{"id":"96b2d9d1-6d6b-4c1a-a31a-6a986a985813","arxiv_id":"1908.05323","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A linear ensemble is uniformly controllable exactly when a matrix built from the control fields evaluated at the drift's preimages has full rank at every parameter value.","lead":"Engineers often need one shared control signal to drive a whole population of similar systems at once. This paper gives a rank test that tells when such an ensemble of linear systems can be steered to any desired state.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4's continuous-triangularization premise fails for A(β)=[[0,1],[β,0]] on [0,1], so the multidimensional reduction is not established.","rationale":"The scalar theorem appears sound: a measure-annihilation argument supports the rank condition, so the main objection is not there. The single load-bearing flaw is the multidimensional reduction in Section III, specifically Theorem 4's reliance on a continuous triangularization of every continuous A(β) with real eigenvalues. The concrete family A(β)=[[0,1],[β,0]] on [0,1] has real eigenvalues but admits no continuous triangularization: continuity forces T(0)=0 while similarity to the nonzero nilpotent A(0) forbids it. Consequently, the proof of (i)⇔(ii) does not apply to eigenvalue coalescence where the Jordan structure changes, and no limiting argument is provided. This does not refute the scalar results, nor does it prove the rank-test conclusion false; it means the multidimensional theorem is not established as stated. Since the reader already marked the paper CONDITIONAL on essentially this gap, I retain that verdict rather than moving to REJECT: adding a no-coalescence hypothesis or a genuine limiting argument may repair the theorem.","tokens_in":29143,"tokens_out":27844,"duration_ms":316889,"concrete_test":"Attempt to construct a continuous invertible P(β) for A(β)=[[0,1],[β,0]] on K=[0,1] such that P^{-1}AP is upper triangular. The topological argument shows this is impossible: as β→0+, continuity forces the triangular form to tend to the zero matrix, whereas at β=0 it must be similar to the nonzero nilpotent A(0)=[[0,1],[0,0]]. A successful repair of Theorem 4 should either rule out this case by an explicit hypothesis or supply a limiting argument that recovers the claimed equivalence at the coalescence point.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B and Theorem 4 require transforming a continuous A(β) with real eigenvalues into an upper triangular T(β) via a continuous invertible P(β). This is not always possible. Take K=[0,1] and A(β)=[[0,1],[β,0]]. The eigenvalues are ±√β, which are real for all β∈K. If a continuous P(β) existed with T=P^{-1}AP upper triangular, then for β>0 the eigenvalues are distinct, so T(β) is diagonal with entries ±√β. Hence T(β)→0 as β→0+. By continuity T(0)=0. But T(0) is similar to A(0)=[[0,1],[0,0]], a nonzero nilpotent matrix whose upper-triangular form cannot be the zero matrix. Contradiction. Thus the theorem's opening construction is impossible for this ordinary continuous family. The proof of (i)⇔(ii) in Theorem 4 therefore does not cover eigenvalue coalescence where the Jordan structure changes, and no alternative argument or limiting construction is supplied in the paper. The scalar Theorem 1 is not affected by this objection, but the multidimensional claims that depend on Theorem 4 inherit the gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies uniform ensemble controllability of time-invariant linear ensemble systems on C(K,R^n), where K is a compact subset of R. The authors introduce the idea of separating points and define an Ensemble Controllability Gramian. The main scalar result, Theorem 1, characterizes controllability of the multi-input scalar system by the condition that the control fields b_i restricted to each level set of the drift a span all continuous functions on that level set; Corollary 1 turns this into a finite-dimensional rank condition. The paper then extends the framework to multi-dimensional systems with diagonalizable drift and real eigenvalues (Theorems 2 and 3), to Jordan-block drift (Proposition 2 and Corollary 2), and finally to general real-eigenvalue matrices via triangularization (Theorem 4). The overarching claim is that ensemble controllability on an infinite-dimensional function space can be checked by finite-dimensional controllability tests on the reparameterized systems.","tokens_in":29384,"tokens_out":26794,"duration_ms":302083,"significance":"If the results are correct, the paper offers a clean and potentially useful algebraic criterion for a broad class of linear ensemble systems: the scalar condition 'separating points on every fiber' and the associated rank test for the Ensemble Controllability Gramian are natural and go beyond earlier