{"id":"d6ef4d3e-9ce8-4b78-8eae-b21d2f00344a","arxiv_id":"1908.05125","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors derive new regressor extensions that they claim relax persistence-of-excitation requirements and deliver finite-time estimation with alertness, but one proof and the simulation formula do not match the claims.","lead":"This paper proposes new constructions for the DREM parameter estimator, claiming convergence under excitation conditions weaker than persistence of excitation, a guaranteed transient performance improvement, and a finite-time estimator that stays alert to time-varying parameters. Readers outside adaptive control may care because it targets a long-standing obstacle: excitation requirements in online parameter estimation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 5's guarantee of transient improvement is false: the proof's det inequality does not control det^2, and an admissible scalar example makes the proposed d(t) converge slower, not faster.","rationale":"The reader's weakest assumption correctly identifies the sign issue in Prop 5; an explicit counterexample confirms it lands. This is the most load-bearing concern because the abstract's first headline contribution is the 'guaranteed' transient improvement, and the theorem as stated is false, not just imprecisely proved. I checked the other advertised results: Prop 3's weak-PE statement appears true (a scalar φ taking 1 every third sample gives Δ PE with K=2 while φ is not PE for K≤2 for \\bar K=2), so only the proof of (18) is loose; Prop 7's algebraic identity is correct, although the Section VI-B simulation incorrectly uses \\hat θ_i(0) instead of \\hat θ_i(t-T_D). Since the false Prop 5 can likely be repaired by choosing d(t) with the sign of det Φ0 (or by restricting to positive determinants), the appropriate outcome is revision rather than outright rejection; the reader's CONDITIONAL verdict is unchanged.","tokens_in":12099,"tokens_out":16132,"duration_ms":165631,"concrete_test":"Run the scalar CT example: φ(t)=-1, H1=1/(p+1), γ=1, common initial condition, and compare the exact errors \\tilde θ_0(t)=exp(-∫_0^t(1-e^{-s})^2 ds)\\tilde θ(0) and \\tilde θ_N(t)=exp(-(1-e^{-2t})/2)\\tilde θ(0). At t=0 both coincide; after the crossing near t≈1.6 (where t+2e^{-t}=2), the d=-1 estimator has the larger error and tends to e^{-1/2}\\tilde θ(0), while the d=0 estimator converges to zero. This directly falsifies the universal inequality |\\tilde θ_0|>|\\tilde θ_N| claimed in Proposition 5.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 5 is the basis for the abstract's advertised 'quantifiable transient performance improvement.' Its proof establishes only det Φ(t) = det Φ0(t) + |adj Φ0(t) φ(t)|^2 > det Φ0(t), and then treats a larger algebraic determinant as a larger decay rate in (25). The PEE (11) depends on Δ^2 = det^2, so the inference is valid only when det Φ0 and det Φ have the same sign and |det| increases; the paper never establishes either condition. The flaw is not merely a missing argument. For the scalar case m=1 with φ(t) = -1 and H1 = 1/(p+1), the choice (28) gives d(t) = -1. Then Φ0(t) = e^{-t} - 1 and Φ(t) = e^{-t}, so Δ_N^2 = e^{-2t} while Δ_0^2 = (1-e^{-t})^2. The exact solutions of (11) are \\tilde θ_0(t) = \\tilde θ(0) exp(-γ∫_0^t (1-e^{-s})^2 ds) → 0 and \\tilde θ_N(t) = \\tilde θ(0) exp(-γ(1-e^{-2t})/2) → \\tilde θ(0)e^{-γ/2} ≠ 0. Hence |\\tilde θ_N(t)| > |\\tilde θ_0(t)| for all sufficiently large t, contradicting the claimed universal inequality. The advertised guarantee therefore fails as stated; a sign-compensated choice of d(t) might restore it, but the paper needs revision.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies dynamic regressor extension and mixing (DREM) parameter estimators for linear regression models. Its main contributions are: (i) a unified presentation of continuous-time and discrete-time DREM; (ii) two new extended regressor constructions, one based on a sliding window in discrete time, claimed to achieve convergence under excitation strictly weaker than persistency of excitation (Prop. 3), and one adding a feedforward term d(t) to the mixing operator, claimed to guarantee a quantifiable transient performance improvement (Prop. 5); and (iii) a new finite-time convergent (FTC) estimator based on a sliding-window regressor that retains alertness to time-varying parameters (Prop. 7). The paper also shows that Kreisselmeier's regressor extension is a particular case of the proposed generalized operator. Simulations are presented for the transient-improvement and FTC claims.","tokens_in":12415,"tokens_out":17261,"duration_ms":150038,"significance":"If the weaker-than-PE convergence result and the alert FTC scheme were correct, they would be valuable for adaptive control and identification, where PE is a major bottleneck. The generalized LTV operator framework and the explicit determinant identities are useful conceptual tools, and the proofs of Props. 