Koopman Generator Decomposition for Port-Hamiltonian Systems

arXiv:2603.17849 · eess.SY, cs.SY, math.DS · Submitted 2026-03-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Koopman Generator Decomposition for Port-Hamiltonian Systems".

Dev: Here's the gist of the paper. We study how the vector-field structure of nonlinear port-Hamiltonian systems is reflected in the infinitesimal Koopman generator The generator admits a natural bracket decomposition into a conservative interconnection-bracket derivation, a dissipative metric-bracket derivation, and an input-port derivation This paper asks how the pH decomposition of the vector field appears at the level of the Koopman generator The positive object is instead the metric, or Dirichlet, bracket

f, f: R = ∇f⊺R∇f ≥ 0 The resulting lifted models are finite-dimensional portHamiltonian surrogates that are passive and support damping injection in the lifted coordinates Koopman Generator Decomposition The input-parametrized generator decomposes as Kuf = ∇f⊺

(J − R)∇H + Gu: This decomposition is given by Kuf = KJ −

f, H: R + ∇f⊺Gu The conservative component is formally skew-adjoint on a test space whenever the conservative flow preserves the reference measure and the relevant boundary terms vanish The dissipative component is not claimed to be a positive operator on arbitrary observables; rather, the positive semidefinite object is the metric bracket

f, f: R = ∇f⊺R∇f ≥ 0 The Hamiltonian observable satisfies KuH = −

H, H: R + y⊺u ≤ y⊺u Finite-Dimensional Structure The weak Galerkin matrices show mass-metric skewness of the conservative contribution and positive semidefiniteness of the dissipative bracket matrix The conservative weak matrix is skew while the dissipative metric-bracket matrix is positive semidefinite The lifted pH surrogate is passive in lifted coordinates, while plant-level conclusions require residual, injectivity, and output-consistency assumptions The empirical Dirichlet matrix can be estimated by DbR = bNsamp/N Xsamp k=1

09: Passivity and Control The closed-loop storage satisfies ˙He = −(P z)⊺D(P z) − ye⊺Kdye + ye⊺v ≤ ye⊺v If the pair (Acl, Cdiss) is detectable, then the equilibrium z = 0 of the surrogate with v = 0 is asymptotically stable The damping-injection law u = −Kdye + v penalizes all nonzero components of the surrogate output ye Structural Distinctions The distinction is visible even for the identity dictionary ψi(x) = xi The Dirichlet matrix DR remains positive semidefinite because it represents the metric bracket, whereas CR generally has no sign guarantee The linear pH example shows that with state coordinates and the correct storage metric P = Q, the lifted pH surrogate is exact and has zero residual The pendulum example shows that the Hamiltonian identity KuH = −

H, H: R + y⊺u holds exactly, while finite dictionaries can still have nonzero closure residuals

Title and authors: Rosa: So we're looking at this paper titled "Koopman Generator Decomposition for Port-Hamiltonian Systems". It looks like they're trying to see how the internal structure of these nonlinear systems shows up in something called the infinitesimal Koopman generator.

Dev: Right. And what they found is that this generator breaks down into three specific parts: a conservative interconnection-bracket derivation, a dissipative metric-bracket derivation, and an input-port derivation. It’s trying to map the structure of these physical systems onto the mathematical tools we use for analyzing their long-term behavior in the Koopman framework.

Taro: I'm curious about what this decomposition means practically when you're actually trying to control or predict a system that behaves like a port-Hamiltonian system. Does it give you a clearer picture of where the energy is going?

Rosa: Well, they focus on how the decomposition looks at the positive object, which they say is actually not the conservative part itself, but this metric bracket involving f R. They argue that this metric bracket f R is positive semidefinite because it equals gradient f transpose R gradient f, which means it's always greater than or equal to zero.

Dev: That non-negativity of the metric bracket is what they use to prove something important about the energy balance for the Hamiltonian observable. They find that this leads to an identity where the Hamiltonian observable Ku H equals minus that metric bracket plus some term involving input and output, which is less than or equal to zero.

Taro: So it’s showing that even though the dissipative part isn't necessarily positive on all observables, there's this specific metric structure that keeps things bounded in a certain way. Is this what lets them build these finite-dimensional models?

