Pith. sign in

REVIEW 3 major objections 4 minor 28 references

A two-layer quantum circuit encodes any matrix chain product in depth independent of chain length.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 05:56 UTC pith:QJEYRFOI

load-bearing objection The K-independent depth claim is real but only for the raw state-encoding unitary; the paper discloses the signal-weight caveat in Section 4.3, yet the abstract and Table 1 still overstate it. the 3 major comments →

arxiv 2607.13191 v1 pith:QJEYRFOI submitted 2026-07-14 quant-ph

Two-Tower Quantum Matrix Chain Multiplication: Trading Qubits for Depth

Giacomo Antonioli , Anna Bernasconi , Alessandro Berti , Gianna M. Del Corso , Alessandro Poggiali This is my paper
classification quant-ph MSC 81P6868Q12 PACS 03.67.Lx
keywords quantum matrix multiplicationmatrix chain productstate preparationQRAMcircuit depthamplitude encodingquantum subroutinesmatrix-chain-vector
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that multiplying a chain of K arbitrary real matrices can be encoded as amplitudes of a quantum state in circuit depth polylogarithmic in the matrices' dimensions and independent of K, using O(sum log P_k) qubits, provided each matrix can be loaded efficiently from quantum-accessible memory. If true, this would give a building block that turns matrix-chain multiplication into a shallow state-preparation problem, helping downstream algorithms that only need aggregate quantities like norms, traces, or inner products. The chief caveat is the signal weight: for well-conditioned matrices the amplitude carried by the useful subspace decays roughly as N^{-(K-1)}, so extracting the product with constant probability costs repeated amplitude amplification that reintroduces K dependence.

Core claim

Theorem 5 states that for any chain M^(0)...M^(K-1), there is a unitary U_W = OP(R) OP(L) that maps |0...0> to the Frobenius-normalized chain product on the outer registers, with every shared index contracted to |0>, plus an orthogonal residual. The construction interleaves state-preparation operators: the even-indexed matrices are loaded in parallel in one layer, the odd-indexed ones are contracted in parallel in the second, using the adjoint state-preparation operator to convert shared-register basis states into matrix-entry amplitudes. For even chains, the last matrix is loaded with a modified operator that keeps the output column register alive. The depth is O(max_k polylog(P_k P_{k+1}))

What carries the argument

The contraction mechanism: for a matrix M, the adjoint state-preparation operator SP(M)^dagger acts on a basis state |i,j> as SP(M)^dagger |i,j> = (m_ij/||M||_F) |0,0> + |perp>, projecting onto the contracted state with amplitude proportional to the matrix entry. This single identity turns a superposition over shared indices into the inner-product sums that define the chain product, and it is what lets all K matrices be handled in two parallel layers. A second operator, MSP, is the even-chain variant that contracts only the shared register while preserving the output column register.

Load-bearing premise

The headline constant-depth result holds for the raw quantum state; if the goal is to read out the product or feed it to an algorithm that needs the contracted subspace with constant probability, the exponentially small signal weight for well-conditioned chains forces amplitude amplification, reintroducing dependence on K.

