Pith. sign in

REVIEW 2 major objections 5 minor 61 references

Preconditioning transformations of adjoint systems for evolution equations

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adjoint systems for evolution equations can be preconditioned by symplectic transformations without losing the adjoint's ability to backpropagate objective derivatives, and that a scale preconditioner turns an…

desk verdict A genuinely new symplectic framework for adjoint preconditioning with a solid theoretical core; the numerical payoff is real but only demonstrated under a manually tuned scale parameter. read the letter →

arxiv 2505.21705 v1 pith:QIIDPVEC submitted 2025-05-27 math.OC cs.NAmath.NA

Brian K. Tran , Ben S. Southworth , Hannah F. Blumhoefer , Samuel Olivier This is my paper
classification math.OCcs.NAmath.NA MSC 49K1565P10
keywords adjointsystemspreconditioningsymplecticgeometryfiberwisedualitypairingscaleradiationdiffusioninverseproblemsadjoint-basedoptimization
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's aim is to give a principled way to precondition adjoint systems of evolution equations, analogous to preconditioning gradient descent, so that gradient-based optimization of differential-equation-constrained problems can be stabilized and accelerated. It develops two families of transformations: those induced by transforming the state dynamics, and those that reshape only the adjoint equation by changing the duality pairing, including state-dependent pairings. The central theoretical claim is that every such transformation is a symplectomorphism of the adjoint Hamiltonian system, and hence preserves the adjoint-variational conservation law $d\langle \xi,\delta u\rangle_{P(u)}/dt=0$, which is what guarantees the preconditioned adjoint still backpropagates the derivative of the objective. The paper then specializes the duality-pairing construction to scale-separated coupled evolution equations and demonstrates it on an inverse problem for the radiation diffusion equations: scale-preconditioned adjoint descent converges in 10–15 iterations with an accurate final-time wavefront, where naive gradient descent diverges at any practical step size. If the framework is right, preconditioning can be moved from the optimizer to the adjoint dynamics themselves, with a guarantee that the gradient returned at the initial time is still the true derivative.

What carries the argument

The load-bearing object is the adjoint Hamiltonian system on $T^*U$: the Hamiltonian $H(t,u,p)=\langle p,f(t,u)\rangle$ with the canonical symplectic form, whose Hamiltonian vector field is the cotangent lift of the state vector field $f$. Preconditioning is a map on $T^*U$; the paper shows each class of maps is an extended symplectic isomorphism (for state-space transformations) or a symplectic isomorphism between different symplectic forms (for duality-pairing transformations), so the Hamiltonian structure and its conserved pairing survive. For the state-dependent pairing $P(u)$, the pullback symplectic form acquires an extra term that is naturally written with connection Christoffel symbols $\Gamma^\beta_{\nu\gamma}(u)=(P(u)^{-1})^\alpha_\nu D_\gamma P^\beta_\alpha(u)$, so the fiberwise-preconditioned adjoint equation takes the invariant form $D\xi/Dt=-[Df(t,u)]^*_{P(u)}\xi$. The conserved quantity $d\langle \xi,\delta u\rangle_{P(u)}/dt=0$ is what carries the argument: it is the precise statement that the preconditioned adjoint still backpropagates derivatives.

What would settle it

Take the one-dimensional system $\dot u=u^2$, choose $P(u)=u^2+1$, and integrate the fiberwise-preconditioned adjoint equation (2.29b) together with the variational equation (2.29c) using the induced scheme (3.20); then compute $\langle \xi,\delta u\rangle_{P(u)}$ at each time step. If the discrete pairing drifts systematically rather than staying constant to round-off, Proposition 2.4 and the induced-scheme conservation property would be refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that the backpropagation property of adjoint systems is a symplectic invariant, not a fragile numerical accident. For a general evolution equation $\dot u=f(t,u)$ with adjoint variable $p$, the adjoint system is Hamilton's equations for $H=\langle p,f(t,u)\rangle$ on the cotangent bundle; the fact that $\langle p,\delta u\rangle$ is conserved along adjoint and variational flows is a consequence of the symplecticity of the Hamiltonian flow. The paper proves that each of its preconditioning transformations—a diffeomorphism of the state space, a global change of duality pairing, or a fiberwise state-dependent pairing $P(u)$—is an extended symplectic isomorphism, and therefore the preconditioned adjoint system satisfies an analogous conservation law. In the fiberwise case (Proposition 2.4) the law is $d\langle \xi,\delta u\rangle_{P(u)}/dt=0$, where the derivative of $P(u)$ along the flow enters the adjoint equation as a connection term that precisely cancels the derivative of the pairing. The paper's numerical claim is that, for a semi-discretized radiation diffusion inverse problem, a constant block-diagonal scale preconditioner makes this theoretical structure pay off: gradient descent with the scale-preconditioned adjoint converges in 10–15 iterations and reproduces the wavefront, whereas naive gradient descent is unstable for any practical step size.

