemical potentials obtained from the RBCFP method converge
faster than one from the FCFP method.
One interesting observation in Fig. 5 was that when we used the same damping parameter for the
FCFP method as in Fig. 4, it could not reach the given precision. We numerically found a = 0.24
as the nearly optimal value to reach the precision. However, the RBCFP implementations with the
same damping parameter still converge well. This might be attributed to the fact that the Jacobian
at n∗ is defined by the polynomial matrix function, rather than the Fermi-Dirac function, and the
upper bound of damping parameters for the FCFP is altered. For more rigorous results, we leave
this observation to future work.

FIG. 5: Comparison of the FCFP and RBCFP methods with block sizes m = 5, 25, 75 using the
Chebyshev approximation method in Eq. (24) for system BaTiO3. The degree of the method is
500. The x-axis labels the number of coordinate evaluations. The y-axis is the error of the
electron density. The FCFP method runs with damping parameter a = 0.24, but the RBCFP
method with a = 0.95.

V. SUMMARY AND DISCUSSIONS

We proposed an algorithm for the density-functional theory with complexity that scales linearly
with the dimension of the density update F (n) which is often much less than the number of electrons.
Therefore, this can be considered as linear/sublinear scaling, which compared to the cubic scaling
in classical algorithms, is a significant reduction.
The first natural question is whether the current algorithm can be improved to have a better
dependence on the dimension of the density update F (n). There are quantum algorithms that
offer quadratic speedup, e.g., the gradient estimation approach by Huggins et al. [41] for estimating
 19

FIG. 6: Comparison of the FCFP and RBCFP methods with block sizes m = 4, 12, 72 using the
Chebyshev approximation method in Eq. (24) for system H2 O-sheet. The degree of the method is
500. The x-axis labels the number of coordinate evaluations. Left: the y-axis is the error of the
electron density. Right: the values of the chemical potential µ during the iterations. The FCFP
method runs with damping parameter a = 0.4, but the RBCFP method with a = 0.99. The
damping parameter for the chemical potential in Eq. (39) is η = 0.1.

multiple observables. But our formulation in Eq. (24) is based on a single observable with multiple
density operators. It is not yet clear whether these algorithms can be applied.
As pointed out in Section III B, the degree of the polynomial in the approximation of the Fermi-
Dirac function increases considerably for lower temperature values. This is due to the fact that in
this regime the Fermi-Dirac function approaches a step function, which is discontinuous. In this
regime, ℓ must be proportional to β̂/ϵ, and the overall complexity increases significantly.
Another common practice in DFT calculations is to exclude core electrons and incorporate their
effects by using pseudopotentials. Although it is not clear whether this practice is needed in a
quantum algorithm, it is still of theoretical interest to explore how such potentials can be block
encoded into UH . These issues will be explored in separate works.

ACKNOWLEDGEMENT

XL’s research is supported by the National Science Foundation Grants DMS-2111221. CW ac-
knowledges support from National Science Foundation grant CCF-2238766 (CAREER). Both XL
and CW were supported by a seed grant from the Institute of Computational and Data Science
(ICDS) and the National Science Foundation Grants CCF-2312456.

VI. APPENDICES

A. Relation between the max norm and 2-norm of square matrix

We show that for any square matrix A ∈ CN ×N ,

∥A∥max ≤ ∥A∥2 , (40)

 20

where ∥A∥max := maxi,j |Aij | and ∥A∥2 is the 2-norm of A. Let A = U ΣV † be the singular value
decomposition of A. Denote xi = U † ei and yi = V † ei for each i ∈ [N ]. By the Cauchy-Schwarz
inequality, we observe that

N
|x†i Σyj |2
X
2
|Aij | = |eTi Aej |2 = =| σk (x̄i )k (yj )k |2
k=1
! ! (41)
X X
2 2
≤ σk |(xi )k | σk |(yj )k | ≤ ∥A∥22 ∥xi ∥22 ∥yj ∥22 = ∥A∥22 ,
k k

which proves the statement.

