{"id":"e97ca5d0-8e45-469e-bab6-9f0d5ca3e0ee","arxiv_id":"1908.04001","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"It provides LMI conditions for stabilizing Markov jump linear systems with exponentially distributed mode-observation delay and time-varying state delay, with mixed H2/H∞ performance guarantees.","lead":"This paper designs feedback controllers for systems that switch randomly between modes, when the controller only learns the current mode after a random delay. It reshapes the delayed system into a standard form and gives matrix inequalities intended to guarantee stability and mixed H2/H∞ performance.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LMI (3) in Theorem 1 is infeasible as written: both first and second LMIs have an I_n block in the (4,4) diagonal position, impossible for a negative definite matrix; the numerical example cannot be valid.","rationale":"The central claim of the paper is that the LMI system (3) provides a sufficient condition for designing a mixed H2/H-infinity controller. For this claim to hold, the LMI must be a valid sufficient condition and must be satisfiable in some nontrivial cases. The displayed LMIs in (3) contain an obvious obstruction: the (4,4) block is I_n. Negative definiteness of a symmetric matrix implies negative definiteness of every principal submatrix; since I_n is positive definite, the LMI is infeasible. This is not a subtle proof gap but a direct algebraic contradiction. The numerical example's claim to satisfy (3) therefore cannot be correct as stated, so the paper's only demonstration of the design method is invalid. Even if one treats I_n as an OCR error for -I_n, the proof's derivation from (17) remains unjustified: the transformation changes both the (4,4) block from -Qhat to ±I_n and the (4,1) block from Y^{-T}Z^T B^T to Z^T B^T, yet the proof's 'indefinite block' addition does not correspond to any valid matrix operation. Thus the central theorem is not established. The reader's verdict of REJECT with high correctness risk is appropriate. I do not adopt the reader's weakest assumption (exponential delay) as the primary concern because even granting Assumption 1, the LMI infeasibility defeats the result. The exponential assumption is a modeling limitation, not an internal inconsistency. Additionally, Proposition 1's transition rate formula appears erroneous for observation events, which would further invalidate the generator used in the LMIs; this reinforces the rejection. In summary, the LMI infeasibility is the single most load-bearing concern, and its resolution would require either a corrected LMI with a valid derivation or new numerical evidence.","tokens_in":12341,"tokens_out":18658,"duration_ms":164321,"concrete_test":"Analytically: in the first LMI of (3), set the free vector to e_4 ⊗ v (v ≠ 0). The quadratic form equals v^T I_n v = ||v||^2 > 0, which violates negative definiteness; hence the LMI is infeasible. To check the numerical example's validity, run an SDP solver (e.g., MATLAB/YALMIP with Sedumi) on (3) using the data of Section 5; the solver will return infeasible. If the authors respond that I_n is a typo for -I_n, re-derive the Schur complement from (17) with the corrected block and verify whether the resulting LMI is equivalent; the proof's added block must be replaced by a valid congruence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is that the LMI system (3) in Theorem 1, as displayed, cannot be satisfied. In the first LMI (the 5x5 block) and the second LMI (the 6x6 block), the (4,4) diagonal block is I_n, the n-by-n identity matrix. A symmetric matrix M satisfying M<0 must have every principal submatrix negative definite, so in particular the (4,4) block must be negative definite. Since I_n>0, for any nonzero vector v the quadratic form with v in the 4th block and zeros elsewhere equals v^T I_n v = ||v||^2 > 0, contradicting M<0. Hence the LMI is infeasible for all choices of Y_j, Z_j, epsilon, Lambda. This makes Theorem 1 vacuous as a design procedure and directly contradicts the numerical example in Section 5, which claims to find Y1,Y2,Z1,Z2,epsilon,Lambda satisfying (3). The proof of Theorem 1 attempts to justify replacing -Qhat_{kij} (the original (4,4) block from the Schur complement of (17)) with I_n by adding an 'indefinite block' with -Y^T Qhat Y - I_n; but this operation is not a congruence or Schur complement, and the sum of matrices does not have the displayed form. If the intended block was -I_n, the stated LMI still does not