Wall-Chebyshev Projector

• Wall-Chebyshev projector is a polynomial spectral projector that approximates an idealized wall function to isolate the ground state by filtering out excited-state components.
• It is constructed using Chebyshev polynomial expansions on a rescaled Hamiltonian spectrum and applied in deterministic configuration-interaction, stochastic quantum Monte Carlo, and quantum algorithms.
• The method offers optimal convergence properties with improved computational efficiency, making it a valuable tool in large-scale quantum simulations and electronic structure calculations.

The wall-Chebyshev projector is a polynomial spectral projector obtained by approximating an idealized “wall” function with Chebyshev polynomials after an affine rescaling of the Hamiltonian spectrum. In the many-body formulations summarized in the literature, it is the infinite-imaginary-time limit of the exponential projector and is designed to retain the ground-state component while suppressing all excited-state components. The construction appears in deterministic configuration-interaction projector methods, in stochastic projector-based quantum Monte Carlo, and in quantum algorithms for ground-state preparation; closely related Chebyshev polynomial projectors also arise in interval filtering for singular-value problems and in occupied-subspace projection for Kohn–Sham density functional theory (Zhang et al., 2016, Zhao et al., 2024, Filip et al., 1 Aug 2025).

1. Definition and spectral normalization

In imaginary-time propagation, one starts from

$\Psi(\beta)=e^{-\beta(\hat H-S)}\Psi(0),$

and, with a shift chosen such that $E_0\le S<E_1$, the limit

$\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$

annihilates all components with $E>E_0$ while leaving the ground-state component intact. For any trial state with nonzero overlap onto the true ground state, $\hat W\Psi(0)\propto\Psi_0$ (Zhao et al., 2024).

The projector is expressed in polynomial form after mapping the Hamiltonian spectrum to $[-1,1]$. One common rescaling is

$$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$

while a closely related form used in quantum algorithms is

$\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$

Under these maps, the ground-state eigenvalue is placed at the endpoint or discontinuity of the wall function, and the remaining spectrum is placed in the complementary region (Zhao et al., 2024, Filip et al., 1 Aug 2025).

The wall function itself is written in slightly different but equivalent normalized forms, depending on the spectral map. In one convention,

$$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$

and in another,

$$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$

This suggests that the defining feature is not the particular location of the discontinuity, but the use of an endpoint-localized step-like function that singles out the ground-state eigenvalue after rescaling (Zhao et al., 2024, Filip et al., 1 Aug 2025).

2. Chebyshev-series construction

The construction employs Chebyshev polynomials of the first kind,

$E_0\le S<E_1$0

or equivalently $E_0\le S<E_1$1. Because these polynomials are orthogonal on $E_0\le S<E_1$2 with weight $E_0\le S<E_1$3, they furnish near-minimax polynomial approximations for analytic functions on that interval (Zhang et al., 2016).

In the deterministic projector formulation, the wall-Chebyshev generator is derived as the $E_0\le S<E_1$4 limit of an $E_0\le S<E_1$5th-degree Chebyshev approximation to the exponential propagator. With

$E_0\le S<E_1$6

the finite-$E_0\le S<E_1$7 exponential-Chebyshev approximation is

$E_0\le S<E_1$8

and using $E_0\le S<E_1$9, one obtains the wall limit

$\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$0

In the normalized endpoint formulation, the same projector appears as

$\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$1

The coefficients are therefore explicit and require no numerical fitting (Zhang et al., 2016, Zhao et al., 2024).

The literature assigns two distinct kinds of performance statements to this construction. In projector analysis, the wall-Chebyshev generator is described as optimal in the sense of maximizing the asymptotic convergence factor

$\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$2

for a given polynomial degree $\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$3 (Zhang et al., 2016). In stochastic QMC analysis, the uniform approximation error at the first excitation energy is reported to scale asymptotically as $\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$4, so achieving $\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$5 requires $\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$6 (Zhao et al., 2024). These statements concern different quantities: one addresses asymptotic ground-state filtering, the other pointwise approximation of the wall function.

3. Deterministic projector configuration interaction

A deterministic many-body realization appears in projector configuration interaction (PCI), which combines projection onto the ground state with a path-filtering truncation scheme and is formulated as a deterministic version of full configuration interaction quantum Monte Carlo (FCIQMC). In this setting, one works in a determinant basis $\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$7 and propagates

$\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$8

For a linearized projector step, the spawning amplitudes are

$\hat W=\lim_{\beta\to\infty}e^{-\beta(\hat H-S)}$9

and path filtering imposes a single spawning-threshold $E>E_0$0 through

$E>E_0$1

Diagonal amplitudes are kept exact, coefficients are updated from the retained amplitudes, and the state is renormalized (Zhang et al., 2016).