What would settle it

Simulate the K=3 or K=4 circuit for random well-conditioned matrices and compare the amplitudes on the subspace |p,0,...,0,q> with W_pq/(prod ||M^(k)||_F): a mismatch at machine precision would falsify correctness; measuring the two-qubit-gate depth as K grows would falsify the constant-depth claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Setting the last matrix to a column vector gives a matrix-chain-vector routine that encodes M^(0)...M^(K-2) v in polylog depth with no output column register.
  • The raw output can feed amplitude-estimation-based routines for norms, traces, variances, graph-matrix powers, linear-system solvers, and kernel evaluations without tomography.
  • Total gate count remains Theta(K polylog(N)); the gain is depth, not total work.
  • The signal weight Omega_W = ||W||_F^2 / prod ||M^(k)||_F^2 controls how much of the state is useful; amplitude amplification boosts it at cost O(1/sqrt Omega_W).
  • For normal matrices, Omega_W for A^K approaches (sigma_1^2 / ||A||_F^2)^{K-1}, so peaked spectra keep the signal strong while flat spectra dilute it.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If a downstream algorithm can consume the raw state without post-selecting the contracted subspace, the K-independent depth is a genuine end-to-end advantage; otherwise the amplification overhead restores a K dependence that should be counted in comparisons with classical depth.
  • A statevector-to-block-encoding conversion, which the paper leaves open, would let intermediate Two-Tower outputs be multiplied further by QSVT-based methods, potentially giving a hybrid that controls both depth and signal decay.
  • The same two-layer contraction idea could be applied to products of unitaries or to tensor-network contractions, where contractions over multiple bond indices might be parallelized in the same way.
  • The classical preprocessing to build the quantum-accessible tree costs O(PR) per matrix, so the asymptotic depth saving assumes the data tree is already built; a fair end-to-end comparison should include that one-time cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper presents Two-Tower Matrix Multiplication, a quantum subroutine that prepares a quantum state whose amplitudes encode the product W = M^(0)...M^(K-1) of K distinct matrices. The construction is a generalization of the authors' earlier two-matrix subroutine [15]: even-indexed matrices are loaded with state-preparation operators SP(M^(2i)) in one parallel layer, odd-indexed matrices are contracted with adjoint operators SP(M^(2j+1))^† in a second parallel layer, and for even K the last matrix is handled by an MSP operator. The authors prove correctness for odd and even K (Theorem 5), derive depth O(max_k polylog(P_k P_{k+1})) independent of K, qubit count O(sum_k log P_k), total gate count Theta(K polylog), and give a chain-vector corollary plus Qiskit/QCLAB implementations. They also explicitly define a signal weight Omega_W = ||W||_F^2 / prod ||M^(k)||_F^2 and note that amplitude amplification costs O(1/sqrt(Omega_W)).

Significance. The contraction mechanism (Eq. 3) and the odd/even correctness proofs are clean, and the depth/qubit counts follow directly from the two-layer parallel structure. If the result is understood as a raw state-preparation primitive in the QRAM model, it is a genuine and nontrivial improvement over block-encoding-based chain multiplication, whose depth typically grows with K. The availability of open-source implementations and the explicit proof for arbitrary K are strengths. However, the advertised 'constant-depth' claim does not automatically extend to any end-to-end computation: the useful component of the output has squared weight Omega_W, and for well-conditioned N x N factors Omega_W ~ N^{-(K-1)}, so post-selection or amplitude amplification introduces a substantial K-dependent overhead. The resource comparison in Table 1 also omits the classical QRAM storage needed for the input matrices. The significance is therefore conditional: the formal state-preparation result is sound, but its practical reach is narrower than the abstract suggests.

