Collective steering in finite time: controllability on $\text{GL}^+(n,\mathbb{R})$

Mahmoud Abdelgalil; Tryphon T. Georgiou

REVIEW 3 major objections 4 minor 19 references

Collective steering in finite time: controllability on $\text{GL}^+(n,\mathbb{R})$

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Under the Kalman rank condition, a single broadcast time-varying gain can steer any swarm of identical linear systems to any same-orientation configuration in any prescribed positive time; no continuous universal feedback formula exists.

desk verdict Completes Brockett's program with a likely-correct main theorem, but the written proof has three repairable gaps that need a referee's attention. read the letter →

arxiv 2411.18766 v1 pith:F4TQAI7A submitted 2024-11-27 math.OC cs.SYeess.SY

Mahmoud Abdelgalil , Tryphon T. Georgiou This is my paper · ORCID

classification math.OCcs.SYeess.SY MSC 93B0593C1522E1549J15

keywords collectivesteeringcommonfeedbackgainbilinearcontrolsystemstrongcontrollabilitygenerallineargroupKalmanrankconditiontopologicalobstructionsymmetricpositivedefinitefactorization

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

Collective steering asks whether a swarm of n identical linear agents can be repositioned at will by broadcasting the same time-varying feedback gain to every agent. The paper's central claim is that the answer is yes exactly when the underlying pair (A,B) satisfies the classical Kalman rank condition: the induced right-invariant flow on the space of positive-determinant matrices is controllable and, more strongly, can be steered from the identity to any target in any prescribed positive time. The same result carries a negative counterpart: there is no universal feedback formula that is continuous in the target specification, because the state space is topologically non-contractible. If correct, the result reduces a seemingly geometric swarm-control problem to a textbook algebraic condition and gives an explicit five-segment construction for the control law.

What carries the argument

The load-bearing device is a constant-gain reset: for any chosen time step $t_s>0$, controllability of $(A,B)$ lets one pick $K_c$ so that $A_c=A+BK_c$ has purely imaginary eigenvalues in $(2\pi i/t_s)\mathbb{Z}$, forcing $e^{A_c t_s}=I$. With this reset, the linearized flow $\dot{\Phi}_t=A_c\Phi_t+BU_t$ is analyzed through the Gramian $W_t=\int_0^t e^{-A_c\tau}BB^\top e^{-A_c^\top\tau}d\tau$, a positive definite matrix that encodes the energy needed to steer in each direction. The paper defines a class of factors $\Phi_k$ such that $W_{t_s}^{-1/2}\Phi_k W_{t_s}^{1/2}$ is symmetric positive definite; for these, the optimal trajectory $\Phi_t^\star=e^{A_c t}(I+W_t W_{t_s}^{-1}(\Phi_k-I))$ stays inside $\mathrm{GL}^+(n,\mathbb{R})$, so the optimal input rewrites as a feedback gain $K_t=U_t^\star(\Phi_t^\star)^{-1}+K_c$. Concatenating five such factors — possible because every positive-determinant matrix is a product of five symmetric positive definite matrices — yields strong controllability. The obstruction half uses the polar decomposition $\mathrm{GL}^+(n,\mathbb{R})\simeq\mathrm{Sym}^+(n)\times\mathrm{SO}(n)$ and the non-contractibility of $\mathrm{SO}(n)$: a continuous universal steering formula would give a homotopy contracting the state manifold, which is impossible.

What would settle it

Compute, for the double-integrator pair with $t_s\to0$, the five factors $\Phi_k$ obtained from the symmetric positive definite factorization of a target such as the rotation matrix $e^{(\pi/4)\Omega}$, and check numerically that each segment $\Phi_{k,t}^\star=e^{A_c t}(I+W_t W_{t_s}^{-1}(\Phi_k-I))$ stays nonsingular over $[0,t_s]$ and that $K_t=U_t^\star(\Phi_t^\star)^{-1}+K_c$ is integrable. The theorem is falsified if any determinant crosses zero or any gain develops a non-integrable singularity; Section 5 presents this example as evidence that the construction works for every $t_s>0$.