follow from (17) because the (4,1) block changes from Y^{-T}Z^T B^T to Z^T B^T without a valid transformation. Thus the central sufficient condition is either infeasible or unproven. Additionally, Proposition 1's transition rate formula q_{(i1,j1),(i2,j2)} = 1(j1=j2)lambda_{i1 i2} + 1(i1=i2=j2)g_{j1 j2} appears incorrect for observation transitions: from a delayed state (i1 != j1), an observation should move to (j1,j1) with rate g_{i1 j1}, but the indicator 1(i1=i2=j2) cannot hold. This would corrupt the generator tilde S used in the LMIs. Both issues independently invalidate the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies state-feedback control for continuous-time Markov jump linear systems subject to an unknown time-varying state delay and a random delay in the observation of the mode. Under Assumption 1, which makes the mode-observation delay exponentially distributed, the pair (r(t), r̃(t)) is claimed to be a time-homogeneous Markov process, reducing the closed-loop system to a standard delayed Markov jump linear system. The main result, Theorem 1, presents LMI conditions for computing state-feedback gains that guarantee weak delay-dependent stochastic stability and prescribed mixed H2/H∞ performance levels; a numerical example is given to illustrate the design.","tokens_in":12765,"tokens_out":12005,"duration_ms":119001,"significance":"The remodeling idea in Section 3 is genuinely interesting: the exponential assumption is the natural memoryless condition that lets the mode-observation delay be absorbed into an enlarged Markov chain, and Proposition 1 is credible. If Theorem 1 were correct, the LMI design would be a useful tool for asynchronous switching with delayed mode information. However, the paper provides no code and only a simulation, and the central LMI system is infeasible as written; the proof contains a gap in the claimed equivalence. The paper's contribution is therefore not currently supported.","major_comments":[{"comment":"Both matrix LMIs in (3) have I_n in the (4,4) diagonal block. Since a negative-definite symmetric matrix must have all principal submatrices negative definite, and I_n is positive definite, no choice of Y_j, Z_j, ε, and Λ can satisfy either LMI. The numerical example in Section 5 therefore cannot be valid. If the intended block was -I_n or -hat Q_{kij}, the displayed statement still does not follow from the proof, because the proof replaces the (4,4) entry -hat Q_{kij} with I_n by adding a block containing -Y_j^T hat Q_{kij} Y_j - I_n; this is not a congruence or Schur-complement operation and does not preserve equivalence. In addition, the (4,1) block changes from Y_j^{-T} Z_j^T B_i^T in the intermediate LMI to Z_j^T B_i^T in (3) without a stated transformation.","section":"Theorem 1, Eq. (3)"},{"comment":"The statement 'Since ||x(t)||^2 ≥ ||x(t+ϑ)||^2 for some ϑ ∈ R_+ and all -τ ≤ ϑ ≤ 0' is false for a general trajectory; for example, a trajectory with increasing norm on the interval can have ||x(t+ϑ)|| > ||x(t)||. The bound V(x(t),t,k) ≤ x^T(t)P_k x(t) + σ||x(t)||^2 with σ = τ λ_max(Q_k), and the subsequent derivation of exponential decay in (10), rely on this false inequality. Thus the proof of weak delay-dependent stochastic stability and the H2 bound in Proposition 2 is not established.","section":"Proposition 2, inequality (9)"},{"comment":"The step 'from which we obtain' replaces a Schur-complement LMI whose (4,4) block is -hat Q_{kij} with the LMI displayed in (3), whose (4,4) block is I_n, by adding a block containing -Y_j^T hat Q_{kij} Y_j - I_n. This operation is not a congruence transformation or a Schur complement, and the displayed 3x3 block with a zero entry is not negative definite as claimed. Moreover, the (4,1) block changes from Y_j^{-T} Z_j^T B_i^T to Z_j^T B_i^T with no stated transformation. Therefore the sufficiency argument connecting Propositions 2 and 3 to the LMIs in (3) is not established.","section":"Section 4.1, proof of Theorem 1"}],"minor_comments":[{"comment":"The word 'Control' appears duplicated in the arXiv title; the running header uses 'Mixed H2/H∞ Control of Delayed Markov Jump Linear Systems'.","section":"Title"},{"comment":"The sentence 'Therefore, if the second LMIs of (3) are satisfied' should refer to the first matrix LMI, and similarly 'the third LMIs' should refer to the second matrix LMI; the numbering is confusing.","section":"Section 4, proof of Theorem 1"},{"comment":"The third displayed inequality in (3) is not a valid block matrix: the lower