Block Encoding: Sparse Matrices

• Block encoding is a unitary embedding that normalizes a sparse matrix into A/α to enable quantum singular value transformation and simulation.
• It encompasses both oracle-based approaches and explicit circuit synthesis, balancing broad applicability with efficient resource usage.
• Recent work leverages periodic-diagonal and banded structures to achieve exact, low-ancilla block encodings with linear or near-linear gate complexity.

Block encoding is the standard unitary embedding used to expose a matrix $A$ to quantum singular value transformation, qubitization, Hamiltonian simulation, and quantum linear-system algorithms: one seeks a larger unitary $U$ whose top-left block is $A/\alpha$, so that polynomial or spectral transformations can be implemented indirectly on $A$. In the sparse-matrix setting, the subject divides into two main lines. One is the generic sparse-access model, which assumes location and value oracles and yields broad asymptotic theorems. The other is explicit circuit synthesis for structured sparse matrices, where one replaces black-box access by direct reversible arithmetic, permutations, shift operators, or specialized diagonal constructions. Recent work on sparse matrices with periodic diagonal structure is a representative example of the second line, showing how a narrow but practically important structured class can admit exact low-ancilla block encodings with poly$(n)$ or even linear gate complexity in the number of qubits (Camps et al., 2022, Zecchi et al., 11 Feb 2026).

1. Formal setting and normalization

For $A\in\mathbb{C}^{2^n\times 2^n}$, a unitary $U$ on $n+a$ qubits is an $(\alpha,a,\varepsilon)$-block encoding when the projected top-left block approximates $A/\alpha$; in standard notation,

$U$0

Exact variants set $U$1, and some constructions are presented directly as

$U$2

The normalization factor $U$3 is central because downstream costs in QSVT, Hamiltonian simulation, and linear-system routines scale with the encoded matrix $U$4, rather than with $U$5 itself (Camps et al., 2022).

In sparse-matrix algorithms, one therefore distinguishes two questions. The first is whether a block encoding exists under a given input model. The second is whether the resulting $U$6, ancilla count, and gate depth are favorable enough to make the embedding useful. This distinction is especially important because sparse-query theorems, explicit circuit constructions, and approximate compiler-style encodings optimize different resources. A plausible implication is that “block encoding of sparse matrices” is not a single method class, but an interface whose meaning depends strongly on the assumed access model.

2. Generic sparse-access block encoding

The standard background result for sparse matrices is oracle-based. If $U$7 is $U$8-row-sparse and $U$9-column-sparse, and one has row-location, column-location, and entry oracles,

$A/\alpha$0

together with an entry oracle for $A/\alpha$1, then one can implement an $A/\alpha$2-block-encoding of $A/\alpha$3. In the symmetric $A/\alpha$4-sparse case this gives normalization $A/\alpha$5. This is the usual baseline against which explicit sparse constructions are compared (Nguyen et al., 2022).

That framework is broad, but it is intentionally abstract. It measures oracle complexity rather than gate-level synthesis cost, and it does not exploit repeated values, shift structure, Fourier structure, or arithmetic regularity unless those are encoded into the oracles. Constructive work has therefore emphasized a complementary problem: given a structured sparse matrix with enough pattern or algebraic regularity, can one build the state preparation, permutations, reflections, or SELECT circuitry explicitly rather than postulating them as black boxes (Camps et al., 2022)?

Paradigm	Structural assumption	Representative property
Sparse-access oracle model	Row/column/value oracles	$A/\alpha$6-block-encoding (Nguyen et al., 2022)
Arithmetic repeated-value circuits	Arithmetic description, $A/\alpha$7 distinct values	Base $A/\alpha$8 (Sünderhauf et al., 2023)
Dictionary-based sparse encoding	Dictionary items $A/\alpha$9, injective $A$0	Depth $A$1, $A$2 (Yang et al., 2024)
Coherent-permutation synthesis	Arbitrary sparse index maps	Explicit gate-level MCX compression and amplitude reordering (Setty, 29 Aug 2025)
Periodic-diagonal structured sparse matrices	Cyclic shifts plus periodic diagonals	Exact one-ancilla sinusoidal diagonal block encoding (Zecchi et al., 11 Feb 2026)

3. Explicit circuit synthesis beyond the oracle model

A major research direction treats sparse matrices whose nonzero pattern and values admit direct reversible logic. One approach uses an arithmetic description of sparsity and repeated values. In that setting, matrix entries are grouped into $A$3 distinct nonzero values $A$4, with row and column sparsities $A$5, and explicit reversible oracles compute row or column positions from compact labels. The resulting base block encoding has subnormalization

$A$6

while a PREP/UNPREP variant gives

$A$7

For suitable Toeplitz, tridiagonal, and Laplacian families, this reduces flag-qubit count according to sparsity and data-loading cost according to repeated values (Sünderhauf et al., 2023).