Watch

Extended reading notes

Core claim

The paper establishes the equivalence: $(A,B)$ satisfies the Kalman rank condition (the block matrix $[B\ AB\ \cdots\ A^{n-1}B]$ has full row rank) if and only if the bilinear system $\dot{\Phi}_t=(A+BK_t)\Phi_t$ is controllable on $\mathrm{GL}^+(n,\mathbb{R})$ — the invertible $n\times n$ real matrices with positive determinant — and if and only if it is strongly controllable, meaning every $\Phi_{\mathrm{fn}}\in\mathrm{GL}^+(n,\mathbb{R})$ and every $t_{\mathrm{fn}}>0$ admit an integrable gain $K_\cdot:[0,t_{\mathrm{fn}}]\to\mathbb{R}^{m\times n}$ taking the flow from $I$ to $\Phi_{\mathrm{fn}}$. The constructive proof fixes a step $t_s=t_{\mathrm{fn}}/5$, chooses a constant gain $K_c$ with $e^{(A+BK_c)t_s}=I$, and factors the target into five symmetric positive definite matrices in the metric of the Gramian $W_{t_s}$; each factor is then traversed by the optimal open-loop segment, which stays invertible, and the gains are read off in feedback form. The paper also shows that the optimal open-loop controller can fail to be expressible as a feedback law, repairs an earlier proposed sufficient condition by replacing a norm bound with a symmetrized one, and proves that a continuous universal feedback formula cannot exist since $\mathrm{GL}^+(n,\mathbb{R})$ is not contractible. For the group of orientation-preserving diffeomorphisms, it obtains a partial reachability result under a contraction condition.

Load-bearing premise

The whole construction assumes that for every controllable pair $(A,B)$ and every chosen time step $t_s>0$ there is a constant gain $K_c$ with $e^{(A+BK_c)t_s}=I$; the paper invokes this standard linear-systems fact without proof, and the stitching argument collapses if it ever failed.

Editorial extensions

If this is right

For any controllable $(A,B)$, a swarm of $n$ identical agents can be sent from one nondegenerate configuration to any other configuration with the same orientation in arbitrarily short time using only a broadcast time-varying gain; the final time does not constrain feasibility.
The proof is constructive: every target is reached by at most five concatenated optimal segments, so the result yields an explicit piecewise-smooth feedback law rather than a bare existence statement.
Because a continuous universal formula is ruled out, any practical implementation must either switch between local charts, accept discontinuities in the gain as a function of the target, or use a non-feedback selection rule.
Strong controllability of the $\mathrm{GL}^+(n,\mathbb{R})$ flow implies controllability of the Lyapunov equation on positive definite matrices, so the same Kalman condition governs steering of covariance matrices.
For orientation-preserving diffeomorphisms, the contraction-condition result indicates how the same stitching idea extends beyond linear maps, albeit with a Lipschitz smallness hypothesis rather than a rank condition.

Reading between the lines

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

The uniform five-factor bound suggests that control effort for a fixed target concentrates as $t_{\mathrm{fn}}\to0$; a testable extension is to characterize the minimal $\sup_t\|K_t\|$ as a function of $t_{\mathrm{fn}}$ and the target's distance from the identity.
The symmetrized condition (16) gives a quantitative ball of targets reachable by a single optimal feedback segment; a natural next question is whether iterating the optimal map rather than stitching five factors yields shorter or less aggressive paths.
The topological obstruction applies to any state space with nontrivial fundamental group; for $\mathrm{SO}(n)$ targets one expects branch cuts analogous to angle unwrapping in two dimensions.
For the diffeomorphism setting, the contraction condition (28) is only sufficient; the paper leaves open whether a rank-like condition characterizes reachability on $\mathrm{Diff}^+(\mathbb{R}^n)$ and whether the five-factor factorization has an infinite-dimensional analogue.

