Control theory and splitting methods

1. Problem Statement and Motivation

The paper addresses the construction of numerical splitting methods for evolution equations of the form:
$$\dot{x}(t) = f_0(x(t)) + f_1(x(t))$$
where the flows associated with $f_0$ and $f_1$ are easily integrable. The authors focus specifically on forward splitting methods, where the coefficients associated with the flow of $f_0$ (denoted $\alpha_i$) must be positive ($\alpha_i > 0$). This constraint arises in non-reversible dynamics, such as the heat equation or stochastic differential equations, where backward time integration is ill-posed or unstable.

The Core Challenge:

• Order Barrier: It is a well-known result that for real-valued coefficients with $\alpha_i > 0$ and $\beta_i \in \mathbb{R}$, the maximum achievable order of a splitting method is 2.
• Open Questions:
  • Can arbitrary high-order methods be constructed if complex coefficients are allowed for $f_1$ (while keeping $\alpha_i > 0$)?
  • What are the precise algebraic obstructions preventing higher orders in the real case?
  • How do these numerical limitations relate to small-time local controllability (STLC) in control theory?

2. Methodology: The Control-Theoretic Link

The authors establish a deep isomorphism between numerical splitting schemes and control-affine systems.

• The Interpretation: A splitting method is interpreted as a trajectory of the control system:
  $$\dot{x}(t) = f_0(x(t)) + u(t)f_1(x(t))$$
  where the control $u(t)$ is a finite sum of Dirac masses (impulsive controls). The coefficients of the splitting method correspond to the time durations ($\alpha_i$) and impulse amplitudes ($\beta_i$).
• The Goal: Finding a splitting method of order $N$ is equivalent to finding a control $u$ such that the flow of the time-dependent vector field $f_0 + u(t)f_1$ approximates the flow of $f_0 + f_1$ up to order $O(T^{N+1})$.
• The Magnus System: The authors introduce a formal differential equation on a free nilpotent Lie algebra (the Magnus system). They prove that the existence of a splitting method of order $N$ is equivalent to the controllability of this Magnus system.
  • Specifically, they utilize the Lie Algebra Rank Condition (LARC) and tools from geometric control theory (such as the extension method and tangent vector analysis) to analyze the reachability of the system.
• Key Tools:
  • Chen-Fliess Series and Magnus Expansion: Used to relate the solution of the control system to formal Lie series.
  • Hall Sets and Coordinates of the Second Kind: Used to decompose the state space and analyze the specific contributions of different Lie brackets (commutators).
  • Positivity Arguments: Crucial for handling the constraint $\alpha_i > 0$, which prevents simple time-reversal arguments used in driftless systems.

3. Key Contributions and Results

The paper provides rigorous proofs for several conjectures and establishes new existence/non-existence results.

A. Arbitrary Order with Complex Coefficients (Theorem 1.11 & 1.12)

• Result: For any order $N$, there exists a splitting method with coefficients in $(\mathbb{R}_+, \mathbb{C})$. That is, $\alpha_i > 0$ (real) and $\beta_i \in \mathbb{C}$ (complex).
• Control Theory Analogue: The system $\dot{x} = f_0 + u(t)f_1$ with complex-valued controls is small-time locally controllable (STLC) if and only if the complex Lie algebra generated by $f_0$ and $f_1$ spans the space ($Lie_{\mathbb{C}}(f_0, f_1)(0) = \mathbb{C}^d$).
• Significance: This resolves an open question from [8], proving that complex coefficients remove the order barrier for forward methods, provided the vector fields are holomorphic.

B. Order Restrictions for Real Coefficients (Theorem 1.13, 1.15, 1.17)

• The "Bad" Bracket: The authors identify the Lie bracket $W_1 = [f_1, [f_1, f_0]]$ (and its generalizations $W_j$) as the primary obstruction to high-order splitting with real coefficients.
• Result:
  • The maximal order for $(\mathbb{R}_+, \mathbb{R})$ methods is 2.
  • If one allows the flow of $W_1$ (or assumes $W_1=0$), the maximal order becomes 4.
  • Generally, if $W_1, \dots, W_{N-1}$ vanish (or are available as flows), the maximal order is $2N$.
• Mechanism: The obstruction arises because the coordinate associated with $W_1$ in the Magnus expansion is governed by a positive-definite quadratic form of the control. Since $\alpha_i > 0$, the control cannot oscillate sufficiently to cancel this drift, unlike in driftless systems where time-reversal allows cancellation.

C. High-Order Methods via Commutators and Degeneracies (Theorems 1.16, 1.18, 1.19)

• Commutator Flows: If the numerical scheme is allowed to explicitly use flows of specific commutators (e.g., $e^{t[f_1, [f_1, f_0]]}$), arbitrary high-order methods exist in $(\mathbb{R}_+, \mathbb{R})$.
• Degeneracies: If the specific system satisfies structural degeneracies (e.g., $[f_1, [f_1, f_0]] = 0$), high-order methods exist without needing extra commutator flows.
• Conjecture Proven: The paper proves the conjecture from [19] that for $N \in \mathbb{N}^*$, there exists an $(\mathbb{R}_+, \mathbb{R})$ splitting method of order $2N$ involving $X_0, X_1$ and the commutators $W_1, \dots, W_{N-1}$.

D. New Control Theory Results

• The paper proves that for drift systems with complex controls, the Lie Algebra Rank Condition is sufficient for STLC (Theorem 1.12). This contrasts with the real case, where LARC is necessary but not sufficient due to "bad" brackets.
• It establishes a new sufficient condition for STLC in systems with drift and real controls when specific "bad" brackets are neutralized by adding commutator controls (Theorem 1.19).

4. Technical Significance

1. Unification of Fields: The paper successfully bridges numerical analysis (splitting methods) and geometric control theory. It translates numerical order conditions into controllability problems, allowing the use of powerful control-theoretic tools (like the extension method and tangent vector analysis) to solve numerical existence problems.
2. Resolution of Open Problems: It definitively answers the question of whether arbitrary order forward splitting is possible with complex coefficients (Yes) and characterizes the exact algebraic obstructions in the real case.
3. Algebraic Structure: The authors utilize a specific basis of the free Lie algebra (related to Hall sets and the specific ordering of $M_j$ and $W_j$ brackets) to derive precise error estimates and positivity arguments. This provides a rigorous framework for analyzing "bad" brackets that obstruct controllability.
4. Practical Implications:
   • For stochastic and parabolic PDEs, the results justify the use of complex time steps to achieve high-order accuracy while maintaining stability (forward flows).
   • For Hamiltonian systems, the results clarify when high-order schemes can be built using only real coefficients by exploiting system degeneracies (e.g., specific commutators vanishing).

5. Conclusion

Beauchard, Busnot, and Marbach demonstrate that the limitations of numerical splitting methods are not merely algorithmic but are fundamentally rooted in the controllability properties of the underlying vector fields. By interpreting splitting schemes as control trajectories, they prove that:

• Complexity (complex coefficients) removes the order barrier.
• Structure (vanishing Lie brackets) allows high-order real methods.
• Positivity (forward flows) creates specific quadratic obstructions linked to "bad" Lie brackets.

This work provides a comprehensive theoretical foundation for designing high-order numerical integrators for non-reversible systems.