3 and 7 are self-contained and largely first-principles. However, the transient-improvement guarantee in Prop. 5 is false, and the simulation formula for the new FTC estimator is inconsistent with the theorem. Since the transient-improvement claim is central to the abstract's advertised contribution, the paper cannot be accepted in its present form; the remaining sound ideas are worth pursuing after a substantive revision.","major_comments":[{"comment":"The proof of Proposition 5 establishes only that det Φ(t) = det Φ0(t) + |adj{Φ0(t)}φ(t)|^2 > det Φ0(t), but the parameter error equations (11) and their explicit solutions (25) depend on Δ^2 = det^2, not on det. The step from 'larger det' to 'faster convergence' is therefore valid only if det Φ0(t) ≥ 0 (or, more generally, if |det Φ(t)| > |det Φ0(t)|), which is not established. The claim is in fact false: take m=1, φ(t) = -1, and the LTI part of H equal to 1/(p+1). Then Φ0(t) = e^{-t} - 1 < 0, and (28) gives d(t) = -1, so Φ(t) = e^{-t}. Thus det Φ0(t) < det Φ(t), yet Δ_N^2(t) = e^{-2t} < Δ_0^2(t) = (1-e^{-t})^2 for sufficiently large t. Substituting into the exact solution (25) of (11) gives \\tilde θ_N(t) = \\tilde θ(0) exp(-γ(1-e^{-2t})/2) and \\tilde θ_0(t) = \\tilde θ(0) exp(-γ∫_0^t (1-e^{-s})^2 ds); hence |\\tilde θ_N(t)| > |\\tilde θ_0(t)| for all sufficiently large t, and \\tilde θ_N(t) does not even converge to zero. This contradicts the universal inequality in Proposition 5. A sign-compensated choice of d(t), for example d(t) = -sgn(det Φ0(t)) adj{Φ0(t)}φ(t) with an appropriate gain, might restore a form of the result, but the theorem and its proof require substantial revision.","section":"Section IV-C, Proposition 5 (Eqs. (25), (28), (29))"},{"comment":"The simulation formula for the new finite-time estimator in Section VI-B reads \\hat θ^{FTC-D}_i(t) = [\\hat θ_i(t) - wD_i(t)\\hat θ_i(0)]/(1-wD_i(t)). However, Proposition 7 derives the identity [1-wD_i(t)]θ_i = \\hat θ_i(t) - wD_i(t)\\hat θ_i(t-T_D). Substituting \\hat θ_i(0) for \\hat θ_i(t-T_D) is not an identity for t > T_D, so the simulated scheme is not the estimator of Proposition 7, and its claimed finite-time convergence and alertness do not follow from the theory. The implementation should use \\hat θ_i(t-T_D), and the simulations must be rerun. In addition, the statement of Proposition 7 contains a typo: the right-hand side should have \\hat θ_i(t), not \\hat θ(t).","section":"Section VI-B, formula for \\hat θ^{FTC-D}_i(t)"}],"minor_comments":[{"comment":"In Section II-C, the word 'propostion' should be 'proposition'.","section":"Section II-C"},{"comment":"In Section V-A, first paragraph, 'mutivariable' should be 'multivariable'.","section":"Section V-A"},{"comment":"In the proof of Proposition 3, the displayed equivalence 'Φ(k) > 0, ∀k ⇔ φ(k) ∈ PE with K ≤ \\bar K' is imprecise when \\bar K > K; the condition Φ(k)>0 for all k is equivalent to φ being PE with window size \\bar K, not with an unspecified size K ≤ \\bar K. The intended counterexample is unaffected, but the claim should be restated.","section":"Section III, Proposition 3 proof"},{"comment":"The abstract and introduction advertise a 'unified treatment' of continuous- and discrete-time cases, but Propositions 3, 5, and 7 are each developed for a single time domain (DT for Prop. 3, CT for Props. 5 and 7). Section VII acknowledges that some CT results remain to be derived in DT; the wording should be tempered accordingly.","section":"Abstract and Introduction"},{"comment":"In Section VI-B, the simulation description says the new FTC estimate is 'computed as soon as wD_i(t) < μ_i', but it does not specify what the estimator outputs when wD_i(t) ≥ μ_i; this should be stated for reproducibility.","section":"Section VI-B"},{"comment":"Equation (36) is derived under the standing assumption that Δ(t-T_D)=0 for t det Φ0(t), but the PEE depends on det². For a scalar example with φ(t) = −1 and H1 = 1/(p+1), the proposed d(t) = −1 makes Φ0(t) = e^{−t} − 1 and Φ(t) = e^{−t}. The new det² is e^{−2t}, which decays to zero, while the old det² is (1−e^{−t})², which tends to 1. The error with d(t) converges to a nonzero constant; the error without it converges to zero. That contradicts the universal inequality in Prop 5, so the abstract's second bullet is wrong as stated.