{"id":"d6b9afde-2b6d-41eb-a625-6cc6af422c31","arxiv_id":"2411.14876","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Stochastic exponentials of matrix-valued Lévy processes satisfy SLLN and CLT for log-norm, log-entries, and log-determinant under weak geometric and moment conditions.","lead":"The paper proves laws of large numbers and central limit theorems for the long-time growth of the norm, entries, and determinant of stochastic exponentials of matrix-valued Lévy processes. It connects continuous-time matrix processes with the classical theory of products of random matrices, giving distributional limits under weak assumptions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.20 is false as stated: the conclusion σ²>0 fails for the multiplicative Lévy process X_t = e^{tA} with A = [[α,-β],[β,α]], α>0, β/2π irrational, even though G_X is unbounded and strongly irreducible.","rationale":"The reader's verdict ACCEPT with low correctness risk is untenable because the central claim, Theorem 3.20, is false as stated. The counterexample is a deterministic multiplicative Lévy process whose norm grows at a deterministic exponential rate, so the CLT is degenerate with σ²=0. All hypotheses of the theorem are satisfied: G_X is unbounded (norms grow like e^{αt}) and strongly irreducible (the orbit of any line under the irrational rotations is the full projective line), and the square moment of log M(X_t) is finite. The theorem's conclusion σ²>0 fails. This is not a matter of an outside consensus or a looser condition; it is an internal inconsistency with an explicit example. The source of the error is that unboundedness and strong irreducibility do not imply proximality, and without proximality the norm process can be a.s. deterministic even when the group is non-compact and irreducible. The reader's weakest_assumption noted that proximality is needed for entries (Theorem 3.22) but did not see that it is also needed for the positivity of σ² in the norm CLT, and therefore incorrectly rated the correctness risk as low. Because the central theorem is falsified, the paper cannot be accepted in its current form; the theorem would need to be revised, e.g., by adding proximality to the hypotheses of Theorem 3.20 or by weakening the conclusion to σ²≥0 (and checking which auxiliary results then survive).","tokens_in":28125,"tokens_out":39253,"duration_ms":369017,"concrete_test":"Run the deterministic model L_t = tA with A = [[α,-β],[β,α]], α>0, β/2π irrational. Numerically or symbolically compute X_t = e^{tA}, verify ||X_t||=e^{αt}, so (log||X_t|| - tα)/√t = 0 for all t; hence no σ²>0 exists. This check confirms the theorem's conclusion fails while all hypotheses of Theorem 3.20 hold.","verdict_should_be":"REJECT","load_bearing_attack":"Take L_t = tA with A = [[α, -β],[β, α]] for α>0, β/2π irrational. Then X_t = e^{tA} = e^{αt} R_{βt}. This is a valid stochastic exponential: dX_t = X_t A dt, X_0=I, and det(I+ΔL)=1, so (1.2) holds. The group G_X = {e^{tA}: t∈R} is unbounded (||X_t||=e^{αt}→∞) and acts strongly irreducibly: for any line L, its orbit under {R_{βt}} is all lines because β/2π irrational, so no finite union of lines is invariant. The moment condition E[sup_{s≤1}(log M(X_s))^2] holds since log M(X_s)=αs≤α. Yet log ||X_t|| = αt a.s., so (log||X_t|| - tα)/√t ≡ 0; the CLT holds only with σ²=0, contradicting the claimed σ²>0. The flaw is that unboundedness and strong irreducibility do not imply proximality; here G_X consists of similarities, so the norm is deterministic. The cited discrete-time CLT [4, Thm 1.1] gives σ²≥0 under these hypotheses; proximality is needed for σ²>0. This directly invalidates Theorem 3.20, the paper's central norm CLT.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the long-time behaviour of matrix-valued stochastic exponentials, i.e. multiplicative Lévy processes in GL(R,d), driven by a matrix-valued Lévy process L through dX_t = X_{t-} dL_t. The main results are a strong law of large numbers and a central limit theorem for the logarithmised norm and for logarithms of entries of X_t, Berry-Esseen bounds for the one-point motion, and explicit limit theorems for the logarithmised determinant. The determinant is identified as a one-dimensional stochastic exponential whose characteristics are computed from those of L. The proofs transfer known discrete-time results for products of random matrices to continuous time via