Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

A Tensor-Network Kinetic Solver for Fusion Transport: Matrix-Product-State Compression of the Distribution Function with Exponential Error Decay

Priyanca Ford

Published Sep 7, 2026• 2 citations
Publisher

Abstract

Kinetic descriptions of fusion plasmas are throttled by the curse of dimensionality: a distribution function on a six-dimensional phase space is intractable to store on a dense grid, let alone to advance in time. We show that the kinetic distribution function is strongly compressible in a low-rank tensor-network (matrix-product-state, or tensor-train) representation and that this yields a practical solver on present-day classical hardware, not merely a storage economy. We cast the problem formally: the discretized distribution is a high-order tensor, its matrix-product-state factorization is exact up to a truncation error bounded by the discarded Schmidt weight (Eckart–Young), and the number of retained modes needed for a target accuracy scales with the entanglement entropy of the state. On a one-dimensional-in-space, one-dimensional-in-velocity Bhatnagar–Gross–Krook (BGK) kinetic test problem, discretized on a periodic N_x=128× N_v=64 grid and advanced by operator-split upwind streaming and implicit relaxation, a matrix-product-state truncation reaches a relative-L^2 error of 2.9×10^-6 at bond dimension =8 while storing 0.19× the dense grid, and the error falls exponentially in the bond dimension — from 1.8×10^-1 at =1 to 2.2×10^-10 at =16 and to the double-precision floor (~10^-14) by =24. The velocity moments inherit the same convergence (density error 1.1×10^-6, temperature error 2.8×10^-6 at =8), and the converged distribution carries a Schmidt entanglement entropy below 0.1 nats — the area-law signature that makes it compressible. Every headline quantity is a frozen anchor in the Kronos de-risking register (gate KX-L1-A16). Beyond the anchor we give the general-dimensional machinery that carries the method to the five- and six-dimensional gyrokinetic problem — matrix-product-operator representations of the streaming and collision maps, dimension-linear tensor-train construction by cross interpolation, and a rank-adaptive, conservative dynamical-low-rank integrator that evolves the state on the low-rank manifold without inverting small singular values — together with the derived exponential-to-linear crossover in both storage and per-step arithmetic that motivates it. The result is a classical, deployable-today complement to the fault-tolerant-horizon quantum kinetic route, and we place it as a cross-cutting, verification-first module inside the Kronos autonomous control architecture: an advisory/supervisory reduced-kinetics solver, exercised twin-first beneath a deterministic safety clamp, that hands authority back to the deterministic floor on out-of-distribution inputs. The single honest gate is that the demonstration is a collision-dominated reduced problem, and the rank growth of the fully nonlinear, collisionless six-dimensional gyrokinetic distribution at reactor parameters remains the open question we scope explicitly. Key results (frozen anchors): v = 2 v; t = 3.27 ×10^-3; sigma_1 = 9.4 ×10^-2; sigma_1 = 2.5 ×10^-3; sigma_1 = 4.7 ×10^-5. Methods & codes: CGYRO, TGLF, GENE, Sauter, VQE, ADAPT-VQE. Live verification: 1 gate validator(s) with live-recompute cards (1 reproduced). See the Verification section and data/verification.csv. Related: Paper page · De-risking register · 3D model · Learn more about Kronos Part of the 2026 Kronos publication series; independently re-run and stamped in the Kronos de-risking register (DOI 10.5281/zenodo.22645689). All numerical values are frozen design-point anchors; see the register. Public research artifact. No proprietary, financial, or supply-chain information is included. Previous version: DOI 10.5281/zenodo.22132192. Please cite this version.

