Pith. sign in

REVIEW 2 major objections 4 minor 30 references

Rake-Compress Riccati Recursions for Parallel Scenario-Tree Model Predictive Control

T0 review · 2 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Rake–compress tree contraction solves the branched dual-regularized LQR KKT system exactly, with O(N) work and O(log N) span on any fixed scenario-tree topology.

desk verdict First exact rake-compress factorization for branched LQR with O(log N) span; solid algebra, but the nonlinear MPC demo never verifies the cone condition. read the letter →

arxiv 2608.01332 v1 pith:R6Z435XO submitted 2026-08-02 math.OC cs.RO

João Sousa-Pinto This is my paper · ORCID
classification math.OCcs.RO MSC 93B4090C5565Y05
keywords parallelalgorithmsRiccatiequationsscenario-treemodelpredictivecontroltreecontractiondualregularizationinterior-pointmethodsKKTsystemslogarithmicspan
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that the repeated linear-algebra kernel in scenario-tree model predictive control—a branched, dual-regularized linear-quadratic regulator—can be solved exactly in parallel using rake–compress tree contraction. Two local eliminations, one removing a leaf and one removing a unary node, are closed on a fixed-dimensional representation of conditional quadratic value functions, so contracting any tree to its root factors the KKT system without flattening the tree into paths or forming a dense global system. Reversing the contraction recovers every Riccati coefficient, state, control, and multiplier. The paper proves that, on a reusable topology plan, each factorization and each solve uses $O(N)$ work and storage and $O(\log N)$ span, independent of tree height, balance, and maximum out-degree. If true, every search-direction solve in scenario-tree MPC could run in logarithmic parallel depth instead of depth proportional to tree height.

What carries the argument

The load-bearing object is the conditional quadratic edge function $$\Phi_e(x_i,x_j)=\tfrac12 x_i^TP_e x_i + \sup_\$\lambda$\{\$\lambda$^T(A_e x_i-x_j)-\tfrac12\$\lambda$^TC_e\$\lambda$\}$$ together with the leaf-elimination update $$\tau(\alpha_e,U_j)=P_e+A_e^TU_j(I+C_eU_j)^{-1}A_e$$ and the unary-chain composition $\alpha_{e\ell r}=\alpha_{e_r}\circ_{U_j}\alpha_{e_\ell}$. The update formulas have dimensions fixed by the state and control sizes rather than by the eliminated subtree, and they remain defined when a $C_e$ matrix is singular. Closure of this algebra under both eliminations is what lets rake–compress contraction reduce an arbitrary tree to its root without changing the KKT solution; associativity of the chain composition is what yields logarithmic dependency depth.

What would settle it

Measure the number of primitive dependency levels in a faithful implementation of the paper's scheduler on a chain of $N$ nodes for $N=16,64,256,1024,4096$. The promised $O(\log N)$ span predicts levels near $1+\lceil\log_2 N\rceil$—the paper's own table reports 5, 7, 9, 11, 13—so any linear growth in $N$ would refute Theorem 4.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that the KKT system of a branched dual-regularized LQR on an arbitrary rooted tree is factorized exactly by two algebraic eliminations. A rake removes a nonroot leaf and its parent edge, adding that subtree's condensed contribution to the parent node data; a compress removes a unary node and fuses its two incident edges into one edge by composing conditional quadratics. The edge data at every stage are triples $(A_e,C_e,P_e)$ representing an extended-real quadratic function of the endpoint states, with positive-semidefinite $C_e$ allowed, so exact dynamics constraints are included. The paper proves the rake and compress identities by saddle-point elimination, proves all intermediate data stay well posed under the standing semidefinite assumptions, and proves that reverse expansion recovers the unique KKT solution for any right-hand side. The end-to-end result is an algebraically exact factorization-and-solve with linear work and storage and logarithmic span for fixed local dimensions.

Load-bearing premise

The entire argument assumes that for every edge $R_e \succ 0$, every node $\Delta_i \succeq 0$, and $Q_i - \sum_{e\in\text{out}(i)} M_e R_e^{-1} M_e^T \succeq 0$; if any Newton subproblem in the interior-point usage leaves this cone, the local factorizations can become singular.

