REVIEW 1 major objections 3 minor 3 cited by
On the Capacity of Secure Distributed Batch Matrix Multiplication
T0 review · 1 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read The paper claims that the capacity of secure distributed batch matrix multiplication depends on matrix dimensions, and that for a batch of outer products it is exactly $1 - X/N$, contradicting the previously reported general bound $(1…
desk verdict Real correction of the prior SDMM capacity formula and a solid new PIR connection, but the dimension-independent capacity claims overreach: the proof supplies asymptotic sequences, not the required universal lower bound for rank-deficient matrices. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
Two ingredients carry the argument. First, Lemma 1 converts any SDBMM scheme without $A$ as side information into a multi-message $X$-secure $T$-private information retrieval (MM-XSTPIR) scheme, thinking of $A$ as the stored messages and $B$ as the user's private query, so the new MM-XSTPIR upper bound in Theorem 1 bounds SDBMM rate from above. Second, the achievable schemes use cross-subspace alignment: the servers' stored shares are random linear combinations with poles at distinct constants $f_s + \alpha_n$, so the downloads concentrate the desired products $A_s B_s$ in decodable subspaces while noise terms align in separately decodable subspaces. For the scalar and outer-product regimes, a transformation maps the multiplicative group of the finite field to the additive group modulo $q-1$, realized inside a larger prime field, turning scalar multiplication into scalar addition; this is retrieved with an $X$-secure MDS code. The matrix-invertibility lemma for the cross-subspace decoding matrix is the technical keystone, and the entropy calculations for products of random matrices fix the normalization $H_q(AB)$.
What would settle it
For $N=3$, $X=1$, $K=1$, the paper's capacity formula gives $C(AB,\phi)=1-X/N=2/3$. A concrete check is to search for any valid $X$-secure scheme in this smallest case whose rate exceeds $2/3$, or to prove a converse below $2/3$; either would refute Theorem 2. The same test applies to the general geometric bound in case (38): any explicit scheme with rate above that bound in an intermediate-dimension regime would falsify the converse.
Extended reading notes
Core claim
The paper establishes that the capacity of SDBMM$(AB,\phi)$, where both $A$ and $B$ are held in $X$-secure coded form and the user has no side information, is characterized by Theorem 2. For $N > X$ and $K = 1$ (each product is an outer product of vectors), the capacity is exactly $1 - X/N$. When $K/\min(L,M) \to \infty$, the capacity tends to $0$ for $N \le 2X$ and to $1 - 2X/N$ for $N > 2X$; when $K \le \min(L,M)$ and $\max(L,M)/K \to \infty$, it tends to $1 - X/N$. For intermediate dimensions the paper proves upper bounds $(1 - X/N)\min(L,M,K)/K$ in the regime $2X \ge N > X$ and the geometric-series bound $(1 - X/N)\left(1 + \frac{X}{N-X} + \cdots + \left(\frac{X}{N-X}\right)^{\lfloor K/\min(L,M,K)\rfloor -1}\right)^{-1}$ for $N > 2X$. The dimension-independent capacity, minimized over all matrix dimensions, is $0$ for $N \le 2X$ and $1 - 2X/N$ for $N > 2X$. The paper further argues that the previously reported general two-sided secure capacity $(1 - 2X/N)^+$ cannot be correct, because the $K=1$ case already achieves the larger rate $1 - X/N$, so the true capacity genuinely depends on $L, K, M$.
Load-bearing premise
All converse bounds rely on the MM-XSTPIR upper bound, which assumes perfect $X$-security of the stored data and independence between the user's queries and the server storage; if colluding servers were allowed to learn partial information about $A$ or $B$, the upper bounds would not follow.
Editorial extensions
If this is right
- For batches of outer products of vectors ($K=1$), the capacity is exactly $1 - X/N$, so a user can compute all $S$ products by downloading only the $N/(N-X)$ overhead factor, independent of $L$ and $M$.
- For batches of inner products of long vectors ($L=M=1$, $K \to \infty$), the capacity approaches $1 - 2X/N$ when $N > 2X$, and approaches $0$ when $N \le 2X$; in this regime two-sided security costs twice the overhead.
- The dimension-independent capacity of SDBMM$(AB,\phi)$ is fully settled: it is $0$ for $N \le 2X$ and $1 - 2X/N$ for $N > 2X$.