DOI
10.5281/zenodo.22645767

Related papers

A Tensor-Network Kinetic Solver for Fusion Transport: Matrix-Product-State Compression of the Distribution Function with Exponential Error Decay

Priyanca Ford, P I Ford

Kinetic descriptions of fusion plasmas are throttled by the curse of dimensionality: a distribution function on a six-dimensional phase space is intractable to store on a dense grid, let alone to advance in time. We show that the kinetic distribution function is strongly compressible in a low-rank tensor-network (matrix-product-state, or tensor-train) representation and that this yields a practical solver on present-day classical hardware, not merely a storage economy. We cast the problem formally: the discretized distribution is a high-order tensor, its matrix-product-state factorization is exact up to a truncation error bounded by the discarded Schmidt weight (Eckart–Young), and the number of retained modes needed for a target accuracy scales with the entanglement entropy of the state. On a one-dimensional-in-space, one-dimensional-in-velocity Bhatnagar–Gross–Krook (BGK) kinetic test problem, discretized on a periodic N_x=128× N_v=64 grid and advanced by operator-split upwind streaming and implicit relaxation, a matrix-product-state truncation reaches a relative-L^2 error of 2.9×10^-6 at bond dimension =8 while storing 0.19× the dense grid, and the error falls exponentially in the bond dimension — from 1.8×10^-1 at =1 to 2.2×10^-10 at =16 and to the double-precision floor (~10^-14) by =24. The velocity moments inherit the same convergence (density error 1.1×10^-6, temperature error 2.8×10^-6 at =8), and the converged distribution carries a Schmidt entanglement entropy below 0.1 nats — the area-law signature that makes it compressible. Every headline quantity is a frozen anchor in the Kronos de-risking register (gate KX-L1-A16). Beyond the anchor we give the general-dimensional machinery that carries the method to the five- and six-dimensional gyrokinetic problem — matrix-product-operator representations of the streaming and collision maps, dimension-linear tensor-train construction by cross interpolation, and a rank-adaptive, conservative dynamical-low-rank integrator that evolves the state on the low-rank manifold without inverting small singular values — together with the derived exponential-to-linear crossover in both storage and per-step arithmetic that motivates it. The result is a classical, deployable-today complement to the fault-tolerant-horizon quantum kinetic route, and we place it as a cross-cutting, verification-first module inside the Kronos autonomous control architecture: an advisory/supervisory reduced-kinetics solver, exercised twin-first beneath a deterministic safety clamp, that hands authority back to the deterministic floor on out-of-distribution inputs. The single honest gate is that the demonstration is a collision-dominated reduced problem, and the rank growth of the fully nonlinear, collisionless six-dimensional gyrokinetic distribution at reactor parameters remains the open question we scope explicitly.Key results (frozen anchors): v = 2 v; t = 3.27 ×10^-3; N_v = 128 ×64; sigma_1 = 9.4 ×10^-2; sigma_1 = 2.5 ×10^-3; sigma_1 = 4.7 ×10^-5.Methods & codes: CGYRO, TGLF, GENE, Sauter, VQE, ADAPT-VQE.Live verification: 1 gate validator(s) with live-recompute cards (1 reproduced). See the Verification section and data/verification.csv.Related: Paper page · De-risking register · 3D model · Learn more about KronosPart of the 2026 Kronos publication series; independently re-run and stamped in the Kronos de-risking register (DOI 10.5281/zenodo.22645689).All numerical values are frozen design-point anchors; see the register.Public research artifact. No proprietary, financial, or supply-chain information is included.Note (REPLACE): This record replaces and supersedes DOI 10.5281/zenodo.22132192; please cite this version.

•Sep 9, 2026•0 citations

High-Resolution Tensor-Network Fourier Methods for Exponentially Compressed Non-Gaussian Aggregate Distributions

Juan José Rodríguez-Aldavero, Juan José García-Ripoll

Characteristic functions of weighted sums of independent random variables exhibit low-rank structure in the quantized tensor train (QTT) representation, also known as matrix product states (MPS), enabling up to exponential compression of their fully non-Gaussian probability distributions. Under variable independence, the global characteristic function factorizes into local terms. Its low-rank QTT structure arises from intrinsic spectral smoothness in continuous models, or from spectral energy concentration as the number of components $D$ grows in discrete models. We demonstrate this on weighted sums of Bernoulli and lognormal random variables. In the former, despite an adversarial, incompressible small-$D$ regime, the characteristic function undergoes a sharp bond-dimension collapse for $D \gtrsim 300$ components, enabling polylogarithmic time and memory scaling. In the latter, the approach reaches high-resolution discretizations of $N = 2^{30}$ frequency modes on standard hardware, far beyond the $N = 2^{24}$ ceiling of dense implementations. These compressed representations enable efficient computation of Value at Risk (VaR) and Expected Shortfall (ES), supporting applications in quantitative finance and beyond.