Load-bearing premise

The practical convergence claim assumes that a constant block-diagonal scale preconditioner $S$, whose value is found by sweeping candidates (Figures 6 and 7), is enough to balance the scales of the nonlinear radiation-diffusion adjoint system; the paper offers no automatic rule for choosing $S$, so if every problem needs its own tuning the 10–15 iteration result is a demonstration under favorable parameter selection rather than a robust prediction.

Editorial extensions

If this is right

  • Any of the three preconditioned adjoint systems can replace the standard adjoint inside a gradient-based optimization loop, and the gradient obtained at the initial time remains a true derivative of the terminal objective rather than an approximation.
  • Because the theory admits genuinely nonlinear, state-dependent transformations $P(u)$, adjoint preconditioners can in principle adapt to features that move or change scale over time; the numerical example only exercises the constant-block case.
  • For coupled systems with scale separation, choosing a duality pairing that makes the preconditioned adjoint operator act with the same scaling as the forward operator removes the $O(\alpha^2)$ blow-up that otherwise forces infinitesimal step sizes.
  • The adjoint time-integration scheme induced from the semi-implicit state scheme preserves the discrete pairing identity $\langle p_n,\delta u_n\rangle=\langle p_{n+1},\delta u_{n+1}\rangle$, so structure preservation and practical backward solves are available at the discrete level.
  • In the demonstrated radiation diffusion inverse problem, the expected consequence is stable convergence to an accurate final-time wavefront in 10–15 iterations, where the unpreconditioned adjoint either diverges or makes no progress.

Reading between the lines

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

  • The connection interpretation suggests a diagnostic: since a fiberwise pairing defines a connection on $T^*U\to U$, the curvature or torsion of that connection could predict where a constant scale preconditioner will fail, such as when scale ratios vary strongly in space or time.
  • A natural testable extension is a spatially varying scale preconditioner built from local ratios of the state variables (pointwise $E/T$ in the radiation example) rather than a single global scalar; the paper lists spatially multiscale problems as future work, and the theory in Section 2.2.3 already permits it.
  • The same symplectic machinery may transfer to neural ODE training or other memory-efficient backpropagation settings, where preconditioning the adjoint flow instead of the optimizer could cut iteration counts, but the paper does not test this.
  • The framework's guarantee is continuous-time; the discrete conservation law holds only for the specially induced scheme (3.20), so using an arbitrary adjoint integrator with a preconditioner would forfeit the invariant—this distinction is a caution implied by the paper, not stated as a limitation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper develops a symplectic-geometric framework for preconditioning adjoint systems associated with evolution equations. It distinguishes two classes of transformations: those induced by transforming the state dynamics and those induced by changing the duality pairing, including a fiberwise state-dependent generalization. The central theoretical results are Propositions 2.1, 2.2, and 2.4, which state that each class of preconditioned adjoint system satisfies an adjoint-variational conservation law, interpreted as preservation of the backpropagation property of the adjoint equation. The paper then introduces a scale preconditioner for coupled evolution equations with large scale separation and applies it to an inverse problem for the radiation diffusion equations, reporting convergence in 10-15 iterations.

Significance. If the framework is correct, it provides a clean geometric unification of adjoint preconditioning and a useful conservation-law perspective that is naturally compatible with the authors' earlier work on symplectic adjoint systems. The theoretical derivations are mostly direct coordinate calculations, and the symplectic interpretation, particularly the connection formulation in (2.32)-(2.33), is an appealing contribution. The numerical claim is plausible but, as presented, is a favorable hyperparameter demonstration rather than a robust prediction: the scale preconditioner is selected by sweeping candidate values, no automatic selection rule is given, and the chosen values are not reported. The advertised application therefore needs additional work before the practical significance claimed in the abstract is fully supported.