Editorial extensions

If this is right

  • A single factorization can serve many right-hand sides, so in interior-point scenario-tree MPC every Newton search-direction solve inherits $O(N)$ work and $O(\log N)$ span for fixed state and control dimensions.
  • The standard scenario-MPC convention of one control per information node is covered without duplicating decisions: the paper's exact lifting has at most $2N-1$ nodes and preserves linear size and logarithmic span.
  • Positive-semidefinite dual regularization, including the unregularized case, is admitted by the algebra, so the solver applies both to interior-point Newton systems and to pure dynamics-constrained LQ problems.
  • The dependency depth is independent of tree height, balance, and maximum out-degree, so a chain with a long trunk, a high-degree star, and an irregular comb all reduce in the same logarithmic number of primitive levels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the contraction schedule is determined solely by the parent array, a topology plan should also be reusable across a whole MPC run and across any problem instance that keeps the same branching structure but changes the dynamics or costs; the paper states reuse only while the parent array is unchanged, but the separation of quadratic coefficients from right-hand sides makes extension immedia
  • The logarithmic span is an algorithmic dependency bound, not a promise of wall-clock speedup at every size; the reported crossover, where the GPU is slower at small trees and faster at larger trees, suggests the practical benefit appears only when enough parallelism exists to amortize launch and scheduling overhead.
  • The readiness-weighted sibling reduction uses an alphabetic prefix code keyed to input arrival levels; the same scheduling recipe could apply to other associative tree accumulations whose inputs become ready at different times, not only Riccati data.
  • For nonlinear scenario-tree MPC, the practical bottleneck may move from the linear solve to keeping Newton subproblems inside the standing-convexity cone; combining this solver with an explicit Hessian-modification or regularization rule would be a natural next step.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper presents a parallel solver for the KKT systems arising in dual-regularized branched linear-quadratic regulator (LQR) problems on scenario trees. The method eliminates controls locally, then applies rake and compress tree-contraction operations to conditional quadratic data, and reverses the contraction to recover all Riccati coefficients, states, controls, and multipliers. The main theoretical results (Lemma 1, Propositions 1–5, Theorems 1–4) state that, under the standing convexity assumptions (Definition 1), the contraction identities are exact, the solver is equivalent to the original KKT system, and with a reusable topology plan each solve on an N-node tree uses O(N) work and storage and O(log N) span, independent of tree height, balance, and out-degree. An exact linear-size lifting converts the node-control scenario-MPC form to the edge-control form. The paper validates the LQR solver against dense KKT systems, reports contraction-depth measurements, benchmarks CPU vs. GPU performance, and demonstrates the full stack on a nonconvex autonomous-driving scenario-tree MPC problem solved with the companion primal-dual interior-point package.

Significance. If the conditional claims hold, this is a significant contribution to parallel optimal control: it gives a work-efficient, algebraically exact factorization and solve for tree-structured LQR with logarithmic span independent of the tree's shape, together with an explicit and reusable contraction schedule. The paper is unusually complete on the linear-algebra side: proofs are self-contained, there are no fitted parameters, the dense-KKT validation is convincing, and the three MIT-licensed JAX packages make the method reproducible. The main caveat is that the correctness proofs are conditional on Definition 1, and the paper's end-to-end scenario-tree MPC demonstration does not verify that the nonconvex interior-point subproblems satisfy that condition; this does not undermine the LQR core but does limit the application claim as currently stated.