Rosa: Exactly. They use these bracket identities to motivate building these finite-dimensional weak Galerkin and data driven lifted models. The big idea is that when the measure used for the conservative part of the flow is conservative for the Hamiltonian interconnection flow, then their conservative contribution ends up being skew in that mass metric, and they see a positive semidefiniteness with this dissipative bracket matrix.

Dev: That means they can create these surrogate models that are passive and support damping injection directly in those lifted coordinates. It’s not just theory anymore; it’s about creating a model that respects the underlying physics of energy dissipation when you lift the system's dynamics into a higher space.

Title and authors: Taro: From my side, I wonder how robust these lifted models are when things go wrong in reality, like when the world misbehaves and we need graceful recovery. The paper mentions that to get those plant level conclusions, you need residual and injectivity assumptions. What does that translate to for a real-world autonomous system?

Rosa: It means the lifted pH surrogate is passive in those lifted coordinates, but getting those final plant level conclusions depends on making sure the residual is small enough and that the model mapping is injective. They also mention using an empirical Dirichlet matrix, which they estimate by taking samples and calculating a specific matrix involving these measurements.

Dev: The way they frame the improvements is really helpful because it shows how to use this structure-preserving lifting for actual control design. They show that for the closed loop storage, you can satisfy an inequality where the time derivative of the Hamiltonian observable is bounded by terms related to that metric bracket and input.

Taro: So if we can guarantee that this stored energy is nonincreasing using a damping injection law, what does that imply for stability when we are dealing with these lifted coordinates?

Rosa: If you have a pair of storage and dissipation matrices, and you make sure the detection condition is met, then the equilibrium point of the surrogate model with no external input becomes asymptotically stable when you set the input to zero. The damping injection law they propose is something like minus Kdye plus v, where Kd is positive definite.

Dev: That specific damping injection law ensures that that surrogate storage quantity is nonincreasing, which directly leads to asymptotic stability of that zero point if the system meets certain detectability conditions and we use no external inputs. It’s a way to guarantee convergence by penalizing all nonzero components of the surrogate output.

Taro: I like how they distinguish between the conservative and dissipative parts using these algebraic structures. They show that even for a simple identity dictionary, where the dictionary is just x itself, the Dirichlet matrix remains positive semidefinite because it represents that metric bracket, whereas the first order derivation matrix generally doesn't have a sign guarantee.

Title and authors: Rosa: That distinction is key because it allows us to prioritize stability analysis on that symmetric metric bracket instead of relying on the first order approximation which can be messy. It provides a mathematically sound way to separate the conservative dynamics from the dissipative ones when you are learning dynamics from data.

Dev: And they use this structural distinction in their finite dimensional weak Galerkin matrices, showing that the conservative matrix is skew and the dissipative metric bracket matrix is positive semidefinite. That’s a very concrete algebraic property we can rely on when we are building these models, separating them from just any approximate first order derivation matrices.

Taro: So for someone who only cares about how to build a stable model, this paper suggests focusing on that specific positive definite metric bracket rather than the general dissipative matrix approximation. That makes sense in terms of what we need for control synthesis.

Rosa: Exactly. And they show an illustrative example with state coordinates and the correct storage metric P equal to Q, and in that case, the lifted pH surrogate is exact, meaning there's zero residual error when you use that specific dictionary. It replicates the nonlinear dynamics perfectly within those state coordinates without approximation errors related to how you build the generator structure.

Dev: That exact representation in state coordinates is a big deal because it means no modeling error from the generator itself, just from the finite-dimensional lifting process, which they show can be zero under those conditions. It’s a very strong result for identification accuracy.

Taro: The pendulum example they use also shows that for that Hamiltonian identity involving Ku H equals minus bracket plus input to hold exactly, it works perfectly when you use state coordinates and the correct storage metric P equal to Q, which is good because it confirms the structure holds in a concrete physical case.

Rosa: It really ties everything together by showing how this mathematical decomposition translates into tangible improvements for system identification and control. The empirical Dirichlet matrix estimation method they show can be used to estimate that critical symmetric metric bracket when you only have snapshots of the system data.

Dev: So, to wrap up the results, the main idea is using these bracket identities to construct lifted models that are passive in their new coordinates and support guaranteed damping injection for stability. This gives us a framework for robust control synthesis based on those lifted coordinates.