major comments (3)
  1. [Abstract; Table 1; Section 4.3] The headline 'depth independent of K' is stated for the raw output state of U_W, but the paper itself shows in Section 4.3 that the useful subspace has squared weight Omega_W = ||W||_F^2 / prod ||M^(k)||_F^2, and that boosting Omega_W to a constant costs O(1/sqrt(Omega_W)) amplitude-amplification rounds. For well-conditioned N x N factors Omega_W ~ N^{-(K-1)}, so that overhead grows as Omega(N^{(K-1)/2}). Thus any end-to-end procedure that needs the chain product W, or that uses the product encoding in a downstream algorithm, has a K-dependent cost that is omitted from the abstract and from Table 1's 'O(polylog(N))' execution depth. The claim should be explicitly qualified as applying to raw state preparation, and the amplification cost should be included in the resource accounting or clearly stated as a separate stage.
  2. [Algorithm 1; Figures 2-3 with Definitions 2 and 4] There is a systematic reversal of register order. Definition 2 fixes SP(M) to map |0>^p|0>^r to sum m_{i,j}|i>^p|j>^r, i.e. first register = row index, second = column. Definition 4 accordingly places SP(M^(2i)) on (l_i,r_i) and SP(M^(2j+1))^† on (r_j,l_{j+1}). However Algorithm 1 line 2 says 'apply SP(M^(k)) on registers (q_{k+1}, q_k)' for even k, and the Figure 3 caption describes OP(L) acting on pairs (r_i,l_i) and OP(R) on (l_{j+1},r_j). Read literally, these reversed pairs load M^T rather than M; tracking the K=3 case gives W^T (or a reversed product order), not W as claimed in Theorem 5. Please standardize the register ordering in the pseudocode and circuit captions to match Definitions 2-4, or explicitly explain a different convention and prove the output state under that convention.
  3. [Table 1; Section 2.2] The memory comparison is incomplete. The 'Memory/qubits' row lists classical O(K N^2) words versus Theta(K log N) qubits, but the proposed algorithm operates in the QRAM model, where all K N^2 matrix entries are stored in a classical quantum-accessible data structure; the QRAM also requires O(K N^2) classical preprocessing. The qubit count includes only circuit registers, as a footnote says, but the table does not state the accompanying QRAM storage size. As presented, the row overstates the memory compression of the full quantum system. Please report circuit-register qubits and QRAM storage separately, or clearly state that the classical data structure is excluded from the resource comparison.
minor comments (4)
  1. [Section 3.1, Figure 2] The caption says MSP(B) acts on registers (s,n) and SP(A) on (n,m). For A in R^{M x N} and B in R^{N x S}, Definition 2/3 imply SP(A) should act on (m,n) and MSP(B) on (n,s). The caption appears to swap the roles of n and s, which is confusing given the register-order issue noted above.
  2. [Section 1] The statement that classical unbounded-parallelism matrix chain multiplication has depth O(log K log N) is plausible under the arithmetic-circuit model with constant-time scalar operations, but it should be accompanied by a citation or a short justification; matrix multiplication itself is commonly quoted as O(log^2 N) depth in NC.
  3. [Section 4.3] The transition 'for large K is dominated by the largest singular value: Omega_W ~ (sigma_1^2/||A||_F^2)^{K-1}' is only accurate in the limit where the second singular value is small; for a generic normal matrix with sigma_1>sigma_2, Omega_W tends to 1, not to a decaying exponential. Please state the precise asymptotic regime or replace the displayed formula with the more careful ratio.
  4. [Throughout] The paper uses both 'P_k' and 'p_k = log P_k' and at times writes O(polylog(P_k P_{k+1})) and at other times O(polylog(N)) for square chains. This is fine, but a short notation reminder near Table 1 would help.

Circularity Check

0 steps flagged

No significant circularity: Theorem 5 is proved directly from the contraction identity Eq. (3); the main self-citation [15] supplies the base-case circuit but is not load-bearing.

full rationale

The central derivation is self-contained. Theorem 5's correctness is proven by direct calculation: OP(L) prepares all even-indexed matrices and OP(R) contracts each shared register via SP(M)^†|i,j> = (m_ij/||M||_F)|0,0> + |⊥> (Eq. 3), which follows immediately from Definition 2 and unitarity. The chain-product sums emerge by linearity from this identity, not from any fitted parameter or prior result assumed as equivalent to the target. The MSP operator used for even K is not postulated as an independent oracle; it is defined as V(M)^†U(M) and its action is derived from SP. The only substantive self-citation is [15], which supplies the original two-matrix circuit that the paper rearranges into OP(L)/OP(R); for K=2 the paper states the rearrangement explicitly, and for general K the proof in Section 4.1 does not cite [15] as a theorem. Thus the self-citation is not load-bearing. Section 4.3's signal-weight analysis is an honest limitation: Ω_W = ||W||_F^2 / ∏||M^(k)||_F^2 may decay as N^{-(K-1)}, and boosting it costs O(1/√Ω_W), reintroducing K-dependence for end-to-end use; this qualifies the raw-state depth claim but is not circular. Section 6 likewise flags the missing statevector-to-block-encoding conversion as an open problem rather than silently assuming it. Within the stated QRAM model, the polylog depth per SP is an external assumption, not a consequence of the target result.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

The central result is parameter-free. It rests on the QRAM model, power-of-two dimensions, and standard row-loading unitaries. No target quantity is fitted, and no new physical entities such as particles, forces, or dimensions are introduced; MSP is a defined operator built from SP.