major comments (2)
  1. [Section 3.3, Eqs. (3.11)-(3.13)] The scaling condition nu/sigma = alpha/beta is inconsistent with the block scalings displayed in (3.11)-(3.13). Reading the matrix as L = [[A,B],[C,D]] with B ~ O(beta) and C ~ O(alpha), the off-diagonal entries of P^{-1} L^* P are P1^{-1} C P2 and P2^{-1} B P1, which scale as (nu/sigma) alpha and (sigma/nu) beta. Requiring these to match the target scales O(beta) and O(alpha) gives nu/sigma = beta/alpha, not alpha/beta. With the printed ratio, the two blocks scale as alpha^2/beta and beta^2/alpha, so the general analysis and Example 3.1, which effectively uses the opposite off-diagonal convention, cannot both be correct as written. Please reconcile the block convention and the ratio; as it stands, the informal derivation does not justify the advertised scale-preconditioning rule.
  2. [Section 4, Figures 6 and 7] The numerical headline of 10-15 gradient-based optimization iterations is not reproducible from the manuscript. The scale preconditioner S is chosen by sweeping candidate values, the actual S values used in Figures 6-9 are not reported, Remark 3.1 explicitly limits the demonstration to constant multiples of the identity, and Section 3.3 supplies no automatic rule for selecting S. In addition, the step size gamma=0.1 is fixed without any sensitivity study. Please report the S values and gamma, describe how S was selected, and provide either a selection rule or an experiment showing that the convergence count is robust over a range of S and gamma. The current Data Availability statement, which offers data only upon reasonable request, does not permit independent verification of the numerical results.
minor comments (5)
  1. [Section 4, page 28] There is a typo in 'thick Marhsak wave'; it should read 'Marshak wave'.
  2. [Abstract and Section 4] The abstract states convergence in '10-15 iterations' while the text says 'O(10) iterations'; please make the counts consistent and state explicitly which figures support the 10-15 range.
  3. [Section 4, Figures 6 and 7] The figure legends say only 'several values of the scale preconditioner'; adding a table with the exact S values would materially improve reproducibility and would help the reader connect the swept values to the convergence curves.
  4. [Section 4] The claim that naive gradient descent is 'unstable for any practical gradient descent step size' is stronger than what the experiments show, since only gamma=0.1 and gamma approximately 1e-21 are tested; please soften the claim or provide a broader stability sweep.
  5. [Data Availability Statement] The heading contains a typo: 'A vailability' should be 'Availability'.

Circularity Check

1 steps flagged · score 3.0 of 10

The symplectic derivation is self-contained; the numerical headline is partly a fitted result because the scale preconditioner is selected by sweeping on the same problem.

  1. fitted input called prediction [Section 4, around eq. (4.5) and Figures 6-7]
    "Due to the large scale-separation between the energyE(t f )∼ O(1024) and temperatureT(t f )∼ O(103), we use the block-diagonal scale preconditioner described in Section 3.3 and compare various choices of values for the scale preconditioner. ... As can be seen in Figure 6 and Figure 7, by utilizing a suitable scale preconditioner, the gradient-based optimization exhibits rapid, stable convergence inO(10) iterations."

    The abstract's headline 'converges in 10-15 iterations' is reported after the free scale parameter has been chosen by comparing convergence plots on that same inverse problem. Section 3.3 explicitly says 'We will proceed informally' and gives no automatic rule for selecting S, and Remark 3.1 restricts the demonstration to constant multiples of the identity; the paper does not report the S values used. The performance claim is therefore conditional on a parameter selected to make the displayed runs converge; it is a favorable hyperparameter demonstration rather than an independent prediction of the framework.

full rationale

The central geometric results are derived, not assumed: Proposition 2.4 is verified by a direct calculation that cancels the extra DP(u) term against the derivative of the fiberwise pairing, and Propositions 2.1-2.2 follow from cotangent-lift and symplectic-isomorphism arguments that are standard. The self-citations to [51] and [53] supply background (cotangent lifts, the basic conservation law) and a uniqueness statement for the discrete adjoint integration scheme; none is the load-bearing premise of the new backpropagation theorem, and the key conservation law is proved in the text rather than merely imported. The only partial circularity is in the numerical demonstration: the scale preconditioner S is not derived by a stated rule for the radiation diffusion example, but is swept across candidate values (Figures 6-7) and then the resulting convergence is reported as the method's behavior. Section 3.3's scaling analysis is explicitly informal, and Remark 3.1 notes the numerical test uses only constant multiples of the identity, so the 10-15 iteration claim is not robustly parameter-free. The data availability statement only offers data 'upon reasonable request,' further limiting independent verification of the S selection. This affects the advertised application's empirical status, not the theoretical derivation. Overall score 3 reflects a mostly self-contained derivation with one empirically fitted input in the headline numerical claim.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The framework's theoretical core rests on standard symplectic geometry and differentiability assumptions. The numerical demonstration adds domain assumptions about the radiation diffusion model and stability of the discretization, plus a heuristic scaling analysis for the scale preconditioner. No new physical entities are introduced.

free parameters (2)
  • scale preconditioner S = several values swept (exact optimal value not given in text)
    The numerical experiments compare various scale-preconditioner values (Figures 6 and 7); the convergence claim is for a value selected from this sweep, so S is fitted to achieve the reported result.
  • gradient descent step size gamma = 0.1
    Used as a constant step size across all runs in Section 4; chosen by hand for the demonstration.
assumptions (5)
  • domain assumption All maps are Frechet differentiable (Section 2.1).
    Stated at the start of Section 2.1; the derivations require differentiability of the vector field, cost function, and duality-pairing map.
  • standard math The canonical symplectic form on T*U and the Hamiltonian flow property are used.
    Standard results from symplectic geometry, invoked in Section 1.1 and Proposition 1.1.
  • domain assumption The radiation diffusion model (4.1) and the given physical parameters accurately describe the thick Marshak wave.
    The numerical example depends on the model, the DG discretization (MFEM), and the specified opacities and diffusion coefficients.
  • domain assumption The induced adjoint time-integration scheme (3.20) is stable and accurate for the radiation diffusion problem.
    The scheme is guaranteed to satisfy the discrete conservation law (3.21) by construction, but stability and accuracy for this stiff nonlinear problem are not proven in the paper.
  • ad hoc to paper The informal scaling analysis of the general block two-by-two system (Section 3.3) justifies the choice of scale preconditioner via P1 ~ O(sigma), P2 ~ O(nu) with nu/sigma = alpha/beta.
    The paper says 'We will proceed informally in this discussion'; the scaling argument is a heuristic that underlies the scale-preconditioning prescription, but is not a theorem.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Preconditioning transformations of adjoint systems for evolution equations." pith.science (2026). https://pith.science/paper/QIIDPVEC