skeletons, and use the stochastic-logarithm/determinant Itô calculus for the determinant part.","tokens_in":28446,"tokens_out":8202,"duration_ms":81028,"significance":"If repaired, the paper is a substantive contribution: it connects the theory of stochastic exponentials of matrix-valued Lévy processes with the modern discrete-time product-of-random-matrices toolbox, gives explicit sufficient conditions on the driving Lévy characteristics for the geometric assumptions, and provides a clean explicit representation of the determinant process. The determinant part, in particular Theorem 4.1 and Corollary 4.2, is a useful and carefully derived structural result. The transfer technique based on Lemma 3.14 is sound, and the moment conditions are stated in terms of the driving process. However, the central norm CLT as stated is false, so the paper cannot be accepted in its present form.","major_comments":[{"comment":"Theorem 3.20 is false as stated because the positivity assertion σ²>0 is not a consequence of the hypotheses. Take L_t = tA with A = [[α, -β], [β, α]], α>0 and β/2π irrational. Then X_t = e^{tA} = e^{αt} R_{βt}, condition (1.2) holds, G_X is unbounded and strongly irreducible, and E[sup_{0≤s≤1}(log M(X_s))²] < ∞, yet log‖X_t‖ = αt, so the CLT (3.10) holds only with σ²=0. The cited discrete-time result [4, Thm. 1.1] gives σ²≥0 under these hypotheses; positivity requires an additional proximality (or non-similarity) assumption. The theorem should be corrected either by adding proximality to the hypotheses or by weakening the conclusion to σ²≥0 and explicitly discussing the degenerate similarity case, with the proof adjusted accordingly.","section":"§3.4, proof of Theorem 3.20 and §1 Introduction"},{"comment":"The informal statement in the Introduction that the assumptions 'basically only exclude cases where X is contained in the set of similarity matrices' is too strong: the formal hypotheses of Theorem 3.20 (unboundedness plus strong irreducibility) do not exclude the unbounded similarity group appearing in the counterexample above. If the theorem is repaired by adding proximality, the surrounding remarks — in particular Remark 3.21 and the way Theorem 3.20 is invoked in Theorem 3.22 — should be revisited so that the hypotheses under which σ²>0 is asserted are consistent throughout.","section":"§3.4, proof of Theorem 3.20 and §1 Introduction"}],"minor_comments":[{"comment":"The statement writes E[sup_{0≤s≤t} M(X_t)^ε]; the argument of M should be X_s, not X_t.","section":"§3.2, Lemma 3.11"},{"comment":"The theorem states 'for all n ≥ 1' but the bound is in t; it should read 'for all t ≥ 1'. In the proof, the sentence 'Applying (3.13), that is [33, Thm. 2.1]' refers to the discrete-time result that is being transferred, not to the continuous-time statement being proved; please rephrase to avoid the appearance of circularity.","section":"§3.5, Theorem 3.23"},{"comment":"In the definitions of T1(x) and T2(x), the symbol x is used both as the threshold and as the running variable in the set; renaming the threshold, for example r, would remove the ambiguity.","section":"§4.2, Theorem 4.7"}],"recommendation":"major_revision","confidential_remarks":"The counterexample to Theorem 3.20 is decisive, but the issue is localized to the positivity claim and the missing proximality condition. The remaining results — especially the determinant representation and asymptotics, and the Berry-Esseen bounds under condition (i-p) — appear sound, and the theorem can be repaired within the scope of the manuscript. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a referee, but not as is. The norm CLT, Theorem 3.20, is false as stated. Let L_t = tA with A = [[α,-β],[β,α]], α>0, β/2π irrational. Then X_t = e^{tA} = e^{αt}R_{βt} is the stochastic exponential, (1.2) holds, G_X is unbounded and strongly irreducible (dense rotations), and the second moment condition is satisfied. But log||X_t|| = αt, so the CLT holds only with σ²=0, contradicting the asserted σ²>0. The proof invokes Benoist–Quint [4, Thm 1.1], which requires proximality; 'unbounded plus strongly irreducible' does not imply proximality, and this group consists of similarities.