- The previous claim that the general two-sided secure capacity is $(1 - 2X/N)^+$ is false, because the $K=1$ outer-product scheme already achieves $1 - X/N > (1 - 2X/N)^+$ whenever $X>0$ and $N>X$.
- For the one-sided secure setting SDBMM$(B,A)$, the capacity is $1 - X/N$ independent of matrix dimensions, matching the prior result, and the paper supplies an achievable scheme that works for all $K < L$ cases.
Reading between the lines
- Beyond the paper: if the scalar-multiplication-to-addition transformation can be generalized from $K=1$ to square matrices, the capacity for $L=K=M>1$ may exceed the geometric-series upper bounds, since the $K=1$ case already breaks the previously accepted answer.
- Beyond the paper: the SDBMM-to-MM-XSTPIR connection means that any future improvement to MM-XSTPIR capacity, or to the related private computation problem, will automatically give tighter converses for these matrix multiplication problems; conversely, better SDBMM schemes would provide new lower bounds for those PIR problems.
- Beyond the paper: the dimension-dependence of capacity has a practical reading: a system designer can change the communication cost of security by reorganizing a computation into outer products, inner products, or square blocks, because these regimes have genuinely different capacities.
- Beyond the paper: the scalar scheme's rate approaches $1 - X/N$ only as the field size $q \to \infty$, since the zero-indicator overhead and the prime-field embedding vanish only in that limit; for small $q$, the exact finite-field capacity is not addressed and could be strictly lower.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies secure distributed batch matrix multiplication (SDBMM), where sequences of matrix products A_s B_s are retrieved from N servers that store the matrices A and B in X-secure coded form. It defines rate as the ratio of the entropy of the desired products to the average download, and capacity as the supremum over schemes and batch sizes. The authors prove upper bounds by connecting SDBMM to a new multi-message X-secure T-private information retrieval (MM-XSTPIR) problem and proving an upper bound on the latter. They prove capacity results and asymptotic characterizations for several variants: SDBMM(AB,φ), SDBMM(B,A), SDBMM(B,B), SDBMM(B,φ), and SDBMM(AB,B). Achievability uses a general download-A-and-B scheme, a cross-subspace alignment scheme, and a monomorphic transformation turning scalar multiplication into scalar addition. The paper also claims that the previously reported two-sided secure SDMM capacity (1-2X/N)^+ is false in general, and it states dimension-independent capacities as corollaries.
Significance. If the main results hold, the paper makes a substantial contribution: it introduces the MM-XSTPIR connection, gives the first capacity characterizations for secure batch matrix multiplication in several parameter regimes, exactly settles the outer-product case K=1, asymptotically settles the long inner-product case, and provides a counterexample to a previously published converse. The achievable schemes are explicit and the converse proofs are largely self-contained. The main caveat is that one of the advertised corollaries, the dimension-independent capacity of SDBMM(AB,φ), is not established by the supplied arguments.
major comments (1)
- [Section 3.4, Corollary 1] The lower-bound half of Corollary 1 is not proved. By the definition in Section 2.1, the dimension-independent capacity is the infimum of C(AB,φ) over all L,K,M. For N>2X, the claimed value 1-2X/N therefore requires C(AB,φ) ≥ 1-2X/N for every triple (L,K,M). Theorem 2 cases (34) and (35) are asymptotic statements: they give sequences of dimensions with K/min(L,M)→∞ along which capacity tends to 0 or to 1-2X/N. These sequences show only that the infimum is no larger than the stated values; they do not show that no finite dimension has smaller capacity. The only lower bounds proved for finite dimensions are the general scheme of Section 5.1 and the CSA scheme of Section 5.2. When K<min(L,M), Lemma 2 gives H_q(AB)=LK+KM-K^2, so the CSA scheme achieves (1-2X/N)(LK+KM-K^2)/(LM), which is strictly below 1-2X/N, and the general scheme achieves (1-X/N)(LK+KM-K^2)/(LK+KM), which is also below 1-2X/N for the family L=M=n, K=n-1 with n≥3 (e.g., N=5, X=1 gives rates 2(n+1)/(5n) and (3/5)(n^2-1)/n^2, both < 3/5 for every finite n). Remark 3.4 supplies a lower bound only for K≥min(L,M), and Section 6 explicitly leaves the square-matrix case open. Thus the lower-bound half of Corollary 1 is unsupported as written; the authors must either provide a universal lower bound for K<min(L,M) or revise the corollary to a weaker, asymptotic statement.
minor comments (3)
- [Theorem 2] The arrow notation 'C(AB,φ)→...' is used without explicitly stating the order of limits; the theorem should state that the limit is taken with q→∞ and then with the indicated dimension ratio tending to infinity (or jointly, as appropriate).