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.

Desk Editor's Note

read the letter

This paper finishes Brockett's program by proving that the Kalman rank condition is necessary and sufficient for both controllability and strong controllability of the bilinear system on GL+(n,R). I think the main theorem is correct, but the proof as written is not fully there. Three things need fixing.

What's genuinely new and good: Theorem 3's equivalence is new, and the use of Ballantine's five-factor SPD decomposition to get strong controllability is a clever and clean construction. The correction of Brockett's norm argument in Section 2.2 is careful and right—the symmetrized condition (16) does repair the flaw. The topological obstruction for continuous feedback laws in Section 4 is a nice addition.

Now the soft spots, in order of severity. First, the existence of Kc with e^{Ac ts}=I is asserted without proof. It's true—pole placement can put the eigenvalues at distinct imaginary multiples of 2π/ts—but since the entire stitching construction rests on it, it should be stated as a lemma. Second, in Theorem 3 the implication (ii)=> (i) is dismissed as trivial. It isn't. One needs the invariant-subspace argument: the controllable subspace R is A-invariant and contains Im(B), so quotient dynamics are independent of K. Any reachable Φ must map R into itself and have quotient form e^{Āt}π, which excludes generic matrices. The claim is true but the argument is missing. Third, Theorem 1's proof misapplies Lemma 2. Lemma 2 bounds ||W^{1/2} Δ W^{-1/2}||, but Proposition 1's condition needs ||W^{-1/2} Φ_k W^{1/2} - I||. These two are similar, but not equal in norm, and the proof doesn't bridge that. This one is also fixable—you can choose the factorization so the correct norm is small—but as written it's a real gap.

Bottom line: the ideas are solid and the main result is almost certainly correct. The paper is for geometric control and ensemble-control people, and it deserves a serious referee. Send it out, but the referee should ask for the pole-placement lemma, a written necessity proof, and a corrected application of Lemma 2.

Referee Report

3 major / 4 minor

Summary. The paper studies the right-invariant bilinear control system Phi_dot = (A + B K) Phi on GL^+(n,R), which models the collective steering of n identical linear agents by a broadcast time-varying feedback gain. The main results are Theorem 1, proving controllability under the Kalman rank condition, and Theorem 3, asserting the equivalence of the Kalman rank condition, controllability, and strong controllability in arbitrary finite time. The proof strategy repairs a norm-monotonicity lemma of Brockett through a symmetrized condition, decomposes the target matrix into symmetric positive definite factors, and concatenates motion primitives obtained from optimal open-loop segments that remain in GL^+(n,R). The paper also gives a topological obstruction to the existence of a universal feedback formula that is continuous in the terminal data and presents partial results for dynamics on orientation-preserving diffeomorphisms.

Significance. If the main result is correct, it gives a clean and usable characterization: for identical linear agents, the Kalman rank condition is exactly equivalent to strong controllability of the collective flow on GL^+(n,R), and steering can be done arbitrarily fast. The symmetrized condition in Proposition 1 is a genuine repair of Brockett's lemma, and the use of symmetric positive definite factors as motion primitives in Proposition 2 is elegant. The constructive proof is parameter-free and does not rely on numerical fitting. The topological obstruction in Section 4 is a useful complement. However, two load-bearing arguments are missing from the manuscript, and the illustrative example in Section 5 contains a demonstrably incorrect factor, so the paper cannot be accepted in its present form.

major comments (3)