B. Proof of Theorem 16

We define the fixed-point mapping as follows

NI
X
F̂ (nt+1 ) = ⟨rj |Pξt,j pℓ (Ht )Pξt,j |rj ⟩uj . (42)
j=1

At iteration t, uj is the j-th standard basis vector, Pξt,j stands for the measurement projector of
Ht corresponding to index j and rj is the grid point. In other words, the quantity in Eq. (42) is
an unbiased estimate for the update of the electron density.
Now we consider the iteration defined as,

nt+1 = (1 − a)nt + aF̂ (nt ). (43)

Define the characteristic function It that values 1 if the first p-th iterates stay in Bγ (n∗ ) and
otherwise 0. We denote by Et := E[·||It = 1], the expectation conditioned on an event that It = 1.
Let et = nt − n∗ be the error between the current iterate and the fixed point.
The fixed-point iteration in Eq. (43) yields a recursive inequality as follows,

Et [∥et+1 ∥2w It+1 ] ≤ Et [∥et+1 ∥2w It ]

h  i
≤ (1 − a)2 ∥et ∥2w + 2a(1 − a)Et et , F̂ (nt ) − n∗ + a2 Et [∥F̂ (nt ) − n∗ ∥2w ]
w
(44)

The middle term of the right-hand side in Eq. (44) is simplified as

 
h  i NI
X
Et et , F̂ (nt ) − n∗ = et , ⟨rj |pℓ (Ht )|rj ⟩uj − n∗  . (45)
w
j=1
w
 21

To simplify this inequality, we first observe that for every j ∈ [M ],

h
2 i
Et [(uj , F̂ (nt ))2 ] = Et ⟨rj |Pξt,j pℓ (Ht )Pξt,j |rj ⟩
NI
X
= (⟨rj |Pξ pℓ (Ht )Pξ |rj ⟩)2 · ∥Pξ |rj ⟩∥22
ξ=1
NI
X
= pℓ (λt,ξ )2 · ∥Pξ |rj ⟩∥42
ξ=1 (46)
 2
XNI
≤ pℓ (λt,ξ ) · ∥Pξ |rj ⟩∥22  pℓ (x) ≈ f (x) > 0
ξ=1

2
= Et [⟨rj |Pξt,j pℓ (Ht )Pξt,j |rj ⟩]
 h i2
= Et (uj , F̂ (nt ))

By this result, the last term of the right hand side in Eq. (44) can be estimated as

NI
X  2 
Et [∥F̂ (nt ) − n∗ ∥2w = w(j)Et uj , F̂ (nt ) − n∗
j=1
NI
X h i
= w(j)Et (uj , F̂ (nt ))2 − 2(uj , F̂ (nt ))(uj , n∗ ) + (uj , n∗ )2
j=1
NI
X  h i2 
2
≤ w(j) Et uj , F̂ (nt ) − 2Et [(uj , F̂ (nt ))](uj , n∗ ) + (uj , n∗ ) (47)
j=1
NI
X  h i 2
= w(j) Et (uj , F̂ (nt )) − (uj , n∗ )
j=1
NI
X
=∥ ⟨rj |pℓ (Ht )|rj ⟩uj − n∗ ∥2w ,
j=1

where w(j)’s are defined in Eq. (27). Therefore, we can reduce Eq. (44) to

  2
XNI
Et [∥et+1 ∥2w It+1 ] ≤ (1 − a)et + a  ⟨rj |pℓ (Ht )|rj ⟩uj − n∗  = ∥F (nt ) − n∗ ∥2w ≤ c2 ∥et ∥2w ,
j=1
w
(48)
where c is defined in Definition 14. This proves Theorem 16.
 22

C. Convergence of the RBCFP when m = 1

We recall the RBCFP method in Eq. (33) with m = 1, namely,

X
F̂R (nt ) = (ukt , F̂ (nt ))ukt + (uk , nt )uk , (49)
k̸=kt

where kt denotes the index sampled at iteration i. Specifically, the estimated component is expressed
as

(ukt , F̂ (nt )) = ⟨rkt |Pξt pℓ (Ht )Pξt |rkt ⟩, (50)

which is the kt -th component of the full coordinate estimation in Eq. (42). Here ξt ∈ [NI ] denotes
the index corresponding to measurement.
Noticing the randomness of the RBCFP method from sampling index, we observe that
  
X
Et [F̂R (nt )] = Ekt Eξt (ukt , F̂ (nt ))ukt + (uk , nt )uk 
k̸=kt
 
X (51)
= Ekt ⟨rkt |pℓ (Ht )|rkt ⟩ukt + (uk , nt )uk  remove quantum noise
k̸=kt

NI − 1
= Ekt [⟨rkt |pℓ (Ht )|rkt ⟩ukt ] + nt .
NI

Similar to the mixing scheme in Eq. (43), we consider the following iteration,

nt+1 = (1 − a)nt + aF̂R (nt ). (52)

