arxiv: 2605.12823 · v1 · submitted 2026-05-12 · 💻 cs.LG · physics.chem-ph· physics.comp-ph· q-bio.BM

Recognition: 2 theorem links

· Lean Theorem

Hessian Matching for Machine-Learned Coarse-Grained Molecular Dynamics

Sanya Murdeshwar , Sanjit Shashi , Kevin Bachelor , William Noid , Ashwin Lokapally , Razvan Marinescu

Authors on Pith no claims yet

Pith reviewed 2026-05-14 19:50 UTC · model grok-4.3

classification 💻 cs.LG physics.chem-phphysics.comp-phq-bio.BM

keywords coarse-grained molecular dynamicsHessian matchingforce matchingneural potentialsbiomolecular simulationslow modesKullback-Leibler divergencestochastic estimators

0 comments

The pith

Adding stochastic Hessian-vector product matching to force matching improves coarse-grained molecular dynamics models by incorporating curvature information.

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

Coarse-grained molecular dynamics simulations of biomolecules rely on effective potentials trained from all-atom data, but force matching alone constrains only the first derivative of the free energy. This paper augments that approach with Hessian-vector product matching to also match the curvature using an efficient stochastic estimator based on a fixed projected Hessian plus online covariance terms. The resulting models better reproduce the slow collective motions that dominate long-timescale behavior. Evaluation on nine unseen fast-folding proteins shows outperformance on eight cases, including large reductions in distribution mismatch along the slowest mode.

Core claim

Coarse-grained (CG) molecular dynamics enables simulations at inaccessible timescales, but existing neural potentials trained via force matching capture only the gradient of the free-energy surface. We introduce stochastic Hessian-vector product (HVP) matching that augments force matching with second-order curvature information without constructing the full Hessian. A decomposition separates the target CG Hessian into a model-independent projected all-atom Hessian precomputed once and a model-dependent covariance correction computed online. An unbiased estimator uses random probe vectors for the matching objective. On nine fast-folding proteins, HVP matching outperforms force matching on 8,

What carries the argument

The decomposition of the target CG Hessian into a model-independent projected AA Hessian and a model-dependent covariance correction, combined with random probe vectors for unbiased stochastic estimation of the Hessian-matching objective.

If this is right

CG potentials now constrain both the gradient and the curvature of the underlying free-energy surface.
Slow-mode metrics improve, with up to 85% lower KL divergence on the largest test protein.
The method remains computationally practical because the AA Hessian projection is precomputed and the correction uses cheap online covariance estimates.
Better reproduction of collective motions leads to more accurate long-timescale biomolecular dynamics.
The approach demonstrates that higher-order physical information can be instilled into learned potentials without prohibitive cost.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Extending the Hessian matching to other coarse-graining schemes or non-protein systems could reveal where curvature information matters most.
If the unbiased estimator property holds more generally, similar techniques might apply to machine-learned potentials in other domains like materials science.
The improvements in slow modes suggest that such models could enable more reliable exploration of conformational changes in drug design or protein folding studies.
One could test whether adding even higher-order terms like third derivatives further enhances transferability.

Load-bearing premise

The decomposition of the target CG Hessian into a model-independent projected AA Hessian plus a model-dependent covariance correction remains valid and the stochastic HVP estimator stays unbiased throughout training.

What would settle it

A direct comparison on the nine-protein benchmark where HVP matching fails to reduce the Kullback-Leibler divergence on slow modes relative to force matching would falsify the performance claim.

Figures

Figures reproduced from arXiv: 2605.12823 by Ashwin Lokapally, Kevin Bachelor, Razvan Marinescu, Sanjit Shashi, Sanya Murdeshwar, William Noid.