[Section 2.2, Eq. (14); proofs of Theorems 1 and 3] The existence of a constant gain K_c with e^{(A+BK_c)t_s} = I is asserted without proof, with the text saying only that this is guaranteed by the Kalman rank condition. Since every motion primitive in Theorems 1 and 3 depends on this assertion, it should be stated as a lemma and proved or cited. A proof is available by pole placement: assign the controllable pair a real spectrum consisting of distinct purely imaginary eigenvalues in (2 pi i / t_s) Z, so that A+BK_c is diagonalizable over C and e^{(A+BK_c)t_s} = I. As written, the construction is not fully verifiable from the text.
[Theorem 3, proof] The proof dismisses the implications (iii) => (ii) and (ii) => (i) as trivial. The first is indeed definitional, but the second is not. When the Kalman rank condition fails, the controllable subspace R is a proper A-invariant subspace containing Im B; writing pi for the quotient map onto R^n / R, every trajectory of (5) satisfies pi Phi_t = e^{bar A t} pi, where bar A is induced by A on the quotient. Hence any reachable Phi_fn must preserve R and satisfy this quotient condition for some t, which excludes generic elements of GL^+(n,R). This argument, or a citation, should be supplied before the equivalence is claimed.
[Section 5, definitions of S3 and the factorization S3 S2 S1 = Phi_fn] The matrix S3 = (S2 S1^2 S2)^(-1/2) S2 S1 is the orthogonal polar factor of S2 S1 and is not symmetric positive definite in general. Consequently S3 S2 S1 has the same singular values as S2 S1, which are not all 1 for a 45-degree rotation, so the claimed factorization and the corresponding segment Phi_3 do not satisfy condition (19) as written. The illustrative example needs a corrected positive definite factorization or should be removed; this does not affect the main theorems, but as stated the example is erroneous.

minor comments (4)

[Proof of Proposition 1, Eq. (18)] The displayed factor contains a typographical double plus sign, reading '(I + + ... )'; the extra '+' should be removed.
[Proof of Theorem 1] The deduction from Lemma 2 should explicitly state that Lemma 2 is applied to the inverse Gramian W_{ts}^{-1}, or equivalently to the transposed increment (Phi_k - I)^T; otherwise the displayed bound on ||W_{ts}^{-1/2}(Phi_k - I) W_{ts}^{1/2}|| does not follow from the version of the lemma applied to W_{ts}.
[Example 1] For odd n, the matrix -I does not belong to GL^+(n,R), so the sentence 'The case of odd n is not substantially different' should either be expanded with a valid terminal matrix or removed.
[Section 6, Proposition 6] The contraction assumption on psi should specify the norm with respect to which psi is a contraction, since the proof uses the Euclidean norm in the displayed estimate.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the controllability proofs are built from external factorization and standard Grammian/pole-placement facts, with the sole self-citation non-load-bearing.

full rationale

The central claims are not circular. Theorem 1 stitches motion primitives whose invertibility is guaranteed by Proposition 1, which rests on Grammian monotonicity and the choice e^{Ac ts}=I; the existence of such Kc is asserted from the Kalman rank condition via standard pole placement, and while unproved in the text it is an external, standard fact that does not encode the target controllability. Theorem 3's strong controllability is built from Proposition 3, which factors any Phi_fn into five symmetric positive definite factors using Ballantine's external Theorem 2, and then applies Proposition 2; no fitted parameter or data is involved. The only self-citation is in Section 3, where the sentence 'This fact follows from [1, Proposition 4]' is immediately followed by 'A rather remarkable result on such a factorization appeared much earlier in [4-6]', and the actual proof invokes Theorem 2, so the self-citation is not load-bearing. The concluding remarks cite [1] only for motivation and anticipated future work. The paper does have verification gaps: the pole-placement assertion underlying (14') is not proved, and the necessity directions in Theorem 3 are dismissed as 'trivial' even though (ii)⇒(i) requires a written argument about the uncontrollable subspace. These are omitted proofs, not circular reductions: no equation is defined in terms of the target result, and no fitted quantity is renamed as a prediction.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the standing Kalman assumption, on the existence of a time-reset feedback Kc (true, but unproved in the text), and on Ballantine's external factorization theorem. No fitted constants or invented entities appear.