axioms (4)
  • domain assumption QRAM-based state preparation can load a P x R matrix into a (p+r)-qubit state in depth O(polylog(PR)) using Theta(log(PR)) qubits.
    Section 2.2, Definition 2; all depth and qubit bounds are stated within this model.
  • domain assumption All matrix dimensions are powers of two and all entries are real.
    Definition 1; required for register allocation and for the real-amplitude contraction identity Eq. (3).
  • standard math For any real matrix M, there exists a unitary SP(M) with first column equal to the Frobenius-normalized vectorized entries, and a unitary MSP(M) = V(M)^dagger U(M) with the stated action on |i>|0>.
    Section 2.2, Definitions 2 and 3; the row-loading construction follows [26] and the existence of unitary completions is standard linear algebra.
  • standard math The vectorization identity vec(AXB) = (B^T tensor A) vec(X) (Horn and Johnson, Lemma 4.3.1).
    Section 3.2.3; used as motivation for the construction, not as a step in the correctness proof.

pith-pipeline@v1.3.0-alltime-deepseek · 16032 in / 22359 out tokens · 210157 ms · 2026-08-02T05:56:15.011093+00:00 · methodology

0 comments
read the original abstract

Matrix chain multiplication -- computing $\mathcal{W} = M^{(0)}\cdots M^{(K-1)}$ where $M^{(k)} \in \mathbb{R}^{P_k \times P_{k+1}}$ -- arises in scientific computing, machine learning, and graph analysis. Despite the importance of this problem, for chains of distinct matrices, the classical number of operations grows linearly with the chain length $K$ and polynomially in the matrix dimensions. We present \emph{Two-Tower Matrix Multiplication}, a quantum subroutine that encodes the product $\mathcal{W}$ of the $K$ matrices into a quantum state in circuit depth $\mathcal{O}(\max_{k} \mathrm{polylog} (P_k P_{k+1}))$, which is independent of~$K$ within the QRAM-based state-preparation model, whereas the qubit count is $\mathcal{O}\bigl(\sum_{k} \log P_k \bigr)$; the total gate count remains linear in $K$, so the gain is in the circuit depth. The construction interleaves state-preparation operators across two layers; within each layer, all operators act on disjoint registers and execute in parallel. This subroutine can be specialized for the chain-vector case, which computes the product of $K-1$ matrices applied to a vector. We prove the correctness of the subroutine for all $K$ and provide two implementations using the Qiskit and QCLAB frameworks. The subroutine is applicable to any downstream quantum algorithm that operates on a matrix encoded in the statevector, including norm estimation, graph-matrix powers, linear system solving, and quantum machine learning kernels.

Figures

Figures reproduced from arXiv: 2607.13191 by Alessandro Berti, Alessandro Poggiali, Anna Bernasconi, Giacomo Antonioli, Gianna M. Del Corso.

Figure 1
Figure 1. Figure 1: Unitary matrices corresponding to (a) SP(M), (b) SP(M) † , and (c) MSP(M) shown schematically for 2 × 2 matrix M. In SP(M), the first column (blue) contains the vectorised, Frobenius-normalised entries of M; the remaining columns (grey, marked ∗) are fixed by unitarity and play no role in the algorithm. In SP(M) † , the first row (red) contains the normalised entries of M; the remaining rows are irrelevant… view at source ↗
Figure 2
Figure 2. Figure 2: Two equivalent circuits for computing C = AB, with A ∈ R M×N, B ∈ R N×S , m = log M, n = log N, s = log S . (a) Original circuit from Bernasconi et al. [15], using four state-preparation operators (V(A), U(A) † , V(B T ), U(B T )) and performing two register swaps. (b) The same circuit reorganised into the Two-Tower two-layer structure OP(L) and OP(R): layer OP(L) applies SP(A) = U(A) V(A) on registers (n,… view at source ↗
Figure 3
Figure 3. Figure 3: Two-Tower circuit for a chain of K matrices. (a) K odd: registers are ordered top-to-bottom as (r(K−1)/2, l(K−1)/2, . . . , r0, l0). OP(L): all SP(M(2i) ) act simultaneously on disjoint register pairs (ri , li). OP(R): all SP(M(2 j+1)) † act simultaneously on the interleaved pairs (lj+1,rj). (b) K even: an additional output register l K 2 appears at the top. The last matrix M(K−1) uses the MSP state prepar… view at source ↗
Figure 4
Figure 4. Figure 4: Two-Tower circuit for D = ABC with A ∈ R M×N, B ∈ R N×S , C ∈ R S×T , and m = log M, n = log N, s = log S , t = log T. OP(L): SP(C) on registers (t, s) and SP(A) on registers (n, m) execute in parallel. OP(R): SP(B) † on registers (s, n) contracts the shared summation index. The output state |ψ2⟩ encodes D in registers (t, m) (outer boundaries) with the internal registers (s, n) contracted to |0⟩. than as … view at source ↗
Figure 5
Figure 5. Figure 5: Chain-vector circuit for w = ABCv with A ∈ R M×N, B ∈ R N×S , C ∈ R S×T , v ∈ R T×1 . Since PK = 1, the output register vanishes and the MSP decomposition of the last factor reduces to SP(v) † on the t-qubit boundary register alone. The product w = ABCv ∈ R M is encoded in the m-qubit register, with all shared registers (t, s, n) contracted to |0⟩. which for large K is dominated by the largest singular val… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 5 linked inside Pith