necessary or sufficient conditions. The reparameterization viewpoint and the connection to the previous linear-parameter-variation results in [23] are valuable. The paper is also clearly written and the examples are illuminating. However, the multidimensional claims currently rest on several proof gaps, and the proof of the scalar Theorem 1 itself contains an unjustified inference, so the significance is conditional on a rigorous repair of these arguments.","major_comments":[{"comment":"The sufficiency proof contains an unjustified inference. After showing that L0|Ki = C(Ki,R) and that A|Ki is contained in L0|Ki, the text concludes 'A⊆L0' (page 9). This does not follow: a function whose restriction to each set of a partition lies in the restriction of a linear space need not itself belong to that global linear space. At best one could hope for A being contained in the closure of L0 in the uniform topology, but that containment is not shown. Additionally, the proof invokes Proposition 1 for the multi-input system restricted to Ki, although Proposition 1 is a single-input statement. Since Theorem 1 and its Corollary 1 underpin the multidimensional results, this gap is load-bearing and must be repaired, for example by proving directly that the closure of L0 is a subalgebra using the fiberwise rank condition.","section":"Section II-B, proof of Theorem 1"},{"comment":"The multidimensional proofs are sketched to a degree that is not sufficient for the claims made. Lemma 1 is proved by asserting the existence of a bijective parameterization ψ and then checking the scalar Gramian, but the appendix construction only treats the two-dimensional case and the equivalence between controllability of the original two-dimensional system on C(K,R^2) and controllability of the scalar system on C(K′,R) is not fully justified. Theorem 2's induction step (page 18) is a single paragraph: for n=k+1, the proof forms a vector with the scalarized first k states and the (k+1)-st state, but it does not verify that the resulting two-dimensional system satisfies the hypotheses of Lemma 1, nor that the induction hypothesis transfers to the coupled system with shared control U. Theorem 3's proof ends with 'the rest of the proof directly follows the same case discussed in Theorem 2' (page 19), which does not address the coupling among the n scalar sub-ensembles. These are central gaps for the main multidimensional characterization.","section":"Section III-A, Lemma 1 and Theorems 2–3"},{"comment":"Theorem 4 assumes that A(β) can be transformed to an upper triangular matrix T(β), but the paper does not prove that the transformation P(β) can be chosen continuously in β. A pointwise triangularization with a discontinuous P would not preserve uniform ensemble controllability on C(K,R^n), since the induced map on the function space would not be a bounded isomorphism. Note that the example A(β)=[[0,1],[β,0]] from the stress-test is not an actual counterexample: T(β)=[[√β,1],[0,−√β]] is continuous on [0,1] and similar to A(β) for every β. Still, the general existence of a continuous real Schur form is not automatic when eigenvalue multiplicities change, and the manuscript supplies no proof. The proof also relies on Corollary 2, whose statement is itself not rigorous: it refers to both J and Λ and never states that the similarity P(β) is continuous. Theorem 4 therefore needs an explicit lemma on continuous triangularization (or an explicit assumption) before the equivalence (i)⇔(ii) is established.","section":"Section III-B, Theorem 4"}],"minor_comments":[{"comment":"The phrase 'This means enables the characterization' is ungrammatical and should be rewritten.","section":"Abstract"},{"comment":"The word 'reparameteriezed' in Remark 4 is a typo; it should be 'reparameterized'.","section":"Remark 4"},{"comment":"The notation in Corollary 2 is inconsistent: it introduces P^{-1}AP=Λ, but then writes the transformed equation as dY/dt = JY + B̃U, and later refers to Z without defining it. The intended system should be stated clearly.","section":"Corollary 2"},{"comment":"The statement of Theorem 4 says 'This system is uniformly ensemble controllable' in part (i) without clearly identifying whether 'this system' is the original X-system or the transformed Y-system; since the theorem is formulated after a triangularizing transformation, the relation back to the original system should be explicit.","section":"Theorem 4"},{"comment":"The claim that the results extend to Lp_n(K,μ) because C(K,R^n) is dense in Lp_n(K,μ) needs more justification: density of a subset of states does not by itself transfer an approximation property of a particular control-linear reachable set from the uniform norm to the Lp norm.","section":"Remark 6"}],"recommendation":"major_revision","confidential_remarks":"The scalar