\n\nWhat the paper does well: the unified CT/DT presentation is clean. Prop 4 (KRE as a particular case of the generalized operator) is a nice observation with a correct direct calculation. The finite-time alert estimator in Prop 7 is a genuinely useful construction: the sliding-window identity with \\hatθ_i(t−T_D) is correct, and the reset mechanism preserves the FTC property when new excitation arrives. That contribution is worth keeping. The weak-excitation result in Prop 3 is plausible and likely correct, but the proof of (18) is a logical non-implication chain rather than an explicit counterexample; it needs a concrete φ to be rigorous.\n\nTwo more soft spots. The simulation in VI-B defines the new FTC estimator using \\hatθ_i(0) instead of \\hatθ_i(t−T_D), which is what Prop 7 requires. That is a substantive mismatch, not just a typo, because the experimental demonstration is part of the paper's evidence. And the transient-performance simulation in VI-A only shows improvement for the chosen regressor; since the theorem is false, it cannot be taken as validation of the claimed guarantee.\n\nThis paper comes from the group that built DREM, and the ideas are substantive. The finite-time alert estimator alone deserves a serious referee, and the weak-excitation result, once supplied with a proper counterexample, could be a real contribution. But the central transient-performance claim is false and needs to be either replaced with a sign-compensated version or withdrawn. I would send this to peer review with the expectation of major revision: fix Prop 5, correct the simulation formula, and tighten Prop 3's proof.","headline":"The finite-time alert estimator and the KRE unification are real contributions, but the advertised transient-performance guarantee in Prop 5 is wrong, so the paper needs major revision before it can be trusted.","tokens_in":12996,"tokens_out":6988,"would_cite":false,"duration_ms":63104,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C40","93E10","93E12"],"pacs":[],"model":"deepseek-v4-flash","headline":"DREM parameter estimators converge under excitation strictly weaker than persistence of excitation, and a new regressor choice provably speeds their transients.","keywords":["dynamic regressor extension and mixing","DREM","parameter estimation","linear regression model","persistence of excitation","finite-time convergence","adaptive control","discrete-time estimation"],"falsifier":"Run the continuous-time DREM estimator of Proposition 5 with filter parameters chosen so that $\\det\\{\\Phi_0(t)\\}<0$ for some interval while $\\varphi(t)\\neq 0$, apply $d(t)=\\operatorname{adj}\\{\\Phi_0(t)\\}\\varphi(t)$, and plot $|\\tilde\\theta_i(t)|$ against the $d=0$ case: if the parameter-error absolute value is not pointwise smaller, the claimed transient improvement fails for that sign.","tokens_in":11874,"feed_emoji":"📈","tokens_out":8711,"duration_ms":85453,"temperature":0.7,"pith_summary":"This paper strengthens dynamic regressor extension and mixing (DREM), a technique that converts one vector linear regression into m scalar regressions by multiplying through by the adjugate of an extended regressor matrix. It contributes two new designs for that matrix. The first, built from delayed regressor samples, makes the determinant's non-square-summability rather than persistence of excitation of the original regressor the condition for convergence, and this condition is genuinely weaker in discrete time. The second adds a feedforward term chosen as $\\operatorname{adj}\\{\\Phi_0\\}\\varphi$ to the filtering operators and guarantees, through Sylvester's determinant formula, that every component of the parameter error is pointwise smaller than without the feedforward. The paper also gives a sliding-window finite-time estimator that keeps its finite-time property after parameter jumps, because its weighting signal can grow when new excitation arrives.","feed_headline":"DREM estimators converge with weaker excitation and faster transients","feed_subtitle":"Determinant-based regressor extension speeds transients and keeps finite-time tracking after parameter jumps.","key_machinery":"The load-bearing object is the extended regressor matrix $\\Phi=H[\\varphi^\\top]$ built by applying $m$ scalar linear operators to the regressor, together with its determinant $\\Delta=\\det\\{\\Phi\\}$. Mixing multiplies the extended