- [Remark 4] The phrase 'presents a contradiction that calls into question the converse bound in [9]' would be more precise as 'provides a counterexample to the converse bound in [9]', since the issue is with a claimed theorem in the prior work, not with the logical framework of the present paper.
- [Section 5.4.1] In the rate computation near Eq. (159), the term 2SN log_q(2) accounts for secret-sharing the zero-indicator bits; this is correct but could use one sentence explaining that each indicator is secret-shared across all N servers, so the total download of indicators is 2SN bits.
Circularity Check
No significant circularity: the capacity bounds are derived from the SDBMM/MM-XSTPIR definitions and independent entropy arguments; the cited CSA invertibility lemma is a tool, not the target result.
full rationale
The paper's converse chain is self-contained. Lemma 1 reduces SDBMM to MM-XSTPIR by instantiating A as messages and B as queries; this is a standard scheme-specialization argument, not an assumption of the capacity being proved. Theorem 1's upper bound is proved from the PIR storage/query/answer definitions via Lemmas 4-7, which use only entropy inequalities, Han's inequality, and the security/privacy constraints. No parameter is fitted to data and no target capacity value is inserted as an input. On the achievability side, the general scheme and the CSA scheme give explicit downloads; their rates are evaluated using Lemma 2, whose entropy formulas are proved in Appendix B from random-matrix rank arguments. The only same-author citation that plays a constructive role is [15], used for the invertibility of the Cauchy-style matrix M_N in Lemma 11; that lemma is a parameter-free linear-algebra statement with a proof sketch adapted from [15] and is not equivalent to any SDBMM capacity claim. The possible gap in Corollary 1 (the limit cases (34)-(35) give only an upper bound on the dimension-independent infimum unless a universal lower bound for K < min(L,M) is supplied) is a rigor/correctness concern about an inferred claim, not a circularity: the paper does not define its conclusion into its assumptions. Overall, no derivation step reduces to its own inputs.
Assumptions & free parameters
assumptions (5)
- domain assumption The matrices A and B are uniform, independent, and stored via X-secure secret sharing (Eq. (1)-(7)).
- standard math Han's inequality is used to average the converse bounds over all colluding sets (Section 4, Eq. (73), (90)).
- standard math For a random square matrix over F_q, the probability of being singular tends to 0 as q->infty (Lemma 8, citing [33]).
- standard math The multiplicative group of a finite field is cyclic, so scalar multiplication can be mapped to addition modulo q-1 (Section 5.4.1).
- standard math For any integer nu>1 there exists a prime p with nu < p < 2nu (Bertrand-Chebyshev), yielding 2(q-1) < p < 4(q-1) (Section 5.4.1).