Title and authors: Taro: I just want to say that focusing on the positive definite Dirichlet matrix DR as a tool for stability analysis, instead of the first order approximation CR which has no sign guarantee, is a practical shift for how we approach modeling these systems.

Rosa: It is a practical shift because it gives you something concrete to work with that has known properties, whereas that other term can be unpredictable in its sign. We're looking at this paper now, "Koopman Generator Decomposition for Port-Hamiltonian Systems".

Dev: And the main implication is that if you use these structure preserving lifted pH surrogates, you get guaranteed nonincrease of your surrogate storage when you apply that damping injection law u equals minus Kdye plus v.

Taro: If the pair of detection and dissipation matrices is detectable, then the equilibrium point of that surrogate model with zero input becomes asymptotically stable when you set the input to zero. That’s a nice stability guarantee.

Rosa: So, we've seen how they decompose the generator into conservative, dissipative, and input-port parts, and how they use that to build these finite dimensional lifted models for port hamiltonian systems.

Dev: And the key improvements are that we can rely on the metric bracket being positive semidefinite for stability analysis and that we get a structure preserving lift with damping injection guarantees in the lifted coordinates.

Taro: The paper shows us how to distinguish between the conservative weak matrix being skew and that dissipative metric-bracket matrix being positive semidefinite, which gives us a better mathematical framework for modeling.

Rosa: And they show it works exactly in state coordinates when you use the correct storage metric P equals Q, meaning zero residual error for identification purposes under those conditions.

Dev: So we've covered how this decomposition helps us build passive and controllable lifted models using these specific algebraic properties derived from the Koopman generator.

Taro: That’s a lot of material to digest about how to handle the conservative versus dissipative structure in these nonlinear systems when you are building data driven models.

Rosa: And that's where we wrap up on this paper, "Koopman Generator Decomposition for Port-Hamiltonian Systems". It gives us tools to move from abstract dynamics to concrete, stable surrogate models in a way that respects the underlying physics of energy flow.

The paper's summary: Rosa: So, this paper breaks down how the internal structure of these nonlinear port-Hamiltonian systems shows up in that infinitesimal Koopman generator.

Dev: Exactly what they did is look at the generator and say it splits into three parts: a conservative part, a dissipative part, and an input-port part.

Rosa: And they focus on how this decomposition looks when you look at the positive object, which they say isn't just the conservative part itself but something called the metric bracket involving f R.

Dev: That metric bracket is always greater than or equal to zero because it equals gradient f transpose R gradient f, which means it’s non-negative.

Taro: So this non-negativity of that specific bracket is what they use to prove something important about the energy balance for the Hamiltonian observable.

Rosa: Right. They find that this leads to an identity where the Hamiltonian observable, Ku H, equals minus that metric bracket plus some term involving input and output, which is less than or equal to zero.

Dev: That’s the core finding. It means even though the dissipative part isn't necessarily positive on every single observable you pick randomly, there’s this specific metric structure that keeps things bounded in a certain way.

Taro: And because of that, they use these bracket identities to motivate building these finite-dimensional weak Galerkin and data driven lifted models.

Rosa: The big idea is that when the measure used for the conservative part of the flow is conservative for the Hamiltonian interconnection flow, then their conservative contribution ends up being skew in that mass metric, and they see a positive semidefiniteness with this dissipative bracket matrix.

Dev: That means they can create these surrogate models that are passive and support damping injection directly in those lifted coordinates. It’s not just theory anymore; it’s about creating a model that respects the underlying physics of energy dissipation when you lift the system's dynamics into a higher space.

Taro: I wonder how robust these lifted models are when things go wrong in reality, like when the world misbehaves and we need graceful recovery. The paper mentions that to get those plant level conclusions, you need residual and injectivity assumptions.

Rosa: It means the lifted pH surrogate is passive in those lifted coordinates, but getting those final plant level conclusions depends on making sure the residual is small enough and that the model mapping is injective. They also mention using an empirical Dirichlet matrix, which they estimate by taking samples and calculating a specific matrix involving those measurements.

Dev: The way they frame the improvements is really helpful because it shows how to use this structure-preserving lifting for actual control design. They show that for the closed loop storage, you can satisfy an inequality where the time derivative of the Hamiltonian observable is bounded by terms related to that metric bracket and input.

The paper's summary: Taro: So if we can guarantee that this stored energy is nonincreasing using a damping injection law, what does that imply for stability when we are dealing with these lifted coordinates?