vector equation $Y=\\Phi\\theta$ by the adjugate matrix, which by $\\operatorname{adj}\\{M\\}M=\\det\\{M\\}I$ yields $m$ independent scalar regressions $Y_i=\\Delta\\theta_i$; the gradient estimators on these scalar regressions have parameter error dynamics $\\dot{\\tilde\\theta}_i=-\\gamma_i\\Delta^2\\tilde\\theta_i$ in continuous time, so performance is governed entirely by $\\Delta^2$. The paper's new results are mechanisms for shaping $\\Delta$: a delay-window operator that turns $\\Delta$ into a sum of rank-one regressor products, relaxing persistence of excitation; a feedforward gain $d(t)=\\operatorname{adj}\\{\\Phi_0(t)\\}\\varphi(t)$ that adds a positive quadratic term to $\\det\\{\\Phi\\}$, improving transients; and a sliding-window weighting $w^D(t)$ built from the last $T_D$ seconds of $\\Delta^2$, enabling reset-free finite-time convergence.","core_discovery":"The central discovery is that both the transient speed and the excitation requirements of DREM estimators are controlled by the scalar determinant $\\Delta=\\det\\{\\Phi\\}$ of the extended regressor matrix, and that this determinant can be shaped by choosing the free operator $H$. In discrete time, taking $H$ to be a window of delayed regressor samples yields $\\Phi(k)=\\sum_{j=k+1}^{k+\\bar K}\\varphi(j-(1+\\bar K))\\varphi^\\top(j-(1+\\bar K))$, so $\\varphi(k)\\in PE$ implies $\\Delta(k)\\notin\\ell^2$ and $\\Delta(k)\\in PE$, while a decaying scalar example shows $\\Delta(k)\\notin\\ell^2$ without $\\varphi(k)\\in PE$; thus DREM converges under strictly weaker excitation than gradient or least-squares estimators. In continuous time, choosing the feedforward gain $d(t)=\\operatorname{adj}\\{\\Phi_0(t)\\}\\varphi(t)$ in the general LTV operator makes $\\det\\{\\Phi(t)\\}=\\det\\{\\Phi_0(t)\\}+|\\operatorname{adj}\\{\\Phi_0(t)\\}\\varphi(t)|^2$, so the squared determinant and hence the exponential decay rate in the scalar parameter error equations is pointwise larger than for $d=0$. Finally, replacing the memoryless weighting $w(t)$ by $w^D(t)=\\exp(-\\gamma\\int_{t-T_D}^{t}\\Delta^2(s)\\,ds)$ yields the identity $[1-w^D]\\theta=\\hat\\theta(t)-w^D\\hat\\theta(t-T_D)$, from which a finite-time estimate can be formed that revives when excitation reappears.","pith_inferences":["Editorial inference: if the determinant sign condition in Proposition 5 can be enforced or replaced by $|\\det|$, the transient-improvement guarantee would extend to arbitrary sign; a natural test is to rerun the simulation with $\\Phi_0$ chosen with negative determinant and compare error curves.","Editorial inference: the delay-window construction of Proposition 3 suggests a continuous-time analogue where convergence is guaranteed by an integral-over-window condition on $\\Delta$ rather than by PE; the paper leaves this extension to future work.","Editorial inference: the same sliding-window $w^D$ mechanism could be applied to other DREM-based adaptive controllers and observers wherever parameter jumps or intermittent excitation are expected, since it removes the need for resets.","Editorial inference: because the determinant $\\Delta$ is scalar, it could serve as a real-time excitation monitor or as an excitation-injection target in composite adaptive control; this is not discussed in the paper."],"forward_implications":["In discrete time, DREM converges whenever $\\Delta(k)$ is not square-summable, a condition strictly weaker than $\\varphi(k)$ being persistently exciting; for example $\\varphi(k)=(k+1)^{-1/4}$ is not PE but yields $\\Delta(k)=(k+1)^{-1/2}\\notin\\ell^2$.","If $\\Delta$ is persistently exciting, convergence is exponential, and this is also weaker than $\\varphi$ being PE in the qualified window sense of Proposition 3.","The feedforward choice $d(t)=\\operatorname{adj}\\{\\Phi_0(t)\\}\\varphi(t)$ makes each parameter error component strictly smaller at every time than the same DREM estimator without feedforward.","The new finite-time estimator uses $w^D(t)=\\exp(-\\gamma\\int_{t-T_D}^{t}\\Delta^2(s)\\,ds)$; when $\\Delta$ grows over a window of length $T_D$, $w^D$ grows, so the finite-time property is regained without resetting the estimator, allowing tracking of time-varying parameters.","The unified continuous- and discrete-time treatment makes the determinant $\\Delta$ the single quantity to monitor for both classes of estimators."],"supporting_citations":[{"why":"Supplies the original continuous-time DREM construction, including the adjugate mixing that produces scalar regressions and the gradient estimator on