Cite this review
Pith. "Pith review of On the Capacity of Secure Distributed Batch Matrix Multiplication." pith.science (2026). https://pith.science/paper/SUHELTLO
@misc{pith2026190806957,
author = {Pith},
title = {Pith review of: On the Capacity of Secure Distributed Batch Matrix Multiplication},
year = {2026},
howpublished = {\url{https://pith.science/paper/SUHELTLO}},
note = {Machine review of arXiv:1908.06957}
}
abstract
The problem of secure distributed batch matrix multiplication (SDBMM) studies the communication efficiency of retrieving a sequence of desired matrix products ${\bf AB}$ $=$ $({\bf A}_1{\bf B}_1,$ ${\bf A}_2{\bf B}_2,$ $\cdots,$ ${\bf A}_S{\bf B}_S)$ from $N$ distributed servers where the constituent matrices ${\bf A}=({\bf A}_1, {\bf A}_2, \cdots, {\bf A}_S)$ and ${\bf B}=({\bf B}_1, {\bf B}_2,\cdots,{\bf B}_S)$ are stored in $X$-secure coded form, i.e., any group of up to $X$ colluding servers learn nothing about ${\bf A, B}$. It is assumed that ${\bf A}_s\in\mathbb{F}_q^{L\times K}, {\bf B}_s\in\mathbb{F}_q^{K\times M}, s\in\{1,2,\cdots, S\}$ are uniformly and independently distributed and $\mathbb{F}_q$ is a large finite field. The rate of an SDBMM scheme is defined as the ratio of the number of bits of desired information that is retrieved, to the total number of bits downloaded on average. The supremum of achievable rates is called the capacity of SDBMM. In this work we explore the capacity of SDBMM, as well as several of its variants, e.g., where the user may already have either ${\bf A}$ or ${\bf B}$ available as side-information, and/or where the security constraint for either ${\bf A}$ or ${\bf B}$ may be relaxed. We obtain converse bounds, as well as achievable schemes for various cases of SDBMM, depending on the $L, K, M, N, X$ parameters, and identify parameter regimes where these bounds match. A remarkable aspect of our upper bounds is a connection between SDBMM and a form of private information retrieval (PIR) problem, known as multi-message $X$-secure $T$-private information retrieval (MM-XSTPIR). Notable features of our achievable schemes include the use of cross-subspace alignment and a transformation argument that converts a scalar multiplication problem into a scalar addition problem, allowing a surprisingly efficient solution.
Figures
Forward citations
Cited by 3 Pith papers
-
$X$-secure $T$-private Information Retrieval from MDS Coded Storage with Byzantine and Unresponsive Servers
A cross-subspace alignment scheme with layered interference cancellation achieves rate 1-(Kc+X+T+2B-1)/(N-U) for X-secure T-private retrieval from MDS-coded storage with U unresponsive and B Byzantine servers, improvi...
-
Private and Secure Distributed Matrix Multiplication with Flexible Communication Load
Secure generalized PolyDot codes give a flexible recovery-threshold and communication-load trade-off for private and secure distributed matrix multiplication.
-
Secure Coded Multi-Party Computation for Massive Matrix Operations
A secure multi-party computation scheme for matrix polynomials uses polynomial sharing and claims large worker savings, but its transpose procedure is wrong, breaking the arbitrary-polynomial result.
Reference graph
Works this paper leans on
-
[1]
High-dimensional co ded matrix multiplication,
K. Lee, C. Suh, and K. Ramchandran, “High-dimensional co ded matrix multiplication,” in Proc. of IEEE International Symposium on Information Theory (ISIT) , June 2017, pp. 2418– 2422
work page 2017
-
[2]
Polynomial cod es: An optimal design for high- dimensional coded matrix multiplication,
Q. Yu, M. Maddah-Ali, and S. Avestimehr, “Polynomial cod es: An optimal design for high- dimensional coded matrix multiplication,” in Proc. of Advances in Neural Information Pro- cessing Systems , Dec. 2017, pp. 4403–4413
work page 2017
-
[3]
On the optimal recovery threshold of coded matrix multiplication,
S. Dutta, M. Fahim, F. Haddadpour, H. Jeong, V. Cadambe, a nd P. Grover, “On the optimal recovery threshold of coded matrix multiplication,” IEEE Transactions on Information Theory, vol. 66, no. 1, pp. 278–301, 2019
