Tensor methods for the computation of MTTA in large systems of loosely interconnected components
Pith reviewed 2026-05-25 09:03 UTC · model grok-4.3
The pith
Splitting local and synchronization transitions yields a linearly convergent MTTA algorithm that becomes quadratic and scales to billions of states in tensor-train format.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
By decoupling local and synchronization transitions, the authors formulate an algorithm for the MTTA that converges linearly and can be accelerated to quadratic convergence, while the same decoupling permits all matrices and vectors to be represented in tensor-train format, enabling numerical solution of problems with up to billions of states.
What carries the argument
The decoupling of local and synchronization transitions, which produces both the convergent iteration and the low-rank tensor-train structure.
If this is right
- The algorithm converges linearly to the exact MTTA value.
- A simple modification upgrades the convergence rate to quadratic.
- Tensor-train storage keeps memory and arithmetic costs independent of the full state-space size.
- Systems with up to billions of states become numerically tractable.
Where Pith is reading between the lines
- The same local-synchronization split could be used to compute absorption probabilities or other transient quantities in the same model class.
- Quadratic convergence removes the iteration count penalty that appears when the linear rate is close to one.
- Tensor-train techniques developed here may transfer directly to steady-state analysis of loosely coupled continuous-time Markov chains.
Load-bearing premise
The components must be loosely interconnected so that separating local and synchronization transitions both produces fast convergence and keeps the tensor-train ranks small.
What would settle it
Applying the method to a system whose components are tightly coupled and observing either failure of linear convergence or tensor-train ranks that grow linearly with the number of components.
Figures
read the original abstract
We are concerned with the computation of the mean-time-to-absorption (MTTA) for a large system of loosely interconnected components, modeled as continuous time Markov chains. In particular, we show that splitting the local and synchronization transitions of the smaller subsystems allows to formulate an algorithm for the computation of the MTTA which is proven to be linearly convergent. Then, we show how to modify the method to make it quadratically convergent, thus overcoming the difficulties for problems with convergent rate close to $1$. In addition, it is shown that this decoupling of local and synchronization transitions allows to easily represent all the matrices and vectors involved in the method in the tensor-train (TT) format - and we provide numerical evidence showing that this allows to treat large problems with up to billions of states - which would otherwise be unfeasible.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops an iterative method for computing mean time to absorption (MTTA) in continuous-time Markov chains composed of loosely interconnected components. By separating local transitions from synchronization transitions, it derives a linearly convergent algorithm that is then accelerated to quadratic convergence; the resulting operators admit low-rank tensor-train representations, enabling numerical solution of systems with up to 10^9 states.
Significance. If the stated convergence proofs and TT-rank bounds hold under the modeling assumptions, the work supplies a practical route to MTTA computation for state spaces that are otherwise intractable, together with explicit linear and quadratic rates and reproducible numerical evidence on billion-state instances. These elements constitute a concrete advance for reliability and performance analysis of large stochastic systems.
minor comments (3)
- [Abstract] Abstract, line 4: 'convergent rate' should read 'convergence rate'.
- [§5] §5, Table 2: the reported TT-ranks for the largest instances are given only as upper bounds; explicit per-iteration ranks would strengthen the reproducibility claim.
- [§3.2] Notation for the synchronization matrix S is introduced in §2 but first used in §3.2 without a forward reference; a brief reminder would aid readability.
Simulated Author's Rebuttal
We thank the referee for the positive assessment of the manuscript and the recommendation to accept.
Circularity Check
No significant circularity
full rationale
The derivation begins from standard continuous-time Markov chain (CTMC) modeling of loosely interconnected components and proceeds by splitting local versus synchronization transitions to obtain a linearly convergent iteration for mean-time-to-absorption (MTTA); the iteration is then algebraically modified to quadratic convergence, after which the resulting operators are shown to admit low-rank tensor-train representations. None of these steps reduces by construction to a fitted parameter, a self-referential definition, or a load-bearing self-citation; the convergence proofs and TT-rank bounds rest on the explicit separation of transition classes and on the Kronecker structure of the infinitesimal generator, both of which are external to the algorithm itself. The abstract and manuscript description therefore remain self-contained against external CTMC and tensor-algebra benchmarks.