Rosa: If you have a pair of storage and dissipation matrices, and you make sure the detection condition is met, then the equilibrium point of the surrogate model with no external input becomes asymptotically stable when you set the input to zero. The damping injection law they propose is something like minus Kdye plus v, where Kd is positive definite.

Dev: That specific damping injection law ensures that that surrogate storage quantity is nonincreasing, which directly leads to asymptotic stability of that zero point if the system meets certain detectability conditions and you use no external inputs. It’s a way to guarantee convergence by penalizing all nonzero components of the surrogate output.

Taro: I like how they distinguish between the conservative and dissipative parts using these algebraic structures. They show that even for a simple identity dictionary, where the dictionary is just x itself, the Dirichlet matrix remains positive semidefinite because it represents that metric bracket, whereas the first order derivation matrix generally doesn't have a sign guarantee.

Rosa: That distinction is key because it allows us to prioritize stability analysis on that symmetric metric bracket instead of relying on the first order approximation which can be messy. It provides a mathematically sound way to separate the conservative dynamics from the dissipative ones when you are learning dynamics from data.

Dev: And they use this structural distinction in their finite dimensional weak Galerkin matrices, showing that the conservative matrix is skew and the dissipative metric bracket matrix is positive semidefinite. That’s a very concrete algebraic property we can rely on when we are building these models, separating them from just any approximate first order derivation matrices.

Taro: So for someone who only cares about how to build a stable model, this paper suggests focusing on that specific positive definite metric bracket rather than the general dissipative matrix approximation. That makes sense in terms of what you need for control synthesis.

Rosa: Exactly. And they show an illustrative example with state coordinates and the correct storage metric P equals Q, and in that case, the lifted pH surrogate is exact, meaning there's zero residual error when you use that specific dictionary. It replicates the nonlinear dynamics perfectly within those state coordinates without approximation errors related to how you build the generator structure.

Dev: That exact representation in state coordinates is a big deal because it means no modeling error from the generator itself, just from the finite dimensional lifting process, which they show can be zero under those conditions. It’s a very strong result for identification accuracy.

Taro: The pendulum example they use also shows that for that Hamiltonian identity involving Ku H equals minus bracket plus input to hold exactly, it works perfectly when you use state coordinates and the correct storage metric P equals Q, which is good because it confirms the structure holds in a concrete physical case.

The paper's summary: Rosa: It really ties everything together by showing how this mathematical decomposition translates into tangible improvements for system identification and control. The empirical Dirichlet matrix estimation method they show can be used to estimate that critical symmetric metric bracket when you only have snapshots of the system data.

Dev: So, to wrap up the results, the main idea is using these bracket identities to construct lifted models that are passive in their new coordinates and support guaranteed damping injection for stability. This gives us a framework for robust control synthesis based on those lifted coordinates.

Taro: I just want to say that focusing on the positive definite Dirichlet matrix DR as a tool for stability analysis, instead of the first order approximation CR which has no sign guarantee, is a practical shift for how you approach modeling these systems.

Rosa: It is a practical shift because it gives you something concrete to work with that has known properties, whereas that other term can be unpredictable in its sign. We're looking at this paper now, "Koopman Generator Decomposition for Port-Hamiltonian Systems".

Dev: And the main implication is that if you use these structure preserving lifted pH surrogates, you get guaranteed nonincrease of your surrogate storage when you apply that damping injection law u equals minus Kdye plus v.

Taro: If the pair of detection and dissipation matrices is detectable, then the equilibrium point of that surrogate model with zero input becomes asymptotically stable when you set the input to zero. That’s a nice stability guarantee.

Rosa: So, we've seen how they decompose the generator into conservative, dissipative, and input-port parts, and how we use that to build these finite dimensional lifted models for port hamiltonian systems.

Dev: And the key improvements are that we can rely on the metric bracket being positive semidefinite for stability analysis and that we get a structure preserving lift with damping injection guarantees in the lifted coordinates.

Taro: The paper shows us how to distinguish between the conservative weak matrix being skew and that dissipative metric-bracket matrix being positive semidefinite, which gives us a better mathematical framework for modeling.

Rosa: And they show it works exactly in state coordinates when you use the correct storage metric P equals Q, meaning zero residual error for identification purposes under those conditions.