work page 2019
-
[4]
S. Dutta, Z. Bai, H. Jeong, T. Low, and P. Grover, “A Unified Coded Deep Neural Network Training Strategy Based on Generalized PolyDot Codes for Ma trix Multiplication,” arXiv preprint arXiv:1811.10751, 2018
arXiv 2018
-
[5]
Straggler mitigation in distributed matrix multiplication: Fundamental limits and optimal coding,
Q. Yu, M. A. Maddah-Ali, and A. S. Avestimehr, “Straggler mitigation in distributed matrix multiplication: Fundamental limits and optimal coding,” IEEE Transactions on Information Theory, vol. 66, no. 3, pp. 1920–1933, 2020
work page 1920
-
[6]
Lagrange coded computing: Optimal design for resiliency, security, and privacy,
Q. Yu, S. Li, N. Raviv, S. M. M. Kalan, M. Soltanolkotabi, a nd S. A. Avestimehr, “Lagrange coded computing: Optimal design for resiliency, security, and privacy,” in The 22nd Interna- tional Conference on Artificial Intelligence and Statistics . PMLR, 2019, pp. 1215–1225
work page 2019
-
[7]
Secure Distributed Computing with St raggling Servers Using Polynomial Codes,
H. Yang and J. Lee, “Secure Distributed Computing with St raggling Servers Using Polynomial Codes,” IEEE Transactions on Information Forensics and Security , vol. 14, no. 1, pp. 141–150, Jan. 2019
work page 2019
-
[8]
On the Capacity of Secure Dist ributed Matrix Multiplication,
W.-T. Chang and R. Tandon, “On the Capacity of Secure Dist ributed Matrix Multiplication,” arXiv preprint arXiv:1806.00469 , 2018
arXiv 2018
Show all 33 references
-
[9]
On the Capacity and Straggler-Robustness of Dis- tributed Secure Matrix Multiplication,
J. Kakar, S. Ebadifar, and A. Sezgin, “On the Capacity and Straggler-Robustness of Dis- tributed Secure Matrix Multiplication,” IEEE Access, vol. 7, pp. 45 783–45 799, 2019
2019
-
[10]
Gasp co des for secure distributed matrix multiplication,
R. G. D’Oliveira, S. El Rouayheb, and D. Karpuk, “Gasp co des for secure distributed matrix multiplication,” IEEE Transactions on Information Theory , vol. 66, no. 7, pp. 4038–4050, 2020. 33
2020
-
[11]
Private and s ecure distributed matrix multiplication with flexible communication load,
M. Aliasgari, O. Simeone, and J. Kliewer, “Private and s ecure distributed matrix multiplication with flexible communication load,” IEEE Transactions on Information Forensics and Security , vol. 15, pp. 2722–2734, 2020
2020
-
[12]
Pr ivate Information Retrieval,
B. Chor, O. Goldreich, E. Kushilevitz, and M. Sudan, “Pr ivate Information Retrieval,” in Proceedings of the 36th Annual Symposium on Foundations of Co mputer Science , 1995, pp. 41–50
1995
-
[13]
Multi-message Private Infor mation Retrieval: Capacity Results and Near-optimal Schemes,
K. Banawan and S. Ulukus, “Multi-message Private Infor mation Retrieval: Capacity Results and Near-optimal Schemes,” IEEE Transactions on Information Theory , vol. 64, no. 10, pp. 6842 – 6862, 2018
2018
-
[14]
Single-Server Multi-Message Pri vate Information Retrieval with Side Information,
S. Li and M. Gastpar, “Single-Server Multi-Message Pri vate Information Retrieval with Side Information,” arXiv preprint arXiv:1808.05797 , 2018
2018 arXiv
-
[15]
Cross subspace alignmen t and the asymptotic capacity of x -secure t -private information retrieval,
Z. Jia, H. Sun, and S. A. Jafar, “Cross subspace alignmen t and the asymptotic capacity of x -secure t -private information retrieval,” IEEE Transactions on Information Theory , vol. 65, no. 9, pp. 5783–5798, Sep. 2019
2019
-
[16]
X-secure T -private Information Retrieval from MDS Coded Storage with Byzantine and Unresponsive Servers,
Z. Jia and S. A. Jafar, “ X-secure T -private Information Retrieval from MDS Coded Storage with Byzantine and Unresponsive Servers,” IEEE Transactions on Information Theory , 2020, DOI: 10.1109/TIT.2020.3013152
2020
-
[17]
Cross subspace alignment codes for coded distribu ted batch computation,
——, “Cross subspace alignment codes for coded distribu ted batch computation,” IEEE Trans- actions on Information Theory , vol. 67, no. 5, pp. 2821–2846, 2021