major comments (2)
  1. [Section X and Section VIII] The end-to-end claim that the rake–compress solver parallelizes search-direction solves in scenario-tree MPC requires each Newton subproblem to satisfy Definition 1, Eq. (3); the paper neither proves that primal-dual-lipa maintains this condition nor reports any check that it held during the 53 iterations of the driving example. For the nonconvex kinematic dynamics (v cos ψ, v sin ψ), the Lagrangian Hessian can be indefinite, and the dual regularization Δ_i enters only the (2,2) block of the KKT matrix, not the primal Schur complement in Eq. (3). The paper should report the minimum eigenvalue of Qi − Σ Me Re^{-1} Me^T across all Newton iterations and nodes, or specify and verify a Hessian modification that enforces (3), or restrict the Introduction and Conclusion claims to Newton subproblems that are known to satisfy Definition 1.
  2. [Section II, Definition 1] The standing convexity assumption in Eq. (3) is strictly stronger than well-posedness of the LQR problem: a one-step chain with Q=0, M=1, R=1 has a nonsingular KKT matrix but violates (3). The abstract's phrase 'allows positive-semidefinite dual regularization, including the unregularized case' therefore overstates the scope. The paper should state explicitly that (3) is not necessary, describe the class of problems covered, and indicate what happens (e.g., indefinite P_i or singular factors) when it is violated, so that users of the released software can detect the failure.
minor comments (4)
  1. [Section III, first paragraph] The phrases 'Arakeeliminates' and 'Acompresseliminates' are missing spaces; please correct to 'A rake eliminates' and 'A compress eliminates'.
  2. [Section IX.A] The description of the 26 test instances does not state how many, if any, had singular Ce or zero Δi; since these are allowed by Definition 1 and are the cases the LQR solver is claimed to handle, a brief breakdown would strengthen the validation.
  3. [Section X] The reported final residual (9.44e-8) and barrier parameter (7.50e-11) are given without the solver's termination tolerance; please state the stopping criterion and confirm the residual is consistent with it.
  4. [Section VII, Eq. (51)] The nested ceilings in D_N are correct but slightly hard to read; a parenthetical explanation of why the outer ceiling is needed (ℓ(z) is an integer while the bound is not) would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the parallel Riccati derivation, exactness proofs, and complexity bounds are self-contained; the self-citation to [16] is contextual, and the cone condition is a stated assumption rather than a fitted input.

full rationale

The paper's central derivation is not circular. The rake and compress updates in Section IV-A (Eqs. (15)-(17)) are proved directly from the saddle function in Eq. (1) through Lemma 1, and the reverse expansion is proved in Theorem 1 from the partial-elimination invariant; no update formula is imported as a black box. The complexity theorems (Lemmas 3-5 and Theorem 4) are proved for the paper's exact deterministic schedule, and the paper explicitly states that Miller-Reif and Gazit-Miller-Teng theorems cannot be applied directly and that Section VII proves the needed bounds. The KKT equivalence (Theorems 1-3) and the node-control lifting (Proposition 2) are proven from the stated assumptions. There are no fitted parameters, and the numerical validation compares against an independently assembled dense KKT system. The self-citation to [16] supplies motivation and the interior-point context, but the chain algebra it cites is re-derived in this paper, so it is not load-bearing. The only caveat, which is a correctness gap rather than circularity, is that the correctness theorems are explicitly conditional on Definition 1, Eq. (3), and the nonlinear example's interior-point Newton subproblems are not shown to satisfy that cone condition; this affects the breadth of the end-to-end claim but does not make the derivation equivalent to its inputs.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No free parameters are fitted to data, and no physical or unexplained entities are introduced. The load-bearing assumptions are the convexity cone of Definition 1 and the CREW PRAM dependency model behind the O(log N) span claim. The conditional quadratic edge triples are internal mathematical representations, not free postulates.

assumptions (4)
  • domain assumption Standing convexity: Re > 0, Delta_i >= 0, and Qi - sum_{e in out(i)} Me Re^{-1} Me^T >= 0 for every node i.
    Definition 1, Eq. (3). This drives KKT nonsingularity and the well-posedness of every local factorization and contraction update.
  • standard math The chain conditional-value composition of Särkkä and García-Fernández [7] is associative and can be used at each unary node.
    Section I-A and IV-A. The paper re-proves the needed identities in Lemma 1, so this is background support rather than an unverified black box.
  • domain assumption Complexity is measured on a CREW PRAM with unlimited parallelism and bounded-arity primitive operations; actual JAX/XLA hardware span may be larger.
    Section VII-B and VIII state that the span is an algorithmic dependency bound, not a one-device wall-clock guarantee; JAX/XLA may lower one logical level to multiple kernels.
  • standard math Shannon-Fano-Elias coding gives an order-preserving prefix-free binary tree used for readiness-weighted reductions.
    Lemma 4, Section VII-B; this is a standard information-theoretic construction from Cover and Thomas [28].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rake-Compress Riccati Recursions for Parallel Scenario-Tree Model Predictive Control." pith.science (2026). https://pith.science/paper/R6Z435XO

