{"id":"7efa8ea8-e8ed-4b36-94ea-4d6b0c938606","arxiv_id":"1908.07684","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For Ito stochastic LQ control with indefinite state and control weights, mean-square stabilizability is shown to be equivalent, under exact detectability, to solvability of a generalized algebraic Riccati equation.","lead":"This paper proves a necessary and sufficient condition for stabilizing an Ito stochastic system with indefinite cost weights: under a detectability assumption, stabilizability is equivalent to the existence of a maximal solution of a generalized algebraic Riccati equation. The result relaxes earlier positive-definiteness requirements and gives an explicit optimal feedback gain.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sufficiency proof of Theorem 2 depends on an unstated exact-observability lemma for the reduced pair (A22,C22,Q22^{1/2}); exact observability is stronger than the detectability in Assumption 1, and no derivation is supplied.","rationale":"The reader's verdict pinpoints the right spot. I re-read the sufficiency argument in Section 3: the steps up to eq. (33) follow from Lyapunov theory, and the logical jump is the sentence invoking Lemma 3 in [Qi et al., 2017] to conclude exact observability of the reduced subsystem (A22, C22, Q22^{1/2}). If that lemma is true, the chain works: exact detectability transferred to the closed-loop pair by Lemma 2, combined with Z2 > 0, rules out nonzero unobservable states, because the Lyapunov equation (28) would make E[x2' Z2 x2] constant on an unobservable trajectory, contradicting decay. But the paper neither states the lemma nor proves the transfer, and the cited source is not reproduced in the text. I do not see a different, more serious flaw; the stability of the remaining block, the input-to-state bound from [Hinrichsen et al.], and the optimal-cost identity are standard. Because the missing lemma is central and unstated, CONDITIONAL remains the appropriate verdict.","tokens_in":12206,"tokens_out":13994,"duration_ms":134200,"concrete_test":"State and prove Lemma 3 of [Qi et al., 2017] in the notation of this paper, checking it against eqs. (13)-(15), (28), and (33). Concretely: assume E[y' F_hatP(0,T) y] = 0 for nonzero y and derive from (33) that Q22^{1/2} x2(t) = 0 on [0,T]; then use eq. (28) with Z2 > 0 to show E[x2(t)' Z2 x2(t)] is constant, while Lemma 2 plus Assumption 1 forces E|x2(t)|^2 -> 0. A contradiction unless y = 0. If this derivation uses only Assumption 1, the gap is expository; if it needs any additional observability or positivity condition such as Q22 > 0, then Theorem 2's hypotheses are insufficient and the sufficiency proof needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The sufficiency proof of Theorem 2 is coherent through eq. (33), but the sentence 'Following from Lemma 3 in [Qi et al., 2017]' asserts exact observability of the reduced triple (A22, C22, Q22^{1/2}) from the standing assumptions. Exact observability is strictly stronger than the exact detectability stipulated in Assumption 1, and the positive semidefinite matrix Z2 > 0 in eq. (28) alone does not imply it; one must also rule out non-decaying unobservable modes of (A22, C22). The paper gives no proof of this implication, and Lemma 2, which would transfer exact detectability from (A, C, Q_hatP^{1/2}) to the transformed pair (\\bar A, \\bar C, \\bar Q_hatP^{1/2}), is also deferred with only a reference. This is the least secure link in the argument: if Lemma 3 of [Qi et al., 2017] requires an extra condition, or if the transfer in Lemma 2 fails, the positivity of F_hatP(0, T) in eqs. (35)-(36) does not follow, and the mean-square stabilization conclusion of Theorem 2 collapses. The gap is likely repairable, since detectability plus Z2 > 0 should forbid nonzero unobservable initial states (E[x2' Z2 x2] would be constant on such trajectories), but as submitted the theorem depends on unstated results.