From this iteration, we have

Et [∥et+1 ∥2w It+1 ] ≤ Et [∥et+1 ∥2w It ]

h  i h i
= (1 − a)2 ∥et ∥2w + 2a(1 − a)Et et , F̂R (nt ) − n∗ + a2 Et ∥F̂R (nt ) − n∗ ∥2w .
w
(53)

By the observation (Eq. (51)), we first simplify the middle term of the right hand side in Eq. (53)
as follows
h  i  h i
Et et , F̂R (nt ) − n∗ = et , Et F̂R (nt ) − n∗
w w
NI − 1
= ∥et ∥2w + (et , Ekt [⟨rkt |pℓ (Ht )|rkt ⟩ukt − n∗ (kt )ukt ])w (54)
NI
NI − 1
= ∥et ∥2w + Ekt [w(kt )et (kt ) (⟨rkt |pℓ (Ht )|rkt ⟩ − n∗ (kt ))] .
NI
 23

We estimate the last term of the right hand side in Eq. (53) as follows,

h i
Et ∥F̂R (nt ) − n∗ ∥2w
 
 2 X
= Et w(kt ) F̂ (nt ) − n∗ , ukt + w(k)(et , uk )2 
k̸=kt
  2  N − 1
= Et w(kt ) F̂ (nt ) − n∗ , ukt +
I
∥et ∥2w (55)
NI
  h i 2  N − 1
I
≤ Ekt w(kt ) Eξt (ukt , F̂ (nt )) − (ukt , n∗ ) + ∥et ∥2w similar to Eq. (47)
NI
h i N −1
2 I
= Ekt w(kt ) (⟨rkt |pℓ (Ht )|rkt ⟩ − (ukt , n∗ )) + ∥et ∥2w remove quantum noise.
NI

Note that the first term ∥et ∥2w is expressed as

NI − 1
∥et ∥2w = ∥et ∥2w + Ekt [w(kt )e(kt )2 ]. (56)
NI

To put together all results, we can simplify Eq. (53) as

NI − 1 h
2
i
Et [∥et+1 ∥2w It+1 ] ≤ ∥et ∥2w + Ekt w(kt ) [(1 − a)et (kt ) + a(⟨rkt |pℓ (Ht )|rkt ⟩ − n∗ (kt ))]
NI
  2
NI
NI − 1 1 X
≤ ∥et ∥2w + (1 − a)et + a  ⟨rj |pℓ (Ht )|rj ⟩uj − n∗ 
NI N j=1
w
NI − 1 c2
≤ ∥et ∥2w + ∥et ∥2w
NI NI
1 − c2
 
≤ 1− ∥et ∥2w ,
NI
(57)

which completes the proof of the convergence of the RBCFP with m = 1.

D. Proof of Theorem 19

The key idea for proving Theorem 19 is not very different from the proof in Appendix VI C.
The main difference is that the error analysis (Eq. (53)) now involves the term F̂R,m for a given
m ∈ [NI ] in the middle and last terms.
 24

We first observe that

  
h i Xm X
Et F̂R,m (nt ) = E{kt,j }m
j=1
E{ξt,j }m  (ukt,j , F̂ (nt ))ukt,j +
j=1
(uk , nt )uk 
j=1 k̸∈{kt,j }m
j=1
 
Xm X
= E{kt,j }m  ⟨rkt,j |pℓ (Ht )|rkt,j ⟩ukt,j + (uk , nt )uk 
j=1
j=1 k̸∈{kt,j }m
j=1
  (58)
m NI NI −1



X m − m−1
= E{kt,j }m  ⟨rkt,j |pℓ (Ht )|rkt,j ⟩ukt,j  + nt
NI
j=1


j=1 m
 
m
 ⟨rkt,j |pℓ (Ht )|rkt,j ⟩ukt,j  + NI − m nt ,
X
= E{kt,j }m
j=1
j=1
NI

where the expectations E{kt,j }mj=1

and E{ξt,j }m
j=1
are performed with respect to the index sampling
and quantum noise, respectively. From this result, we achieve a slight modification of the recursive
inequality in Eq. (57) as follows
 
m
NI − m X
Et [∥et+1 ∥2w It+1 ] ≤ ∥et ∥2w + E{kt,j }m  w(kt,j )c2kt,j et (kt,j )2 
NI j=1
j=1

NI − m mc2 (59)
≤ ∥et ∥2w + ∥et ∥2w
NI NI
m(1 − c2 )
 
≤ 1− ∥et ∥2w ,
NI

where the first inequality can be verified as Eq. (55). This concludes the proof of Theorem 19.