@misc{pith2026260801332,
  author       = {Pith},
  title        = {Pith review of: Rake-Compress Riccati Recursions for Parallel Scenario-Tree Model Predictive Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R6Z435XO}},
  note         = {Machine review of arXiv:2608.01332}
}
abstract

Scenario-tree model predictive control (MPC) represents future information by a rooted tree and optimizes a nonanticipative policy over that tree. Numerical methods for solving the resulting nonlinear program typically compute their search directions through a sequence of branched linear-quadratic regulator (LQR) subproblems. The standard tree Riccati recursion requires linear work but has a dependency chain proportional to tree height. We present an algebraically exact parallel solver based on rake-compress tree contraction. After independent local control condensation, its two operations act on node and edge data that represent conditional quadratic functions. A rake eliminates a leaf and its parent edge, adding their reduced contribution to the parent-node data. A compress eliminates a unary node and replaces its two adjacent edges by one edge, using the same conditional-value composition as parallel Riccati methods on a chain. Together they contract an arbitrary rooted tree to its root; reversing the contraction recovers every Riccati coefficient, state, control, and multiplier. Given a reusable topology plan, a solve with $N$ nodes and fixed state and control dimensions has $O(N)$ arithmetic work and storage and $O(\log N)$ span, independently of tree height, balance, and maximum out-degree. The formulation allows positive-semidefinite dual regularization, including the unregularized case, and an exact linear-size lifting covers the standard scenario-MPC convention of one control per information node. We prove the contraction identities and equivalence to the Karush-Kuhn-Tucker (KKT) system. Three MIT-licensed JAX packages implement the bidirectional contraction, the dual-regularized LQR solver, and a user-facing primal-dual interior-point solver for tree-structured optimal control.

Figures

Figures reproduced from arXiv: 2608.01332 by the authors.

Figure 1
Figure 1. figure 1 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 1
Figure 1. The two rake–compress primitives. Orange nodes are removed and blue nodes remain in the tree. Reverse expansion later reconstructs each removed node from the data saved during contraction. (a) Chain ρ C1 C2 C1 R3 R2 R1 three structural rounds (b) Star ρ R1 R1 R1 R1 R1 R1 one structural round (c) Balanced binary tree ρ R2 R2 R1 R1 R1 R1 two structural rounds (d) Irregular comb ρ C1 R3 R1 C1 R2 R1 R1 three structural … view at source ↗
Figure 2
Figure 2. Representative forward contraction schedules. Orange nodes are raked, violet nodes are compressed, and the subscript is the structural round. Within a round, rakes precede compressions. The root remains after the displayed removals, and reverse expansion follows the labels in decreasing dependency order. nodes raked and compressed in structural round t; nodes with the same label can be processed concurrently. This s… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Irregular information tree in the autonomous-driving example. Leaves are ordered depth first; line width is proportional to the square root of scenario probability, white circles mark branching nodes, and colored dots mark leaves. L, h, and D denote the numbers of leav…
Figure 3
Figure 3. Figure 3: Bird’s-eye schematic and optimized ego-policy paths. At stage 4, the adjacent vehicle yields or merges, or the lead vehicle brakes. Later branches resolve the adjacent-vehicle reaction, road surface, and cyclist entry. Vehicle glyphs and annotations are schematic; the …

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 28 canonical work pages

  1. [7]

    Temporal parallelization of dynamic programming and linear quadratic control,

    S. Särkkä and Á. F. García-Fernández, “Temporal parallelization of dynamic programming and linear quadratic control,”IEEE Transactions on Automatic Control, vol. 68, no. 2, pp. 851–866, 2023

    work page 2023

  2. [14]

    A parallel structure exploiting factorization algorithm with applications to model predictive control,

    I. Nielsen and D. Axehill, “A parallel structure exploiting factorization algorithm with applications to model predictive control,” in54th IEEE Conference on Decision and Control, 2015, pp. 3932–3938

    work page 2015

  3. [15]

    Parallelizing LQR computation through endpoint-explicit Riccati recursion,

    F. Laine and C. Tomlin, “Parallelizing LQR computation through endpoint-explicit Riccati recursion,” in58th IEEE Conference on Decision and Control, 2019, pp. 1395–1402

    work page 2019

  4. [25]

