Linear Scaling Block-Encoding Architecture

• Linear scaling block-encoding architecture is a method that embeds non-unitary matrices into unitary operators, achieving near-optimal resource usage for quantum circuits.
• It leverages techniques such as Gray-code ordering, combinatorial optimization, and the LCU framework to minimize gate counts and ancilla qubits in sparse and structured matrices.
• These architectures underpin advanced quantum algorithms—including QSVT, adiabatic solvers, and ladder operator schemes—ensuring efficient simulation and improved scaling in practical applications.

A linear scaling block-encoding architecture refers to a class of constructions for embedding generally non-unitary operators (matrices) into unitaries in such a way that the resources—particularly quantum gate count, circuit depth, or classical pre-processing—scale linearly (or nearly linearly) with the size or sparsity of the matrix or system. These architectures have enabled a variety of quantum algorithms, from linear system solvers and Hamiltonian simulation to error correction and coding, to reach practical performance levels or optimal asymptotic scaling.

1. Definition and General Principles

Let $A \in \mathbb{C}^{N \times N}$ be a matrix with specified structure (sparse, banded, displacement-structured, MPO, etc.). An $(\alpha, a, \epsilon)$-block-encoding is a unitary $U \in \mathbb{C}^{2^{n+a} \times 2^{n+a}}$ such that

$$U = \begin{pmatrix} A/\alpha & * \ * & * \end{pmatrix}$$

when the $a$-qubit ancilla are in $|0\rangle^{\otimes a}$, and $\|A/\alpha\| \leq 1$, $\|A - \alpha \tilde A\| < \epsilon$ for $\tilde A = \langle 0^a| U |0^a\rangle$ (Kuklinski et al., 2024). The quantum circuit complexity (number of one- and two-qubit gates, circuit depth, ancillae) and classical overhead compose the fundamental tradeoffs. Linear scaling block-encodings specifically achieve resource overheads $O(N)$, $O(s N)$, or $O(L)$, depending on the context.

2. Sparse and Structured Matrix Block-Encoding

For general $s$-sparse matrices, linear scaling block-encoding circuits exploit circuit-level optimizations, combinatorial assignment, and permutation networks to minimize the overhead of multi-controlled gates and amplitude reordering (Setty, 29 Aug 2025). The method decomposes the input matrix into diagonal offsets or data items, compresses control logic using combinatorial optimization (Hungarian assignment), and performs amplitude reordering through coherent permutation operators. The final step uses compressed multi-controlled-X (MCX) and rotation gates to realize the required matrix rows and columns, achieving total gate count $O(s \log N)$, total depth $O(s + \log N)$, and qubit count $O(\log N + \log s)$. Empirical results on tridiagonal and structured sparse matrices confirm these scalings.

For unstructured sparse matrices, the S-FABLE and LS-FABLE schemes leverage Gray-code ordering and controlled $R_y$ rotation pruning to compress the number of elementary gates required (Kuklinski et al., 2024). S-FABLE applies a Gray-permuted Hadamard conjugation, obtaining $O(N)$ significant rotations and $O(N \log N)$ CNOTs for $N \times N$ matrices with $O(N)$ nonzero entries. LS-FABLE circumvents the $O(N^2 \log N)$ classical preprocessing by direct, thresholded mapping from sparse matrix elements to gate angles, with small controlled error. Both methods demonstrate empirical linear scaling in gate complexity and error bounds as a function of system size.

3. Block-Encoding for Matrix Product Operator and Structured Hamiltonians

In tensor network contexts, especially matrix product operators (MPOs), each local MPO tensor—of bond dimension $\chi$—is block-encoded as the upper-left block of a $(D+2)$-qubit unitary, with $D = \lceil\log_2 \chi\rceil$ (Nibbi et al., 2023). The complete Hamiltonian block-encoding is synthesized by composing $O(L)$ such unitaries for a system of length $L$. The total gate count is $O(L \chi^2)$ and the ancilla qubit count is $L+D$. This one-site-at-a-time compositionality ensures scaling linear in $L$ for fixed $\chi$, outperforming standard LCU/Pauli expansion approaches which generally incur $O(L^2)$ or worse.

In displacement-structured settings, such as circulant, Toeplitz, or Hankel systems, block-encodings are constructed by decomposing the matrix into a linear combination of displacement matrices, each efficiently implementable as a quantum shift or permutation (Wan et al., 2019). The overall resource scaling is $O(n)$, where $n$ is the matrix size, under a QRAM data-access model; quadratic speedup over classical scaling is maintained in a black-box model.

