Relationship Between Controllability Scoring and Optimal Experimental Design
Pith reviewed 2026-05-16 02:43 UTC · model grok-4.3
The pith
Finite-time controllability scores in linear networks match D- and A-optimality criteria from optimal experimental design via additive Gramian decomposition.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
We establish a structural connection between finite-time controllability scoring and approximate optimal experimental design (OED): the finite-time controllability Gramian decomposes additively across nodes, yielding an affine matrix model of the same form as the information-matrix model in OED. This yields a direct correspondence between the volumetric controllability score (VCS) and D-optimality, and between the average energy controllability score (AECS) and A-optimality, implying that the classical D/A invariance gap has a direct analogue in controllability scoring. By contrast, controllability scoring generically admits a unique optimizer, unlike approximate-OED formulations. Finally, a
What carries the argument
The finite-time controllability Gramian with its additive decomposition across nodes, which produces an affine matrix model matching the information matrix in OED.
If this is right
- Volumetric controllability score corresponds directly to D-optimality.
- Average energy controllability score corresponds directly to A-optimality.
- Controllability scoring problems generically possess a unique optimizer.
- Source-like nodes without negative self-loops are progressively downweighted by AECS as the horizon lengthens.
Where Pith is reading between the lines
- Algorithms developed for D- and A-optimal design could be repurposed to compute controllability scores without deriving new solvers.
- The long-horizon downweighting suggests that control strategies for large networks may need explicit time-scale adjustments for source nodes.
- Similar additive decompositions might link controllability to other optimality criteria such as E-optimality.
- Viewing controllability as an experimental-design problem opens the possibility of joint design of inputs and measurements in one optimization.
Load-bearing premise
The finite-time controllability Gramian for linear networked systems admits an additive decomposition across nodes.
What would settle it
Compute the finite-time controllability Gramian for a small linear network with known A and B matrices and verify whether it equals the sum of per-node contributions; mismatch would falsify the additive model.
Figures
read the original abstract
Controllability scores provide control-theoretic centrality measures that quantify the relative importance of state nodes in networked dynamical systems. We establish a structural connection between finite-time controllability scoring and approximate optimal experimental design (OED): the finite-time controllability Gramian decomposes additively across nodes, yielding an affine matrix model of the same form as the information-matrix model in OED. This yields a direct correspondence between the volumetric controllability score (VCS) and D-optimality, and between the average energy controllability score (AECS) and A-optimality, implying that the classical D/A invariance gap has a direct analogue in controllability scoring. By contrast, we point out that controllability scoring generically admits a unique optimizer, unlike approximate-OED formulations. Finally, we uncover a long-horizon phenomenon with no OED counterpart: source-like state nodes without a negative self-loop can be increasingly downweighted by AECS as the horizon grows. Two numerical examples corroborate this long-horizon downweighting behavior.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper claims a structural equivalence between finite-time controllability scoring and approximate optimal experimental design (OED) for linear networked systems. The finite-time controllability Gramian admits an exact additive decomposition across input nodes, producing an affine matrix-valued map identical in form to the information matrix of approximate OED. This yields direct correspondences between the volumetric controllability score (VCS) and D-optimality and between the average energy controllability score (AECS) and A-optimality. The manuscript further asserts that controllability scoring generically possesses a unique optimizer (unlike OED) and identifies a long-horizon downweighting effect for source-like nodes without negative self-loops under AECS, supported by spectral analysis and two numerical examples.
Significance. If the derivations hold, the work supplies a clean, parameter-free bridge between control-theoretic centrality measures and classical OED criteria, allowing transfer of optimality conditions and algorithms across domains. The explicit additive decomposition from linearity and the identification of the long-horizon phenomenon (absent in OED) constitute genuine contributions. Credit is due for grounding the correspondences in the spectral form of the Gramian and for providing numerical corroboration of the downweighting behavior.
major comments (2)
- [§3] §3 (main theorem on Gramian decomposition): the statement that W_T admits an exact additive decomposition W_T = sum_i x_i W_T^{(i)} is correct under the discrete-time linear assumption, but the manuscript does not explicitly state the discrete-time restriction or contrast it with the continuous-time case where the integral form may alter the affine structure; this assumption is load-bearing for the claimed OED equivalence.