Axiom & Free-Parameter Ledger
axioms (1)
- standard math Standard properties of continuous-time Markov chains and the integral formula for mean time to absorption
Reference graph
Works this paper leans on
-
[1]
Berman, A., Plemmons, R. J., 1994. Nonnegative matrices in the ma th- ematical sciences. Vol. 9 of Classics in Applied Mathematics. Society f or Industrial and Applied Mathematics (SIAM), Philadelphia, PA, revise d reprint of the 1979 original. URL https://doi.org/10.1137/1.9781611971262
-
[2]
Multigrid methods combined with low-rank approximation for tensor structured Markov chains
Bolten, M., Kahl, K., Kressner, D., Macedo, F., Sokolovi´ c, S., 201 6. Multi- grid methods combined with low-rank approximation for tensor stru ctured markov chains. arXiv preprint arXiv:1605.06246
work page internal anchor Pith review Pith/arXiv arXiv
-
[3]
Approximation of 1 /x by exponential sums in [1 ,∞ )
Braess, D., Hackbusch, W., 2005. Approximation of 1 /x by exponential sums in [1 ,∞ ). IMA journal of numerical analysis 25 (4), 685–697
work page 2005
- [4]
-
[5]
Kronecker based matrix repres entations for large Markov models
Buchholz, P., Kemper, P., 2004. Kronecker based matrix repres entations for large Markov models. Springer, pp. 256–295
work page 2004
- [6]
-
[7]
The numerical range as a spectral set
Crouzeix, M., Palencia, C., 2017. The numerical range as a spectr al set. arXiv preprint arXiv:1702.00668
work page internal anchor Pith review Pith/arXiv arXiv 2017
-
[8]
A multilinear sin gu- lar value decomposition
De Lathauwer, L., De Moor, B., Vandewalle, J., 2000. A multilinear sin gu- lar value decomposition. SIAM Journal of Matrix Analysis and Applicat ions 21 (4), 1253–1278. URL https://doi.org/10.1137/S0895479896305696
-
[9]
Tensor rank and the ill-posedness of t he best low-rank approximation problem
de Silva, V., Lim, L.-H., 2008. Tensor rank and the ill-posedness of t he best low-rank approximation problem. SIAM Journal of Matrix Analy sis and Applications 30 (3), 1084–1127. URL https://doi.org/10.1137/06066518X
-
[10]
Demmel, J. W., 1997. Applied numerical linear algebra. Society for Indus- trial and Applied Mathematics (SIAM), Philadelphia, PA. URL https://doi.org/10.1137/1.9781611971446
-
[11]
Dolgov, S. V., Khoromskij, B. N., Oseledets, I. V., 2012. Fast so lution of parabolic problems in the tensor train/quantized tensor train fo rmat with initial application to the Fokker-Planck equation. SIAM Journal of Scientific Computing 34 (6), A3016–A3038. URL https://doi.org/10.1137/120864210 21
-
[12]
Dolgov, S. V., Savostyanov, D. V., 2014. Alternating minimal ene rgy meth- ods for linear systems in higher dimensions. SIAM Journal on Scientifi c Computing 36 (5), A2248–A2271
work page 2014
-
[13]
Donatelli, S., 1993. Superposed stochastic automata: A class o f stochastic petri nets with parallel solution and distributed state space. Perf ormance Evaluation 18 (1), 21–36
work page 1993
-
[14]
Hierarchical singular value decomposition of tensors
Grasedyck, L., 2010. Hierarchical singular value decomposition of tensors. SIAM Journal on Matrix Analysis and Applications 31 (4), 2029–2054
work page 2010
-
[15]
Computation of best l∞ exponential sums for 1 /x by remez algorithm