@misc{pith2026250521705,
  author       = {Pith},
  title        = {Pith review of: Preconditioning transformations of adjoint systems for evolution equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QIIDPVEC}},
  note         = {Machine review of arXiv:2505.21705}
}
read the original abstract

Achieving robust control and optimization in high-fidelity physics simulations is extremely challenging, especially for evolutionary systems whose solutions span vast scales across space, time, and physical variables. In conjunction with gradient-based methods, adjoint systems are widely used in the optimization of systems subject to differential equation constraints. In optimization, gradient-based methods are often transformed using suitable preconditioners to accelerate the convergence of the optimization algorithm. Inspired by preconditioned gradient descent methods, we introduce a framework for the preconditioning of adjoint systems associated to evolution equations, which allows one to reshape the dynamics of the adjoint system. We develop two classes of adjoint preconditioning transformations: those that transform both the state dynamics and the adjoint equation and those that transform only the adjoint equation while leaving the state dynamics invariant. Both classes of transformations have the flexibility to include generally nonlinear state-dependent transformations. Using techniques from symplectic geometry and Hamiltonian mechanics, we further show that these preconditioned adjoint systems preserve the property that the adjoint system backpropagates the derivative of an objective function. We then apply this framework to the setting of coupled evolution equations, where we develop a notion of scale preconditioning of the adjoint equations when the state dynamics exhibit large scale-separation. We demonstrate the proposed scale preconditioning on an inverse problem for the radiation diffusion equations. Naive gradient descent is unstable for any practical gradient descent step size, whereas our proposed scale-preconditioned adjoint descent converges in 10-15 gradient-based optimization iterations, with highly accurate reproduction of the wavefront at the final time.

Figures

Figures reproduced from arXiv: 2505.21705 by the authors.

Figure 1
Figure 1. Schematic of the state dynamics for u(·) evolving on the base space U and backpropagation through the adjoint equation for p(·) evolving in reverse time on the fibers of T ∗U over u(·). In particular, the adjoint system on U × U ∗ associated with the problem (2.9) is given by the state dynamics on U with a specified initial condition, together with the adjoint equation on U ∗ with a specified terminal condition, i.e… view at source ↗
Figure 2
Figure 2. Schematic of the mapping arising from the preconditioning of the adjoint system induced by a transformation of the state dynamics. The base space and the fibers are transformed under this map. Finally, since T ∗L is an extended symplectic automorphism, the following proposition immediately follows. Let δue be the solution to the variational equation, given by linearizing the transformed state dynamics (2.12), i.e., … view at source ↗
Figure 3
Figure 3. Schematic of the mapping arising from the preconditioning of the adjoint system induced by a duality pairing. Only the fibers are transformed under this map, while the base space is left invariant. This yields an analogous conservation law to Proposition 2.1. Proposition 2.2. Let u satisfy the state dynamics (2.10a), let δu satisfy the associated variational equation, and let ξ satisfy the preconditioned adjoint equ… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Snapshots of the evolution of E (left) and T (right) of the thick Marshak wave with a constant equilibrium initial state. Now, we consider the following inverse problem, which is to reconstruct the initial conditions E(0) and T(0) such that the radiation energy and tem…
Figure 5
Figure 5. Figure 5: The observed values E∗ compared to E(tf ) produced from the constant equilibrium initial conditions (left) and T ∗ compared to T(tf ) produced from the constant equilibrium initial conditions (right). We solve this optimization problem using an adjoint-based gradient a…
Figure 6
Figure 6. Figure 6: E component of the cost function (left) and T component of the cost function (right) versus iteration, for several values of the scale preconditioner, using the orthogonal projection [PITH_FULL_IMAGE:figures/full_fig_p030_6.png]
Figure 7
Figure 7. Figure 7: E component of the cost function (left) and T component of the cost function (right) versus iteration, for several values of the scale preconditioner, using the E⃗ -coordinate oblique projection. As can be seen in [PITH_FULL_IMAGE:figures/full_fig_p031_7.png]
Figure 8
Figure 8. Figure 8: The initialized and reconstructed initial conditions, E(0) (left) and T(0) (right) [PITH_FULL_IMAGE:figures/full_fig_p031_8.png]
Figure 9
Figure 9. Figure 9: Comparison of the unperturbed, observed and reconstructed values at the final time for E (left) and T (right). Remark 4.1. It is interesting to note that the initial states, shown in [PITH_FULL_IMAGE:figures/full_fig_p032_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 43 canonical work pages

