{"id":"ad44e76d-1bca-419f-82ea-0f9061fa690a","arxiv_id":"1908.08357","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A single-path-space probability construction for impulse-controlled continuous Markov processes with randomized impulses, including Markov and independent-cycles policy subclasses.","lead":"This paper builds a rigorous probability model for impulse control of continuous Markov processes, where each intervention can randomly relocate the state. It proves the controlled process can be defined on a single path space and identifies policy classes that yield a Markov process or independent cycles.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.3 only realizes cycle-local policies: the construction evaluates τ_{k+1} on the current cycle path, while Definition 2.3 allows arbitrary {F_{t−}}-stopping times; equality τ_k(X)=τ̃_k is never proved for k≥2.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the induction in Theorem 3.3 substitutes the current coordinate path for the entire past, so the construction is not shown to realize arbitrary admissible {F_{t−}}-stopping-time policies. This is not a stylistic or presentation issue; it affects the truth of Theorem 3.3 as stated. A concrete deterministic policy depending on the first cycle can make the constructed intervention interval differ from the prescribed one. The issue is addressable by restricting the admissible class or by adding a cycle-locality hypothesis and proving the correspondence, and the paper's later sections already work with such subclasses. Therefore the appropriate verdict remains conditional rather than rejection: the constructive framework has real content, but the general existence theorem needs qualification before it can be taken as a fully general foundation.","tokens_in":32516,"tokens_out":11348,"duration_ms":107124,"concrete_test":"Instantiate the model with E=R, fundamental process a Brownian motion, and Z constant. Define τ_1(ω)=1 if ω(1/2)≥0 and τ_1(ω)=2 otherwise; define τ_2(ω)=τ_1(ω)+1. These are admissible {F_{t−}}-stopping times, and the intended policy has second cycle duration exactly 1. Now compute the product-space construction: τ̃_1=τ_1(ω_0), while τ̃_2=τ_2(ω_1)=τ_1(ω_1)+1. Hence τ̃_2−τ̃_1=τ_1(ω_1)+1−τ_1(ω_0), which is not identically 1 and, because ω_1 is an independent fresh Brownian path, has non-degenerate distribution. This violates the stated property that the interval [τ_1,τ_2) under P^{(τ,Z)}_ν has length τ_2−τ_1=1. The same failure persists if one tries to repair the algebra by adding τ̃_1 to τ̃_2. The only repair is to impose cycle-locality in Definition 2.3 and prove τ_k(X̃)=τ̃_k under that hypothesis.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The proof of Theorem 3.3 builds the intervention times inductively on the product space: equations (3.4)–(3.6) define τ̃_2(e_1)=τ_2(ω_1), and the induction step defines τ̃_{k+1} similarly from the fresh coordinate path ω_k. Thus the constructed measure P^{(τ,Z)}_ν realizes a policy in which the kth decision depends only on the current cycle path, not on previous cycles or on absolute clock time. However, Definition 2.3 allows any {F_{t−}}-stopping times on the single path space, which may depend on the full past. The proof never establishes τ_k(X̃)=τ̃_k for k≥2; the observation after (3.5) only checks consistency of τ_1 when replacing ω_0 by ω_1. Corollary 3.2 (Galmarino/Courrège–Priouret) does not bridge this gap, because the pasted path and the fresh coordinate path do not agree on the history before the current cycle. Consequently, the central claim of Theorem 3.3(a) is not proved for the stated class of admissible policies; it appears to hold only for policies that are already cycle-local, such as the Markov policies of Definition 4.1 or the independent-cycles policies of Definition 5.1. The paper should either restrict Definition 2.3 to such policies, or add and prove an explicit hypothesis ensuring τ_k of the pasted path depends only on the current cycle path.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a single-path-space model for impulse control of continuous strong Markov processes with random effects. The main construction (Theorem 3.3) builds, for a nominal impulse policy and an initial distribution, a probability measure on the path space under which the coordinate process evolves as the fundamental Markov process between interventions, with the post-intervention state drawn from a policy-dependent kernel. The paper then isolates a class of Markov nominal impulse policies for which the controlled process is Markov and forms a Markov family, and a class of independent-cycles policies for which renewal arguments can be applied. The (s,S) inventory ordering policy is given as an example.","tokens_in":32812,"tokens_out":18929,"duration_ms":194155,"significance":"If the central construction is repaired, the paper addresses a genuine modelling problem: impulse control with randomized post-intervention