Hackbusch, W., 2019. Computation of best l∞ exponential sums for 1 /x by remez algorithm. Computing and Visualization in Science 20 (1-2), 1–1 1
work page 2019
-
[16]
Higham, N. J., 2008. Functions of matrices. Society for Indust rial and Ap- plied Mathematics (SIAM), Philadelphia, PA, theory and computation . URL https://doi.org/10.1137/1.9780898717778
-
[17]
Hillar, C. J., Lim, L.-H., 2013. Most tensor problems are np-hard. Journal of the ACM (JACM) 60 (6), 45
work page 2013
-
[18]
A Compositional Approach to Performance Mo delling
Hillston, J., 1996. A Compositional Approach to Performance Mo delling. Cambridge University Press, New York, NY, USA
work page 1996
-
[19]
Kazeev, V. A., Khoromskij, B. N., 2012. Low-rank explicit QTT re presen- tation of the Laplace operator and its inverse. SIAM Journal of Ma trix Analysis and Applications 33 (3), 742–758. URL https://doi.org/10.1137/100820479
-
[20]
Kolda, T. G., Bader, B. W., 2009. Tensor decompositions and app lications. SIAM Rev. 51 (3), 455–500. URL https://doi.org/10.1137/07070111X
-
[21]
Low-rank tensor methods fo r communicat- ing markov processes
Kressner, D., Macedo, F., 2014. Low-rank tensor methods fo r communicat- ing markov processes. In: International Conference on Quantit ative Eval- uation of Systems. Springer, pp. 25–40
work page 2014
-
[22]
Algorithm 941: htucker–a Matlab tool- box for tensors in hierarchical Tucker format
Kressner, D., Tobler, C., 2014. Algorithm 941: htucker–a Matlab tool- box for tensors in hierarchical Tucker format. ACM Trans. Math. Software 40 (3), Art. 22, 22. URL https://doi.org/10.1145/2538688
-
[23]
Lubich, C., Oseledets, I. V., Vandereycken, B., 2015. Time integ ration of tensor trains. SIAM Journal of Numerical Analysis 53 (2), 917–94 1. URL https://doi.org/10.1137/140976546
-
[24]
Computing performability measures in markov chains by means of matrix functions
Masetti, G., Robol, L., 2018. Computing performability measures in markov chains by means of matrix functions. arXiv preprint arXiv:1803.06322. 22
-
[25]
Masetti, G., Robol, L., Chiaradonna, S., Di Giandomenico, F., 2019 . Stochastic evaluation of large interdependent composed models th rough kronecker algebra and exponential sums. In: Application and Theo ry of Petri Nets and Concurrency. pp. 47–66
work page 2019
-
[26]
MathWorks, 2018. MATLAB R2018a. The Mathworks, Inc., Nat ick, Mas- sachusetts
work page 2018
-
[27]
Oseledets, I., Dolgov, S., Kazeev, V., Lebedeva, O., Mach, T., 20 19. MAT- LAB TT-Toolbox. https://github.com/oseledets/TT-Toolbox
-
[28]
Oseledets, I. V., 2011. Tensor-train decomposition. SIAM Jou rnal of Scien- tific Computing 33 (5), 2295–2317. URL https://doi.org/10.1137/090752286
-
[29]
Oseledets, I. V., Dolgov, S. V., 2012. Solution of linear systems a nd matrix inversion in the tt-format. SIAM Journal on Scientific Computing 34 (5), A2718–A2739
work page 2012
- [30]
-
[31]
Trivedi, K. S., Bobbio, A., 2017. Reliability and Availability Engineering : Modeling, Analysis, and Applications. Cambridge University Press
work page 2017
-
[32]
Trivedi, K. S., Bobbio, A., 2017. Reliability and Availability Engineering : Modeling, Analysis, and Applications. Cambridge University Press. 23
work page 2017
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.