row contains X - 1/λ_max(L_k^{-1}) with no (2,2) entry. It should be written as a scalar inequality -Λ + X^2 λ_max(L_k^{-1}) ≤ 0 or as a proper 2x2 Schur complement.","section":"Theorem 1, Eq. (3)"},{"comment":"The numerical example sets X = 2 but does not specify the full initial function φ on [-τ,0]; since the stability and performance bounds depend on φ(0) and on the integral defining X, the simulation is not fully reproducible.","section":"Section 5"},{"comment":"The paper assumes δ_+ ∈ (0,1], but if δ_+ = 1 then hat Q_{kij} = (1-δ_+)Q_{kij} = 0, and the (2,2) block in the LMI (4) is zero, so the LMI cannot be negative definite. The boundary case δ_+ = 1 should be excluded or treated separately.","section":"Assumption 1 and Proposition 2"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an arXiv preprint. The central design theorem is not usable as stated because the LMI system (3) is infeasible, and the proof connecting Propositions 2 and 3 to (3) contains a non-equivalent matrix manipulation. These are load-bearing errors, not mere presentation issues. If the authors can provide a corrected LMI formulation and a complete proof, a fresh submission may be warranted, but I do not recommend inviting revision of the present version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core modeling idea in this paper is genuinely new and worth knowing. The authors embed the mode process r and its delayed observation \\tilde r into a joint process s=(r,\\tilde r) and show that, under Assumption 1 (exponential mode-observation delays), s is a time-homogeneous Markov chain. That reduction lets them write the closed-loop system as a standard delayed MJLS and then apply off-the-shelf LMI machinery. As far as I can tell, this exponential-delay trick is not in the prior work they cite, and it is a clean, useful observation.\n\nThe problem is the main theorem. The displayed LMI system (3) in Theorem 1 has an I_n block in the (4,4) diagonal position of both the second and third LMIs. A negative definite matrix cannot have a positive definite principal sub-block, so these LMIs are infeasible for any finite dimensions. The numerical example that claims to satisfy (3) cannot be right. The proof attempts to justify the I_n by adding an indefinite block with -Y^T \\hat Q Y - I_n, but that does not turn the (4,4) entry into I_n; the block matrix algebra does not work. This is not a minor typo; the sufficient condition is central to the design claim.\n\nThere is also a real gap in Proposition 2's stability argument. The line \"Since ||x(t)||^2 >= ||x(t+theta)||^2...\" is false in general for a continuous signal over an interval, and the subsequent bound on V(x(t),t,k) relies on it. That needs fixing too.\n\nTo be fair, the transition rate formula in Proposition 1 that the stress-test questioned is actually plausible. From a delayed state (i1,j1), the observation should move to (i1,i1) at rate g_{j1,i1}, and the indicator 1(i1=i2=j2) captures exactly that when (i2,j2)=(i1,i1). So that particular criticism does not hold up.\n\nOverall: the modeling contribution is real, but the main result is currently either infeasible or unproven, and the numerical example looks fabricated relative to the stated LMI. This is not publishable as it stands. It deserves a serious referee only because the exponential-delay embedding could be repaired and the LMI conditions rewritten; the idea is worth a resubmission. I would not cite the theorem, but I might cite the modeling trick once it is properly established.\n\nRecommendation: send to peer review if you want an expert to confirm the infeasibility and help the authors see the fix; otherwise a desk reject with encouragement to revise is defensible. For now, treat the LMI results as unverified.","headline":"The exponential-mode-delay remodeling is a genuinely useful idea, but the main LMI theorem is infeasible as written and the proof doesn't fix it; the paper needs major correction.","tokens_in":13389,"tokens_out":5555,"would_cite":false,"duration_ms":51754,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E15","93C30","93C23","93D21"],"pacs":[],"model":"deepseek-v4-flash","headline":"A Markov jump linear system with exponentially distributed mode-observation delay can be remodeled as a standard delayed Markov jump linear system, and a set of LMIs then yields feedback gains meeting prescribed H2 and H∞ bounds.","keywords":["Markov jump linear systems","mode-observation delay","state delay","mixed H2/H∞ control","linear matrix