Within this framework, the wall-Chebyshev projector replaces the linearized imaginary-time propagator by the infinite-step optimal Chebyshev generator. The path-filtering threshold is the sole truncation parameter, and the resulting error in energy is reported to scale roughly linearly in $E>E_0$2; values $E>E_0$3 in Hartree units are stated to be sufficient for chemical accuracy, and extrapolation $E>E_0$4 may be carried out by fitting energies to a low-order power law in $E>E_0$5 (Zhang et al., 2016).

The deterministic PCI study reports benchmark calculations on $E>E_0$6 at equilibrium and stretched geometries, where chemical accuracy is achieved with wave functions containing less than $E>E_0$7 of the full CI space. It also reports computations on the ground state of $E>E_0$8 with up to quaduple-$E>E_0$9 basis sets and wave functions as large as 200 million determinants, allowing direct comparison with FCIQMC and density matrix renormalization group (DMRG). The wall-Chebyshev generator is noted to require only two vectors of storage and no three-vector recurrence, which the authors identify as advantageous for large-scale deterministic CI implementations (Zhang et al., 2016).

4. Stochastic projector-based quantum Monte Carlo

A stochastic implementation of an approximate wall projector has been introduced for determinant-based QMC, specifically in initiator FCIQMC and multi-reference coupled-cluster Monte Carlo (MR-CCMC). In this setting, the projector is realized through the same truncated Chebyshev expansion,

$\hat W\Psi(0)\propto\Psi_0$0

with coefficients chosen so that $\hat W\Psi(0)\propto\Psi_0$1 over $\hat W\Psi(0)\propto\Psi_0$2. Applying the three-term recurrence in operator form,

$\hat W\Psi(0)\propto\Psi_0$3

one obtains the projected state

$\hat W\Psi(0)\propto\Psi_0$4

Each recurrence requires a single application of $\hat W\Psi(0)\propto\Psi_0$5, so an $\hat W\Psi(0)\propto\Psi_0$6th-order projection costs $\hat W\Psi(0)\propto\Psi_0$7 Hamiltonian applications (Zhao et al., 2024).

The stochastic algorithm is also expressed as a product of linear factors,

$\hat W\Psi(0)\propto\Psi_0$8

with Chebyshev nodes

$\hat W\Psi(0)\propto\Psi_0$9

and effective time steps $[-1,1]$0. In MR-CCMC, newly spawned determinants are screened by a metric-tree (BK-tree) search to determine whether they lie within the prescribed excitation level of the reference space, and symmetry screening may be applied at reference selection (Zhao et al., 2024).

The convergence analysis states that the asymptotic convergence factor is

$[-1,1]$1

with $[-1,1]$2 growing linearly with $[-1,1]$3. Relative to the largest stable linear propagation step $[-1,1]$4, the $[-1,1]$5th-order wall-Chebyshev projector is reported to provide a theoretical speed-up of $[-1,1]$6 in the number of projector steps. The same study states that, in practice, the Chebyshev weights avoid blooms, and observed speed-ups are larger; for $[-1,1]$7, the wall-Chebyshev projector often reduces the required Hamiltonian applications by an order of magnitude or more to reach a given statistical error (Zhao et al., 2024).

The reported benchmarks are correspondingly sharp. For $[-1,1]$8 with a $[-1,1]$9 CAS and MR-CCMCSD, a linear projector with $$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$0 required $$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$1 Hamiltonian applications to equilibrate to $$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$2 walkers, whereas the second-order Chebyshev projector ($$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$3) reached the same walker population with only 66 Hamiltonian applications. For $$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$4, the first-order linear projector required $$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$5 iterations to stabilize the shift and energy, while the fifth-order wall-Chebyshev projector ($$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$6) converged in $$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$7 iterations and reduced wall-clock time from $$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$8 h on 12 cores to $$\tilde H=\frac{2\hat H-(E_{N-1}+E_0)}{E_{N-1}-E_0},$$9 h on 6 cores for the same statistical error bar. The non-parallelity errors in the $\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$0 binding curve, relative to DMRG, are reported to drop by a factor of $\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$1 under wall-Chebyshev propagation (Zhao et al., 2024).

5. Quantum algorithm for ground-state preparation

A quantum-algorithmic realization uses the same wall-Chebyshev series as a ground-state projector that is efficiently implemented as a product of Hamiltonian operators. In this formulation,

$\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$2

with

$\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$3

When $\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$4 is not known exactly, one instead starts from an estimate $\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$5, obtains an upper spectral estimate $\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$6 from a Gershgorin circle bound,

$\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$7

chooses

$\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$8

and works with

$\tilde H = 2\,(H-E_0)/R -1,\qquad R=E_{\max}-E_0.$9

This guarantees that the true spectrum lies in $$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$0 (Filip et al., 1 Aug 2025).