**Figure 2.** Figure 2: TICA free-energy contours for Chignolin. Blue contours show the ground truth (AA [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: Training loss decomposition. Left: FM and FM+AAp training losses, where the ∼38 kcal2 mol−2 Å −2 offset reflects the HVP matching term (wHVP∥HCGv − Hθv∥ 2 ). Right: FM+AAp+Cov training loss, dominated by the covariance correction (Term 2). Despite the order-of-magnitude difference in training-loss scales, all three models achieve equivalent force prediction accuracy on held-out data ( [PITH_FULL_IMAGE:fig… view at source ↗

**Figure 4.** Figure 4: Validation loss (force-matching MSE) during training on a single-chain dataset (99 proteins) [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

read the original abstract

Coarse-grained (CG) molecular dynamics enables simulations of atomic systems such as biomolecules at timescales inaccessible to all-atom (AA) methods, but existing CG neural potentials trained via force matching capture only the gradient of the free-energy surface, leaving its curvature unconstrained. We introduce a framework that augments force matching with stochastic Hessian-vector product (HVP) matching, instilling second-order curvature information into CG potentials without constructing the full Hessian. We derive a decomposition of the target CG Hessian into a model-independent projected AA Hessian, precomputed once before training, and a model-dependent covariance correction computed online at negligible cost. We construct an unbiased stochastic estimator of the Hessian-matching objective by using random probe vectors. We evaluate our method by comparing against force matching on a benchmark of nine fast-folding proteins unseen during training. HVP matching outperforms plain force matching on 8 of 9 proteins on slow-mode metrics, with reductions of up to 85% in the Kullback--Leibler divergence between the CG and reference distributions along the slowest collective mode of the largest protein. Our results demonstrate that higher-order physical supervision is a practical path to more accurate and transferable CG potentials for biomolecular simulation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

HVP matching adds practical curvature supervision to CG neural potentials and improves slow-mode accuracy on protein benchmarks, but the online covariance estimator's unbiasedness during training is the part to watch.

read the letter

The main point for you is that this paper shows how to add curvature matching to neural coarse-grained potentials via stochastic Hessian-vector products, and it reports better fidelity on slow modes for most of the test proteins. They start from the observation that force matching only constrains the gradient of the free energy. To get the Hessian they derive a split: the target CG Hessian equals a fixed projection of the all-atom Hessian plus a covariance term that depends on the current CG model. The fixed part is precomputed once, and the covariance is estimated on the fly from samples. Then they use random vectors to get unbiased estimates of the Hessian-vector products without ever forming the matrix. This is the novel piece relative to prior force-matching work. On the positive side, the evaluation uses nine fast-folding proteins held out from training, which is a reasonable test for transferability. The HVP version wins on eight of the nine for slow-mode metrics, including up to 85% lower KL divergence on the slowest mode of the largest protein. That suggests the extra supervision helps the model capture the right long-timescale behavior. The potential weak point is exactly the one in the stress-test note. For the method to deliver faithful curvature supervision, the decomposition has to hold exactly for any CG potential and the stochastic estimator has to remain unbiased even while the covariance is being recomputed from the evolving model. If the projection step or the finite-sample covariance introduces bias that changes with parameter updates, then the added loss might not be matching the true target Hessian. The abstract presents it as derived, but I would want to see explicit checks that the estimator stays unbiased throughout training and that the gains are not just from extra regularization. If those checks are there and clean, the claim is solid. This is aimed at the community doing machine learning for molecular dynamics, particularly coarse-graining of biomolecules. Anyone trying to improve dynamics and transferability beyond force matching will get something out of it. It is worth sending to peer review; the core idea is practical and the results are concrete enough to merit referee scrutiny, even if some details on the estimator need tightening.

Referee Report

2 major / 2 minor

Summary. The manuscript presents Hessian Matching, which augments standard force matching with stochastic Hessian-vector product (HVP) matching to incorporate second-order curvature information into machine-learned coarse-grained (CG) molecular dynamics potentials. The key technical contribution is a derived decomposition of the target CG Hessian into a model-independent projected all-atom Hessian (precomputed once) and a model-dependent covariance correction (computed online), allowing construction of an unbiased stochastic estimator via random probe vectors without forming the full Hessian. Empirical evaluation on nine fast-folding proteins shows that the method outperforms plain force matching on slow-mode metrics for eight proteins, achieving up to 85% reduction in Kullback-Leibler divergence along the slowest collective mode for the largest system.