4. Explicit Linear Block-Encoding for Canonical Operators

For banded matrices with periodic diagonal or Toeplitz structure, explicit block-encodings using the LCU framework leverage the periodicity to efficiently extract the real and imaginary parts of phase operators (Zecchi et al., 11 Feb 2026). A single-ancilla circuit suffices for diagonal cases, with $O(n)$ single-qubit phase gates and $O(n)$ two-qubit gates, while a general banded case with LCU requires $O(\text{poly}(n))$ gates but still offers a polynomial scaling compared to the exponential cost of generic dense-matrix block-encodings. Applications to quantum singular value transformation (QSVT)-based solvers for differential equations demonstrate optimal dependence on system parameters.

For local finite-difference discretizations of canonical operators (e.g., the Laplacian), efficient explicit block-encoding circuits have been constructed that utilize only a fixed number of ancillas (2 for 1D, $2+\lceil\log_2 D\rceil$ for $D$-dimensional grids), and achieve Clifford+$T$ gate and depth $O(D \, n)$ for $n$-qubit data registers per spatial dimension (Sturm et al., 2 Sep 2025). The normalization factor is exactly the spectral norm for power-of-two $D$, with success probabilities matching best theoretical values. No black-box oracles are required.

5. Quantum Algorithmic Applications and Impact

Linear scaling block-encoding architectures are critical for state-of-the-art quantum algorithms in linear system solving, simulation, and quantum signal processing:

• Quantum Discrete Adiabatic Linear Solver (BEES-QDALS): A paradigm combining block-encoding of adiabatic Hamiltonians, an eigenvalue separator (repeated squaring to isolate the ground state), and a first-order approximation to Hamiltonian evolution, achieving $O(\kappa)$ complexity versus $O(\kappa^2)$ in prior approaches for condition number $\kappa$ (Wu et al., 2024). Direct circuit-level block-encoding eliminates costly Hamiltonian simulation, with measured fidelity advantages.
• Quantum Singular Value Transformation (QSVT)/QSP: Block-encoding is the bottleneck subroutine for amplitude amplification, matrix inversion, and time evolution. Efficient block-encodings enable optimal (in input parameters) execution of polynomial transformations, and several linear-scaling strategies are now available for physical Hamiltonians of high relevance (Kane et al., 2024).
• Ladder Operator Block-Encoding (LOBE): Direct block-encodings of second-quantized ladder operators, with normalization matching spectral norms and resource usage $\mathcal{O}(N_f + N_b \log M)$ for $N_f$ fermions, $N_b$ bosons (cutoff $M$) (Simon et al., 14 Mar 2025). Dramatic improvements over Pauli-basis expansion approaches are observed, especially as locality increases.

6. Algorithmic and Resource Scaling Summary

The various architectures can be summarized by the following dominant scaling behaviors:

Context	Gate Count	Ancilla Qubits	Comments
Generic Sparse	$O(s \log N)$	$O(\log N + \log s)$	$s$-sparse, arbitrary pattern (Setty, 29 Aug 2025)
S-FABLE	$O(N)$ rotations	$O(\log N)$	Dense→sparse via Hadamard (Kuklinski et al., 2024)
MPO	$O(L \chi^2)$	$L + D$	$\chi$ = bond dim, $L$ = sites (Nibbi et al., 2023)
Periodic Diag	$O(n)$	$1$	Band-diagonal, cyclic shift (Zecchi et al., 11 Feb 2026)
Canonical Diff	$O(D n)$	$2 + \lceil \log_2 D \rceil$	Stencil/Laplacian, $n$ per axis (Sturm et al., 2 Sep 2025)
Ladder Op.	$O(N_f+N_b\log M)$	$O(1)$	Fermion/boson, $k$-local terms (Simon et al., 14 Mar 2025)

This scaling directly determines the feasibility of implementing QSVT, adiabatic solvers, Hamiltonian simulation, and advanced error-correcting codes at physically relevant problem sizes.

7. Concluding Remarks and Applications

The design of linear scaling block-encoding architectures has fundamentally redefined the practicality of several foundational quantum algorithms. By leveraging matrix structure—sparsity, periodicity, low displacement rank, tensor network locality, or algebraic symmetries—these methods achieve optimal or near-optimal scaling in circuit resources. Demonstrated applications span quantum linear system solving with favorable $\kappa$ dependence (Wu et al., 2024), sparse or structured Hamiltonian simulation, PDE solution via QSVT, and efficient implementation of error-correcting codes with minimal per-symbol overhead (Matsuzono et al., 2014).

Contemporary architectures frequently combine novel circuit-layer decompositions (Gray-code, coherent permutation), explicit algebraic manipulations, and hierarchical composition. Continued refinement and benchmarking, especially at the interface of algorithm design and hardware constraints, are expected to drive further advances in both theory and realization of quantum computational tasks at scale.