states, formulated on a single path space rather than on a countable product of path spaces. The proposed subclasses (Markov policies and independent-cycles policies) are natural and potentially useful for long-term average control problems, and the (s,S) example illustrates a practically relevant setting. The paper is also careful in its use of strong Markov families, Galmarino/Courrège–Priouret characterizations, and the Ionescu Tulcea extension theorem. However, the main existence theorem is currently proved only for a narrower class of policies than the one stated in Definition 2.3, and a few supporting arguments need tightening before the results can be accepted as stated.","major_comments":[{"comment":"The construction defines the second intervention by ~τ2(e1) = τ2(ω1), i.e., the original policy functional τ2 is evaluated on the fresh coordinate path ω1, not on the pasted path T1(e1); the induction step repeats this by evaluating τ_{k+1} on ω_k. Equality ~τ_k = τ_k(~X) is asserted in the final paragraph of the proof but is not proved for k ≥ 2, and it is false in general under Definition 2.3. For example, if τ1 ≡ 5 and τ2 ≡ 10 (both are admissible {F_{t−}}-stopping times), the construction gives ~τ2 = 10 in the time scale of the second coordinate, so the pasted process intervenes at absolute time 15 rather than 10. The observation after (3.5) and Corollary 3.2 only compare ω1 with ω0 on [0, τ1(ω0)); they do not compare T1(e1) with ω1 on the second cycle. The theorem can be repaired either by restricting Definition 2.3 to cycle-local policies, as is effectively done in Definitions 4.1 and 5.1, or, for the general class, by defining ~τ_{k+1}(e_k) = τ_{k+1}(T_k(e_k)) and using Corollary 3.2 to prove consistency with the final pasted path. As written, the central claim of Theorem 3.3(a) is not established for the stated class of admissible policies.","section":"Theorem 3.3 proof, transition kernel measurability"},{"comment":"In the verification that P1 is a transition kernel, the proof asserts that ω0 ↦ ∫_E P_{v1}(F1) Q1(ω0, dv1) is F0-measurable because v1 ↦ P_{v1}(F1) is universally measurable and can be uniformly approximated by simple functions. Universal measurability does not imply Borel measurability, and composing a non-Borel universally measurable function with the F0-measurable maps (~Y1, ~Z1) need not produce an F0-measurable function (a Dirac-type kernel makes this explicit). The Ionescu Tulcea step therefore requires either a Borel-measurable transition family, for example under a Feller-type assumption, or a completed σ-algebra with a correspondingly amended statement of the theorem. This is a technical gap, but it affects the proof of existence in Theorem 3.3.","section":"Theorem 3.3 proof, transition kernel measurability"},{"comment":"The theorem states that {P^{θ_s(τ,Z)}_x : s ≥ 0, x ∈ E} is a Markov family. Under the standard definition cited from Karatzas and Shreve (Definition 2.5.11), a member P_{s,x} must satisfy P_{s,x}(X(s) = x) = 1. The measures constructed here satisfy P^{θ_s(τ,Z)}_x(X(0) = x) = 1, since they are measures on paths indexed from time 0. Equation (4.19) is a useful policy-shift relation, but it does not by itself place the collection in the cited definition. The proof should either define the measures on time-shifted paths, e.g. by composing with the shift operator θ_s, or reformulate the claim as a time-inhomogeneous Markov family with policy shifts.","section":"Theorem 4.9"}],"minor_comments":[{"comment":"The proof of Theorem 3.1 is deferred with the statement that the proof of Courrège and Priouret 'remains valid with the slight modification'. Since Corollary 3.2 is used in the main construction, the authors should either give a complete proof of the left-continuous version or cite a source where it is proved.","section":"Theorem 3.1"},{"comment":"The notation in the induction step overloads ω_k, which is used both for the coordinate path in Ω_k and for the pasted path T_k(e_k). In particular, equations (3.4)–(3.6) are hard to parse because ω1 is used in both senses. Introducing separate notation for the pasted path would clarify the argument and help expose the consistency issue raised above.","section":"Theorem 3.3 proof, notation"},{"comment":"In the induction step, the phrase '~Y_{k+1} ∈ E_k corresponds to ω_k(τ_{k+1}−) ∈ E' appears to contain a typo: ~Y_{k+1} should take values in E, not in E_k.","section":"Induction step, text around Eq. (3.9)"},{"comment":"The proof of Theorem 4.6 repeatedly uses the phrase that the collections {F_{j−1}} and {G_j(e_{j−1})} are 'sufficiently large' to conclude equality of conditional expectations. This should be replaced by an explicit monotone-class or π–λ argument, since the current wording is too terse for a proof of this complexity.","section":"Theorem 4.6 proof"}],"recommendation":"major_revision","confidential_remarks":"The main issue in Theorem 3.3 is serious but appears to be repairable