A key structural identity is the factorization

$$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$1

with nodes

$$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$2

This leads directly to an implementation by linear combinations of unitaries (LCU): each factor is block-encoded, ancillas are prepared and uncomputed around a SELECT operation for the Pauli terms of $$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$3, and post-selection realizes $$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$4. The total number of Hamiltonian-oracle calls is $$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$5, and if $$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$6, the total gate count is stated as $$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$7 with $$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$8 ancilla overhead (Filip et al., 1 Aug 2025).

The error analysis is expressed in terms of the convergence factor

$$\mathrm{wall}(x)= \begin{cases} \infty,&x<0,\ 1,&x=0,\ 0,&x>0, \end{cases}$$9

which yields suppression of the first excited component according to

$$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$0

The resulting asymptotic scaling is

$$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$1

for query complexity, and

$$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$2

for gate complexity. With amplitude amplification, the query count acquires an additional factor $$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$3, where the success probability is

$$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$4

The analysis further states that in practice $$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$5 (Filip et al., 1 Aug 2025).

The reported numerical benchmarks emphasize robustness to inaccurate ground-state-energy estimates. For hydrogen chains in STO-3G, the polynomial order $$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$6 needed for 1 mHartree accuracy is tabulated for known and unknown $$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$7, and the wall-Chebyshev projector maintains moderate orders in both regimes. For the two-site Hubbard model at $$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$8, when $$W(x)= \begin{cases} \infty,&x\le -1,\ 1,&x=-1,\ 0,&x>-1. \end{cases}$$9 is known the wall-Chebyshev and eigenstate-filter projectors converge in $E_0\le S<E_1$00–20 while step-function and ITE projectors need $E_0\le S<E_1$01–100; if $E_0\le S<E_1$02 is unknown and set to Hartree–Fock, only wall-Chebyshev and ITE still converge. The same study states that eigenstate-filter and step-function projectors both rely critically on knowing $E_0\le S<E_1$03 to within $E_0\le S<E_1$04, whereas wall-Chebyshev remains the most slowly damped mode even when the shift is inaccurate (Filip et al., 1 Aug 2025).

6. Relation to other Chebyshev spectral projectors

The wall-Chebyshev projector belongs to a broader class of Chebyshev spectral filters, but the target subspace and approximation objective vary substantially across applications. In FEAST-style singular-value computation, the relevant object is not a ground-state wall projector but a step-function projector onto an interval $E_0\le S<E_1$05. There, the step function

$E_0\le S<E_1$06

is approximated by a Chebyshev–Jackson expansion

$E_0\le S<E_1$07

with Jackson damping factors $E_0\le S<E_1$08 used to suppress Gibbs oscillations. The resulting approximate spectral projector is symmetric positive semi-definite with eigenvalues in $E_0\le S<E_1$09, admits pointwise error bounds of order $E_0\le S<E_1$10, and is used in a FEAST SVDsolver without shifted linear systems; convergence is governed by the spectral gap of the approximate projector, with subspace convergence rate $E_0\le S<E_1$11 (Jia et al., 2022).

In large-scale Kohn–Sham density functional theory, the relevant object is the occupied-subspace projector rather than a ground-state-only projector. The two-level Chebyshev filter based complementary subspace method (CS2CF) uses an outer Chebyshev filter on the full Hamiltonian to compute a basis $E_0\le S<E_1$12 for the occupied subspace and an inner Chebyshev filter on the projected Hamiltonian $E_0\le S<E_1$13 to compute only the $E_0\le S<E_1$14 fractionally occupied states. The density projector in the projected space is written as

$E_0\le S<E_1$15

so that the full density matrix is

$E_0\le S<E_1$16

Because this avoids full diagonalization of $E_0\le S<E_1$17, the dominant $E_0\le S<E_1$18 bottleneck is reduced to

$E_0\le S<E_1$19

For an 8,000-atom bulk silicon system with 32,000 electrons on 34,560 cores, the reported timings are approximately 9 s for the outer ALB update, 2 s for the Hamiltonian-vector update, and 40 s for the CS2CF subspace solve, for a total SCF time of approximately 51 s; a full ELPA diagonalization of the $E_0\le S<E_1$20 DG matrix is reported to take approximately 650 s per SCF on the same cores. The same implementation yields 1.0 ps of ab initio molecular dynamics in approximately 28 h of wall time, with an average SCF step wall time of 51 s (Banerjee et al., 2017).

These related methods clarify a common misconception. “Chebyshev projector” does not denote a single algorithmic object. In the wall-Chebyshev setting, the polynomial approximates the infinite-imaginary-time wall that isolates the ground state; in Chebyshev–Jackson FEAST, it approximates an interval indicator; in CS2CF, it constructs occupied-subspace and complementary-subspace projectors for electronic structure. The common element is the use of explicitly controlled polynomial filtering on a spectrally rescaled operator, but the spectral target, convergence criterion, and implementation pathway differ materially across these settings (Jia et al., 2022, Banerjee et al., 2017).