{"id":"cc2b07ab-7022-45f4-9801-d643061f6b8e","arxiv_id":"2607.05320","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Causal separation of asset returns by declared drivers forces a diagonal-plus-low-rank covariance and a closed-form projected Markowitz solution with uniqueness, invariance, and certified approximate-separation bounds.","lead":"A single screening-off condition on drivers yields a full static mean-variance theory: diagonal-plus-low-rank risk, a closed-form projected Markowitz portfolio, and exact sensitivity bounds. It matters because it turns the choice of conditioning information into a testable primitive rather than an unexamined input to covariance regularization.","discovery_kind":"first_principles","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged A4 perimeter.","rationale":"The reader's strongest claim is precisely the formal content that follows from the separation condition alone. That content is standard measure-theoretic and convex-optimization reasoning; the only place where an extra structural hypothesis is required is the causal-semantics subsection, which the paper and the reader both treat as optional. Because the load-bearing concern for the causal half is already correctly identified (A4(ii)+(iv)), and because no independent inconsistency appears in the portfolio-theory half, the appropriate stress-test outcome is an honest non-finding. The recommended concrete check simply reconfirms that the derivation does not smuggle A4 into the non-causal theorems. Verdict therefore remains CONDITIONAL with the same scope the reader already assigned: accept the mathematics once A4 and the existence of small real-market ε* are treated as open empirical inputs.","tokens_in":26573,"tokens_out":583,"duration_ms":5240,"concrete_test":"Independently re-derive the decision-node covariance (9) and the projector form of M in Theorem 4.1 from Definitions 2.4 and 3.3 plus the tower property alone, without invoking Assumption 2.16 (A4) or any causal graph; if both objects still follow and the machine-precision residuals of Table 3 remain at floating-point zero, the formal core is secure and the reader's CONDITIONAL verdict needs no further adjustment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mathematical claim (separation ⇒ diagonal-plus-low-rank Qt via the tower, closed-form projected Markowitz with M = Q^{-1/2} P Q^{-1/2}, uniqueness of the minimal sufficient separator as a σ-algebra, and first-order sensitivity under approximate separation) is internally consistent under the stated primitives. The proofs of Theorems 2.14, 3.1, 4.1 and Proposition 5.1 rely only on the graphoid intersection property (A2), the two-window screening definition, and standard KKT/Woodbury algebra; they do not require A4. A4 is needed solely for the causal soundness/identification/intervention package (Props. 2.17–2.20), which the paper already delimits as optional and which the reader correctly isolates. Experiments E1–E7 are confirmatory on a DGP that satisfies the normal form by construction and therefore cannot falsify the derivation; they also cannot create a new load-bearing hole. The classical-factor-model limit is acknowledged by the paper itself. Consequently no additional soft spot in the strongest claim is more load-bearing than the one already named.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper takes a single structural condition—causal separation: mutual conditional independence of asset returns given a declared driver set’s realized path through the investment horizon, plus asset-past screening (S0)—and derives the static mean–variance theory it induces. Separation yields, via an exact tower decomposition, a diagonal-plus-low-rank decision-node covariance; the constrained program admits a closed-form projected Markowitz solution in which Q⁻¹ is replaced by its projection onto the constraint-compatible subspace. The paper proves uniqueness of the minimal sufficient separator as a σ-algebra, invariance under separator equivalence and driver reparametrization, a conditional efficient frontier, an exact Hansen–Jagannathan gap in the shadow prices of the geometry, regularization by the idiosyncratic floor, an O(n(m+p)²) two-stage solver, and first-order sensitivity under approximate separation. Causal semantics (soundness, identification, interventional invariance, dependence floor of latent confounding) are stated under an explicit structural margin A4 and delimited as optional for the portfolio theory. Seven reproducible synthetic experiments check identities at machine precision and quantify estimation risk, robustness, intervention invariance, and scaling.","tokens_in":26884,"tokens_out":1661,"duration_ms":26818,"significance":"If the results hold, the paper supplies a clean, self-contained static theory that answers a prior question the covariance-regularization literature leaves open: conditional on what information should moments be computed? The selection theory (existence, uniqueness as information set, invariance), the exact tower identification of the low-rank block as response to driver innovations, the projected solution with explicit HJ gap and conditioning floor, and the certified-tolerance sensitivity bounds are genuine contributions. Strengths that should count in the assessment include: (i) machine-precision verification of five structural identities (Table 3); (ii) a fully reproducible package with fixed seeds and a pre-specified field protocol; (iii) honest delimitation of the causal package under A4 and of the synthetic scope; (iv) an exact linear-cost solver that never forms the n×n matrix. The classical-factor-model limit is acknowledged rather than hidden. The work is complementary to, and narrower than, the author’s prior sensitivity/PDE-control line, isolating the conditioning condition itself.","major_comments":[{"comment":"The abstract and title lead with “causal separation,” but the portfolio theory of Sections 3–5 requires only the observational screening-off property (Def. 2.4) and regularity A1–A3; the causal package (Props. 