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies infinite-horizon mean-square stabilization for linear Itô stochastic systems with indefinite state and control weighting matrices, described by (1) and cost (2). It introduces a generalized algebraic Riccati equation (GARE, eq. (5)) and an associated set P of matrices defined by an LMI plus a kernel condition. The main claims are Theorem 1, stating that under mean-square stabilizability and P nonempty the generalized differential Riccati equation converges as t → −∞ to a maximal solution of the GARE, and Theorem 2, giving a necessary and sufficient stabilization condition in terms of existence of a maximal solution, together with the optimal stabilizing controller (20) and optimal cost E[x0' Pbar x0]. The proof technique decomposes the GARE solution into a positive semidefinite SARE solution plus an element of P, thereby reducing the indefinite problem to a definite one. A two-dimensional numerical example illustrates the result.","tokens_in":12550,"tokens_out":9483,"duration_ms":83405,"significance":"If correct, the result extends earlier work by Rami et al. on indefinite stochastic LQ control by allowing the matrix R + D'PD to be singular and by providing a stabilization characterization, not just an optimal-control solvability condition. The decomposition of a GARE solution into an SARE solution plus an element of the LMI set P is a conceptually interesting device, and the paper is clearly structured. The main contributions are weakened regularity assumptions and a stabilization theorem in a setting where only exact detectability is assumed. However, the rigor of the proof is currently undermined by an under-derived maximality claim in Theorem 1 and by an imported exact-observability condition in the sufficiency proof of Theorem 2 that is stronger than the standing assumptions; these gaps are load-bearing for the central result.","major_comments":[{"comment":"The maximality claim is under-derived. The proof fixes an arbitrary P_hat in P, constructs Z_P_hat(t,T) and its limit bar Z_P_hat, and defines bar P = bar Z_P_hat + P_hat. This construction shows only that bar P ≥ P_hat for that particular P_hat. To satisfy Definition 1, one must show bar P ≥ P_tilde for every P_tilde in P. The sentence 'for arbitrary P_hat and bar Z_P_hat ≥ 0, it is easy to verify that bar P ≥ P_hat' merely restates the construction; it does not compare bar P with an arbitrary element of P. The proof needs an additional argument, such as showing that the limiting object is independent of the chosen P_hat, or a comparison argument that bar P dominates each P_tilde in P.","section":"Section 3, proof of Theorem 1, final paragraph"},{"comment":"The sufficiency proof imports exact observability of the reduced triple (bar A22, bar C22, bar Q_P22^{1/2}) from 'Lemma 3 in [Qi et al., 2017]' without stating or proving that lemma. Exact observability is strictly stronger than the exact detectability in Assumption 1, and the paper gives no derivation of this property from the standing assumptions. The step is load-bearing: the positivity of bar F_P(0,T) in (34)-(36) depends on exact observability, and without it the proof that (bar A22, bar C22) is mean-square stable collapses. Either this lemma must be stated and proved from Assumption 1, or an additional assumption must be introduced. The same concern applies to Lemma 2, whose proof is omitted with only a reference to [Zhang et al., 2004]; that lemma is used to transfer exact detectability to the transformed system and is not self-contained.","section":"Section 3, proof of Theorem 2, after eq. (33)"},{"comment":"The decomposition bar P = bar Z_P_hat + P_hat with bar Z_P_hat ≥ 0 is asserted for an arbitrary P_hat in P as soon as the GARE has a solution bar P. This assertion is not justified by the GARE alone; it requires the maximality of bar P, i.e., bar P ≥ P_hat, which is part of the theorem statement but is not used explicitly in the proof. Without maximality, bar Z_P_hat = bar P - P_hat need not be positive semidefinite, and the Lyapunov function candidate V(t,x(t)) = E[x'(t) bar Z_P_hat x(t)] in (22) may not be valid. The proof should explicitly write bar Z_P_hat = bar P - P_hat ≥ 0 by maximality and then verify that it satisfies the SARE (13). This is repairable, but as written the sufficiency proof rests on an unstated use of the maximality hypothesis.","section":"Section 3, proof of Theorem 2, first paragraph of sufficiency"}],"minor_comments":[{"comment":"The title contains 'ItO' where 'Itô' is intended; the running text uses 'Itˆo', which should be typeset consistently.","section":"Throughout"},{"comment":"There is a typo: 'whther' should read 'whether'.","section":"Proof of Theorem 1, paragraph beginning 'Next, we mainly investigate'"},{"comment":"The notation 'Ker(R+D' P_hat D)⊆(KerB∩KerD)' should be written as 'Ker(R+D' P_hat D) ⊆ Ker(B) ∩ Ker(D)' for clarity, since KerB and KerD are not standard notation.","section":"Definition of the set P, Section 3"},{"comment":"The assertion that (38) implies both lim_{t→∞} E[bar u'(t)bar u(t)] = 0 and ∫_0^∞ E[bar u'(t)bar u(t)]dt < ∞ is not fully justified. Equation (38) only gives the limit