E. Stability of FCFP and RCFP methods

We denote the probability filtration Fj = σ(nt |t ≤ j), which is defined due to the randomness of
quantum noise up to time j. Define a characteristic function as
(
j−1
1, if {nt }t=1 ⊂ Bγ (n∗ )
Ij = (60)
0, otherwise.

Let Xj be a stochastic process defined as

Xj = ∥ej ∥2w Ij (61)

We note that Ij is Fj−1 measurable and Xj is Fj -measurable. In the following analysis, we assume
that the initial guess n0 is given as a deterministic vector in Bγ (n∗ ) in Definition 14, i.e., where
the fixed-point function is contractive.
 25

By definition of Xj , we observe that

 
2
P {nj ̸∈ Bγ (n∗ ) for some j ∈ [J]|n0 } ≤ P sup Xj > γ |n0 . (62)
1≤j≤J

Define Ej−1 as the conditional expectation on the filtration Fj−1 given n0 , then

Ej−1 [Xj ] = E ∥ej ∥2w Ij = 1, n0 P {Ij = 1|n0 }

 
(63)

By definition of Ij and the technical result, we have

E ∥ej ∥2w Ij = 1, n0 ≤ c2 ∥ej−1 ∥2w .

 
(64)

From this, we obtain that

Ej−1 [Xj ] ≤ c2 ∥ej−1 ∥2w Ij−1 = c2 Xj−1 ≤ Xj−1 , (65)

which yields a supermartingale,

E[Xj |n0 ] ≤ E[Xj−1 |n0 ]. (66)

Finally, using Markov’s inequality, we arrive at

 
2 X0
P {nj ̸∈ Bγ (n∗ ) for some j ∈ [J]|n0 } ≤ P sup Xj > γ |n0 ≤ , (67)
1≤j≤J γ2

which proves the stability of FCFP.

Due to the similar property as in Eq. (59), a similar result can be obtained for the RBCFP
method as follows,
X0
P {nj ̸∈ Bγ (n∗ ) for some j ∈ [J]|n0 } ≤ . (68)
γ2

[1] Daniel S Abrams and Seth Lloyd. Simulation of many-body fermi systems on a universal quantum
computer. Physical Review Letters, 79(13):2586, 1997.
[2] Alán Aspuru-Guzik, Anthony D Dutoi, Peter J Love, and Martin Head-Gordon. Simulated quantum
computation of molecular energies. Science, 309(5741):1704–1707, 2005.
[3] Haim Avron and Sivan Toledo. Randomized algorithms for estimating the trace of an implicit symmetric
positive semi-definite matrix. Journal of the ACM (JACM), 58(2):1–34, 2011.
[4] Ryan Babbush, Peter J Love, and Alán Aspuru-Guzik. Adiabatic quantum simulation of quantum
chemistry. Scientific reports, 4(1):6603, 2014.
[5] Ryan Babbush, Nathan Wiebe, Jarrod McClean, James McClain, Hartmut Neven, and Garnet Kin-Lic
Chan. Low-depth quantum simulation of materials. Physical Review X, 8(1):011044, 2018.
[6] Roi Baer, Daniel Neuhauser, and Eran Rabani. Self-averaging stochastic kohn-sham density-functional
theory. Physical review letters, 111(10):106402, 2013.
[7] Thomas E Baker and David Poulin. Density functionals and kohn-sham potentials with minimal
wavefunction preparations on a quantum computer. Physical Review Research, 2(4):043238, 2020.
[8] Thomas L. Beck. Real-space mesh techniques in density-functional theory. Reviews of Modern Physics,
72(4):1041–1080, October 2000.
 26