part of the paper is promising and the Ensemble Controllability Gramian is a useful idea, but the proof of Theorem 1 has a real gap and the multidimensional reductions are too sketchy in their present form. The referee should not treat the stress-test example A=[[0,1],[β,0]] as a valid counterexample to continuous triangularization, because a continuous upper triangular form exists for that family; nonetheless, the general existence of a continuous triangularizing transformation is unproved and should be addressed. I would support publication after a rigorous revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me save you some time. The scalar result is the real contribution: Theorem 1/Corollary 1 gives a clean finite-rank test for uniform ensemble controllability of multi-input scalar linear ensembles, and I think it is basically correct, though the proof is not. The multidimensional part is a survey-style extension that rests on an unstated continuous-triangularization assumption, and Theorem 4 does not establish it.\n\nThe stress-test example sent with the report does not actually kill Theorem 4. For A(β)=[[0,1],[β,0]] on [0,1], a continuous triangularization does exist: take P=[[1,0],[√β,1]], which gives T=[[√β,1],[0,-√β]], continuous at β=0. So that specific objection should be withdrawn. But the general worry stands: no proof or reference is given that every continuous matrix family with real eigenvalues on a compact set admits a continuous P with P⁻¹AP upper triangular. Without that, Theorem 4 and Corollary 2 are conditional on an assumption that is not stated. This is a real gap, and it affects exactly the part of the paper that claims to cover non-diagonalizable systems.\n\nThe proof of Theorem 1 also has a hole. The paper claims A⊆\\overline{L0} follows from local restrictions to the injective branches. That inference is not valid as written; you cannot glue local approximations from a linear span without bump functions inside L0. The theorem is very likely salvageable by a direct argument: the full-rank Gramian yields continuous coefficients c_i(η), and polynomial approximation in a then puts ξ into \\overline{L0}. But the paper does not supply that argument.\n\nWhat is genuinely new and useful: the separating-points framing, the Ensemble Controllability Gramian, and the explicit rank condition for multi-input scalar ensembles with non-injective drift. Examples 2, 3, and 5 are concrete and illuminating. The citation to [23] for the linear-parameter-variation case is appropriate; the paper checks consistency rather than deriving its main claim from it. No data-fitting or circularity issues.\n\nThe induction in Lemma 1 and Theorem 2 is sketched, but the pieces seem recoverable. Theorem 3's block-diagonal normal form is plausible. For a referee, the main work is to fix Theorem 1's proof and to make the triangularization assumption in Theorem 4 explicit (or prove it for a restricted class like analytic families).\n\nWho this is for: people working on ensemble control of linear systems, quantum ensembles, neuron populations, and robot swarms. It deserves peer review; I would send it out, expecting major revision on the multidimensional sections. The scalar theorem is worth publishing on its own.","headline":"The scalar Gramian rank test is a genuinely useful result and likely correct, but the multidimensional part leans on an unproven continuous-triangularization assumption, and the proof of the scalar theorem is too sloppy as written.","tokens_in":29879,"tokens_out":11637,"would_cite":true,"duration_ms":117780,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93B05","93C05","41A10","46E15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that uniform ensemble controllability of time-invariant linear ensembles with real eigenvalues is equivalent to finite-dimensional controllability tests on reparameterized individual systems, captured by the rank of an…","keywords":["ensemble controllability","separating points","Stone-Weierstrass theorem","Ensemble Controllability Gramian","linear ensemble systems","uniform ensemble controllability","polynomial approximation","reparameterization"],"falsifier":"Exhibit a continuous family $A(\\beta)$ on a compact interval with real eigenvalues for which no continuous global triangularization exists, and choose $B(\\beta)$ so that the induced reparameterized diagonal system is controllable for every eigenvalue tuple. If the original ensemble is not uniformly ensemble controllable on $C(K,\\mathbb{R}^n)$—for instance, if the closure of the Lie algebra generated by the drift and control fields is not all of $C(K,\\mathbb{R}^n)$—then the paper's Theorem 4 equivalence is false; a direct check on such a two-dimensional crossing example would settle the