  1. [1]

    Natural gradient works efficiently in learning.Neural Computation, 10(2): 251–276, 1998

    Shun-ichi Amari. Natural gradient works efficiently in learning.Neural Computation, 10(2): 251–276, 1998. doi: 10.1162/089976698300017746. (Page 2, 7)

    doi:10.1162/089976698300017746 1998

  2. [2]

    Anderson, J

    R. Anderson, J. Andrej, A. Barker, J. Bramwell, J.-S. Camier, J. Cerveny, V. Dobrev, Y. Du- douit, A. Fisher, Tz. Kolev, W. Pazner, M. Stowell, V. Tomov, I. Akkerman, J. Dahm, D. Med- ina, and S. Zampini. MFEM: A modular finite element methods library.Computers & Math- ematics with Applications, 81:42–74, 2021. doi: 10.1016/j.camwa.2020.06.009. (Page 27)

    doi:10.1016/j.camwa.2020.06.009 2021

  3. [3]

    Arnold, Franco Brezzi, Bernardo Cockburn, and Donatella Marini

    Douglas N. Arnold, Franco Brezzi, Bernardo Cockburn, and Donatella Marini. Discontinuous Galerkin methods for elliptic problems. In Bernardo Cockburn, George E. Karniadakis, and Chi-Wang Shu, editors,Discontinuous Galerkin Methods, pages 89–101, Berlin, Heidelberg,

    work page

  4. [4]

    Ascher, Steven J

    Uri M. Ascher, Steven J. Ruuth, and Raymond J. Spiteri. Implicit-explicit Runge-Kutta methods for time-dependent partial differential equations.Applied Numerical Mathematics, 25(2):151–167, 1997. ISSN 0168-9274. doi: https://doi.org/10.1016/S0168-9274(97)00056-1. Special Issue on Time Integration. (Page 24)

    doi:10.1016/s0168-9274(97)00056-1 1997

  5. [5]

    Mirror descent and nonlinear projected subgradient methods for convex optimization.Operations Research Letters, 31(3):167–175, 2003

    Amir Beck and Marc Teboulle. Mirror descent and nonlinear projected subgradient methods for convex optimization.Operations Research Letters, 31(3):167–175, 2003. ISSN 0167-6377. doi: https://doi.org/10.1016/S0167-6377(02)00231-6. (Page 2, 6)

    doi:10.1016/s0167-6377(02)00231-6 2003

  6. [6]

    A. M. Bloch.Nonholonomic Mechanics and Control. Interdisciplinary Applied Mathematics. Springer New York, NY, 2015. doi: 10.1007/978-1-4939-3017-3. (Page 2)

    doi:10.1007/978-1-4939-3017-3 2015

  7. [7]

    Boscarino

    S. Boscarino. High-Order Semi-implicit Schemes for Evolutionary Partial Differential Equa- tions with Higher Order Derivatives.Journal of Scientific Computing, 96(1):11, 2023. ISSN 0885-7474. doi: 10.1007/s10915-023-02235-0. (Page 24)

    work page doi:10.1007/s10915-023-02235-0 2023

  8. [8]

    Boscarino, R

    S. Boscarino, R. B¨ urger, P. Mulet, G. Russo, and L. M. Villada. Linearly Implicit IMEX Runge–Kutta Methods for a Class of Degenerate Convection-Diffusion Problems.SIAM Journal on Scientific Computing, 37(2):B305–B331, 2015. ISSN 1064-8275. doi: 10.1137/ 140967544

    work page 2015

Show all 61 references
  1. [9]

    Boscarino, F

    S. Boscarino, F. Filbet, and G. Russo. High Order Semi-implicit Schemes for Time Dependent Partial Differential Equations.Journal of Scientific Computing, 68(3):975–1001, 2016. ISSN 0885-7474. doi: 10.1007/s10915-016-0168-y. (Page 24)

    2016 doi

  2. [10]

    Buvoli and B

    T. Buvoli and B. S. Southworth. A new class of Runge–Kutta methods for nonlinearly parti- tioned systems.arXiv preprint: 2401.04859, 2023. (Page 23, 24) 34 PRECONDITIONING TRANSFORMATIONS OF ADJOINT SYSTEMS FOR EVOLUTION EQUATIONS

    2023 arXiv

  3. [11]

    Buvoli, B

    T. Buvoli, B. K. Tran, and B. S. Southworth. Multirate Runge–Kutta for nonlinearly parti- tioned systems.arXiv preprint: 2504.03257, 2025. (Page 24)

    2025 arXiv

  4. [12]

    D. G. Cacuci. Sensitivity theory for nonlinear systems. I. Nonlinear functional analysis ap- proach.J. Math. Phys., 22(12):2794–2802, 1981. (Page 2)