stat.MLCross-listedarXiv:2603.23106•Mar 25, 2026•0 citations

[SUPERSEDED] The Tensor-Network Kinetic Solver - Classical, Deployable Today

Priyanca Ford

This record has been superseded. The current, updated version is A Tensor-Network Kinetic Solver for Fusion Transport: Matrix-Product-State Compression of the Distribution Function with Exponential Error Decay (DOI 10.5281/zenodo.22645768). Please read and cite the current version. This page stays online so existing citations keep resolving. The kinetic distribution function is highly compressible in a low-rank tensor-network representation, and that yields a practical classical solver, not a storage trick. On a 1D1V BGK test, a matrix-product-state truncation reaches relative-L² error 2×10−⁴ at rank 8 using ~0.19× the dense storage, with the error falling exponentially in rank. Unlike the fault-tolerant-horizon quantum route, this runs today.

•Aug 28, 2026•0 citations

A priori Assessment of Tensor-Network Encoding for Isotropic Turbulent Flows

Massen Esmaeili, Hirad Alipanah, Robert Pinkston +5 more

Tensor networks (TNs), originally developed for simulating many-body quantum systems, provide a systematic framework for approximating high-dimensional fields. This is achieved by factorizing the field into interconnected tensors with small bond dimensions, thereby restricting the correlations captured across field bipartitions. Belonging to the family of TNs, the matrix product state (MPS) ansatz is utilized here as a reduced-order modeling framework to construct truncated representations of isotropic turbulent flow data. Two direct numerical simulation (DNS) datasets are considered: the hydrodynamic field of an incompressible three-dimensional flow, and a conserved Fickian scalar in a similar flow. Each field is encoded as an MPS through a sequence of singular value decompositions (SVDs) in which small singular values are discarded. The truncated representation is contracted back to the full grid, and the resulting reconstructed field is compared against DNS. An interleaved ordering of the spatial tensor indices of the transport variables is applied prior to decomposition in order to localize the dominant inter-tensor correlations. Velocity reconstructions achieve $99.8\%$ fidelity using only $5\%$ of the original DNS memory, while the scalar field reaches the same fidelity at $15\%$ memory usage. A wide range of lower- and higher-order statistics, including velocity gradients, dissipation, and structure functions, are systematically examined. At these compression levels, the total kinetic energy and the scalar energy are both recovered within $0.2\%$ relative error, while the mean dissipation and mean scalar dissipation remain within approximately $10\%$ of the DNS generated values. These findings support the suitability of MPS for scalable reduced-order analysis of complex turbulent datasets and motivate further exploration of TN-based methods in computational turbulence.

physics.flu-dynCross-listedarXiv:2608.28869•Sep 1, 2026

Quantum-inspired method for solving the Vlasov-Poisson equations

Erika Ye, Nuno Loureiro

Kinetic simulations of collisionless (or weakly collisional) plasmas using the Vlasov equation are often infeasible due to high-resolution requirements and the exponential scaling of computational cost with respect to dimension. Recently, it has been proposed that matrix product state (MPS) methods, a quantum-inspired but classical algorithm, can be used to solve partial differential equations with exponential speed-up, provided that the solution can be compressed and efficiently represented as a MPS within some tolerable error threshold. In this work, we explore the practicality of MPS methods for solving the Vlasov-Poisson equations for systems with one coordinate in space and one coordinate in velocity, and find that important features of linear and nonlinear dynamics, such as damping or growth rates and saturation amplitudes, can be captured while compressing the solution significantly. Furthermore, by comparing the performance of different mappings of the distribution functions onto the MPS, we develop an intuition of the MPS representation and its behavior in the context of solving the Vlasov-Poisson equations, which will be useful for extending these methods to higher-dimensional problems.