  1. [1]

    G. H. Golub, C. F. Van Loan, Matrix computations, 3rd Edition, Johns Hopkins University Press, Baltimore, MD, 1996

    1996

  2. [2]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, S. Lloyd, Quantum machine learning, Nature 549 (7671) (2017) 195–202

    2017

  3. [3]

    Janzing, P

    D. Janzing, P. Wocjan, BQP-complete problems concern- ing mixing properties of classical random walks on sparse graphs, arXiv preprint quant-ph/0610235 (2006)

    Pith/arXiv arXiv 2006

  4. [4]

    N. A. Nghiem, T.-C. Wei, Quantum algorithm for esti- mating largest eigenvalues, Physics Letters A 488 (2023) 129138

    2023

  5. [5]

    D. Bini, M. Capovani, F. Romani, G. Lotti, o(n2.7799) complexity for n×n approximate matrix multiplication, Information Processing Letters 8 (1979) 234–235. doi: 10.1016/0020-0190(79)90113-3

    doi:10.1016/0020-0190(79)90113-3 1979

  6. [6]

    Alman, V

    J. Alman, V . V . Williams, A refined laser method and faster matrix multiplication, in: Proceedings of the 2021 ACM-SIAM Symposium on Discrete Algorithms (SODA), SIAM, 2021, pp. 522–539

    2021

  7. [7]

    R. Duan, H. Wu, R. Zhou, Faster matrix multiplication via asymmetric hashing, in: Proceedings of the 64th Annual IEEE Symposium on Foundations of Computer Science (FOCS), IEEE, 2023, pp. 2129–2138

    2023

  8. [8]

    A. W. Harrow, A. Hassidim, S. Lloyd, Quantum algorithm for linear systems of equations, Physical Review Letters 103 (15) (2009) 150502

    2009

  9. [9]

    Wossnig, Z

    L. Wossnig, Z. Zhao, A. Prakash, Quantum linear sys- tem algorithm for dense matrices, Physical Review Letters 120 (5) (2018) 050502. doi:10.1103/PhysRevLett. 120.050502

    doi:10.1103/physrevlett 2018

  10. [10]

    Chakraborty, A

    S. Chakraborty, A. Gilyén, S. Jeffery, The Power of Block- Encoded Matrix Powers: Improved Regression Tech- niques via Faster Hamiltonian Simulation, in: 46th In- ternational Colloquium on Automata, Languages, and Pro- gramming (ICALP 2019), V ol. 132 of Leibniz International Proceedings in Informatics (LIPIcs), Schloss Dagstuhl – Leibniz-Zentrum für Info...

    doi:10.4230/lipics.icalp.2019.33 2019

  11. [11]

    D. Fang, L. Lin, Y . Tong, Time-marching based quantum solvers for time-dependent linear dif- ferential equations, Quantum 7 (2023) 955. doi:10.22331/q-2023-03-20-955. URL http://dx.doi.org/10.22331/ q-2023-03-20-955

    doi:10.22331/q-2023-03-20-955 2023

  12. [12]

    Li, P.-L

    X. Li, P.-L. Zheng, C. Pan, F. Wang, C. Cui, X. Lu, Faster quantum subroutine for matrix chain multiplication via Chebyshev approximation, Scientific Reports 15 (1) (2025) 28559

    2025

  13. [13]