- [§5] §5 (uniqueness of optimizer): the claim that controllability scoring generically admits a unique optimizer is asserted without a precise characterization of the set of systems for which this holds (e.g., when the Gramian map is strictly concave); a counter-example or sufficient condition would be needed to substantiate the contrast with approximate OED.
minor comments (3)
- [Abstract] The abstract refers to 'source-like state nodes without a negative self-loop' but the precise definition of 'source-like' and the sign condition on the self-loop appear only later; move the definition to the problem statement.
- [Numerical examples] Figure 2 (numerical example): the horizon axis scaling and the precise network adjacency matrix used are not stated in the caption, making reproduction difficult.
- [§2] Notation: the selection vector x is introduced without an explicit statement that its entries lie in [0,1] with sum equal to the number of inputs; this is standard in approximate OED but should be recorded once.
Simulated Author's Rebuttal
We thank the referee for the careful reading, positive evaluation, and constructive suggestions. We address each major comment below and will revise the manuscript accordingly to improve clarity and rigor.
read point-by-point responses
-
Referee: [§3] §3 (main theorem on Gramian decomposition): the statement that W_T admits an exact additive decomposition W_T = sum_i x_i W_T^{(i)} is correct under the discrete-time linear assumption, but the manuscript does not explicitly state the discrete-time restriction or contrast it with the continuous-time case where the integral form may alter the affine structure; this assumption is load-bearing for the claimed OED equivalence.
Authors: We agree that the additive decomposition and resulting OED equivalence rely on the discrete-time setting. The manuscript focuses on finite-time controllability for discrete-time linear networked systems (as is conventional for exact Gramian decompositions over finite horizons), but we did not explicitly flag this restriction or contrast it with continuous time. In the revision we will add a clear statement at the beginning of §3 that all results assume discrete-time dynamics, and we will include a brief remark noting that the continuous-time Gramian (defined via matrix exponential integral) preserves additivity under linearity but yields a different parameterization that does not directly map to the standard approximate-OED information matrix form. This clarification will be incorporated without altering the core derivations. revision: yes
-
Referee: [§5] §5 (uniqueness of optimizer): the claim that controllability scoring generically admits a unique optimizer is asserted without a precise characterization of the set of systems for which this holds (e.g., when the Gramian map is strictly concave); a counter-example or sufficient condition would be needed to substantiate the contrast with approximate OED.
Authors: We accept that the generic uniqueness claim would benefit from a more precise supporting statement. The uniqueness follows from the strict concavity of the log-det (VCS) and trace (AECS) objectives once the controllability Gramian is positive definite, which holds for almost all controllable linear systems with distinct node dynamics or full-rank input matrices. In the revision we will add a short paragraph in §5 giving a sufficient condition (positive-definiteness of the Gramian together with linear independence of the node-wise Gramian contributions) and noting that nongeneric counterexamples (identical node dynamics producing flat directions) are measure-zero. This strengthens the contrast with approximate OED, where combinatorial selection routinely permits multiple optima even for full-rank information matrices. revision: yes
Circularity Check
No significant circularity identified
full rationale
The paper's central derivation follows directly from the definition of the finite-time controllability Gramian W_T = sum_{k=0}^{T-1} A^k B B^T (A^T)^k. With B formed by concatenating per-node input vectors b_i, linearity immediately yields the exact additive decomposition W_T = sum_i x_i W_T^{(i)} as an algebraic identity, without any fitting, ansatz, or external uniqueness theorem. The VCS-D and AECS-A correspondences are then obtained by matching this affine matrix form to the standard information-matrix model in approximate OED; both optimality criteria are applied to identical mathematical objects. The long-horizon downweighting for source nodes is obtained from the explicit spectral expression of each W_T^{(i)} and corroborated by numerical examples. No load-bearing step reduces to a self-citation, a fitted parameter renamed as prediction, or a renaming of a known empirical pattern. The derivation is therefore self-contained against the system equations.
Axiom & Free-Parameter Ledger
axioms (1)
- domain assumption Finite-time controllability Gramian admits additive decomposition across nodes in linear networked systems
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.