\n\nThe paper still has solid parts. The skeleton reduction in Lemma 3.14 is sound under the stated moments. The determinant analysis is the highlight: Theorem 4.1 and Corollary 4.2 give a clean Itô-formula representation of det X_t as a one-dimensional stochastic exponential, and the CLT via Doney–Maller fits. The sufficient geometric conditions in Section 3.1 (positive-definite diffusion, compound Poisson with (i-p) semigroup) are useful and are presented as sufficient. The entry and Berry–Esseen theorems already assume proximality, so they are unaffected.\n\nThe fix is local: add proximality (or condition (i-p)) to Theorem 3.20, matching Theorems 3.22–3.23 and the cited discrete-time CLT. The SLLN half stays. Minor typos like 'sceleton' should be cleaned. I would not accept the paper in this state, but I would definitely send it to review: the determinant representation and the skeleton method are good enough that the author's response should be obtained. The stress-test counterexample is correct, and the reader's high-confidence ACCEPT overrates the soundness. Bring it to reading group if you want to see a clean counterexample to a plausible-looking theorem.\n\nRecommendation: major revision, with proximality added to Theorem 3.20 and the σ²=0 case discussed.","headline":"Norm CLT in Theorem 3.20 is overbroad—needs proximality—but the determinant representation and skeleton method are solid.","tokens_in":28961,"tokens_out":4755,"would_cite":false,"duration_ms":45390,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G51","60J57","60H10","60B15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves strong laws of large numbers and central limit theorems for the norm, entries, and determinant of matrix-valued stochastic exponentials, with the determinant handled through an explicit one-dimensional…","keywords":["central limit theorem","law of large numbers","stochastic exponential","matrix-valued Lévy process","multiplicative Lévy process","products of random matrices","Lyapunov exponent","general linear group"],"falsifier":"Take $L_t=\\operatorname{diag}(B_t^1,B_t^2)$ with two independent standard Brownian motions. Then $G_X$ preserves the coordinate axes, so the paper's strong-irreducibility hypothesis fails, and a direct calculation gives $(\\log\\|X_t\\|-\\lambda t)/\\sqrt{t}\\Rightarrow \\max(Z_1,Z_2)$ rather than $N(0,\\sigma^2)$; this shows the geometric condition is doing real work. Conversely, finding a process that satisfies the stated geometric and moment assumptions but whose normalized log norm does not converge to a centered Gaussian would refute Theorem 3.20.","tokens_in":27929,"feed_emoji":"📈","tokens_out":10793,"duration_ms":101326,"temperature":0.7,"pith_summary":"Matrix-valued stochastic exponentials are the natural multiplicative analogues of matrix exponentials: given a Lévy process $L$ of $d\\times d$ matrices, $X$ solves $dX_t=X_{t-}\\,dL_t$, $X_0=I$. This paper proves that, under a non-singularity condition on the jumps and a geometric condition on the group generated by $X$, the logarithms of the norm and of $\\|yX_t\\|$ obey a strong law of large numbers and a central limit theorem: $(1/t)\\log\\|X_t\\|\\to\\lambda$ almost surely and $(\\log\\|X_t\\|-t\\lambda)/\\sqrt{t}\\Rightarrow N(0,\\sigma^2)$ with $\\sigma^2>0$. Individual entries require an additional proximality (dominant-eigenvalue) condition, while the determinant is treated separately: $\\det(X_t)$ is shown to be a one-dimensional stochastic exponential of an explicit Lévy process, giving explicit SLLN and CLT for $\\log|\\det X_t|$ and, where moments allow, Berry-Esseen bounds. The value is that the growth rates of the norm, entries, and volume of a random matrix flow are described by parameters computable from the driving Lévy process and its stationary projective measure.","feed_headline":"Log-norm of matrix stochastic exponentials is Gaussian at large times","feed_subtitle":"New limit theorems cover the norm, entries, and determinant of solutions to dX = X dL under weak moment assumptions.","key_machinery":"The argument runs through discrete skeletons: for any step $h>0$, the sequence $X_{nh}$ is a product of i.i.d. random matrices, so the established SLLN and CLT for products of random matrices apply along the skeleton. Lemmas 3.13 and 3.14 show the intervening small-time fluctuations $\\sup_{0\\le s\\le1}\\log M(X_{n,n+s})$ are asymptotically negligible under the second-moment condition, which lets the continuous-time limit follow from the discrete one by Slutsky-type arguments. The geometric input is the group $G_X$ generated by the support of the time-$E$ value of $X$ for an independent