2021
-
[18]
The Capacity of Private Computat ion,
H. Sun and S. A. Jafar, “The Capacity of Private Computat ion,” IEEE Transactions on Information Theory, vol. 65, no. 6, pp. 3880–3897, June 2019
2019
-
[19]
T. M. Cover and J. A. Thomas, Elements of Information Theory . Wiley, 2006
2006
-
[20]
The Capacity of Private Informat ion Retrieval,
H. Sun and S. A. Jafar, “The Capacity of Private Informat ion Retrieval,” IEEE Transactions on Information Theory , vol. 63, no. 7, pp. 4075–4088, July 2017
2017
-
[21]
The Capacity of Robust Private Information Retrie val with Colluding Databases,
——, “The Capacity of Robust Private Information Retrie val with Colluding Databases,” IEEE Transactions on Information Theory , vol. 64, no. 4, pp. 2361–2370, April 2018
2018
-
[22]
The Capacity of Symmetric Private Information Ret rieval,
——, “The Capacity of Symmetric Private Information Ret rieval,” IEEE Transactions on Information Theory, vol. 65, no. 1, pp. 322–329, January 2019
2019
-
[23]
The Capacity of Private Infor mation Retrieval from Coded Databases,
K. Banawan and S. Ulukus, “The Capacity of Private Infor mation Retrieval from Coded Databases,” IEEE Transactions on Information Theory , vol. 64, no. 3, pp. 1945–1956, 2018
1945
-
[24]
Private Function R etrieval,
M. Mirmohseni and M. A. Maddah-Ali, “Private Function R etrieval,” arXiv preprint arXiv:1711.04677, 2017
2017 arXiv
-
[25]
Private Information Retrieval from Coded Databases with Colluding Servers,
R. Freij-Hollanti, O. Gnilke, C. Hollanti, and D. Karpu k, “Private Information Retrieval from Coded Databases with Colluding Servers,” SIAM Journal on Applied Algebra and Geometry , vol. 1, no. 1, pp. 647–664, 2017. 34
2017
-
[26]
Private Information Retrieval f rom MDS Coded Data with Colluding Servers: Settling a Conjecture by Freij-Hollanti et al
H. Sun and S. A. Jafar, “Private Information Retrieval f rom MDS Coded Data with Colluding Servers: Settling a Conjecture by Freij-Hollanti et al.” IEEE Transactions on Information Theory, vol. 64, no. 2, pp. 1000–1022, February 2018
2018
-
[27]
Tow ard Optimal Secure Distributed Storage Systems with Exact Repair,
R. Tandon, S. Amuru, T. C. Clancy, and R. M. Buehrer, “Tow ard Optimal Secure Distributed Storage Systems with Exact Repair,” IEEE Transactions on Information Theory, vol. 62, no. 6, pp. 3477–3492, 2016
2016
-
[28]
The capacity of pri vate information retrieval from uncoded storage constrained databases,
M. A. Attia, D. Kumar, and R. Tandon, “The capacity of pri vate information retrieval from uncoded storage constrained databases,” IEEE Transactions on Information Theory , vol. 66, no. 11, pp. 6617–6634, 2020
2020
-
[29]
Fundamental Limits o f Cache-Aided Private Informa- tion Retrieval With Unknown and Uncoded Prefetching,
Y. Wei, K. Banawan, and S. Ulukus, “Fundamental Limits o f Cache-Aided Private Informa- tion Retrieval With Unknown and Uncoded Prefetching,” IEEE Transactions on Information Theory, vol. 65, no. 5, pp. 3215–3232, May 2019
2019
-
[30]
Private information retrieval with side information,
S. Kadhe, B. Garcia, A. Heidarzadeh, S. El Rouayheb, and A. Sprintson, “Private information retrieval with side information,” IEEE Transactions on Information Theory , vol. 66, no. 4, pp. 2032–2043, 2019
2019
-
[31]
The Capacity of Cache Aided Private Informa tion Retrieval,
R. Tandon, “The Capacity of Cache Aided Private Informa tion Retrieval,” arXiv preprint arXiv:1706.07035, 2017
2017 arXiv
-
[32]
The capacity of t-priv ate information retrieval with private side information,
Z. Chen, Z. Wang, and S. A. Jafar, “The capacity of t-priv ate information retrieval with private side information,” IEEE Transactions on Information Theory , vol. 66, no. 8, pp. 4761–4773, 2020
2020
-
[33]
How Often do Determinants Over Finite F ields Vanish?
W. Waterhouse, “How Often do Determinants Over Finite F ields Vanish?” Discrete Mathe- matics, no. 65, pp. 103–104, 1987. 35
1987
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.