A closely related low-depth formulation is dictionary-based sparse block encoding. Here the sparse matrix is represented by dictionary items $A$8, support sets $A$9, and injective maps $(n)$0. The resulting non-Hermitian sparse construction uses PREP, UNPREP, and a column oracle $(n)$1, and achieves subnormalization

$(n)$2

When compiled into the $(n)$3 gate set, the protocol has circuit depth $(n)$4 with $(n)$5 ancillary qubits for a $(n)$6 sparse matrix with $(n)$7 non-zero elements. The construction is phrased as an LCU-style PREP–SELECT–UNPREP architecture and is demonstrated on graph Laplacians and discrete differential operators (Yang et al., 2024).

For arbitrary sparse matrices, gate-level implementation remains difficult because the index mapping that moves amplitudes from a default diagonal placement into the true sparse pattern is costly. A recent framework addresses this by decomposing the index map into shifts, deletes, and inserts, then using coherent permutation operators to reorder amplitudes while preserving superposition. The central optimization problem is a linear assignment problem with cost matrix

$(n)$8

solvable in $(n)$9 by the Hungarian algorithm, and used to choose amplitude permutations that compress families of MCX gates and improve nearest-neighbor realizability (Setty, 29 Aug 2025).

Approximate compiler-style methods also exist for unstructured sparse matrices. “Sparse-FABLE” applies FABLE to the Hadamard-conjugated matrix $A\in\mathbb{C}^{2^n\times 2^n}$0 and then conjugates back, while “Lazy Sparse-FABLE” replaces the exact transformed-angle synthesis by

$A\in\mathbb{C}^{2^n\times 2^n}$1

For $A\in\mathbb{C}^{2^n\times 2^n}$2-sparse inputs, the reported empirical scaling is approximately $A\in\mathbb{C}^{2^n\times 2^n}$3 rotation gates and $A\in\mathbb{C}^{2^n\times 2^n}$4 CNOT gates for S-FABLE, with LS-FABLE trading some accuracy for $A\in\mathbb{C}^{2^n\times 2^n}$5 classical preprocessing (Kuklinski et al., 2024).

4. Periodic diagonal structure and banded sparse matrices

A distinct structured subclass consists of sparse matrices whose nonzero entries lie on a small number of diagonals and whose diagonal values vary periodically with the index. For $A\in\mathbb{C}^{2^n\times 2^n}$6, the canonical diagonal building blocks are

$A\in\mathbb{C}^{2^n\times 2^n}$7

together with cyclic shift permutations $A\in\mathbb{C}^{2^n\times 2^n}$8 and $A\in\mathbb{C}^{2^n\times 2^n}$9. In the basic banded case,

$U$0

More generally, the periodic diagonal may be a finite Fourier sum,

$U$1

and shifted periodic bands can be formed by attaching $U$2 or $U$3 to sinusoidal diagonal operators (Zecchi et al., 11 Feb 2026).

The core primitive is a one-ancilla exact block encoding of the cosine diagonal. With

$U$4

the unitary

$U$5

is an exact block encoding of $U$6 with one ancilla and subnormalization $U$7. The state transformation

$U$8

shows that interference with a Hadamard ancilla separates the real and imaginary parts of the diagonal phase unitary $U$9. Because

$n+a$0

the phase operator factorizes as

$n+a$1

so it is implemented by $n+a$2 single-qubit phase gates (Zecchi et al., 11 Feb 2026).

Sparse banded matrices then follow by LCU. For

$n+a$3

one prepares

$n+a$4

In the basic four-term decomposition of $n+a$5, the circuit uses two ancillas for the LCU index and one ancilla internal to $n+a$6, for a total of three ancillas. The resulting exact block form is

$n+a$7

The method is fully explicit for the structured class considered and does not rely on generic sparse-access oracles (Zecchi et al., 11 Feb 2026).

5. Complexity, performance, and applications

The complexity landscape depends strongly on which structural assumptions are used. For periodic diagonal matrices, the exact sinusoidal diagonal encoder has linear gate complexity: $n+a$8 with one implementation using $n+a$9 phase gates and $(\alpha,a,\varepsilon)$0 CNOTs, and an alternative implementation using $(\alpha,a,\varepsilon)$1 phase gates plus two ancilla-controlled $(\alpha,a,\varepsilon)$2-target NOT layers. For the sparse banded periodic construction

$(\alpha,a,\varepsilon)$3

the analyzed gate complexity is $(\alpha,a,\varepsilon)$4, dominated by SELECT and the shift circuits; the paper identifies this as the worst-case banded scenario. For $(\alpha,a,\varepsilon)$5 frequencies on a diagonal, the diagonal-only synthesis has complexity $(\alpha,a,\varepsilon)$6 (Zecchi et al., 11 Feb 2026).