within the manuscript's scope: either the admissible-policy class must be restricted to cycle-local policies, or the construction must evaluate the policy on the pasted path T_k(e_k). The paper's useful subclasses in Sections 4 and 5 are likely to survive the repair, so I do not see grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for the query. Read Helmes–Stockbridge–Zhu carefully. What's genuinely new: they build impulse control with random effects on a single path space for continuous strong Markov processes, rather than countable product spaces, and they identify structural subclasses (Markov nominal policies, independent cycles) that make renewal and Markov-family arguments available. The randomized impulse mechanism via measurable kernels Q(y,z) is a real extension, and the (s,S) inventory example is a nice sanity check. The paper is clearly written and the citations to the modelling literature (Robin, Stettner, Lepeltier–Marchal, Meyer) look right.\n\nNow the soft spot, and it's real. The stress-test note is correct: Theorem 3.3 constructs the measure on the product space by defining τ~_{k+1} from the fresh coordinate path ω_k, i.e. τ~_{k+1}(e_k)=τ_{k+1}(ω_k). That means the realized policy only depends on the current cycle, not the full past. But Definition 2.3 allows arbitrary {F_{t−}}-stopping times on the single path space, which can depend on everything. The proof never shows τ_k(X~)=τ~_k for k≥2; the observation after (3.5) only checks k=1 when ω0 and ω1 agree up to τ1. Corollary 3.2 doesn't bridge this. So the central existence theorem is overstated for the general class. It holds for the cycle-local policies defined later — Markov policies (Section 4) and independent-cycles policies (Section 5) — which are the ones that matter for the paper's applications. But the reader has to live with a mismatch between the general definition and what the construction delivers.\n\nThis is fixable. Either restrict Definition 2.3 to cycle-local policies, or add a condition that τ_k of the pasted path depends only on the current cycle path and prove it. Also, Theorem 3.1 is stated as a 'slight modification' of Courrège–Priouret without proof; that is a minor annoyance, not a flaw, since the result is standard. The independence and iid-cycle arguments in Section 5 look right under the stated subclass.\n\nBottom line: this is a useful paper for stochastic control theorists and applied probabilists, worth a serious referee. The gap is targetable and the authors can close it by narrowing the claim. If they don't, readers should cite it carefully — the general policy class is not realized, only the cycle-local subclasses. I'd send it to review, with a request that the authors fix the scope of Theorem 3.3.","headline":"Solid modelling paper with a genuine but fixable gap: Theorem 3.3 only realizes cycle-local policies, not the full class of admissible stopping times.","tokens_in":33321,"tokens_out":2108,"would_cite":true,"duration_ms":20332,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93E20","60H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single path-space model puts randomized impulse control of continuous Markov processes on one probability space.","keywords":["impulse control","randomized impulse","random effects","Markov impulse policy","Markov family","path space construction","strong Markov process","(s,S) ordering policy"],"falsifier":"Take E=R with a continuous strong Markov process, set the first intervention at τ1≡1, and define the second intervention time by τ2(ω)=2+1_{X(0)∈A}, which is a left-continuous stopping time under the paper's Definition 2.3. Under the constructed measure, the second intervention time is computed from the first-cycle path after shifting, so it depends on the state after the first impulse rather than on X(0). A path with X(0)∈A but different post-impulse state would receive different second intervention times under the original policy and under the constructed process, showing that the theorem as stated does not realize this admissible policy.","tokens_in":32304,"feed_emoji":"🎲","tokens_out":7522,"duration_ms":77596,"temperature":0.7,"pith_summary":"This paper establishes that impulse control with random effects for continuous Markov processes can be modelled on the single path space of càdlàg functions, rather than on the customary countable product of path spaces. For every admissible nominal impulse policy and initial distribution, it constructs a probability measure under which the coordinate process evolves as the underlying strong Markov process between interventions, and at each intervention the next state is drawn from a prescribed distribution conditioned on the pre-intervention state and the nominal impulse. The construction also identifies a class of Markov nominal policies for which the controlled process is Markov and forms a Markov family, and a class of independent-cycles policies for which cycles after the first