arXiv:2205.11990•Physical review. E•Sep 29, 2022•46 citations

Truncated Tensor-Train Reservoirs Are Reservoirs in Their Own Right: A Two-Resource Law for Memory Beyond the State Wall

Takuya Okamura

Compressing the connectivity of a reservoir removes the O(N²) memory wall, but the state vector itself still costs O(N) — the state wall, at N = 4¹⁵ > 10⁹ an unallocatable 8.6 GB. We introduce a reservoir whose state evolves natively in tensor-train (TT) form: TT matrix–vector product, TT input injection, rounding to bond dimension χx, and the hypersphere normalization, which is a scalar division of one core and therefore exactly rank-neutral. When χx reaches the maximal rank of the tensorization the evolution is exact to machine precision; below it, the reservoir is not a degraded copy of the exact one. We show that the exact attractor of sphere-TT dynamics at large N is near-maximally entangled (one-shot truncation error 0.93 at χx = 16, N = 65,536; flat middle-bond Schmidt spectrum), so post-hoc state compression is impossible; the truncated system instead carves out its own low-rank attractor, characterized by a per-step rounding error δ̄ that is measurable at runtime for free. Two quantitative laws organize this regime. First, a decomposition law: δ̄² ≈ (c·(1−β))² + δstock², fitting a 5×3 grid in the orthogonal-mixing weight β and χx within 2–5%; the flow term is injected by the random part of the connectivity, while the floor δstock — measured directly at β = 1, where the rank-1 Kronecker-orthogonal connectivity injects nothing — is the rank cost of the retained memory itself. Second, a two-resource memory law: linear memory capacity follows MC ≈ min{K₀·χx², MC_noise(δ̄)}, with a threshold regime, a quadratic capacity regime (measured MC/χx² constant across χx at fixed β), and a noise-limited regime where δ̄ saturates at the flow floor. Following these laws — raise β and χx together — we demonstrate memory capacity 30–74 (three seeds) at N = 1,073,741,824, where the dense state cannot be allocated, with the entire model (connectivity, state, input coupling) in 861 KB at ~1 s/step on a CPU; the naive configuration scores ≈ 0.1 there. A nonlinear task (NARMA10) reaches r² = 0.64 in the same regime. We state the open problems honestly: the microscopic mechanism of the χ_x² law, the growth of δ̄ with tensor order, and a residual quality-of-dynamics constraint at high β. Code and experiment logs: https://github.com/Volta-0936/sphertt (archived: DOI 10.5281/zenodo.21433254). Companion paper: DOI 10.5281/zenodo.21433291.