exponential $E$: requiring $G_X$ to be unbounded, strongly irreducible, and, for entries, proximal is what activates the random-matrix theorems. For the determinant, the key mechanism is a multivariate Itô computation identifying $\\det(X_t)$ as the one-dimensional stochastic exponential of the explicit Lévy process $\\check L$ above; once that identity is in place, the determinant asymptotics are read off from classical one-dimensional Lévy-process results.","core_discovery":"On the paper's own terms, the central discovery is that the long-time behaviour of a matrix-valued stochastic exponential is governed by the same constants as a discrete-time random matrix product, and that the determinant reduces to an explicitly identifiable one-dimensional Lévy process. Theorem 3.20 states that if the group $G_X$ is unbounded and acts strongly irreducibly on $\\mathbb{R}^d$, and if $\\mathbb{E}[\\sup_{0\\le s\\le1}(\\log M(X_s))^2]<\\infty$ with $M(a)=\\max\\{\\|a\\|,\\|a^{-1}\\|\\}$, then $t^{-1}\\log F(X_t)\\to\\lambda$ almost surely and $(\\log F(X_t)-t\\lambda)/\\sqrt{t}\\Rightarrow N(0,\\sigma^2)$ for $F(a)=\\|a\\|$ or $F(a)=\\|ya\\|$ with $\\|y\\|=1$. Theorem 3.22 adds proximality and an $\\epsilon$-moment condition to get the same CLT for $\\log|\\langle yX_t,z\\rangle|$, hence for individual entries. For the determinant, Theorem 4.1 computes $\\det(X_t)=\\mathcal{E}(\\check L)_t$ with $\\check L_t=\\operatorname{tr}(L_t)+\\frac12\\sum_{m\\neq n}(\\sigma_{(m,m),(n,n)}-\\sigma_{(m,n),(n,m)})t+\\sum_{s\\le t}(\\det(I+\\Delta L_s)-1-\\operatorname{tr}(\\Delta L_s))$, and Theorems 4.5 and 4.7 translate one-dimensional Lévy-process asymptotics into an SLLN and CLT for $\\log|\\det X_t|$ in terms of the characteristics of $L$. Theorem 3.23 provides Berry-Esseen bounds for the norm CLT and for the joint convergence of the radial and directional parts of the one-point motion.","pith_inferences":["The geometric hypotheses are stated through $G_X$ rather than directly through the Lévy triplet of $L$; the paper gives sufficient conditions but not a necessary-and-sufficient criterion, so a natural next step would be to characterize strong irreducibility and proximality of $G_X$ in terms of $L$ beyond the Brownian and compound-Poisson cases.","The determinant identity suggests a general principle: other multiplicative functionals of $X$, such as exterior powers or products of selected singular values, may also be stochastic exponentials of explicit Lévy processes, yielding CLTs for the whole Lyapunov spectrum.","Because $\\lambda$ and $\\sigma^2$ are expressed through the stationary measure of the projective process, one could build simulation-based estimators for these constants from a single long trajectory of the direction process, which the paper does not explicitly develop.","The continuous-time transfer via Lemmas 3.13 and 3.14 is flexible and might extend the discrete-time moderate-deviation or local-limit results cited in the paper to the stochastic-exponential setting; that extension is not claimed here."],"forward_implications":["If the diffusion coefficient matrix of $L$ is positive definite, then $G_X$ automatically satisfies the strong-irreducibility and proximality conditions, so the norm and entry CLTs hold under the stated moment assumptions.","The norm CLT requires only a second logarithmic moment, not an exponential moment, matching the optimal moment conditions known for discrete-time products of random matrices.","The Lyapunov exponent $\\lambda$ and variance $\\sigma^2$ can be obtained as derivatives of $\\Lambda(s)=\\lim_{n\\to\\infty} n^{-1}\\log\\mathbb{E}[\\|X_n\\|^s]$ or by integrating against the unique invariant measure of the projective action, giving concrete routes to estimation.","The growth rate of the determinant is explicit in the Lévy triplet of $L$; when the log-determinant jumps are not integrable, the paper shows the almost-sure growth rate is infinite.","The Berry-Esseen bounds in Theorem 3.23 give a $O(1/\\sqrt{t})$ rate for the normal approximation to $\\log\\|yX_t\\|$, and for the joint law of the direction and radius of the one-point motion."],"supporting_citations":[{"why":"Supplies the strong law of large numbers for products of i.i.d. random matrices used for the discrete skeleton in Theorem 3.20.","marker":"[14]"},{"why":"Supplies the central limit theorem for products of random matrices under the optimal moment condition