[9] Constantine Bekas, Effrosini Kokiopoulou, and Yousef Saad. Computation of large invariant subspaces
using polynomial filtered lanczos iterations with applications in density functional theory. SIAM Jour-
nal on Matrix Analysis and Applications, 30(1):397–418, 2008.
[10] Costas Bekas, Effrosyni Kokiopoulou, and Yousef Saad. An estimator for the diagonal of a matrix.
Applied numerical mathematics, 57(11-12):1214–1229, 2007.
[11] Léon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine
learning. SIAM review, 60(2):223–311, 2018.
[12] DR Bowler and MJ Gillan. An efficient and robust technique for achieving self consistency in electronic
structure calculations. Chemical Physics Letters, 325(4):473–476, 2000.
[13] Gilles Brassard, Peter Hoyer, Michele Mosca, and Alain Tapp. Amplitude amplification and quantum
search algorithms. Journal of Quantum Information and Computation, 1(4):304–320, 2001.
[14] E Braverman, M Israeli, A Averbuch, and L Vozovoi. A fast 3d poisson solver of arbitrary order
accuracy. Journal of Computational Physics, 144(1):109–136, 1998.
[15] Alberto Bressan. Lecture notes on functional analysis. Graduate studies in mathematics, 143, 2012.
[16] Eric Cancès, Gaspard Kemlin, and Antoine Levitt. Convergence analysis of direct minimization and
self-consistent iterations. SIAM Journal on Matrix Analysis and Applications, 42(1):243–274, 2021.
[17] Shuai Chen, Zachary H Aitken, Subrahmanyam Pattamatta, Zhaoxuan Wu, Zhi Gen Yu, David J
Srolovitz, Peter K Liaw, and Yong-Wei Zhang. Simultaneously enhancing the ultimate strength and
ductility of high-entropy alloys via short-range ordering. Nature communications, 12(1):4953, 2021.
[18] Ziang Chen, Yingzhou Li, and Jianfeng Lu. On the global convergence of randomized coordinate
gradient descent for nonconvex optimization. SIAM Journal on Optimization, 33(2):713–738, 2023.
[19] Andrew M Childs, Jin-Peng Liu, and Aaron Ostrander. High-precision quantum algorithms for partial
differential equations. Quantum, 5:574, 2021.
[20] Yat Tin Chow, Tianyu Wu, and Wotao Yin. Cyclic coordinate-update algorithms for fixed-point
problems: Analysis and applications. SIAM Journal on Scientific Computing, 39(4):A1280–A1300,
2017.
[21] K. L. Chung. On a stochastic approximation method. The Annals of Mathematical Statistics, pages
463–483, 1954.
[22] Fabrizio Cleri and Vittorio Rosato. Tight-binding potentials for transition metals and alloys. Physical
Review B, 48(1):22, 1993.
[23] Patrick L Combettes and Jean-Christophe Pesquet. Stochastic quasi-fejér block-coordinate fixed point
iterations with random sweeping. SIAM Journal on Optimization, 25(2):1221–1248, 2015.
[24] Lingxia Cui, Zongmin Wu, and Hua Xiang. Quantum radial basis function method for the poisson
equation. Journal of Physics A: Mathematical and Theoretical, 56(22):225303, 2023.
[25] Yael Cytter, Eran Rabani, Daniel Neuhauser, and Roi Baer. Stochastic density functional theory at
finite temperatures. Physical Review B, 97(11):115207, 2018.
[26] Marcus Elstner, Th Frauenheim, E Kaxiras, G Seifert, and S Suhai. A self-consistent charge density-
functional based tight-binding scheme for large biomolecules. physica status solidi (b), 217(1):357–376,
2000.
[27] Frank Gaitan and Franco Nori. Density functional theory and quantum computation. Physical Review
B, 79(20):205117, 2009.
[28] Julian Gale. Siesta: A linear-scaling method for density functional calculations. In Computational
Methods for Large Systems-Electronic Structure Approaches for Biotechnology and Nanotechnology,
pages 45–75. Wiley & Sons Inc., 2011.
[29] C. J. Garcı́a-Cervera, J. Lu, and W. E. A sub-linear scaling algorithm for computing the electronic
structure of materials. Communications in Mathematical Sciences, 5(4):999–1026, 2007.
[30] Vikram Gavini, Kaushik Bhattacharya, and Michael Ortiz. Quasi-continuum orbital-free density-
functional theory: A route to multi-million atom non-periodic dft calculation. Journal of the Mechanics
and Physics of Solids, 55(4):697–718, 2007.
[31] Saeed Ghadimi and Guanghui Lan. Stochastic first-and zeroth-order methods for nonconvex stochastic
programming. SIAM Journal on Optimization, 23(4):2341–2368, 2013.
[32] Swarnava Ghosh and Phanish Suryanarayana. Sparc: Accurate and efficient finite-difference formu-
lation and parallel implementation of density functional theory: Isolated clusters. Computer Physics
 27

Communications, 212:189–204, 2017.