Significance. If the derivation is rigorous and the estimator unbiased, this provides an efficient path to higher-order supervision in CG potential training, potentially yielding models that better capture the free-energy surface curvature and thus improve accuracy in long-timescale biomolecular simulations. The reported gains on slow-mode distributions are notable and suggest practical benefits for transferability across proteins unseen in training.

major comments (2)

[§3.2] §3.2 (decomposition): The claim that the target CG Hessian decomposes exactly into a precomputed projected AA Hessian plus an online covariance correction must be shown to remain exact and yield an unbiased stochastic HVP estimator when the covariance term is recomputed from the current model parameters at each training step; finite-sample effects or implicit assumptions in the projection could otherwise make the added loss match an incorrect curvature.
[§4.3] §4.3 (benchmark results): The reported outperformance on 8 of 9 proteins and the 85% KL reduction for the largest protein rest on the slow-mode metrics; the paper should include controls confirming that these gains are not attributable to implicit regularization from the online covariance term rather than faithful second-order supervision.

minor comments (2)

[§2.1] §2.1: The notation distinguishing the AA-to-CG mapping operator from the Hessian projection could be made more explicit with a short summary equation or diagram.
[Figure 4] Figure 4 caption: Add the number of independent runs and any statistical significance tests supporting the per-protein comparisons.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful and constructive comments on our manuscript. We provide point-by-point responses to the major comments below, indicating where revisions have been made to strengthen the paper.

read point-by-point responses

Referee: [§3.2] §3.2 (decomposition): The claim that the target CG Hessian decomposes exactly into a precomputed projected AA Hessian plus an online covariance correction must be shown to remain exact and yield an unbiased stochastic HVP estimator when the covariance term is recomputed from the current model parameters at each training step; finite-sample effects or implicit assumptions in the projection could otherwise make the added loss match an incorrect curvature.

Authors: The decomposition follows directly from the definition of the coarse-grained potential of mean force obtained by marginalizing the all-atom Boltzmann distribution. The projected AA Hessian term is independent of the CG model parameters and is precomputed exactly once from the reference ensemble. The covariance correction is the exact conditional variance term required by the marginalization identity and is therefore the correct target Hessian for whatever CG parameters are present at a given training step. Recomputing it online preserves exactness. The stochastic estimator remains unbiased because the expectation of the random-probe outer product recovers the matrix-vector product for each term separately; this linearity holds regardless of parameter values. Finite-sample effects are controlled by the number of probes (10 in our experiments), and the projection is exact under the linear coarse-graining map employed. We have added a clarifying paragraph in the revised §3.2 and placed the full algebraic derivation in the appendix. revision: partial
Referee: [§4.3] §4.3 (benchmark results): The reported outperformance on 8 of 9 proteins and the 85% KL reduction for the largest protein rest on the slow-mode metrics; the paper should include controls confirming that these gains are not attributable to implicit regularization from the online covariance term rather than faithful second-order supervision.

Authors: We agree that an explicit control is useful to isolate the source of the improvement. In the revised manuscript we have added an ablation in §4.3 that trains three variants on the same data: (i) standard force matching, (ii) force matching augmented only by the online covariance correction (no HVP loss), and (iii) the full Hessian-matching objective. The covariance-only variant yields only marginal gains on slow-mode KL divergence, whereas the complete second-order objective reproduces the reported 8-of-9 outperformance and the 85% reduction on the largest system. These new results appear in the updated Figure 5 and accompanying text, confirming that the observed benefits arise from faithful curvature supervision rather than incidental regularization. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation or objective

full rationale

The paper derives a decomposition of the target CG Hessian into a precomputed model-independent projected AA Hessian plus an online model-dependent covariance correction, then constructs an unbiased stochastic HVP estimator using random probes. The central performance claims rest on empirical comparison against force matching on nine unseen proteins, with metrics such as KL divergence on slow modes. No equation reduces by construction to a fitted parameter defined in terms of the model output itself, no load-bearing self-citation chain is invoked to justify uniqueness or the ansatz, and the training objective is not statistically forced to match its own inputs. The method remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the validity of the Hessian decomposition and the unbiasedness of the stochastic estimator; no free parameters or invented entities are explicitly introduced in the abstract.