Dev: So we've covered how this decomposition helps us build passive and controllable lifted models using these specific algebraic properties derived from the Koopman generator.

Taro: That’s a lot of material to digest about how to handle the conservative versus dissipative structure in these nonlinear systems when you are building data driven models.

Rosa: And that's where we wrap up on this paper, "Koopman Generator Decomposition for Port-Hamiltonian Systems". It gives us tools to move from abstract dynamics to concrete, stable surrogate models in a way that respects the underlying physics of energy flow.

The paper's improvements: Tom: So, we're looking at how this paper suggests we can actually use these mathematical tools to make things work better for real-world systems.

Rosa: The paper points out a few concrete improvements they think are important for using these structure-preserving lifted models in the field.

Dev: They show that if you use state coordinates and you pick the right storage metric P equal to Q, then the lifted pH surrogate becomes exact.

Taro: So there's no modeling error from the generator structure at all, just from how we lift it into a higher space.

Rosa: Right. That’s zero residual error when you use that specific dictionary, which is a huge deal for identification accuracy because it means the model perfectly replicates the nonlinear dynamics in those coordinates.

Dev: And they suggest using these structure-preserving lifted pH surrogates specifically for control design where you can guarantee that the closed loop storage quantity is nonincreasing.

Taro: That means we can use this to ensure the system is passive and that damping injection actually provides a guaranteed nonincrease of the stored energy.

Rosa: Yes, because if you apply a damping injection law like minus K times your output plus some input v, where K is positive definite, then the surrogate storage goes down.

Dev: If you make sure the detection and dissipation matrices are detectable, then that equilibrium point of the surrogate model with zero external input becomes asymptotically stable when you set that input to zero.

Taro: So this means we can implement active damping strategies directly on those lifted models for guaranteed convergence in real-world scenarios.

Rosa: And they also highlight a structural distinction you can use for analysis, where you prioritize the positive definite Dirichlet matrix over the first order approximation because the latter doesn't always give you a sign guarantee.

Dev: That’s practical because it gives us something concrete to work with that has known properties, whereas that other term can be unpredictable in its sign when we are just learning dynamics from data.

Taro: And they show this distinction holds even for the simplest case, like when the dictionary is just the state coordinates itself, which confirms the structure is robust.

Rosa: So to sum up, these improvements give us a way to get exact identification in certain coordinate systems and a way to design controllers that are guaranteed to work passively in those new coordinates.

Dev: The paper gives us concrete rules for building models that respect the physics of energy flow, even when we are using data-driven methods.

Taro: So what this means for autonomy research is that we can build more reliable surrogates because they aren't just approximations; they have these known algebraic properties.

Rosa: And it opens up a path to designing controllers that are guaranteed to be stable in the lifted space, which is exactly what we need when deploying these systems outside the lab.

Conclusion: Rosa: So we’ve looked at how this paper breaks down the Koopman generator into conservative, dissipative, and input parts for port-Hamiltonian systems.

Dev: Basically, they show us how to build these lifted models that are passive in their new coordinates and support guaranteed damping injection in those lifted coordinates.

Taro: It really shows how we can move from abstract dynamics to concrete stable surrogate models while respecting the underlying physics of energy flow.

Rosa: They’ve found that by focusing on the metric bracket being positive semidefinite, we get a structure-preserving lift with damping injection guarantees.

Dev: The key improvement is relying on that metric bracket for stability analysis instead of just looking at the first order approximation which can have sign problems.

Taro: That algebraic distinction between the conservative skew matrix and that positive definite metric bracket gives us a much better way to separate the conservative dynamics from the dissipative ones.

Rosa: And they prove it works exactly in state coordinates when you use a specific storage metric, meaning zero residual error for identification purposes under those conditions.

Dev: So we’ve covered how this decomposition helps us build passive and controllable lifted models using these specific algebraic properties derived from the Koopman generator.

Taro: That’s a lot of material to digest about how to handle the conservative versus dissipative structure in these nonlinear systems when you are building data driven models.

Rosa: And that’s where we wrap up on this paper, "Koopman Generator Decomposition for Port-Hamiltonian Systems". It gives us tools to move from abstract dynamics to concrete, stable surrogate models in a way that respects the underlying physics of energy flow.