    Gilyén, Y

    A. Gilyén, Y . Su, G. H. Low, N. Wiebe, Quantum singular value transformation and beyond: Exponential improve- ments for quantum matrix arithmetics, in: Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing (STOC 2019), ACM, 2019, pp. 193–204. doi:10.1145/3313276.3316366

    arXiv 2019

  14. [14]

    Montanaro, C

    A. Montanaro, C. Shao, Quantum and classical query com- plexities of functions of matrices, in: Proceedings of the 56th Annual ACM Symposium on Theory of Computing, 2024, pp. 573–584

    2024

  15. [15]

    Bernasconi, A

    A. Bernasconi, A. Berti, G. M. Del Corso, A. Poggiali, Quantum subroutine for efficient matrix multiplication, IEEE Access 12 (2024) 116274–116284. doi:10.1109/ ACCESS.2024.3446176

    arXiv 2024

  16. [16]

    Giovannetti, S

    V . Giovannetti, S. Lloyd, L. Maccone, Quantum random access memory, Physical Review Letters 100 (16) (2008) 160501

    2008

  17. [17]

    R. A. Horn, C. R. Johnson, Matrix analysis, Cambridge university press, 2012

    2012

  18. [18]

    Antonioli, A

    G. Antonioli, A. Berti, A. Poggiali, A. Bernasconi, G. M. Del Corso, Outlier detection and other applications of quan- tum matrix multiplication, in: 2025 IEEE International Par- allel and Distributed Processing Symposium Workshops (IPDPSW), IEEE, 2025, pp. 509–518

    2025

  19. [19]

    Bernasconi, A

    A. Bernasconi, A. Berti, G. M. Del Corso, R. Guidotti, A. Poggiali, Quantum subroutine for variance estimation: algorithmic design and applications, Quantum Machine Intelligence 6 (2) (2024) 78

    2024

  20. [20]

    Poggiali, A

    A. Poggiali, A. Bernasconi, A. Berti, G. M. Del Corso, R. Guidotti, et al., Quantum feature selection with variance estimation, in: ESANN, 2023

    2023

  21. [21]

    Poggiali, J

    A. Poggiali, J. Ju, A more efficient quantum circuit for estimating the variance, Quantum Machine Intelligence 8 (1) (2026) 34

    2026

  22. [22]

    M. A. Nielsen, I. L. Chuang, Quantum Computation and Quantum Information, Cambridge University Press, 2010

    2010

  23. [23]

    Javadi-Abhari, M

    A. Javadi-Abhari, M. Treinish, K. Krsulich, C. J. Wood, J. Lishman, J. Gacon, S. Martiel, P. D. Nation, L. S. Bishop, A. W. Cross, B. R. Johnson, J. M. Gambetta, Quantum 10 computing with Qiskit (2024). arXiv:2405.08810, doi: 10.48550/arXiv.2405.08810

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.08810 2024

  24. [24]

    S. Keip, D. Camps, R. Van Beeumen, QCLAB: A mat- lab toolbox for quantum computing, in: 2025 IEEE Inter- national Parallel and Distributed Processing Symposium Workshops (IPDPSW), IEEE, 2025, pp. 1175–1181

    2025

  25. [25]

    Berti, F

    A. Berti, F. Ghisoni, Efficient quantum state prepa- ration with bucket brigade QRAM, arXiv preprint arXiv:2510.16149 (2025)

    Pith/arXiv arXiv 2025

  26. [26]

    Kerenidis, A

    I. Kerenidis, A. Prakash, Quantum Recommendation Sys- tems, in: 8th Innovations in Theoretical Computer Science Conference (ITCS 2017), V ol. 67 of Leibniz International Proceedings in Informatics (LIPIcs), Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2017, pp. 49:1–49:21. doi:10.4230/LIPIcs.ITCS.2017.49

    doi:10.4230/lipics.itcs.2017.49 2017

  27. [27]

    Brassard, P

    G. Brassard, P. Høyer, M. Mosca, A. Tapp, Quantum am- plitude amplification and estimation, Contemporary Math- ematics 305 (2002) 53–74, arXiv:quant-ph/0005055

    Pith/arXiv arXiv 2002

  28. [28]

    C. Cade, A. Montanaro, The quantum complexity of com- puting Schatten p-norms, arXiv preprint arXiv:1706.09279 (2017). 11

    Pith/arXiv arXiv 2017