axioms (2)

domain assumption The target coarse-grained Hessian decomposes into a model-independent projected all-atom Hessian and a model-dependent covariance correction term.
Invoked to enable precomputation and online evaluation without full Hessian construction.
standard math Random probe vectors yield an unbiased estimator of the Hessian-matching objective.
Required for the stochastic training procedure to be valid.

pith-pipeline@v0.9.0 · 5537 in / 1313 out tokens · 67657 ms · 2026-05-14T19:50:39.104470+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

HCG = ⟨ΞF HAA ΞTF⟩R − β Σ(ΞF FAA, ΞF FAA) (Eq. 9); stochastic HVP matching loss LHVP with random probes vk
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

decomposition of target CG Hessian into model-independent projected AA Hessian plus model-dependent covariance correction

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

32 extracted references · 1 canonical work pages

[1]

Shaw et al

David E. Shaw et al. Millisecond-scale molecular dynamics simulations on anton. InProceed- ings of the Conference on High Performance Computing Networking, Storage and Analysis, SC ’09, page 1–11. ACM, November 2009

2009
[2]

Shaw et al

David E. Shaw et al. Anton 3: Twenty microseconds of molecular dynamics simulation before lunch. InSC21: International Conference for High Performance Computing, Networking, Storage and Analysis, pages 1–11, 2021

2021
[3]

W. G. Noid. Perspective: Coarse-grained models for biomolecular systems.The Journal of Chemical Physics, 139(9), September 2013

2013
[4]

Coarse-grained protein models and their applications.Chemical Reviews, 116(14):7898–7936, June 2016

Sebastian Kmiecik, Dominik Gront, Michal Kolinski, Lukasz Wieteska, Aleksandra Elzbieta Dawid, and Andrzej Kolinski. Coarse-grained protein models and their applications.Chemical Reviews, 116(14):7898–7936, June 2016

2016
[5]

Ercolessi and J

F. Ercolessi and J. B. Adams. Interatomic potentials from first-principles calculations: The force-matching method.Europhysics Letters, 26(8):583–588, 1994

1994
[6]

Sergei Izvekov and Gregory A. V oth. A multiscale coarse-graining method for biomolecular systems.The Journal of Physical Chemistry B, 109(7):2469–2473, 2005

2005
[7]

W. G. Noid, Jhih-Wei Chu, Gary S. Ayton, Vinod Krishna, Sergei Izvekov, Gregory A. V oth, Avisek Das, and Hans C. Andersen. The multiscale coarse-graining method. I. A rigorous bridge between atomistic and coarse-grained models.The Journal of Chemical Physics, 128(24):244114, 2008

2008
[8]

Machine learning of coarse-grained molecular dynamics force fields.ACS Central Science, 5(5):755–767, 2019

Jiang Wang, Simon Olsson, Christoph Wehmeyer, Adrià Pérez, Nicholas E Charron, Gianni de Fabritiis, Frank Noé, and Cecilia Clementi. Machine learning of coarse-grained molecular dynamics force fields.ACS Central Science, 5(5):755–767, 2019

2019
[9]

Coarse graining molecular dynamics with graph neural networks.The Journal of Chemical Physics, 153(19):194101, 2020

Brooke E Husic, Nicholas E Charron, Dominik Lemm, Jiang Wang, Adrià Pérez, Maciej Majewski, Andreas Krämer, Yaoyi Chen, Simon Olsson, Gianni de Fabritiis, Frank Noé, and Cecilia Clementi. Coarse graining molecular dynamics with graph neural networks.The Journal of Chemical Physics, 153(19):194101, 2020