    1981

  5. [13]

    Campos, Sina Ober-Bl¨ obaum, and Emmanuel Tr´ elat

    C´ edric M. Campos, Sina Ober-Bl¨ obaum, and Emmanuel Tr´ elat. High order variational in- tegrators in the optimal control of mechanical systems.Discrete and Continuous Dynamical Systems, 35(9):4193–4223, 2015. doi: 10.3934/dcds.2015.35.4193

    2015 doi

  6. [14]

    Y. Cao, S. Li, L. Petzold, and R. Serban. Adjoint sensitivity analysis for differential-algebraic equations: The adjoint DAE system and its numerical solution.SIAM J. Sci. Comput., 24(3): 1076–1089 (14 pages), 2003. (Page 2, 8)

    2003

  7. [15]

    Castor.Radiation Hydrodynamics

    John I. Castor.Radiation Hydrodynamics. Cambridge University Press, 2004. doi: 10.1017/ CBO9780511536182. (Page 26)

    2004

  8. [16]

    Iterative pre-conditioning to ex- pedite the gradient-descent method

    Kushal Chakrabarti, Nirupam Gupta, and Nikhil Chopra. Iterative pre-conditioning to ex- pedite the gradient-descent method. In2020 American Control Conference (ACC), pages 3977–3982, 2020. doi: 10.23919/ACC45564.2020.9147603. (Page 2, 7)

    2020

  9. [17]

    Heizler, Elad Steinberg, Ryan G

    Nitay Derei, Shmuel Balberg, Shay I. Heizler, Elad Steinberg, Ryan G. McClarren, and Mena- hem Krief. The non-equilibrium Marshak wave problem in non-homogeneous media.Physics of Fluids, 36(12):127149, 2024. doi: 10.1063/5.0244247. (Page 28)

    2024 doi

  10. [18]

    Dobrev, Tzanio V

    Veselin A. Dobrev, Tzanio V. Kolev, and Robert N. Rieben. High-order curvilinear finite element methods for Lagrangian hydrodynamics.SIAM Journal on Scientific Computing, 34 (5):B606–B641, 2012. doi: 10.1137/120864672. (Page 11)

    2012 doi

  11. [19]

    The Adjoint Method for Time-Optimal Control Problems.Journal of Computational and Nonlinear Dynamics, 16(2):021003, 2020

    Philipp Eichmeir, Thomas Lauß, Stefan Oberpeilsteiner, Karin Nachbagauer, and Wolfgang Steiner. The Adjoint Method for Time-Optimal Control Problems.Journal of Computational and Nonlinear Dynamics, 16(2):021003, 2020. ISSN 1555-1415. doi: 10.1115/1.4048808. (Page 2, 8)

    2020 doi

  12. [20]

    M. B. Giles and N. A. Pierce. An introduction to the adjoint approach to design.Flow, Turbulence and Combustion, 65:393–415, 2000. (Page 2)

    2000

  13. [21]

    Mirrorless mirror descent: A natural derivation of mirror descent

    Suriya Gunasekar, Blake Woodworth, and Nathan Srebro. Mirrorless mirror descent: A natural derivation of mirror descent. In Arindam Banerjee and Kenji Fukumizu, editors,Proceedings of The 24th International Conference on Artificial Intelligence and Statistics, volume 130 of Pr...

    2021

  14. [22]

    Shampoo: Preconditioned stochastic tensor optimization

    Vineet Gupta, Tomer Koren, and Yoram Singer. Shampoo: Preconditioned stochastic tensor optimization. InInternational Conference on Machine Learning, pages 1842–1850. PMLR,

  15. [23]

    Gustafson and Duggirala K

    Karl E. Gustafson and Duggirala K. M. Rao.Numerical Range: The Field of Values of Linear Operators and Matrices. Universitext. Springer New York, NY, 1997. doi: 10.1007/ 978-1-4613-8498-4. (Page 25)

    1997

  16. [24]

    Springer Series in Com- putational Mathematics

    Ernst Hairer, Gerhard Wanner, and Christian Lubich.Geometric Numerical Integration, Structure-Preserving Algorithms for Ordinary Differential Equations. Springer Series in Com- putational Mathematics. Springer Berlin, Heidelberg, 2006. doi: 10.1007/3-540-30666-8. (Page 25)

    2006 doi

  17. [25]

    Projected gradient descent on Riemannian manifolds with applications to online power system optimization

    Adrian Hauswirth, Saverio Bolognani, Gabriela Hug, and Florian D¨ orfler. Projected gradient descent on Riemannian manifolds with applications to online power system optimization. In 2016 54th Annual Allerton Conference on Communication, Control, and Computing (Allerton), page...