[33] András Gilyén, Yuan Su, Guang Hao Low, and Nathan Wiebe. Quantum singular value transformation
and beyond: exponential improvements for quantum matrix arithmetics. In Proceedings of the 51st
Annual ACM SIGACT Symposium on Theory of Computing, pages 193–204. ACM, 2019.
[34] S. Goedecker. Linear scaling electronic structure methods. Reviews of Modern Physics, 71(4):1085,
1999.
[35] Jürgen Hafner. Ab-initio simulations of materials using vasp: Density-functional theory and beyond.
Journal of computational chemistry, 29(13):2044–2078, 2008.
[36] Eric Hallman and Devon Troester. A multilevel approach to stochastic trace estimation. Linear Algebra
and its Applications, 638:125–149, 2022.
[37] Matthew B Hastings, Dave Wecker, Bela Bauer, and Matthias Troyer. Improving quantum algorithms
for quantum chemistry. arXiv preprint arXiv:1403.1539, 2014.
[38] Geoffroy Hautier, Christopher C Fischer, Anubhav Jain, Tim Mueller, and Gerbrand Ceder. Find-
ing nature’s missing ternary oxide compounds using machine learning and density functional theory.
Chemistry of Materials, 22(12):3762–3767, 2010.
[39] R Haydock, Volker Heine, and MJ Kelly. Electronic structure based on the local atomic environment
for tight-binding bands. Journal of Physics C: Solid State Physics, 5(20):2845, 1972.
[40] P. Hohenberg and W. Kohn. Inhomogeneous electron gas. Physical Review, 136(3B):B864, 1964.
[41] William J Huggins, Kianna Wan, Jarrod McClean, Thomas E O’Brien, Nathan Wiebe, and Ryan
Babbush. Nearly optimal quantum algorithm for estimating multiple expectation values. arXiv preprint
arXiv:2111.09283, 2021.
[42] Hideaki Iiduka. Stochastic fixed point optimization algorithm for classifier ensemble. IEEE Transac-
tions on Cybernetics, 50(10):4370–4380, 2019.
[43] Sai Praneeth Karimireddy, Anastasia Koloskova, Sebastian U Stich, and Martin Jaggi. Efficient greedy
coordinate descent for composite problems. In The 22nd International Conference on Artificial Intel-
ligence and Statistics, pages 2887–2896. PMLR, 2019.
[44] San-Huang Ke, Harold U Baranger, and Weitao Yang. Role of the exchange-correlation potential in
ab initio electron transport calculations. Journal of Chemical Physics, 126(20):201102–201102, 2007.
[45] Taehee Ko and Xiantao Li. Stochastic algorithms for self-consistent calculations of electronic structures.
Mathematics of Computation, 92(342):1693–1728, 2023.
[46] W. Kohn and L. J. Sham. Self-consistent equations including exchange and correlation effects. Physical
Review, 140(4A):A1133–A1138, 1965.
[47] Ruipeng Li, Yuanzhe Xi, Eugene Vecharynski, Chao Yang, and Yousef Saad. A thick-restart lanczos
algorithm with polynomial filtering for hermitian eigenvalue problems. SIAM Journal on Scientific
Computing, 38(4):A2512–A2534, 2016.
[48] Lin Lin. Randomized estimation of spectral densities of large matrices made accurate. Numerische
Mathematik, 136:183–213, 2017.
[49] Lin Lin and Jianfeng Lu. A mathematical introduction to electronic structure theory. SIAM, 2019.
[50] Lin Lin, Jianfeng Lu, and Lexing Ying. Numerical methods for kohn–sham density functional theory.
Acta Numerica, 28:405–539, 2019.
[51] Lin Lin, Yousef Saad, and Chao Yang. Approximating spectral densities of large matrices. SIAM
review, 58(1):34–65, 2016.
[52] Lin Lin and Yu Tong. Optimal polynomial based quantum eigenstate filtering with application to
solving quantum linear systems. Quantum, 4:361, 2020.
[53] Lin Lin and Yu Tong. Heisenberg-limited ground-state energy estimation for early fault-tolerant quan-
tum computers. PRX Quantum, 3(1):010318, 2022.
[54] Lin Lin and Chao Yang. Elliptic preconditioner for accelerating the self-consistent field iteration in
kohn–sham density functional theory. SIAM Journal on Scientific Computing, 35(5):S277–S298, 2013.
[55] Qihang Lin, Zhaosong Lu, and Lin Xiao. An accelerated proximal coordinate gradient method. Ad-
vances in Neural Information Processing Systems, 27, 2014.
[56] Kai-Hsin Liou, Chao Yang, and James R Chelikowsky. Scalable implementation of polynomial filtering
for density functional theory calculation in parsec. Computer Physics Communications, 254:107330,
2020.
 28