assumptions (4)

domain assumption The pair (A,B) satisfies the Kalman rank condition throughout the main theorems.
Imposed as standing assumption in Theorems 1 and 3; the paper proves it is necessary, but the sufficient direction is conditional on it.
standard math For every ts > 0 there exists Kc with e^{(A+BKc)ts} = I.
Asserted in Section 2.2 as guaranteed by the Kalman rank condition; follows from pole placement assigning purely imaginary eigenvalues in 2πi Z, but not proved in the text. Load-bearing for the stitching construction.
standard math Every matrix in GL+(n,R) is a product of at most five symmetric positive definite matrices (Ballantine's theorem).
Theorem 2, cited from [6]; used in Proposition 3 and Theorem 3 to obtain motion primitives satisfying (19).
standard math GL+(n,R) is not contractible for n≥2.
Used in Section 4.2 via π1(GL+(n,R)) = π1(SO(n)) to rule out continuous universal feedback laws.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Collective steering in finite time: controllability on $\text{GL}^+(n,\mathbb{R})$." pith.science (2026). https://pith.science/paper/F4TQAI7A

@misc{pith2026241118766,
  author       = {Pith},
  title        = {Pith review of: Collective steering in finite time: controllability on $\textGL^+(n,\mathbbR)$},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F4TQAI7A}},
  note         = {Machine review of arXiv:2411.18766}
}

read the original abstract

We consider the problem of steering a collection of n particles that obey identical n-dimensional linear dynamics via a common state feedback law towards a rearrangement of their positions, cast as a controllability problem for a dynamical system evolving on the space of matrices with positive determinant. We show that such a task is always feasible and, moreover, that it can be achieved arbitrarily fast. We also show that an optimal feedback control policy to achieve a similar feat, may not exist. Furthermore, we show that there is no universal formula for a linear feedback control law to achieve a rearrangement, optimal or not, that is everywhere continuous with respect to the specifications. We conclude with partial results on the broader question of controllability of dynamics on orientation-preserving diffeomorphisms.

Figures

Figures reproduced from arXiv: 2411.18766 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 18 canonical work pages

[1]

The Holonomy of Optimal Mass Transport: The Gaussian-Linear Case

Mahmoud Abdelgalil and Tryphon T Georgiou. Sub-Riemannian geometry, mixing, and the holonomy of optimal mass transport. arXiv preprint arXiv:2408.14707, 2024

work page Pith review arXiv 2024
[2]

Control on the manifolds of mappings with a view to the deep learning

Andrei Agrachev and Andrey Sarychev. Control on the manifolds of mappings with a view to the deep learning. Journal of Dynamical and Control Systems, 28(4):989–1008, 2022

work page 2022
[3]

Controllability on the group of diffeomorphisms

Andrei A Agrachev and Marco Caponigro. Controllability on the group of diffeomorphisms. In Annales de l’Institut Henri Poincar´e C, Analyse non lin´eaire, volume 26, pages 2503–2509. Elsevier, 2009

work page 2009
[4]

Products of positive definite matrices

Charles Ballantine. Products of positive definite matrices. I. Pacific Journal of Mathematics , 23(3):427–433, 1967

work page 1967
[5]

Products of positive definite matrices

Charles Ballantine. Products of positive definite matrices. II. Pacific Journal of Mathematics, 24(1):7–17, 1968

work page 1968
[6]

Products of positive definite matrices

CS Ballantine. Products of positive definite matrices. III. Journal of Algebra, 10(2):174–182, 1968

work page 1968
[7]

System theory on group manifolds and coset spaces

Roger W Brockett. System theory on group manifolds and coset spaces. SIAM Journal on control, 10(2):265– 284, 1972

work page 1972
[8]

Optimal control of the Liouville equation

Roger W Brockett. Optimal control of the Liouville equation. AMS IP Studies in Advanced Mathematics, 39:23, 2007

work page 2007

Show all 19 references

[9]

Optimal steering of a linear stochastic system to a final probability distribution, Part II

Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. Optimal steering of a linear stochastic system to a final probability distribution, Part II. IEEE Transactions on Automatic Control, 61(5):1170–1180, 2015

2015
[10]

Optimal transport over a linear dynamical system

Yongxin Chen, Tryphon T Georgiou, and Michele Pavon. Optimal transport over a linear dynamical system. IEEE Transactions on Automatic Control, 62(5):2137–2152, 2016

2016
[11]

Topological semigroups, sets of generators, and controllability

Ronald Hirschorn. Topological semigroups, sets of generators, and controllability. Duke Mathematical Journal, 40(4):937 – 947, 1973

1973
[12]

Geometric control theory

Velimir Jurdjevic. Geometric control theory. Cambridge University Press, 1997

1997
[13]

Control systems on Lie groups

Velimir Jurdjevic and H ´ector J Sussmann. Control systems on Lie groups. Journal of Differential equations , 12(2):313–329, 1972

1972
[14]

Ensemble control of finite-dimensional time-varying linear systems

Jr-Shin Li. Ensemble control of finite-dimensional time-varying linear systems. IEEE Transactions on Automatic Control, 56(2):345–357, 2010

2010
[15]

A convexity principle for interacting gases

Robert J McCann. A convexity principle for interacting gases. Advances in Mathematics, 128(1):153–179, 1997

1997
[16]

Some remarks on controllability of the Liouville equation

Maxim Raginsky. Some remarks on controllability of the Liouville equation. arXiv preprint arXiv:2404.14683, 2024

2024 arXiv
[17]

Lecture notes on elementary topology and geometry

Isadore Manuel Singer and John A Thorpe. Lecture notes on elementary topology and geometry. Springer, 2015

2015
[18]

Universal approximation power of deep residual neural networks through the lens of control

Paulo Tabuada and Bahman Gharesifard. Universal approximation power of deep residual neural networks through the lens of control. IEEE Transactions on Automatic Control, 68(5):2715–2728, 2023

2023
[19]

A connected Lie group equals the square of the exponential image

Michael W ¨ustner. A connected Lie group equals the square of the exponential image. Journal of Lie Theory , 13(1):307–309, 2003. 15

2003

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

The Holonomy of Optimal Mass Transport: The Gaussian-Linear Case

Mahmoud Abdelgalil and Tryphon T Georgiou. Sub-Riemannian geometry, mixing, and the holonomy of optimal mass transport. arXiv preprint arXiv:2408.14707, 2024

work page Pith review arXiv 2024

[2] [2]

Control on the manifolds of mappings with a view to the deep learning

Andrei Agrachev and Andrey Sarychev. Control on the manifolds of mappings with a view to the deep learning. Journal of Dynamical and Control Systems, 28(4):989–1008, 2022

work page 2022

[3] [3]

Controllability on the group of diffeomorphisms

Andrei A Agrachev and Marco Caponigro. Controllability on the group of diffeomorphisms. In Annales de l’Institut Henri Poincar´e C, Analyse non lin´eaire, volume 26, pages 2503–2509. Elsevier, 2009

work page 2009

[4] [4]

Products of positive definite matrices

Charles Ballantine. Products of positive definite matrices. I. Pacific Journal of Mathematics , 23(3):427–433, 1967

work page 1967

[5] [5]

Products of positive definite matrices

Charles Ballantine. Products of positive definite matrices. II. Pacific Journal of Mathematics, 24(1):7–17, 1968

work page 1968

[6] [6]

Products of positive definite matrices

CS Ballantine. Products of positive definite matrices. III. Journal of Algebra, 10(2):174–182, 1968

work page 1968

[7] [7]

System theory on group manifolds and coset spaces

Roger W Brockett. System theory on group manifolds and coset spaces. SIAM Journal on control, 10(2):265– 284, 1972

work page 1972

[8] [8]

Optimal control of the Liouville equation

Roger W Brockett. Optimal control of the Liouville equation. AMS IP Studies in Advanced Mathematics, 39:23, 2007

work page 2007