    2016

  18. [26]

    N. H. Ibragimov. Integrating factors, adjoint equations and Lagrangians.Journal of Mathe- matical Analysis and Applications, 318(2):742–757, 2006. (Page 2)

    2006

  19. [27]

    N. H. Ibragimov. A new conservation theorem.J. Math. Anal. Appl., 333(1):311–328, 2007. (Page 2) PRECONDITIONING TRANSFORMATIONS OF ADJOINT SYSTEMS FOR EVOLUTION EQUATIONS 35

    2007

  20. [28]

    Kaasschieter

    E.F. Kaasschieter. Preconditioned conjugate gradients for solving singular systems.Journal of Computational and Applied Mathematics, 24(1):265–275, 1988. doi: 10.1016/0377-0427(88) 90358-5. (Page 27)

    1988 doi

  21. [29]

    Kennedy and Mark H

    Christopher A. Kennedy and Mark H. Carpenter. Additive Runge–Kutta schemes for con- vection–diffusion–reaction equations.Applied Numerical Mathematics, 44(1-2):139–181, 2003. ISSN 0168-9274. doi: 10.1016/s0168-9274(02)00138-1. (Page 24)

    2003 doi

  22. [30]

    Kobayashi and K

    S. Kobayashi and K. Nomizu.Foundations of Differential Geometry Volume I. Wiley Classics Library. Wiley, 1996. (Page 17)

    1996

  23. [31]

    A grey transport acceleration method far time-dependent radiative transfer problems.Journal of Computational Physics, 78(2):459–480, 1988

    Edward Larsen. A grey transport acceleration method far time-dependent radiative transfer problems.Journal of Computational Physics, 78(2):459–480, 1988. ISSN 0021-9991. doi: 10.1016/0021-9991(88)90060-5. (Page 27)

    1988 doi

  24. [32]

    Li and L

    S. Li and L. Petzold. Adjoint sensitivity analysis for time-dependent partial differential equa- tions with adaptive mesh refinement.Journal of Computational Physics, 198(1):310–325, 2004. doi: 10.1016/j.jcp.2003.01.001. (Page 8)

    2004 doi

  25. [33]

    Preconditioned stochastic gradient descent.IEEE Transactions on Neural Networks and Learning Systems, 29(5):1454–1466, 2018

    Xi-Lin Li. Preconditioned stochastic gradient descent.IEEE Transactions on Neural Networks and Learning Systems, 29(5):1454–1466, 2018. doi: 10.1109/TNNLS.2017.2672978. (Page 2, 7)

    2018

  26. [34]

    Maddison, Daniel Paulin, Yee Whye Teh, and Arnaud Doucet

    Chris J. Maddison, Daniel Paulin, Yee Whye Teh, and Arnaud Doucet. Dual space precon- ditioning for gradient descent.SIAM Journal on Optimization, 31(1):991–1016, 2021. doi: 10.1137/19M130858X. (Page 2, 7)

    2021 doi

  27. [35]

    J. E. Marsden and T. S. Ratiu.Introduction to Mechanics and Symmetry. Texts in Applied Mathematics. Springer New York, NY, 2 edition, 1999. doi: 10.1007/978-0-387-21792-5. (Page 4, 11)

    1999 doi

  28. [36]

    J. E. Marsden and M. West. Discrete mechanics and variational integrators.Acta Numer., 10: 317–514, 2001. (Page 25)

    2001

  29. [37]

    New insights and perspectives on the natural gradient method.J

    James Martens. New insights and perspectives on the natural gradient method.J. Mach. Learn. Res., 21(1), 2020. ISSN 1532-4435. (Page 2, 7)

    2020

  30. [38]

    Optimizing neural networks with kronecker-factored approx- imate curvature

    James Martens and Roger Grosse. Optimizing neural networks with kronecker-factored approx- imate curvature. InInternational conference on machine learning, pages 2408–2417. PMLR,

  31. [39]

    Takashi Matsubara, Yuto Miyatake, and Takaharu Yaguchi. The symplectic adjoint method: Memory-efficient backpropagation of neural-network-based differential equations.IEEE Trans- actions on Neural Networks and Learning Systems, pages 1–13, 2023. doi: 10.1109/TNNLS. 2023.324234...

    2023

  32. [40]

    MFEM: Modular finite element methods [Software].mfem.org

    mfem. MFEM: Modular finite element methods [Software].mfem.org. (Page 27)

  33. [41]

    Ox- ford University Press, New York, 1984

    Dimitri Mihalas and Barbara Weibel Mihalas.Foundations of Radiation Hydrodynamics. Ox- ford University Press, New York, 1984. ISBN 0-19-503437-6. (Page 26)

    1984

  34. [42]

    V. T. Nguyen, D. Georges, and G. Besan¸ con. State and parameter estimation in 1-D hyperbolic PDEs based on an adjoint method.Automatica, 67(C):185–191, May 2016. ISSN 0005-1098. (Page 2)

    2016

  35. [43]

    Noack and A

    A. Noack and A. Walther. Adjoint concepts for the optimal control of Burgers equations. Comput Optim Applic, 36:109–133, 2007. doi: 10.1007/s10589-006-0393-7. (Page 8)