zero; the integrability requires an additional argument using the positivity of bar F_P(0,T) and the observability of the reduced subsystem, and this step should be spelled out.","section":"Proof of Theorem 2, after eq. (43)"},{"comment":"The text refers to Fig. 1, Fig. 2, and Fig. 3, but the figures are not included in the manuscript text provided; please ensure the figures are included with captions, and check the reference to 'x(t) (Fig. 3)' since Fig. 3 is labeled 'Optimal states' while Fig. 1 is the state-trajectory plot.","section":"Example, Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper relies on several lemmas from the authors' own prior work, in particular Lemma 3 of [Qi et al., 2017], which is cited as an arXiv preprint and is not stated in the manuscript. The referee was unable to verify the exact-observability claim from the material provided. The editor may wish to request that the authors include a full statement and proof of that lemma, or confirm whether the result is available in a peer-reviewed publication. Apart from this, the main ideas appear plausible and the gaps identified in the report seem repairable, so major revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Useful paper for the stochastic LQ control crowd. The main new thing is a clean decomposition: any solution of the generalized algebraic Riccati equation splits into a PSD matrix satisfying a singular algebraic Riccati equation plus a constant matrix from an LMI set. That lets the authors handle singular R+D'PD and state a necessary and sufficient mean-square stabilization condition with an explicit optimal gain. That is a real extension of Rami et al. (2001), and the convergence argument for the GDRE in Theorem 1 is mostly standard and coherent: monotonicity comes from cost comparison, boundedness from mean-square stabilizability.\n\nThe soft spots are in the proofs, and they are not cosmetic. In Theorem 1, the maximality claim is under-derived. The construction fixes one P_hat in P and produces a limit P_bar with P_bar >= that particular P_hat. To call P_bar maximal you need it to dominate every element of P, and nothing in the proof shows the limit is independent of the chosen P_hat or otherwise global. That is a genuine logical gap, even if I suspect it can be patched.\n\nThe bigger issue is in the sufficiency part of Theorem 2. The proof needs exact observability of the reduced triple (A22bar, C22bar, Q22bar^{1/2}), and it asserts this \"following from Lemma 3 in [Qi et al., 2017]\". But Assumption 1 only gives exact detectability of the original pair, and exact observability is strictly stronger. Lemma 2, which would transfer detectability through the feedback transformation, is stated and then skipped with \"proof similar to ...\". So the positivity of F_hatP(0,T), and hence the mean-square stability of the (A22bar, C22bar) subsystem, rests on an unstated result. This is load-bearing. The gap is likely repairable — with Z2 > 0, an unobservable nonzero initial state would make E[x2'Z2 x2] constant, which should contradict exact detectability — but as submitted the chain is incomplete.\n\nThe numerical example is illustrative only, which is fine for a theory paper. The citation pattern is reasonable; the self-citations are to the papers the lemmas actually come from, so that is not a problem.\n\nWho is this for? Specialists in stochastic LQ and Riccati theory. The decomposition and the iff stabilization criterion will be useful to them if the gaps get filled. I would send it to a serious referee, but I would not accept it as is; the referee should push for a proof of maximality and a real derivation of the exact-observability property, not a pointer to a prior lemma.","headline":"Solid indefinite stochastic LQ stabilization paper with a genuinely useful SARE/LMI decomposition, but two load-bearing proof gaps (maximality of the GARE solution and an imported exact-observability claim) need repair before I'd trust Theorem 2 as stated.","tokens_in":13026,"tokens_out":2251,"would_cite":true,"duration_ms":27774,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E20","93E15","49N10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves a necessary and sufficient mean-square stabilization condition for an Itô stochastic linear system with indefinite state and control weights: if the LMI-defined set $\\mathcal{P}$ is nonempty and an exact detectability…","keywords":["indefinite stochastic LQ control","mean-square stabilization","generalized algebraic Riccati equation","generalized differential Riccati