Dev: The main implication is that if you use these structure-preserving lifted pH surrogates, you get guaranteed nonincrease of your surrogate storage when you apply that damping injection law like minus K times your output plus some input v.

Taro: If the pair of detection and dissipation matrices is detectable, then the equilibrium point of that surrogate model with zero input becomes asymptotically stable when you set the input to zero. That’s a nice stability guarantee for autonomy applications.

Rosa: We’ve seen how they decompose the generator into conservative, dissipative, and input-port parts, and how we use that to build these finite dimensional lifted models for port hamiltonian systems.

Dev: And the key improvements are that we can rely on the metric bracket being positive semidefinite for stability analysis and that we get a structure preserving lift with damping injection guarantees in the lifted coordinates.

Taro: The paper shows us how to distinguish between the conservative weak matrix being skew and that dissipative metric-bracket matrix being positive semidefinite, which gives us a better mathematical framework for modeling.

Rosa: And they show it works exactly in state coordinates when you use the correct storage metric P equals Q, meaning zero residual error for identification purposes under those conditions.

Dev: So we've covered how this decomposition helps us build passive and controllable lifted models using these specific algebraic properties derived from the Koopman generator.

Taro: That’s a lot of material to digest about how to handle the conservative versus dissipative structure in these nonlinear systems when you are building data driven models.

Rosa: And that's where we wrap up on this paper, "Koopman Generator Decomposition for Port-Hamiltonian Systems". It gives us tools to move from abstract dynamics to concrete, stable surrogate models in a way that respects the underlying physics of energy flow.

University of Pennsylvania

eess.SY, cs.SY, math.DS

Submitted: 2026-03-18

Updated: 2026-11-04

Comments: 23 pages, 3 figures. Corrected and updated references; clarified mathematical assumptions and the distinction between surrogate and nonlinear-plant guarantees; expanded technical discussion and examples

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 86/100

The gist: We study how the vector-field structure of nonlinear port-Hamiltonian systems is reflected in the infinitesimal Koopman generator This paper asks how the pH decomposition of the vector field appears…

Key concepts

Koopman Generator Decomposition
This breaks down the system's dynamics generator into three parts: conservative (related to energy conservation), dissipative (related to energy loss), and input-port. It helps understand how the physical structure of a port-Hamiltonian system is mathematically reflected in its learned, finite-dimensional representation.
Metric Bracket [f, f]R
This is a specific mathematical object derived from the system's metric R. Unlike other parts of the generator that might lack a sign guarantee, this bracket is always positive semidefinite (≥ 0). It serves as the reliable measure for identifying the dissipative structure in the system.
Lifted pH Surrogate
This is a simplified, finite-dimensional model created using Koopman theory. These surrogates are designed to be passive and can incorporate damping injection. They allow control strategies to be synthesized directly on this simpler model, ensuring stability under certain conditions.

Terminology

Summary

We study how the vector-field structure of nonlinear port-Hamiltonian systems is reflected in the infinitesimal Koopman generator This paper asks how the pH decomposition of the vector field appears at the level of the Koopman generator

Koopman Generator Decomposition

The infinitesimal Koopman generator admits a natural bracket decomposition into a conservative interconnection-bracket derivation, a dissipative metric-bracket derivation, and an input-port derivation The conservative component is formally skew-adjoint on a test space whenever the conservative flow preserves the reference measure and the relevant boundary terms vanish The dissipative component is not claimed to be a positive operator on arbitrary observables; rather, the positive semidefinite object is the metric bracket [f, f]R = ∇f⊺R∇f ≥ 0 The Hamiltonian observable satisfies KuH = −[H, H]R + y⊺u ≤ y⊺u

Decomposition of Vector Field Structure

For a smooth observable f, the pH dynamics induce the bracket decomposition Kuf = ∇f⊺[(J − R)∇H + Gu] = ∇f⊺J−[f, H]R + ∇f⊺Gu A central distinction is that the dissipative derivation f → [f, H]R is not a positive operator on arbitrary observables The positive object is instead the metric, or Dirichlet, bracket [f, f]R = ∇f⊺R∇f ≥ 0

Finite-Dimensional Structure Preservation

The resulting lifted models are finite-dimensional portHamiltonian surrogates rather than exact copies of the full Koopman generator The main contributions include a Koopman bracket decomposition for pH systems that separates conservative, dissipative, and port derivations These identities motivate structure-preserving lifted port-Hamiltonian surrogates that are passive and support damping injection in the lifted coordinates