    2007 doi

  36. [44]

    Padilla and Andreas H

    Carlos E. Padilla and Andreas H. Von Flotow. Nonlinear strain-displacement relations and flexible multibody dynamics.Journal of Guidance, Control, and Dynamics, 15(1):128–136,

  37. [45]

    N. A. Pierce and M. B. Giles. Adjoint recovery of superconvergent functionals from PDE approximations.SIAM Rev., 42(2):247–264, 2000. (Page 2)

    2000

  38. [46]

    The information geometry of mirror descent.IEEE Transactions on Information Theory, 61(3):1451–1457, 2015

    Garvesh Raskutti and Sayan Mukherjee. The information geometry of mirror descent.IEEE Transactions on Information Theory, 61(3):1451–1457, 2015. doi: 10.1109/TIT.2015.2388583. 36 PRECONDITIONING TRANSFORMATIONS OF ADJOINT SYSTEMS FOR EVOLUTION EQUATIONS (Page 2, 6)

    2015

  39. [47]

    I. M. Ross. A roadmap for optimal control: The right way to commute.Ann. NY Acad. Sci., 1065(1):210–231, 2005. (Page 2)

    2005

  40. [48]

    J. W. Ruge and K. St¨ uben.Multigrid Methods Chapter 4: Algebraic Multigrid, pages 73–130. Society for Industrial and Applied Mathematics, 1987. doi: 10.1137/1.9781611971057.ch4. (Page 27)

    1987 doi

  41. [49]

    Sanders and Nikolaos D

    Brett F. Sanders and Nikolaos D. Katopodes. Adjoint sensitivity analysis for shallow-water wave control.Journal of Engineering Mechanics, 126(9):909–919, 2000. doi: 10.1061/(ASCE) 0733-9399(2000)126:9(909). (Page 8)

    2000 doi

  42. [50]

    J. M. Sanz-Serna. Symplectic Runge–Kutta schemes for adjoint equations, automatic dif- ferentiation, optimal control, and more.SIAM Review, 58(1):3–33, 2016. (Page 2, 5, 8, 9, 10)

    2016

  43. [51]

    B. K. Tran and M. Leok. Geometric methods for adjoint systems.J Nonlinear Sci, 34(25),

  44. [52]

    B. K. Tran and M. Leok. A type II Hamiltonian variational principle and adjoint systems for Lie groups.J Dyn Control Syst, 31(8), 2025. doi: 10.1007/s10883-025-09730-7. (Page 8)

    2025 doi

  45. [53]

    B. K. Tran, B. S. Southworth, and M. Leok. On properties of adjoint systems for evolutionary PDEs.J Nonlinear Sci, 34(95), 2024. doi: 10.1007/s00332-024-10071-1. (Page 3, 11, 13, 25)

    2024 doi

  46. [54]

    B. K. Tran, B. S. Southworth, and T. Buvoli. Order conditions for nonlinearly partitioned Runge–Kutta methods.Electronic Transactions on Numerical Analysis, 63:171–198, 2025. doi: 10.1553/etna vol63s171. (Page 23, 24)

    2025 doi

  47. [55]

    Gradient-based nonlinear model predictive control for systems with state-dependent mass matrix

    Andreas V¨ olz and Knut Graichen. Gradient-based nonlinear model predictive control for systems with state-dependent mass matrix. In2021 60th IEEE Conference on Decision and Control (CDC), pages 1012–1017, 2021. doi: 10.1109/CDC45484.2021.9683175. (Page 11)

    2021

  48. [56]

    Yano and S

    K. Yano and S. Ishihara.Tangent and cotangent bundles: differential geometry. Pure Appl. Math., No. 16. Marcel Dekker, Inc., New York, 1973. (Page 4, 11)

    1973

  49. [57]

    Adahessian: An adaptive second order optimizer for machine learning

    Zhewei Yao, Amir Gholami, Sheng Shen, Mustafa Mustafa, Kurt Keutzer, and Michael Ma- honey. Adahessian: An adaptive second order optimizer for machine learning. Inproceedings of the AAAI conference on artificial intelligence, volume 35, pages 10665–10673, 2021. (Page 2)

    2021

  50. [58]

    Numerical Methods and Algorithms

    Fuzhen Zhang, editor.The Schur Complement and Its Applications. Numerical Methods and Algorithms. Springer New York, NY, 2005. doi: 10.1007/b105056. (Page 24)

    2005 doi

  51. [1992]

    (Page 11)

    doi: 10.2514/3.20810. (Page 11)

    doi

  52. [2000]

    ISBN 978-3-642-59721-3

    Springer Berlin Heidelberg. ISBN 978-3-642-59721-3. (Page 27)

  53. [2024]

    (Page 3, 4, 5, 8, 9, 10, 17)

    doi: 10.1007/s00332-023-09999-7. (Page 3, 4, 5, 8, 9, 10, 17)

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.