equation","singular algebraic Riccati equation","Itô stochastic system","maximal solution","exact detectability"],"falsifier":"Search for matrices $A,B,C,D,Q,R$ satisfying $\\mathcal{P}\\neq\\emptyset$ and Assumption 1 for which the GARE has a maximal solution but the reduced subsystem is not exactly observable (equivalently, the matrix $\\bar{F}(0,T)$ in (34) is singular for some $T>0$). If such a system fails to be mean-square stabilizable by the proposed gain, the theorem is false. If such a system is stabilizable, the exact-observability assumption is not needed and should be derivable from detectability.","tokens_in":11978,"feed_emoji":"🎯","tokens_out":10345,"duration_ms":266871,"temperature":0.7,"pith_summary":"In standard linear-quadratic control, the first question is when an infinite-horizon optimal controller exists at all; this paper answers that question for Itô stochastic systems whose cost matrices $Q$ and $R$ may be indefinite. The paper establishes that mean-square stabilizability is equivalent to the solvability of a generalized algebraic Riccati equation (GARE) by a 'maximal' matrix, provided two structural conditions hold: the LMI-defined set $\\mathcal{P}$ is nonempty and a certain exact detectability condition holds. If those conditions are met and the GARE has such a solution $\\bar{P}$, the stabilizing optimal control is linear state feedback with an explicit gain, and the optimal cost is $E(x_0'\\bar{P}x_0)$. The proof works by decomposing $\\bar{P}$ into a positive semidefinite part solving a singular algebraic Riccati equation plus a constant matrix from $\\mathcal{P}$, which reduces the indefinite-cost problem to a definite one handled by Lyapunov arguments.","feed_headline":"One Riccati equation decides indefinite-cost Itô stabilization","feed_subtitle":"When the generalized algebraic Riccati equation has a maximal solution, the optimal stabilizing controller and cost follow explicitly.","key_machinery":"The central object is the generalized algebraic Riccati equation (5), whose defining feature is that $R+D'PD$ enters through the Moore–Penrose pseudo-inverse and is only required to be positive semidefinite; the companion object is the LMI-defined set $\\mathcal{P}$ and the notion of maximal solution. The argument is carried by the decomposition $\\bar{P}=\\bar{Z}_{\\hat{P}}+\\hat{P}$, which connects the GARE to a singular algebraic Riccati equation (SARE) for $\\bar{Z}_{\\hat{P}}$; via the extended Schur lemma this produces a positive semidefinite matrix $\\bar{Q}_{\\hat{P}}$ and a block-triangular structure in a transformed basis. The Lyapunov candidate $V(t,x)=E[x'\\bar{Z}_{\\hat{P}}x]$, whose decay equals the optimal-cost integrand, then proves mean-square stability of the closed-loop system. Maximality of $\\bar{P}$ is what upgrades stabilization to optimality and identifies the gain $K$ in (20).","core_discovery":"The central discovery is Theorem 2: for system (1) with indefinite $Q,R$, if $\\mathcal{P}\\neq\\emptyset$ and $(A,C,Q_{\\hat{P}}^{1/2})$ is exactly detectable, then the system is mean-square stabilizable if and only if the GARE (5) admits a solution $\\bar{P}$ that is maximal in the partial order defined by $\\mathcal{P}$. In that case the optimal stabilizing controller is $u(t)=Kx(t)$ with $K=-(R+D'\\bar{P}D)^\\dagger(B'\\bar{P}+D'\\bar{P}C)$, and the optimal cost is $J^*=E(x_0'\\bar{P}x_0)$. The proof's key step is the decomposition $\\bar{P}=\\bar{Z}_{\\hat{P}}+\\hat{P}$ with $\\hat{P}\\in\\mathcal{P}$ and $\\bar{Z}_{\\hat{P}}\\ge0$ satisfying the singular algebraic Riccati equation (13); this equivalence reduces indefinite stabilization to the definite case, where the Lyapunov function $E(x'\\bar{Z}_{\\hat{P}}x)$ built from the optimal cost functional establishes mean-square stability of the closed-loop system.","pith_inferences":["The proof's block-triangular form (30)–(31) leaves implicit a reduced-order design: the lower block evolves autonomously, so the controller only needs to shape the upper block; this follows from the proof's structure but is not stated.","A natural next test is whether the exact-observability input can be weakened to detectability; Assumption 1 is detectability, while the sufficiency proof invokes observability of a reduced subsystem, so finding a counterexample or a substitute lemma would settle whether the stated hypotheses are minimal.","Because $\\mathcal{P}\\neq\\emptyset$ and GARE solvability are semidefinite-programming-checkable, the theorem supplies a computational certificate for indefinite LQ stabilization in applications such