•Jul 18, 2026•0 citations
Kinetic descriptions of fusion plasmas are throttled by the curse of dimensionality: a distribution function on a six-dimensional phase space is intractable to store on a dense grid, let alone to advance in time. We show that the kinetic distribution function is strongly compressible in a low-rank tensor-network (matrix-product-state, or tensor-train) representation and that this yields a practical solver on present-day classical hardware, not merely a storage economy. We cast the problem formally: the discretized distribution is a high-order tensor, its matrix-product-state factorization is exact up to a truncation error bounded by the discarded Schmidt weight (Eckart–Young), and the number of retained modes needed for a target accuracy scales with the entanglement entropy of the state. On a one-dimensional-in-space, one-dimensional-in-velocity Bhatnagar–Gross–Krook (BGK) kinetic test problem, discretized on a periodic N_x=128× N_v=64 grid and advanced by operator-split upwind streaming and implicit relaxation, a matrix-product-state truncation reaches a relative-L^2 error of 2.9×10^-6 at bond dimension =8 while storing 0.19× the dense grid, and the error falls exponentially in the bond dimension — from 1.8×10^-1 at =1 to 2.2×10^-10 at =16 and to the double-precision floor (~10^-14) by =24. The velocity moments inherit the same convergence (density error 1.1×10^-6, temperature error 2.8×10^-6 at =8), and the converged distribution carries a Schmidt entanglement entropy below 0.1 nats — the area-law signature that makes it compressible. Every headline quantity is a frozen anchor in the Kronos de-risking register (gate KX-L1-A16). Beyond the anchor we give the general-dimensional machinery that carries the method to the five- and six-dimensional gyrokinetic problem — matrix-product-operator representations of the streaming and collision maps, dimension-linear tensor-train construction by cross interpolation, and a rank-adaptive, conservative dynamical-low-rank integrator that evolves the state on the low-rank manifold without inverting small singular values — together with the derived exponential-to-linear crossover in both storage and per-step arithmetic that motivates it. The result is a classical, deployable-today complement to the fault-tolerant-horizon quantum kinetic route, and we place it as a cross-cutting, verification-first module inside the Kronos autonomous control architecture: an advisory/supervisory reduced-kinetics solver, exercised twin-first beneath a deterministic safety clamp, that hands authority back to the deterministic floor on out-of-distribution inputs. The single honest gate is that the demonstration is a collision-dominated reduced problem, and the rank growth of the fully nonlinear, collisionless six-dimensional gyrokinetic distribution at reactor parameters remains the open question we scope explicitly. Key results (frozen anchors): v = 2 v; t = 3.27 ×10^-3; sigma_1 = 9.4 ×10^-2; sigma_1 = 2.5 ×10^-3; sigma_1 = 4.7 ×10^-5. Methods & codes: CGYRO, TGLF, GENE, Sauter, VQE, ADAPT-VQE. Live verification: 1 gate validator(s) with live-recompute cards (1 reproduced). See the Verification section and data/verification.csv. Related: Paper page · De-risking register · 3D model · Learn more about Kronos Part of the 2026 Kronos publication series; independently re-run and stamped in the Kronos de-risking register (DOI 10.5281/zenodo.22645689). All numerical values are frozen design-point anchors; see the register. Public research artifact. No proprietary, financial, or supply-chain information is included. Previous version: DOI 10.5281/zenodo.22132192. Please cite this version.
Kinetic descriptions of fusion plasmas are throttled by the curse of dimensionality: a distribution function on a six-dimensional phase space is intractable to store on a dense grid, let alone to advance in time. We show that the kinetic distribution function is strongly compressible in a low-rank tensor-network (matrix-product-state, or tensor-train) representation and that this yields a practical solver on present-day classical hardware, not merely a storage economy. We cast the problem formally: the discretized distribution is a high-order tensor, its matrix-product-state factorization is exact up to a truncation error bounded by the discarded Schmidt weight (Eckart–Young), and the number of retained modes needed for a target accuracy scales with the entanglement entropy of the state. On a one-dimensional-in-space, one-dimensional-in-velocity Bhatnagar–Gross–Krook (BGK) kinetic test problem, discretized on a periodic N_x=128× N_v=64 grid and advanced by operator-split upwind streaming and implicit relaxation, a matrix-product-state truncation reaches a relative-L^2 error of 2.9×10^-6 at bond dimension =8 while storing 0.19× the dense grid, and the error falls exponentially in the bond dimension — from 1.8×10^-1 at =1 to 2.2×10^-10 at =16 and to the double-precision