are independent and identically distributed, making classical renewal arguments available for long-term average cost problems. This matters because randomized impulses arise naturally in applications such as inventory management, and a single natural filtration gives the intervention times a simpler, more direct interpretation as stopping times.","feed_headline":"One path space now models randomized impulse control","feed_subtitle":"A new construction makes the controlled process a single coordinate process, unlocking Markov and renewal tools.","key_machinery":"The load-bearing object is the pasting map from the countable product of path spaces to the single path space: after building a product measure through transition kernels that integrate the fundamental process against the random-effect distribution for the next cycle, the map concatenates the coordinate paths at the random intervention times, and its push-forward is the desired single-space measure. Two supporting mechanisms make this work: the path-locality characterization of left-continuous stopping times ensures that evaluating the next intervention time on the new coordinate path does not change the values of earlier intervention times, and the terminal-time condition restricts Markov policies to decisions based on the current cycle, which is what lets the Markov property be proved. The continuity of the fundamental process guarantees that the pre-intervention state appears in the natural filtration and that the pasted paths are càdlàg.","core_discovery":"The central discovery is Theorem 3.3: for a strong Markov process with continuous paths and a given family of post-impulse distributions, any admissible nominal impulse policy determines a family of probability measures on the single space of càdlàg paths. Under the resulting measure, the coordinate process runs as the fundamental Markov process on the first interval, and on each later interval the process again follows the fundamental dynamics, with the post-impulse state conditionally distributed according to the random-effect distribution given the information before the intervention. The measure is built on the countable product of path spaces by applying the Ionescu Tulcea extension theorem to transition kernels that select each new cycle's starting distribution, then pushing forward the product measure through a càdlàg pasting map that splices the coordinate paths at the intervention times. The paper then proves that when the stopping times are terminal times depending only on the current cycle and the impulses depend only on the pre-intervention state, the controlled process is Markov and, together with time-shifted policies, forms a Markov family. For policies whose intervention times are determined within each cycle and whose effect distributions do not depend on the path, the cycles are independent; for the (s,S) ordering policy they are identically distributed after the first cycle.","pith_inferences":["The cycle-local interpretation implicit in the construction suggests that admissible policies using absolute clock time or the full history of earlier cycles are not realized by the theorem as stated; one testable extension is to enlarge the state with a clock or finite memory so that such policies become cycle-local.","Because the random-effect mechanism is just another transition kernel, the same construction should extend to randomized intervention times or relaxed controls by making the next-cycle kernel itself random.","The Markov family is parameterized by time-shifted policies, which points toward viewing randomized impulse control as a family of Markov decision problems on the path space, one for each starting time and state."],"forward_implications":["Randomized impulses can be handled rigorously on one path space: the controller's nominal impulse and the observed pre-impulse state only need to specify a distribution for the new state, not a deterministic value.","For continuous strong Markov processes, any admissible policy's controlled process has a concrete single-space representation, so optimal control questions can be posed with the natural filtration of the controlled process.","Markov nominal policies make the controlled process Markov and the time-shifted family a Markov family, opening dynamic programming and Markov-process tools to randomized impulse control.","Independent-cycles policies, including (s,S) inventory ordering, produce independent cycles; identical stopping rules and effect distributions make cycles after the first identically distributed, allowing renewal theorems for average-cost criteria.","The construction is restricted to continuous paths; extending it to jump processes would require a different treatment of the left-limit state at intervention times."],"supporting_citations":[{"why":"Supplies the path-locality characterization of {Ft−}-stopping times used to keep intervention times consistent across shifted coordinate