as portfolio selection, pollution control, and robust filtering.","The decomposition $\\bar{P}=\\bar{Z}+\\hat{P}$ suggests that the set of infinite-horizon optimal costs may be parameterized by choosing $\\hat{P}\\in\\mathcal{P}$; the paper does not explore this, but it follows from the same GARE–SARE equivalence."],"forward_implications":["When the hypotheses hold, checking stabilizability reduces to verifying $\\mathcal{P}\\neq\\emptyset$ and solving the GARE; no search over feedback gains is needed.","The optimal stabilizing controller is explicit: $u=Kx$ with $K=-(R+D'\\bar{P}D)^\\dagger(B'\\bar{P}+D'\\bar{P}C)$, and every other admissible control pays the nonnegative excess $E\\int_0^\\infty (u-Kx)'(R+D'\\bar{P}D)(u-Kx)\\,dt$.","In the finite-horizon problem, the GDRE solution $P(t,T)$ converges as $t\\to-\\infty$ (equivalently $T\\to\\infty$) to the maximal GARE solution whenever the system is stabilizable.","Indefinite weights no longer block Lyapunov-based analysis: the GARE–SARE equivalence converts the indefinite problem into a definite one with nonnegative cost matrix, so the standard stabilization toolkit applies.","A maximal GARE solution $\\bar{P}\\in\\mathcal{P}$ is simultaneously a stabilizability certificate and the value matrix of the infinite-horizon optimal control problem."],"supporting_citations":[{"why":"Supplies Lemma 1: the finite-horizon GDRE condition for existence of an optimal control, which is the starting point for the infinite-horizon convergence argument.","marker":"[Rami, 2001]"},{"why":"Establishes the earlier asymptotic theory of generalized Riccati equations under strict positive definiteness, which the present paper relaxes to positive semidefiniteness.","marker":"[Rami et al., 2001]"},{"why":"Provides Proposition 1, invoked to prove that exact detectability is preserved under the feedback that defines the reduced system.","marker":"[Zhang et al., 2004]"},{"why":"Supplies Lemma 3, from which the exact observability of the reduced subsystem $\\bar{A}_{22},\\bar{C}_{22},\\bar{Q}_{22}^{1/2}$ is asserted; this is the most load-bearing external input for the sufficiency direction.","marker":"[Qi et al., 2017]"},{"why":"Gives the stability criterion used to conclude mean-square convergence of the full state from convergence of the reduced subsystem and its driving signal.","marker":"[Hinrichsen et al., 1998]"},{"why":"Extended Schur lemma used to show that the matrix $\\bar{Q}_{\\hat{P}}$ is positive semidefinite, which makes the Lyapunov decay argument go through.","marker":"[Albert, 1969]"}],"fun_headline_variants":["Maximal GARE solution stabilizes indefinite Itô systems","GARE solution gives necessary and sufficient stabilization","Indefinite stochastic control: reduced to definite via SARE","Indefinite Itô LQ stabilization: one GARE condition","Decomposition turns indefinite LQ into definite stabilization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The sufficiency proof's load-bearing input is that a certain lower-dimensional subsystem is exactly observable, a stronger property than the exact detectability assumed, and the paper asserts this from a cited lemma rather than deriving it; if that property is not guaranteed, the strict positivity step and the stabilization conclusion collapse.","fun_headline_variants_meta":{"raw":{"variants":["Maximal GARE solution stabilizes indefinite Itô systems","GARE solution gives necessary and sufficient stabilization","Indefinite stochastic control: reduced to definite via SARE","Indefinite Itô LQ stabilization: one GARE condition","Decomposition turns indefinite LQ into definite stabilization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000946,"raw_usage":{"total_tokens":4051,"prompt_tokens":969,"completion_tokens":3082,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":3003}},"tokens_in":585,"tokens_out":3082,"duration_ms":23343,"temperature":1.0,"reasoning_tokens":3003,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:00:13.707104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search for matrices $A,B,C,D,Q,R$ satisfying $\\mathcal{P}\\neq\\emptyset$ and Assumption 1 for which the GARE has a maximal solution but the reduced subsystem is not exactly observable (equivalently, the matrix $\\bar{F}(0,T)$ in (34) is singular for some $T>0$). If such a system fails to be mean-square stabilizable by the proposed gain, the theorem is false. If such a system is stabilizable, the exact-observability assumption is not needed and should be derivable from detectability.","supporting_citations":[],"review_version":1}