    Parallel Branch Model Predictive Control on GPUs

    L. Zhang, C. Lin, and S. Grammatico, “Parallel branch model predictive control on GPUs,”arXiv preprint arXiv:2506.13624, 2025

    work page Pith review arXiv 2025

  5. [1]

    Stochastic MPC for real- time market-based optimal power dispatch,

    P. Patrinos, S. Trimboli, and A. Bemporad, “Stochastic MPC for real- time market-based optimal power dispatch,” in50th IEEE Conference on Decision and Control and European Control Conference, 2011, pp. 7111–7116

    work page 2011

  6. [2]

    Dual stochastic MPC for systems with parametric and structural uncertainty,

    E. Arcari, L. Hewing, M. Schlichting, and M. N. Zeilinger, “Dual stochastic MPC for systems with parametric and structural uncertainty,” inProceedings of the 2nd Conference on Learning for Dynamics and Control, ser. Proceedings of Machine Learning Research, vol. 120. PMLR, 2020, pp. 894–903

    work page 2020

  7. [3]

    Interactive multi-modal motion planning with branch model predictive control,

    Y . Chen, U. Rosolia, W. Ubellacker, N. Csomay-Shanklin, and A. D. Ames, “Interactive multi-modal motion planning with branch model predictive control,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 5365–5372, 2022

    2022

  8. [4]

    Linear quadratic control compu- tation for systems with a directed tree structure,

    A. Zafar, F. Farokhi, and M. Cantoni, “Linear quadratic control compu- tation for systems with a directed tree structure,”IFAC-PapersOnLine, vol. 53, no. 2, pp. 6536–6541, 2020

    work page 2020

Show all 30 references
  1. [5]

    Application of interior-point methods to model predictive control,

    C. V . Rao, S. J. Wright, and J. B. Rawlings, “Application of interior-point methods to model predictive control,”Journal of Optimization Theory and Applications, vol. 99, no. 3, pp. 723–757, 1998

    1998

  2. [6]

    A high- performance Riccati based solver for tree-structured quadratic programs,

    G. Frison, D. Kouzoupis, M. Diehl, and J. B. Jørgensen, “A high- performance Riccati based solver for tree-structured quadratic programs,” IFAC-PapersOnLine, vol. 50, no. 1, pp. 14 399–14 405, 2017

    2017

  3. [8]

    jax-bidirectional-tree-rake-compress: Bidirectional direct rooted-tree rake–compress for JAX,

    J. Sousa-Pinto, “jax-bidirectional-tree-rake-compress: Bidirectional direct rooted-tree rake–compress for JAX,” Project repository, 2026

    2026

  4. [9]

    regularized_lqr_jax: Dual-regularized linear-quadratic control in JAX,

    ——, “regularized_lqr_jax: Dual-regularized linear-quadratic control in JAX,” Project repository, 2026

    2026

  5. [10]

    primal-dual-lipa: Primal–dual interior-point optimization in JAX,

    ——, “primal-dual-lipa: Primal–dual interior-point optimization in JAX,” Project repository, 2026, open-source software

    2026

  6. [11]

    Tree-sparse convex programs,

    M. C. Steinbach, “Tree-sparse convex programs,”Mathematical Methods of Operations Research, vol. 56, no. 3, pp. 347–376, 2002

    2002

  7. [12]

    A Riccati-based primal interior point solver for multistage stochastic programming,

    J. Blomvall and P. O. Lindberg, “A Riccati-based primal interior point solver for multistage stochastic programming,”European Journal of Operational Research, vol. 143, no. 2, pp. 452–461, 2002

    2002

  8. [13]

    A dual Newton strategy for tree-sparse quadratic programs and its implementa- tion in the open-source software treeQP,

    D. Kouzoupis, E. Klintberg, G. Frison, S. Gros, and M. Diehl, “A dual Newton strategy for tree-sparse quadratic programs and its implementa- tion in the open-source software treeQP,”International Journal of Robust and Nonlinear Control, vol. 29, no. 8, pp. 2438–2457, 2019

    2019

  9. [16]