Weak Galerkin Identities

In finite dimensions, a weak Galerkin projection depends on the chosen dictionary and on the Galerkin or sampling measure, and it includes a mass matrix unless the dictionary is orthonormal The conservative contribution is skew in the mass metric when the Galerkin measure is conservative for the interconnection flow and boundary terms vanish, while the dissipative bracket induces a positive semidefinite Dirichlet matrix

Structure-Preserving Lifted Surrogates

The lifted pH surrogate has the form z˙ = (S − D)P z + Bu, S = −S⊺, D = D⊺ ⪰ 0, P = P⊺ ≻ 0 This surrogate is passive in lifted coordinates Corollary 1 states that if the conditions hold, the closed-loop storage satisfies ˙He = −(P z)⊺D(P z) − ye⊺Kdye + ye⊺v ≤ ye⊺v

Control Design for the Surrogate

The control design in this section is formulated for the finitedimensional lifted pH surrogate introduced in Section IV Under the damping-injection law u = −Kdye + v, Kd = K⊺d ⪰ 0, the closedloop storage satisfies ˙He = −(P z)⊺D(P z) − ye⊺Kdye + ye⊺v If the pair (Acl, Cdiss) is detectable, then the equilibrium z = 0 of the surrogate with v = 0 is asymptotically stable

Illustrative Examples

The linear pH example shows that with state coordinates and the correct storage metric P = Q, the lifted pH surrogate is exact and has zero residual The pendulum example shows that the Hamiltonian identity KuH = −[H, H]R + y⊺u holds exactly, while finite dictionaries can still have nonzero closure residuals The Dirichlet matrix DR remains positive semidefinite because it represents the metric bracket, whereas CR generally has no sign guarantee

Conclusion

The finite-dimensional control model used here is the lifted pH surrogate z˙ = (S − D)P z + Bu, S = −S⊺, D = D⊺ ⪰ 0, P = P⊺ ≻ 0 This surrogate is passive, and damping injection gives nonincrease of the surrogate storage; asymptotic stability of z = 0 follows under the stated zerodissipation detectability condition Conclusions in the original state require residual bounds together with injectivity or local observability of the lifting map The MPC formulation is therefore a surrogate-MPC formulation; its closed-loop guarantees depend on standard terminal ingredients The examples illustrate zero closure residual for linear pH systems in suitable coordinates and nonzero closure residual for a nonlinear pendulum dictionary, while confirming the distinction between the first-order derivation matrix CR and the positive Dirichlet matrix DR

REFERENCE

arXiv:2603.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Exact pH representation in state coordinates allows for an exact finite-dimensional Koopman embedding when using a dictionary like z = x, resulting in zero closure residual rN from Section IV. The improved AI system can perfectly replicate nonlinear port-Hamiltonian dynamics within the chosen coordinate space without approximation error related to the generator structure.

  2. Structure-preserving lifted pH surrogates can be used for control design where the closedloop storage satisfies ˙He = −(P z)⊺D(P z) − ye⊺Kdye + ye⊺v ≤ ye⊺v, ensuring the system is passive and damping injection provides guaranteed nonincrease of the surrogate storage. This allows for robust control synthesis based on the lifted coordinates.

  3. Damping injection law u = −Kdye + v, with Kd ≻ 0, ensures that the surrogate storage is nonincreasing, leading to asymptotic stability of the equilibrium z = 0 under zero-dissipation detectability conditions. The AI system can implement active damping strategies directly on the lifted model for guaranteed convergence.

  4. Finite-dimensional models derived from weak Galerkin identities ensure that the conservative weak matrix is skew while the dissipative metric-bracket matrix is positive semidefinite, meaning the AI can rely on these specific algebraic structures for modeling, distinguishing them from approximate first-order derivation matrices. This provides a mathematically sound framework for identifying conservative and dissipative components in learned dynamics.

  5. The use of empirical Dirichlet matrices DR = ∇ψ⊺R∇ψj dν represents the symmetric metric bracket, which is positive semidefinite, whereas the first-order dissipative derivation matrix CR generally has no sign guarantee. The AI system can prioritize the positive definite Dirichlet matrix DR for stability analysis over the potentially nonpositive first-order approximation CR.

Related papers