floor (~10^-14) by =24. The velocity moments inherit the same convergence (density error 1.1×10^-6, temperature error 2.8×10^-6 at =8), and the converged distribution carries a Schmidt entanglement entropy below 0.1 nats — the area-law signature that makes it compressible. Every headline quantity is a frozen anchor in the Kronos de-risking register (gate KX-L1-A16). Beyond the anchor we give the general-dimensional machinery that carries the method to the five- and six-dimensional gyrokinetic problem — matrix-product-operator representations of the streaming and collision maps, dimension-linear tensor-train construction by cross interpolation, and a rank-adaptive, conservative dynamical-low-rank integrator that evolves the state on the low-rank manifold without inverting small singular values — together with the derived exponential-to-linear crossover in both storage and per-step arithmetic that motivates it. The result is a classical, deployable-today complement to the fault-tolerant-horizon quantum kinetic route, and we place it as a cross-cutting, verification-first module inside the Kronos autonomous control architecture: an advisory/supervisory reduced-kinetics solver, exercised twin-first beneath a deterministic safety clamp, that hands authority back to the deterministic floor on out-of-distribution inputs. The single honest gate is that the demonstration is a collision-dominated reduced problem, and the rank growth of the fully nonlinear, collisionless six-dimensional gyrokinetic distribution at reactor parameters remains the open question we scope explicitly.Key results (frozen anchors): v = 2 v; t = 3.27 ×10^-3; N_v = 128 ×64; sigma_1 = 9.4 ×10^-2; sigma_1 = 2.5 ×10^-3; sigma_1 = 4.7 ×10^-5.Methods & codes: CGYRO, TGLF, GENE, Sauter, VQE, ADAPT-VQE.Live verification: 1 gate validator(s) with live-recompute cards (1 reproduced). See the Verification section and data/verification.csv.Related: Paper page · De-risking register · 3D model · Learn more about KronosPart of the 2026 Kronos publication series; independently re-run and stamped in the Kronos de-risking register (DOI 10.5281/zenodo.22645689).All numerical values are frozen design-point anchors; see the register.Public research artifact. No proprietary, financial, or supply-chain information is included.Note (REPLACE): This record replaces and supersedes DOI 10.5281/zenodo.22132192; please cite this version.
Characteristic functions of weighted sums of independent random variables exhibit low-rank structure in the quantized tensor train (QTT) representation, also known as matrix product states (MPS), enabling up to exponential compression of their fully non-Gaussian probability distributions. Under variable independence, the global characteristic function factorizes into local terms. Its low-rank QTT structure arises from intrinsic spectral smoothness in continuous models, or from spectral energy concentration as the number of components DDD grows in discrete models. We demonstrate this on weighted sums of Bernoulli and lognormal random variables. In the former, despite an adversarial, incompressible small-DDD regime, the characteristic function undergoes a sharp bond-dimension collapse for D≳300D \gtrsim 300D≳300 components, enabling polylogarithmic time and memory scaling. In the latter, the approach reaches high-resolution discretizations of N=230N = 2^{30}N=230 frequency modes on standard hardware, far beyond the N=224N = 2^{24}N=224 ceiling of dense implementations. These compressed representations enable efficient computation of Value at Risk (VaR) and Expected Shortfall (ES), supporting applications in quantitative finance and beyond.
This record has been superseded. The current, updated version is A Tensor-Network Kinetic Solver for Fusion Transport: Matrix-Product-State Compression of the Distribution Function with Exponential Error Decay (DOI 10.5281/zenodo.22645768). Please read and cite the current version. This page stays online so existing citations keep resolving. The kinetic distribution function is highly compressible in a low-rank tensor-network representation, and that yields a practical classical solver, not a storage trick. On a 1D1V BGK test, a matrix-product-state truncation reaches relative-L² error 2×10−⁴ at rank 8 using ~0.19× the dense storage, with the error falling exponentially in rank. Unlike the fault-tolerant-horizon quantum route, this runs today.