[57] Miguel AL Marques, Micael JT Oliveira, and Tobias Burnus. Libxc: A library of exchange and
correlation functionals for density functional theory. Computer physics communications, 183(10):2272–
2281, 2012.
[58] R. M. Martin. Electronic Structure: Basic Theory and Practical Methods. Cambridge University Press,
2011.
[59] Michael P Merrick, Karthik A Iyer, and Thomas L Beck. Multigrid method for electrostatic computa-
tions in numerical density functional theory. The Journal of Physical Chemistry, 99(33):12478–12482,
1995.
[60] Raphael A Meyer, Cameron Musco, Christopher Musco, and David P Woodruff. Hutch++: Optimal
stochastic trace estimation. In Symposium on Simplicity in Algorithms (SOSA), pages 142–155. SIAM,
2021.
[61] Phani Motamarri, Sambit Das, Shiva Rudraraju, Krishnendu Ghosh, Denis Davydov, and Vikram
Gavini. Dft-fe–a massively parallel adaptive finite-element code for large-scale density functional theory
calculations. Computer Physics Communications, 246:106853, 2020.
[62] Arkadi Nemirovski, Anatoli Juditsky, Guanghui Lan, and Alexander Shapiro. Robust stochastic ap-
proximation approach to stochastic programming. SIAM Journal on optimization, 19(4):1574–1609,
2009.
[63] Yu Nesterov. Efficiency of coordinate descent methods on huge-scale optimization problems. SIAM
Journal on Optimization, 22(2):341–362, 2012.
[64] Yurii Nesterov and Sebastian U Stich. Efficiency of the accelerated coordinate descent method on
structured optimization problems. SIAM Journal on Optimization, 27(1):110–123, 2017.
[65] Michael A Nielsen and Isaac L Chuang. Quantum Computation and Quantum Information. Cambridge
University Press, 2011.
[66] Julie Nutini, Mark Schmidt, Issam Laradji, Michael Friedlander, and Hoyt Koepke. Coordinate descent
converges faster with the gauss-southwell rule than random selection. In International Conference on
Machine Learning, pages 1632–1641. PMLR, 2015.
[67] Peter JJ O’Malley, Ryan Babbush, Ian D Kivlichan, Jonathan Romero, Jarrod R McClean, Rami
Barends, Julian Kelly, Pedram Roushan, Andrew Tranter, Nan Ding, et al. Scalable quantum simula-
tion of molecular energies. Physical Review X, 6(3):031007, 2016.
[68] R. G. Parr and W. Yang. Density-functional theory of atoms and molecules. Oxford University Press,
1995.
[69] Zhimin Peng, Yangyang Xu, Ming Yan, and Wotao Yin. Arock: an algorithmic framework for asyn-
chronous parallel coordinate updates. SIAM Journal on Scientific Computing, 38(5):A2851–A2879,
2016.
[70] John P Perdew, Kieron Burke, and Yue Wang. Generalized gradient approximation for the exchange-
correlation hole of a many-electron system. Physical review B, 54(23):16533, 1996.
[71] John P Perdew and Yue Wang. Accurate and simple analytic representation of the electron-gas corre-
lation energy. Physical review B, 45(23):13244, 1992.
[72] John P. Perdew and Alex Zunger. Local density-functional theory and its application to atoms and
molecules. Physical Review B, 23(10):5048–5079, 1981.
[73] David Persson, Alice Cortinovis, and Daniel Kressner. Improved variants of the hutch++ algorithm
for trace estimation. SIAM Journal on Matrix Analysis and Applications, 43(3):1162–1185, 2022.
[74] Patrick Rall. Quantum algorithms for estimating physical quantities using block encodings. Physical
Review A, 102(2):022408, 2020.
[75] H. Robbins and S. Monro. A stochastic approximation method. The Annals of Mathematical Statistics,
pages 400–407, 1951.
[76] Ankan Saha and Ambuj Tewari. On the nonasymptotic convergence of cyclic coordinate descent
methods. SIAM Journal on Optimization, 23(1):576–601, 2013.
[77] Grady Schofield, James R Chelikowsky, and Yousef Saad. A spectrum slicing method for the kohn–sham
problem. Computer Physics Communications, 183(3):497–505, 2012.
[78] Gotthard Seifert and Jan-Ole Joswig. Density-functional tight binding—an approximate density-
functional theory method. Wiley Interdisciplinary Reviews: Computational Molecular Science,
2(3):456–465, 2012.
 29