These encodings are intended primarily as front ends for QSVT and related linear-algebra routines. The periodic-diagonal work discusses matrix inversion, exponential propagation, and PDE-derived matrices. For the periodic elliptic problem

$(\alpha,a,\varepsilon)$7

with periodic reaction coefficient $(\alpha,a,\varepsilon)$8, finite differences produce a cyclic tridiagonal matrix that fits the diagonal-plus-shifts form. The same framework also covers advection–diffusion–reaction discretizations with periodic reaction terms and, through shifted sine terms, periodically varying velocity fields (Zecchi et al., 11 Feb 2026).

Numerical studies support the structural advantage of specialized sparse encoders. For the diagonal matrix $(\alpha,a,\varepsilon)$9, the periodic-diagonal construction was compared with FABLE, with the reported result that the specialized circuit scales linearly in $A/\alpha$0, whereas the general-purpose dense block-encoding approach scales exponentially in $A/\alpha$1 for that task. In the elliptic example, reported relative Euclidean errors include $A/\alpha$2, $A/\alpha$3, $A/\alpha$4, and $A/\alpha$5, and in ADR experiments QSVT trajectories track exact classical evolution for sinusoidal, cosine, and multi-mode periodic coefficients (Zecchi et al., 11 Feb 2026).

In QSVT-based sparse matrix inversion, the subnormalization enters through the subnormalized condition number

$A/\alpha$6

so block-encoding choice directly affects the number of QSVT phase factors. A comparative evaluation of ARCSIN, FABLE, and PREPARE-SELECT on 30 sparse test matrices found that PREPARE-SELECT often has better normalization, while ARCSIN and FABLE have much lower classical preprocessing cost. The same study reports that error tolerances of $A/\alpha$7 are more than sufficient in most cases and in some cases $A/\alpha$8 is sufficient, and that the time taken to generate phase factors can impede scaling QSVT to larger test cases (Lapworth, 2024).

6. Scope, misconceptions, and adjacent directions

A persistent misconception is that every sparse-matrix block encoding is interchangeable with every other. The literature instead separates generic sparse-access theorems from narrow but cheaper structure-exploiting constructions. Explicit circuit-design papers repeatedly emphasize that they do not provide universal synthesis for arbitrary sparse matrices; rather, they target matrix families whose nonzero pattern and amplitudes can be generated by simple local rules, permutations, reflections, controlled rotations, cyclic shifts, or arithmetic descriptions (Camps et al., 2022).

The periodic-diagonal banded construction is a clear example of this distinction. It is not a generic arbitrary-sparse-matrix block encoder. Its assumptions include $A/\alpha$9, sparsity expressible using a small number of cyclic shifts, periodic diagonal or band coefficients representable by a small Fourier sum, and natural wrap-around boundary conditions. The method is therefore much narrower than the black-box sparse-access model, but potentially much cheaper when the structure is present. The same paper notes that extension to generic oracle-access models remains open (Zecchi et al., 11 Feb 2026).

Other explicit sparse approaches carry their own tradeoffs. Arithmetic repeated-value encodings reduce data loading to the number $U$00 of distinct values, but require reversible formulas for support sets and row/column maps (Sünderhauf et al., 2023). Dictionary-based schemes achieve depth $U$01 but use $U$02 ancillas (Yang et al., 2024). Coherent-permutation methods bridge theory and implementation for arbitrary sparse matrices, but their main results are structural and hardware-aware rather than closed-form asymptotic gate-count theorems (Setty, 29 Aug 2025). Approximate approaches such as S-FABLE and LS-FABLE are attractive for known unstructured sparse matrices, but they are compiler-style constructions rather than sparse-access-oracle results (Kuklinski et al., 2024).

Finally, sparse matrices should not be conflated with data-sparse dense matrices. Parallel work on hierarchical matrices and semiseparable matrices shows that block encoding also extends beyond entry sparsity to hierarchical compressibility and rank structure, but those settings are dense globally and rely on different access assumptions and factorizations. This suggests that “sparse” and “structured” are overlapping rather than identical categories in block-encoding research (Nguyen et al., 2022, Antonioli et al., 19 Mar 2026).