point.","tokens_in":28956,"feed_emoji":"🎛️","tokens_out":11383,"duration_ms":102622,"temperature":0.7,"pith_summary":"Ensemble control is the problem of steering a continuum of structurally identical systems—one for each parameter value $\\beta$—with a single broadcast input. This paper claims that for time-invariant linear ensembles on a compact parameter interval, uniform ensemble controllability can be decided by finite-dimensional checks: one evaluates controllability of each individual system after a parameter-dependent reparameterization, or checks the rank of a matrix called the Ensemble Controllability Gramian. For a scalar multi-input ensemble, the condition is exactly that the control fields restricted to each preimage of the drift function span all continuous functions on that preimage, equivalently $\\operatorname{rank}(D(\\eta))=\\kappa(\\eta)$ at every drift value $\\eta$. For multidimensional systems with real eigenvalues, the paper reduces ensemble controllability to controllability of the induced reparameterized system at each eigenvalue tuple. If correct, the result turns an infinite-dimensional function-space analysis into a finite-dimensional algebraic test, with direct practical value for population-level control in areas such as neurostimulation, quantum control, and robot swarms.","feed_headline":"Ensemble controllability reduces to a rank test","feed_subtitle":"One broadcast input steering a whole linear ensemble reduces to a finite-dimensional controllability check.","key_machinery":"The load-bearing object is the Ensemble Controllability Gramian $D(\\eta)$, a matrix whose rows are the control row vectors $b(\\beta)$ evaluated at the $\\kappa(\\eta)$ parameter values with $a(\\beta)=\\eta$ (or, in the multidimensional case, at the preimages of each eigenvalue). It carries the argument by turning the density question for the Lie algebra generated by the drift and control fields into a linear-algebra statement: full row rank means the control fields separate every pair of parameter values lying on different injective branches of the drift. The second mechanism is the reparameterization map that rewrites an $n$-dimensional ensemble, coordinate by coordinate, as a one-dimensional ensemble whose drift is built from the eigenvalue functions; this reduces multidimensional controllability to the scalar separating-points criterion. The Stone–Weierstrass theorem supplies the final step: a closed subalgebra of $C(K,\\mathbb{R})$ that separates points and contains constants is dense, so controllability is equivalent to separation.","core_discovery":"At the center is a dictionary between ensemble controllability and separation of points. The paper proves (Theorem 1) that the scalar multi-input ensemble $\\frac{d}{dt}x(t,\\beta)=a(\\beta)x(t,\\beta)+\\sum_{i=1}^m b_i(\\beta)u_i(t)$ is uniformly ensemble controllable on $C(K,\\mathbb{R})$ if and only if $\\operatorname{span}\\{b_1|_{a^{-1}(\\eta)},\\dots,b_m|_{a^{-1}(\\eta)}\\}=C(a^{-1}(\\eta),\\mathbb{R})$ for every $\\eta$ in the range of $a$; because each preimage is finite, this is equivalent to $\\operatorname{rank}(D(\\eta))=\\kappa(\\eta)$, where $D(\\eta)$ is the Ensemble Controllability Gramian. The multidimensional analogue (Theorems 2–4) states that when $A(\\beta)$ is diagonalizable or upper-triangularizable with real eigenvalues, the $n$-dimensional ensemble is uniformly ensemble controllable if and only if the reparameterized system, with each coordinate indexed by its own eigenvalue $\\eta_i=\\lambda_i(\\beta)$, is controllable on $\\mathbb{R}^N$ for every eigenvalue tuple; $N$ is the total number of preimage points. In particular, coincident eigenvalues and non-injective eigenvalue functions do not destroy controllability as long as enough independent controls separate the branches. The paper also shows that a Jordan-block drift requires as many controls as the system dimension, in contrast to the single chain of integrators familiar from finite-dimensional linear control.","pith_inferences":["The rank test suggests a design rule the paper does not spell out: choose $m$ control channels and tune the functions $b_i(\\beta)$ so that their evaluations on each preimage set are linearly independent, turning a verification criterion into a constructive control-design recipe.","A testable extension is to non-diagonalizable or non-real-spectrum ensembles: because the proofs lean on continuous triangularization, systems whose eigenvalue curves cross or whose eigenvectors cannot be chosen continuously may need a condition stated in terms of invariant subspaces rather than individual eigencurves.","The dictionary between separating points and controllability may carry over to nonlinear ensembles whose vector fields generate a separating Lie algebra; if so, the finite-dimensional Gramian test would become a general algebraic condition