used for the discrete skeleton.","marker":"[4]"},{"why":"Provides the Berry-Esseen bounds and the formulas for lambda and sigma-squared used in Theorem 3.23 and Remark 3.21.","marker":"[33]"},{"why":"Provides polynomial-moment Berry-Esseen bounds for the left random walk on the general linear group used in Theorem 3.23.","marker":"[9]"},{"why":"Provides the one-dimensional Lévy-process CLT from which the determinant CLT in Theorem 4.7 is read off.","marker":"[11]"},{"why":"Gives the framework of multiplicative Lévy processes in Lie groups, including the definition and properties of $G_X$ and the invariant measure used throughout.","marker":"[24]"},{"why":"States the strong irreducibility and proximality criteria and the invariant-measure theorem on projective space that the geometric assumptions rely on.","marker":"[7]"},{"why":"Provides the standard one-dimensional Lévy process law of large numbers and CLT facts used for the determinant process.","marker":"[28]"}],"fun_headline_variants":["Matrix exponential logs are Gaussian at large times","CLT for norms, entries, determinant of matrix exponentials","Determinant of stochastic matrix exponential is 1D Lévy","Limit laws for multiplicative Lévy processes in GL(d)","Berry-Esseen bounds for log-norm of matrix exponentials"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the set of values the stochastic exponential can take being direction-mixing in a precise sense: no finite union of proper subspaces of $\\mathbb{R}^d$ is left invariant by $G_X$, and $G_X$ must be unbounded, with proximality for the entry-level results.","fun_headline_variants_meta":{"raw":{"variants":["Matrix exponential logs are Gaussian at large times","CLT for norms, entries, determinant of matrix exponentials","Determinant of stochastic matrix exponential is 1D Lévy","Limit laws for multiplicative Lévy processes in GL(d)","Berry-Esseen bounds for log-norm of matrix exponentials"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000664,"raw_usage":{"total_tokens":3058,"prompt_tokens":997,"completion_tokens":2061,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":1979}},"tokens_in":613,"tokens_out":2061,"duration_ms":14522,"temperature":1.0,"reasoning_tokens":1979,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:46:53.942850+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $L_t=\\operatorname{diag}(B_t^1,B_t^2)$ with two independent standard Brownian motions. Then $G_X$ preserves the coordinate axes, so the paper's strong-irreducibility hypothesis fails, and a direct calculation gives $(\\log\\|X_t\\|-\\lambda t)/\\sqrt{t}\\Rightarrow \\max(Z_1,Z_2)$ rather than $N(0,\\sigma^2)$; this shows the geometric condition is doing real work. Conversely, finding a process that satisfies the stated geometric and moment assumptions but whose normalized log norm does not converge to a centered Gaussian would refute Theorem 3.20.","supporting_citations":[{"cited_title":"Furstenberg and H","cited_arxiv_id":null,"evidence_quote":"Supplies the strong law of large numbers for products of i.i.d. random matrices used for the discrete skeleton in Theorem 3.20."},{"cited_title":"Benoist and J.-F","cited_arxiv_id":null,"evidence_quote":"Supplies the central limit theorem for products of random matrices under the optimal moment condition used for the discrete skeleton."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Berry-Esseen bounds and the formulas for lambda and sigma-squared used in Theorem 3.23 and Remark 3.21."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides polynomial-moment Berry-Esseen bounds for the left random walk on the general linear group used in Theorem 3.23."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the one-dimensional Lévy-process CLT from which the determinant CLT in Theorem 4.7 is read off."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the framework of multiplicative Lévy processes in Lie groups, including the definition and properties of $G_X$ and the invariant measure used throughout."},{"cited_title":"Bougerol and J","cited_arxiv_id":null,"evidence_quote":"States the strong irreducibility and proximality criteria and the invariant-measure theorem on projective space that the geometric assumptions rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the standard one-dimensional Lévy process law of large numbers and CLT facts used for the determinant process."}],"review_version":1}