    Dual-regularized Riccati recursions for interior-point optimal control,

    J. Sousa-Pinto and D. Orban, “Dual-regularized Riccati recursions for interior-point optimal control,”arXiv preprint arXiv:2509.16370, 2025

    2025

  10. [17]

    A dual Newton strategy for scenario decomposition in robust multistage MPC,

    D. Kouzoupis, E. Klintberg, M. Diehl, and S. Gros, “A dual Newton strategy for scenario decomposition in robust multistage MPC,”Inter- national Journal of Robust and Nonlinear Control, vol. 28, no. 6, pp. 2340–2355, 2018

    2018

  11. [18]

    Massively parallelizable proximal algorithms for large-scale stochastic optimal control problems,

    A. K. Sampathirao, P. Patrinos, A. Bemporad, and P. Sopasakis, “Massively parallelizable proximal algorithms for large-scale stochastic optimal control problems,”Optimal Control Applications and Methods, vol. 45, no. 1, pp. 45–63, 2024

    2024

  12. [19]

    A framework for the solution of tree-coupled saddle-point systems,

    C. Hansknecht, B. Heinzelreiter, J. W. Pearson, and A. Potschka, “A framework for the solution of tree-coupled saddle-point systems,” Numerical Linear Algebra with Applications, vol. 32, no. 6, p. e70038, 2025

    2025

  13. [20]

    Parallel tree contraction and its application,

    G. L. Miller and J. H. Reif, “Parallel tree contraction and its application,” in26th Annual Symposium on Foundations of Computer Science, 1985, pp. 478–489

    1985

  14. [21]

    Parallel tree contraction, part 1: Fundamentals,

    ——, “Parallel tree contraction, part 1: Fundamentals,” inRandomness and Computation, ser. Advances in Computing Research, S. Micali, Ed. JAI Press, 1989, vol. 5, pp. 47–72

    1989

  15. [22]

    Optimal tree contraction in the EREW model,

    H. Gazit, G. L. Miller, and S.-H. Teng, “Optimal tree contraction in the EREW model,” inConcurrent Computations: Algorithms, Architecture, and Technology, S. K. Tewksbury, B. W. Dickinson, and S. C. Schwartz, Eds. New York: Plenum Press, 1988, pp. 139–156

    1988

  16. [23]

    Logarithmic-time updates and queries in probabilistic networks,

    A. L. Delcher, A. J. Grove, S. Kasif, and J. Pearl, “Logarithmic-time updates and queries in probabilistic networks,”Journal of Artificial Intelligence Research, vol. 4, pp. 37–59, 1996

    1996

  17. [24]

    Adaptive exact inference in graphical models,

    Ö. Sümer, U. A. Acar, A. T. Ihler, and R. R. Mettu, “Adaptive exact inference in graphical models,”Journal of Machine Learning Research, vol. 12, no. 97, pp. 3147–3186, 2011

    2011

  18. [26]

    Alphabetic minimax trees,

    D. G. Kirkpatrick and M. M. Klawe, “Alphabetic minimax trees,”SIAM Journal on Computing, vol. 14, no. 3, pp. 514–526, 1985

    1985

  19. [27]

    Fast prefix adders for non-uniform input arrival times,

    S. Held and S. Spirkl, “Fast prefix adders for non-uniform input arrival times,”Algorithmica, vol. 77, no. 1, pp. 287–308, 2017

    2017

  20. [28]

    T. M. Cover and J. A. Thomas,Elements of Information Theory, 2nd ed. Hoboken, NJ: Wiley-Interscience, 2006

    2006

  21. [29]

    JAX: Composable transformations of Python+NumPy programs,

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, Y . Katariya, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, and Q. Zhang, “JAX: Composable transformations of Python+NumPy programs,” Software, project repository, 2018

    2018

  22. [30]

    Reachability- based contingency planning against multi-modal predictions with branch MPC,

    M.-K. Bouzidi, B. Derajic, D. Goehring, and J. Reichardt, “Reachability- based contingency planning against multi-modal predictions with branch MPC,” in2025 IEEE Intelligent Vehicles Symposium, 2025, pp. 1086– 1093. João Sousa-PintoJoão Sousa-Pinto has been working on motion p...

    2013

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.