for population controllability.","Because the reparameterized individual systems are finite-dimensional, the result suggests that for linear ensembles the only genuinely infinite-dimensional obstruction is the geometry of the parameter-to-spectrum map, not functional analytic complexity."],"forward_implications":["If the main theorems are correct, a scalar linear ensemble with drift $a$ is uniformly ensemble controllable exactly when the number of controls is at least the maximum number of parameter values sharing a drift value and the Gramian $D(\\eta)$ has full rank at every $\\eta$.","For $n$-dimensional ensembles with real eigenvalues, checking ensemble controllability reduces to checking finite-dimensional controllability of the reparameterized system for each eigenvalue tuple; no infinite-dimensional approximation is needed.","State coupling does not reduce the control budget: an ensemble whose drift is a single Jordan block requires $n$ independent controls, the same number as an ensemble with $n$ coincident real eigenvalues.","The separating-points argument carries over to $L^p_n(K,\\mu)$ and to $C_0(\\Omega,\\mathbb{R}^n)$ because $C(K,\\mathbb{R}^n)$ is dense in those spaces, so the rank conditions extend beyond compact parameter intervals and continuous state spaces.","For the linear-parameter-variation family $\\frac{d}{dt}X=\\beta A X+B U$, the criterion recovers the previously known necessary and sufficient conditions, including the requirement $\\operatorname{rank}(B)=n$ when $0$ lies in the parameter interval."],"supporting_citations":[{"why":"supplies the Stone–Weierstrass theorem and the density of C(K,R) in L^p spaces, the approximation results the proofs build on.","marker":"[11]"},{"why":"supplies the Weierstrass approximation theorem used in the scalar single-input case.","marker":"[1]"},{"why":"provides the earlier linear-parameter-variation ensemble controllability conditions that Example 4 rederives and that the paper generalizes.","marker":"[23]"},{"why":"establishes uniform ensemble controllability for one-parameter families of time-invariant linear systems, the setting extended here.","marker":"[13]"}],"fun_headline_variants":["Ensemble controllability passes a rank test","Infinite ensemble control, finite rank check","Separating points tames ensemble controllability","Ensemble Gramian rank decides controllability","Finite-dimensional test for infinite ensembles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The multidimensional claims rest on the premise that $A(\\beta)$ can be continuously transformed to diagonal or upper-triangular form with the same real eigenvalue functions on the whole parameter set; if such a continuous transformation does not exist, the reparameterization step and hence the equivalence can fail.","fun_headline_variants_meta":{"raw":{"variants":["Ensemble controllability passes a rank test","Infinite ensemble control, finite rank check","Separating points tames ensemble controllability","Ensemble Gramian rank decides controllability","Finite-dimensional test for infinite ensembles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000532,"raw_usage":{"total_tokens":2613,"prompt_tokens":1053,"completion_tokens":1560,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":1495}},"tokens_in":669,"tokens_out":1560,"duration_ms":11186,"temperature":1.0,"reasoning_tokens":1495,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:20:52.300351+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Exhibit a continuous family $A(\\beta)$ on a compact interval with real eigenvalues for which no continuous global triangularization exists, and choose $B(\\beta)$ so that the induced reparameterized diagonal system is controllable for every eigenvalue tuple. If the original ensemble is not uniformly ensemble controllable on $C(K,\\mathbb{R}^n)$—for instance, if the closure of the Lie algebra generated by the drift and control fields is not all of $C(K,\\mathbb{R}^n)$—then the paper's Theorem 4 equivalence is false; a direct check on such a two-dimensional crossing example would settle the point.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Stone–Weierstrass theorem and the density of C(K,R) in L^p spaces, the approximation results the proofs build on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Weierstrass approximation theorem used in the scalar single-input case."},{"cited_title":"L I AND J","cited_arxiv_id":null,"evidence_quote":"provides the earlier linear-parameter-variation ensemble controllability conditions that Example 4 rederives and that the paper generalizes."},{"cited_title":"H ELMKE AND M","cited_arxiv_id":null,"evidence_quote":"establishes uniform ensemble controllability for one-parameter families of time-invariant linear systems, the setting extended here."}],"review_version":1}