them.","marker":"[2]"},{"why":"Supplies the discrete-time DREM estimator and its parameter error equation, which Proposition 3 extends with the delay-window operator.","marker":"[5]"},{"why":"Introduces the extended-regressor construction obtained by filtering both sides of the regression by a single stable operator, shown in the paper to be a particular case of the general operator H.","marker":"[11]"},{"why":"Provides the adjugate identity and Sylvester's determinant formula used in the mixing step and in the proof of the transient-improvement result.","marker":"[12]"},{"why":"Establishes the earlier finite-time DREM estimator under interval excitation, which the paper's Proposition 7 modifies to preserve alertness.","marker":"[8]"},{"why":"Gives the Luenberger-observer interpretation and discusses LTV operators, providing background for the general LTV operator H in Section IV.","marker":"[18]"},{"why":"Supplies the standard gradient estimator and the persistence-of-excitation condition that the paper's new results relax.","marker":"[21]"},{"why":"Shows the relation between regressor PE and determinant PE for LTI system identification, motivating the weaker determinant conditions.","marker":"[4]"}],"fun_headline_variants":["DREM estimators: weaker excitation, faster transients","Exponential DREM without persistent excitation","Finite-time DREM tracking after parameter jumps","Shaping DREM determinant boosts speed, relaxes PE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that increasing the determinant $\\det\\{\\Phi_0(t)\\}$ to $\\det\\{\\Phi_0(t)\\}+|\\operatorname{adj}\\{\\Phi_0(t)\\}\\varphi(t)|^2$ always increases the decay rate $|\\Delta|^2$ in the error equation, which is only guaranteed when the two determinants have the same sign; the paper does not establish that sign condition.","fun_headline_variants_meta":{"raw":{"variants":["DREM estimators: weaker excitation, faster transients","Exponential DREM without persistent excitation","Finite-time DREM tracking after parameter jumps","Shaping DREM determinant boosts speed, relaxes PE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000685,"raw_usage":{"total_tokens":3134,"prompt_tokens":999,"completion_tokens":2135,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":2075}},"tokens_in":615,"tokens_out":2135,"duration_ms":16411,"temperature":1.0,"reasoning_tokens":2075,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:22:53.851076+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the continuous-time DREM estimator of Proposition 5 with filter parameters chosen so that $\\det\\{\\Phi_0(t)\\}<0$ for some interval while $\\varphi(t)\\neq 0$, apply $d(t)=\\operatorname{adj}\\{\\Phi_0(t)\\}\\varphi(t)$, and plot $|\\tilde\\theta_i(t)|$ against the $d=0$ case: if the parameter-error absolute value is not pointwise smaller, the claimed transient improvement fails for that sign.","supporting_citations":[{"cited_title":"Aranovskiy, A","cited_arxiv_id":null,"evidence_quote":"Supplies the original continuous-time DREM construction, including the adjugate mixing that produces scalar regressions and the gradient estimator on them."},{"cited_title":"Belov, R","cited_arxiv_id":null,"evidence_quote":"Supplies the discrete-time DREM estimator and its parameter error equation, which Proposition 3 extends with the delay-window operator."},{"cited_title":"Kreisselmeier, Adaptive observers with exponentia l rate of conver- gence, IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Introduces the extended-regressor construction obtained by filtering both sides of the regression by a single stable operator, shown in the paper to be a particular case of the general operator H."},{"cited_title":"Lancaster and M","cited_arxiv_id":null,"evidence_quote":"Provides the adjugate identity and Sylvester's determinant formula used in the mixing step and in the proof of the transient-improvement result."},{"cited_title":"Gerasimov, R","cited_arxiv_id":null,"evidence_quote":"Establishes the earlier finite-time DREM estimator under interval excitation, which the paper's Proposition 7 modifies to preserve alertness."},{"cited_title":"Ortega, L Praly, S","cited_arxiv_id":null,"evidence_quote":"Gives the Luenberger-observer interpretation and discusses LTV operators, providing background for the general LTV operator H in Section IV."},{"cited_title":"Belov, S","cited_arxiv_id":null,"evidence_quote":"Shows the relation between regressor PE and determinant PE for LTI system identification, motivating the weaker determinant conditions."}],"review_version":1}