paths.","marker":"Courrège and Priouret (1965)"},{"why":"Provides the Ionescu Tulcea extension theorem used to build the probability measure on the countable product of path spaces.","marker":"Neveu (1965)"},{"why":"A product-space construction of impulse-controlled Markov processes that this paper adapts and then pushes down to a single path space.","marker":"Robin (1978)"},{"why":"A product-space model with half-closed cycles and simultaneous interventions, used as a reference point for the new single-space construction.","marker":"Stettner (1983)"},{"why":"An alternative impulse-control model on product spaces whose treatment of jump times and intervention locations motivates the continuous-path simplification.","marker":"Lepeltier and Marchal (1984)"},{"why":"Pioneered the piecing-out construction of Markov processes that underlies pasting coordinate paths at rebirth or intervention times.","marker":"Ikeda et al. (1966)"},{"why":"Provides complete proofs and the single-filtration view of piecing out, which Remark 3.4 connects to the present construction.","marker":"Meyer (1975)"},{"why":"Supplies the definition of strong Markov family and Markov family used to state and prove the Markov policy results.","marker":"Karatzas and Shreve (1988)"},{"why":"Earlier model of impulse control with random consequences, generalized here to the single-path-space, random-effect setting.","marker":"Korn (1997)"}],"fun_headline_variants":["Single path space now models random impulse control","Random impulses on one continuous path","Impulse control with random effects on a single path","Markov family from a single impulse path","One path space for random impulse policies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The construction assumes that every intervention time can be evaluated from the current cycle's path alone, because the next stopping time is applied to the newly pasted coordinate path; policies whose stopping decisions use absolute clock time or the history of earlier cycles are not realized by the model.","fun_headline_variants_meta":{"raw":{"variants":["Single path space now models random impulse control","Random impulses on one continuous path","Impulse control with random effects on a single path","Markov family from a single impulse path","One path space for random impulse policies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000785,"raw_usage":{"total_tokens":3518,"prompt_tokens":1053,"completion_tokens":2465,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":2400}},"tokens_in":669,"tokens_out":2465,"duration_ms":16352,"temperature":1.0,"reasoning_tokens":2400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:44:56.905006+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take E=R with a continuous strong Markov process, set the first intervention at τ1≡1, and define the second intervention time by τ2(ω)=2+1_{X(0)∈A}, which is a left-continuous stopping time under the paper's Definition 2.3. Under the constructed measure, the second intervention time is computed from the first-cycle path after shifting, so it depends on the state after the first impulse rather than on X(0). A path with X(0)∈A but different post-impulse state would receive different second intervention times under the original policy and under the constructed process, showing that the theorem as stated does not realize this admissible policy.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Ionescu Tulcea extension theorem used to build the probability measure on the countable product of path spaces."},{"cited_title":"(1978) Contrˆ ole Impulsionnel des Processus de Markov","cited_arxiv_id":null,"evidence_quote":"A product-space construction of impulse-controlled Markov processes that this paper adapts and then pushes down to a single path space."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A product-space model with half-closed cycles and simultaneous interventions, used as a reference point for the new single-space construction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An alternative impulse-control model on product spaces whose treatment of jump times and intervention locations motivates the continuous-path simplification."},{"cited_title":"and Watanabe, S","cited_arxiv_id":null,"evidence_quote":"Pioneered the piecing-out construction of Markov processes that underlies pasting coordinate paths at rebirth or intervention times."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides complete proofs and the single-filtration view of piecing out, which Remark 3.4 connects to the present construction."},{"cited_title":"and Shreve, S","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of strong Markov family and Markov family used to state and prove the Markov policy results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier model of impulse control with random consequences, generalized here to the single-path-space, random-effect setting."}],"review_version":1}