[79] Bruno Senjean, Saad Yalouz, and Matthieu Saubanère. Toward density functional theory on quantum
computers? SciPost Physics, 14(3):055, 2023.
[80] Abhiraj Sharma and Phanish Suryanarayana. On the calculation of the stress tensor in real-space
kohn-sham density functional theory. The Journal of chemical physics, 149(19):194104, 2018.
[81] J. M. Soler, E. Artacho, J. D. Gale, A. Garcı́a, J. Junquera, P. Ordejón, and D. Sánchez-Portal. The
SIESTA method for ab initio order-N materials simulation. Journal of Physics: Condensed Matter,
14(11):2745, 2002.
[82] Carmen M Tesch, Lukas Kurtz, and Regina de Vivie-Riedle. Applying optimal control theory for
elements of quantum computation in molecular systems. Chemical Physics Letters, 343(5-6):633–641,
2001.
[83] A. Toth, J. A. Ellis, T. Evans, S. Hamilton, C. T. Kelley, R. Pawlowski, and S. Slattery. Local
Improvement Results for Anderson Acceleration with Inaccurate Function Evaluations. SIAM Journal
on Scientific Computing, 39(5):S47–S65, January 2017.
[84] A. Toth and C. T. Kelley. Convergence Analysis for Anderson Acceleration. SIAM Journal on Numer-
ical Analysis, 53(2):805–819, January 2015.
[85] Lloyd N Trefethen. Approximation Theory and Approximation Practice, Extended Edition. SIAM,
2019.
[86] John N Tsitsiklis. Asynchronous stochastic approximation and q-learning. Machine learning, 16:185–
202, 1994.
[87] Eugene Vecharynski, Chao Yang, and John E Pask. A projected preconditioned conjugate gradient
algorithm for computing many extreme eigenpairs of a hermitian matrix. Journal of Computational
Physics, 290:73–89, 2015.
[88] Zaiwen Wen, Chao Yang, Xin Liu, and Yin Zhang. Trace-penalty minimization for large-scale
eigenspace computation. Journal of Scientific Computing, 66:1175–1203, 2016.
[89] J. Wolfowitz. On the stochastic approximation method of Robbins and Monro. The Annals of Mathe-
matical Statistics, 23(3):457–461, 1952.
[90] Stephen J Wright. Coordinate descent algorithms. Mathematical programming, 151(1):3–34, 2015.
[91] Chao Yang, Juan C Meza, Byounghak Lee, and Lin-Wang Wang. Kssolv—a matlab toolbox for solving
the kohn-sham equations. ACM Transactions on Mathematical Software (TOMS), 36(2):1–35, 2009.
[92] Hyobin Yoo, Rebecca Engelke, Stephen Carr, Shiang Fang, Kuan Zhang, Paul Cazeaux, Suk Hyun
Sung, Robert Hovden, Adam W Tsen, Takashi Taniguchi, et al. Atomic and electronic reconstruction
at the van der waals interface in twisted bilayer graphene. Nature materials, 18(5):448–453, 2019.
[93] Yunkai Zhou, Yousef Saad, Murilo L Tiago, and James R Chelikowsky. Self-consistent-field calculations
using chebyshev-filtered subspace iteration. Journal of Computational Physics, 219(1):172–184, 2006.