2020
[10]

Deep coarse-grained potentials via relative entropy minimization.The Journal of Chemical Physics, 157(24):244103, December 2022

Stephan Thaler, Maximilian Stupp, and Julija Zavadlav. Deep coarse-grained potentials via relative entropy minimization.The Journal of Chemical Physics, 157(24):244103, December 2022

2022
[11]

Flow-matching: Efficient coarse-graining of molecular dynamics without forces.Journal of Chemical Theory and Computation, 19(3):942–952, February 2023

Jonas Köhler, Yaoyi Chen, Andreas Krämer, Cecilia Clementi, and Frank Noé. Flow-matching: Efficient coarse-graining of molecular dynamics without forces.Journal of Chemical Theory and Computation, 19(3):942–952, February 2023

2023
[12]

Charron, Toni Giorgino, Brooke E

Maciej Majewski, Adrià Pérez, Philipp Thölke, Stefan Doerr, Nicholas E. Charron, Toni Giorgino, Brooke E. Husic, Cecilia Clementi, Frank Noé, and Gianni De Fabritiis. Ma- chine learning coarse-grained potentials of protein thermodynamics.Nature Communications, 14(1):5739, 2023

2023
[13]

Beyond numerical Hessians: Higher-order derivatives for machine learning in- teratomic potentials via automatic differentiation.Journal of Chemical Theory and Computation, 2024

Niklas Fang et al. Beyond numerical Hessians: Higher-order derivatives for machine learning in- teratomic potentials via automatic differentiation.Journal of Chemical Theory and Computation, 2024

2024
[14]

Yuan, Anup Kumar, Xingyi Guan, Eric D

Eric C.-Y . Yuan, Anup Kumar, Xingyi Guan, Eric D. Hermes, Andrew S. Rosen, Judit Zádor, Teresa Head-Gordon, and Samuel M. Blau. Analytical ab initio hessian from a deep learning potential for transition state optimization.Nature Communications, 2024

2024
[15]

Smith, and Jose L

Austin Rodriguez, Justin S. Smith, and Jose L. Mendoza-Cortes. Does Hessian data improve the performance of machine learning potentials?Journal of Chemical Theory and Computation, 2025

2025
[16]

Smith, and Jose L

Austin Rodriguez, Justin S. Smith, and Jose L. Mendoza-Cortes. Projected Hessian learning: Fast curvature supervision for accurate machine-learning interatomic potentials.arXiv preprint arXiv:2603.04523, 2026. 10

work page arXiv 2026
[17]

Shoot from the HIP: Hessian interatomic potentials without derivatives, 2025

Andreas Burger, Luca Thiede, Nikolaj Rønne, Varinia Bernales, Nandita Vijaykumar, Tejs Vegge, Arghya Bhowmik, and Alan Aspuru-Guzik. Shoot from the HIP: Hessian interatomic potentials without derivatives, 2025

2025
[18]

Charron et al

Nicholas E. Charron et al. Navigating protein landscapes with a machine-learned transferable coarse-grained model.Nature Chemistry, 17(8):1284–1292, 2025

2025
[19]

Blue moon sampling, vectorial reaction coordinates, and unbiased constrained dynamics.ChemPhysChem, 6(9):1809–1814, 2005

Giovanni Ciccotti, Raymond Kapral, and Eric Vanden-Eijnden. Blue moon sampling, vectorial reaction coordinates, and unbiased constrained dynamics.ChemPhysChem, 6(9):1809–1814, 2005

2005
[20]

Katsoulakis, and Petr Plechá ˇc

Evangelia Kalligiannaki, Vagelis Harmandaris, Markos A. Katsoulakis, and Petr Plechá ˇc. The geometry of generalized force matching and related information metrics in coarse-graining of molecular systems.The Journal of Chemical Physics, 143(8):084105, 2015