inequalities","stochastic stability","exponential distribution"],"falsifier":"Re-run the numerical example with the same system matrices and the same gains $K_1,K_2$ but replace the exponential observation delays by a non-exponential distribution, for instance uniform on $[0,1]$, then estimate $\\mathrm{H}_2$ and $\\sup_w\\mathrm{H}_\\infty$ by Monte Carlo simulation; if either exceeds $f_2=15$ or $f_\\infty=17$, the paper's reduction has broken. A cleaner check is to test the Markov property of $s(t)$ directly from simulated sample paths of $(r,\\tilde r)$.","tokens_in":12101,"feed_emoji":"⚙️","tokens_out":7032,"duration_ms":69924,"temperature":0.7,"pith_summary":"This paper establishes a way to design mixed $\\mathrm{H}_2/\\mathrm{H}_\\infty$ state-feedback controllers for Markov jump linear systems in which the controller observes the active mode only after a random delay. The key move is to assume that each mode-observation delay is exponentially distributed and to combine the true mode and the observed mode into one joint process. Because the exponential distribution is memoryless, that joint process is itself a Markov process, and the closed loop becomes a standard delayed Markov jump linear system. The paper then gives a system of linear matrix inequalities whose solutions produce feedback gains with guaranteed stochastic stability and prescribed performance bounds. A numerical example demonstrates the construction for a two-mode system.","feed_headline":"Exponential observation delays make Markov jump control standard","feed_subtitle":"A two-mode example shows LMI-based feedback gains meeting H2 and H∞ bounds despite delayed mode information.","key_machinery":"The load-bearing object is the joint Markov process $s(t)=(r(t),\\tilde r(t))$, which has $N^2$ states and generator $\\tilde S=[\\tilde q_{kk'}]$. Its transition rates express two competing mechanisms: the true mode $r$ jumps at the rates $\\lambda_{i_1i_2}$ while no observation is completed, and the observed mode $\\tilde r$ jumps to the true mode at rate $g_{j_1j_2}$ when an observation is completed. The exponential assumption makes these the only memory the process needs. With $s$ in place, the matrices in the closed loop are rewritten as $\\hat A_s$ and $\\hat B_s \\check K_s$, so $\\Sigma_K$ becomes $\\bar\\Sigma_K$, a standard delayed MJLS. The proof then uses the Lyapunov function $V(x,t,k)=x^\\top P_k x+\\int_{t-\\tau}^t x^\\top(v)Q_k x(v)\\,dv$ and its weak infinitesimal operator to convert the $\\mathrm{H}_2$ and $\\mathrm{H}_\\infty$ inequalities into the LMIs of (3).","core_discovery":"The paper's central claim is Theorem 1: for the closed-loop system $\\Sigma_K$ with control $u(t)=K_{\\tilde r(t)}x(t-\\tau(t))$, if there exist symmetric matrices $Y_j>0$, scalars $\\tau>0$, $\\Lambda>0$, and matrices $Z_j$ satisfying the LMI system (3), then $K_j=Z_jY_j^{-1}$ is a mixed $\\mathrm{H}_2/\\mathrm{H}_\\infty$ controller with $\\mathrm{H}_2\\le f_2$ and $\\sup_w\\mathrm{H}_\\infty\\le f_\\infty$. The supporting structural result is Proposition 1: $s(t)=(r(t),\\tilde r(t))$ is a time-homogeneous Markov process on $\\Theta\\times\\Theta$ with transition rates $q_{(i_1,j_1),(i_2,j_2)}=\\mathbf{1}(j_1=j_2)\\lambda_{i_1 i_2}+\\mathbf{1}(i_1=i_2=j_2)g_{j_1 j_2}$. This reduction, together with the Lyapunov-function arguments in Propositions 2 and 3, is what lets the nonstandard random-delay problem be treated by the standard delayed-MJLS machinery.","pith_inferences":["The exponential assumption is not merely technical; if the observation delay is uniform or deterministic, the pair $(r,\\tilde r)$ is not Markov and the stated LMIs have no formal justification. Replacing the exponential by a phase-type distribution would preserve Markov structure with extra phases and is a natural test of how much the result depends on the assumption.","Because the controller uses the full state $x$ and the observed mode, the same reduction would open the door to output-feedback and observer-based designs, for which the measured output $y$ is already part of the problem statement.","The bound $\\tau+\\Lambda\\le\\min\\{f_2,f_\\infty\\}$ lumps the initial-state and initial-delay energies; optimizing over the Lyapunov matrices instead of fixing $Q_k$ could produce less conservative bounds than the example's $\\tau=7.14$, $\\Lambda=4$.","A direct falsification of the paper's scope would be to simulate the same two-mode example with non-exponential observation delays and check whether the claimed $\\mathrm{H}_2$ and $\\mathrm{H}_\\infty$ bounds still hold; if they fail, the exponential assumption is doing the load-bearing work."],"forward_implications":["If (3) is feasible, the state-feedback gains $K_j=Z_jY_j^{-1}$ render the closed-loop system weakly delay-dependent stochastically stable and make both performance measures satisfy the prescribed bounds.","The remodeling applies to any design method for standard delayed Markov jump linear systems, so stabilization, guaranteed-cost, and other performance objectives can inherit the same reduction.","Existing delay-dependent stability tools, such as the Lyapunov-function argument in Proposition 2, become applicable to systems with random mode-observation delay.","The numerical example shows that for a two-mode system with observation delay rate $g=3$, the LMI conditions return gains $K_1=[-0.7423\\,\\,-0.4074]$ and $K_2=[-0.4397\\,\\,-0.2309]$ that stabilize the system."],"supporting_citations":[{"why":"supplies the continuous-time Markov jump linear system model and the output-feedback baseline the paper builds on.","marker":"[1]"},{"why":"motivates using delayed information of the mode rather than its current value.","marker":"[10]"},{"why":"models random observation of the operation mode, the phenomenon this paper carries into continuous time.","marker":"[11]"},{"why":"provides the stochastic stability framework and weak infinitesimal operator used in Proposition 2.","marker":"[27]"},{"why":"supplies the Lyapunov functional for time-delay systems that the proof differentiates.","marker":"[28]"},{"why":"is the source of the proof development that turns the Lyapunov inequalities into the H2 and H∞ conditions.","marker":"[29]"}],"fun_headline_variants":["Exponential mode delay folded into standard MJLS","Delayed mode observation solved by augmented Markov jump","LMI design for H2/H∞ with delayed Markov mode info","Augmented Markov chain handles random mode delay","Mixed H2/H∞ control despite exponential delay in mode"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Assumption 1, that every mode-observation delay follows an exponential distribution with positive rate, is the load-bearing premise; without the memoryless property, the joint process $(r,\\tilde r)$ would remember how long the current observation has been pending, and the closed loop would not reduce to a standard delayed Markov jump linear system.","fun_headline_variants_meta":{"raw":{"variants":["Exponential mode delay folded into standard MJLS","Delayed mode observation solved by augmented Markov jump","LMI design for H2/H∞ with delayed Markov mode info","Augmented Markov chain handles random mode delay","Mixed H2/H∞ control despite exponential delay in mode"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000688,"raw_usage":{"total_tokens":3110,"prompt_tokens":926,"completion_tokens":2184,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":2108}},"tokens_in":542,"tokens_out":2184,"duration_ms":21501,"temperature":1.0,"reasoning_tokens":2108,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:56:50.180445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the numerical example with the same system matrices and the same gains $K_1,K_2$ but replace the exponential observation delays by a non-exponential distribution, for instance uniform on $[0,1]$, then estimate $\\mathrm{H}_2$ and $\\sup_w\\mathrm{H}_\\infty$ by Monte Carlo simulation; if either exceeds $f_2=15$ or $f_\\infty=17$, the paper's reduction has broken. A cleaner check is to test the Markov property of $s(t)$ directly from simulated sample paths of $(r,\\tilde r)$.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the continuous-time Markov jump linear system model and the output-feedback baseline the paper builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"motivates using delayed information of the mode rather than its current value."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"models random observation of the operation mode, the phenomenon this paper carries into continuous time."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the stochastic stability framework and weak infinitesimal operator used in Proposition 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the Lyapunov functional for time-delay systems that the proof differentiates."},{"cited_title":"536–552 24","cited_arxiv_id":null,"evidence_quote":"is the source of the proof development that turns the Lyapunov inequalities into the H2 and H∞ conditions."}],"review_version":1}