Tensor networks (TNs), originally developed for simulating many-body quantum systems, provide a systematic framework for approximating high-dimensional fields. This is achieved by factorizing the field into interconnected tensors with small bond dimensions, thereby restricting the correlations captured across field bipartitions. Belonging to the family of TNs, the matrix product state (MPS) ansatz is utilized here as a reduced-order modeling framework to construct truncated representations of isotropic turbulent flow data. Two direct numerical simulation (DNS) datasets are considered: the hydrodynamic field of an incompressible three-dimensional flow, and a conserved Fickian scalar in a similar flow. Each field is encoded as an MPS through a sequence of singular value decompositions (SVDs) in which small singular values are discarded. The truncated representation is contracted back to the full grid, and the resulting reconstructed field is compared against DNS. An interleaved ordering of the spatial tensor indices of the transport variables is applied prior to decomposition in order to localize the dominant inter-tensor correlations. Velocity reconstructions achieve 99.8%99.8\%99.8% fidelity using only 5%5\%5% of the original DNS memory, while the scalar field reaches the same fidelity at 15%15\%15% memory usage. A wide range of lower- and higher-order statistics, including velocity gradients, dissipation, and structure functions, are systematically examined. At these compression levels, the total kinetic energy and the scalar energy are both recovered within 0.2%0.2\%0.2% relative error, while the mean dissipation and mean scalar dissipation remain within approximately 10%10\%10% of the DNS generated values. These findings support the suitability of MPS for scalable reduced-order analysis of complex turbulent datasets and motivate further exploration of TN-based methods in computational turbulence.
Kinetic simulations of collisionless (or weakly collisional) plasmas using the Vlasov equation are often infeasible due to high-resolution requirements and the exponential scaling of computational cost with respect to dimension. Recently, it has been proposed that matrix product state (MPS) methods, a quantum-inspired but classical algorithm, can be used to solve partial differential equations with exponential speed-up, provided that the solution can be compressed and efficiently represented as a MPS within some tolerable error threshold. In this work, we explore the practicality of MPS methods for solving the Vlasov-Poisson equations for systems with one coordinate in space and one coordinate in velocity, and find that important features of linear and nonlinear dynamics, such as damping or growth rates and saturation amplitudes, can be captured while compressing the solution significantly. Furthermore, by comparing the performance of different mappings of the distribution functions onto the MPS, we develop an intuition of the MPS representation and its behavior in the context of solving the Vlasov-Poisson equations, which will be useful for extending these methods to higher-dimensional problems.
Compressing the connectivity of a reservoir removes the O(N²) memory wall, but the state vector itself still costs O(N) — the state wall, at N = 4¹⁵ > 10⁹ an unallocatable 8.6 GB. We introduce a reservoir whose state evolves natively in tensor-train (TT) form: TT matrix–vector product, TT input injection, rounding to bond dimension χx, and the hypersphere normalization, which is a scalar division of one core and therefore exactly rank-neutral. When χx reaches the maximal rank of the tensorization the evolution is exact to machine precision; below it, the reservoir is not a degraded copy of the exact one. We show that the exact attractor of sphere-TT dynamics at large N is near-maximally entangled (one-shot truncation error 0.93 at χx = 16, N = 65,536; flat middle-bond Schmidt spectrum), so post-hoc state compression is impossible; the truncated system instead carves out its own low-rank attractor, characterized by a per-step rounding error δ̄ that is measurable at runtime for free. Two quantitative laws organize this regime. First, a decomposition law: δ̄² ≈ (c·(1−β))² + δstock², fitting a 5×3 grid in the orthogonal-mixing weight β and χx within 2–5%; the flow term is injected by the random part of the connectivity, while the floor δstock — measured directly at β = 1, where the rank-1 Kronecker-orthogonal connectivity injects nothing — is the rank cost of the retained memory itself. Second, a two-resource memory law: linear memory capacity follows MC ≈ min{K₀·χx², MC_noise(δ̄)}, with a threshold regime, a quadratic capacity regime (measured MC/χx² constant across χx at fixed β), and a noise-limited regime where δ̄ saturates at the flow floor. Following these laws — raise β and χx together — we demonstrate memory capacity 30–74 (three seeds) at N = 1,073,741,824, where the dense state cannot be allocated, with the entire model (connectivity, state, input coupling) in 861 KB at ~1 s/step on a CPU; the naive configuration scores ≈ 0.1 there. A nonlinear task (NARMA10) reaches r² = 0.64 in the same regime. We state the open problems honestly: the microscopic mechanism of the χ_x² law, the growth of δ̄ with tensor order, and a residual quality-of-dynamics constraint at high β. Code and experiment logs: https://github.com/Volta-0936/sphertt (archived: DOI 10.5281/zenodo.21433254). Companion paper: DOI 10.5281/zenodo.21433291.