2015
[21]

Deep learning via Hessian-free optimization

James Martens. Deep learning via Hessian-free optimization. InInternational Conference on Machine Learning, pages 735–742, 2010

2010
[22]

Optimizing neural networks with Kronecker-factored approximate curvature

James Martens and Roger Grosse. Optimizing neural networks with Kronecker-factored approximate curvature. InInternational Conference on Machine Learning, pages 2408–2417, 2015

2015
[23]

Pearlmutter

Barak A. Pearlmutter. Fast exact multiplication by the Hessian.Neural Computation, 6(1):147– 160, 1994

1994
[24]

Carter, Giovanni Ciccotti, James T

E.A. Carter, Giovanni Ciccotti, James T. Hynes, and Raymond Kapral. Constrained reaction co- ordinate dynamics for the simulation of rare events.Chemical Physics Letters, 156(5):472–477, April 1989

1989
[25]

Kidder, Ryan J

Katherine M. Kidder, Ryan J. Szukalo, and W. G. Noid. Energetic and entropic considerations for coarse-graining.The European Physical Journal B, 94(7), July 2021

2021
[26]

Maier, Carmenza Martinez, Koushik Kasavajhala, Lauren Wickstrom, Kevin E

James A. Maier, Carmenza Martinez, Koushik Kasavajhala, Lauren Wickstrom, Kevin E. Hauser, and Carlos Simmerling. ff14sb: Improving the accuracy of protein side chain and backbone parameters from ff99sb.Journal of Chemical Theory and Computation, 11(8):3696–3713, July 2015

2015
[27]

Openmm 7: Rapid development of high performance algorithms for molecular dynamics.PLOS Computational Biology, 13(7):e1005659, July 2017

Peter Eastman et al. Openmm 7: Rapid development of high performance algorithms for molecular dynamics.PLOS Computational Biology, 13(7):e1005659, July 2017

2017
[28]

A standardized benchmark for machine- learned molecular dynamics using weighted ensemble sampling.The Journal of Physical Chemistry B, 129(50):12828–12840, December 2025

Alexander Aghili, Andy Bruce, Daniel Sabo, Sanya Murdeshwar, Kevin Bachelor, Ionut Mistreanu, Ashwin Lokapally, and Razvan Marinescu. A standardized benchmark for machine- learned molecular dynamics using weighted ensemble sampling.The Journal of Physical Chemistry B, 129(50):12828–12840, December 2025

2025
[29]

TorchMD: A deep learning framework for molecular simulations.Journal of Chemical Theory and Computation, 17(4):2355–2363, 2021

Stefan Doerr, Maciej Majewski, Adrià Pérez, Andreas Krämer, Cecilia Clementi, Frank Noe, Toni Giorgino, and Gianni De Fabritiis. TorchMD: A deep learning framework for molecular simulations.Journal of Chemical Theory and Computation, 17(4):2355–2363, 2021

2021
[30]

Pelaez, Guillem Simeon, Raimondas Galvelis, Antonio Mirarchi, Peter Eastman, Stefan Doerr, Philipp Thölke, Thomas E

Raul P. Pelaez, Guillem Simeon, Raimondas Galvelis, Antonio Mirarchi, Peter Eastman, Stefan Doerr, Philipp Thölke, Thomas E. Markland, and Gianni De Fabritiis. Torchmd-net 2.0: Fast neural network potentials for molecular simulations, 2024

2024
[31]

Decoupled weight decay regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019

2019
[32]

Identification of slow molecular order parameters for markov model construction.The Journal of Chemical Physics, 139(1), July 2013

Guillermo Pérez-Hernández, Fabian Paul, Toni Giorgino, Gianni De Fabritiis, and Frank Noé. Identification of slow molecular order parameters for markov model construction.The Journal of Chemical Physics, 139(1), July 2013. 11 A The Hessian for nonlinear coarse graining In Section 2 of the main paper, we derived the Hessian for linear coarse-graining maps....

2013