2.17–2.20, Thm. 2.19) needs the structural margin A4, especially sink-node returns with no within-horizon contagion (A4(ii)) and completeness of the declared universe for common causes (A4(iv)). The body already delimits this carefully (Remark 2.21). The abstract and introduction should state the same split up front—observational portfolio theory vs. optional causal semantics under A4—so that readers do not over-read the causal claim when fire-sale or feedback channels are present at the horizon.","section":"Abstract; §2.4, Assumption 2.16"},{"comment":"When the factor count is correctly specified, PCA-m is numerically indistinguishable from the structured estimator (Table 1; classical-limit discussion in §1.1). The practical content of the framework therefore rests on three points the paper already has but under-emphasizes: (a) selection without choosing a factor count, with minimality restored by backward elimination (Remark 2.9, Table 4); (b) the dependence-floor diagnostic that detects latent confounding rather than absorbing it (Prop. 2.20, Table 4); (c) interventional distinguishability of causal vs. correlational separators of equal fit (Thm. 2.19, E6/Fig. 5). Elevate these three as the empirical value proposition early in §6, and state explicitly that agreement with correctly-specified PCA is the intended classical limit, not a null result.","section":"§1.1; §6.3, Tables 1–2; §6.7, Table 4 and Fig. 5"}],"minor_comments":[{"comment":"Fig. 4 (E5): the worst-case bound exceeds realized weight displacement by a factor of ~355 at ε=0.1. A short sentence on when the first-order formula (Prop. 5.1) should be preferred as the working diagnostic over Cor. 5.2 would help practitioners; the paper already notes the conservatism but could be more directive.","section":"§6.6, Fig. 4; Remark 5.3"},{"comment":"Table 2: the rebalanced oracle Sharpe of 11.8 is flagged as a simulator artifact and ratios are scale-free, but a one-line note in the table caption that absolute Sharpe levels are not interpretable would reduce misreading.","section":"Table 2"},{"comment":"E1 uses Fisher-z tests on generated regressors (η̂). The caveat is stated; a pointer that the generalised covariance measure of Shah–Peters is the formal replacement (already cited) could be moved from the text into the figure caption for visibility.","section":"§6.2, Fig. 1"},{"comment":"Notation: G^{+h}_t and G_t are introduced in Def. 2.11; a one-line reminder at the start of §3 that the tower is over G_t ⊆ G^{+h}_t would help readers who skip the selection section.","section":"§3.1, Eq. (8)"},{"comment":"Typographical: “Rodríguez Domínguez” appears with varying accent encoding in the author line and self-citations; unify. Also “Sept(A; δ, H)” vs. “Sep_t(A)”—pick one subscript convention.","section":"Title page; Def. 2.4"},{"comment":"Section A Table 5 is useful; add the dependence functional used in each experiment (max residual correlation is the working choice of §5) so the configuration is fully self-contained without the code.","section":"Appendix A, Table 5"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is carefully written and unusually honest about scope (static, synthetic, A4 optional). Self-citation of the author’s prior causal-portfolio line (refs 25–28) is appropriate and the complementarity claim is accurate. Fit for a theory-oriented quant-finance venue is good; a more empirical journal might demand a real-panel run of the field protocol before acceptance. I do not see a load-bearing mathematical error. Minor revision is the right bar: the two major points are clarification and emphasis, not new theorems or new experiments."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple. Take mutual independence of asset returns given the drivers’ realized path through the horizon (plus the asset-past screen S0). From that single condition the paper derives, cleanly, the tower decomposition that forces Qt = diagonal + low-rank, the closed-form projected Markowitz solution with M the projection of Q^{-1} onto the constraint-compatible subspace, uniqueness of the minimal sufficient separator as a σ-algebra, invariance under reparametrization, an exact HJ gap as a quadratic in the shadow prices, the idiosyncratic-floor Lipschitz bounds, and first-order sensitivity under approximate separation. Those derivations are standard probability and KKT algebra under explicit assumptions A1–A3; they do not need the causal package.\n\nWhat is actually new is the selection theory (existence of the optimal approximate separator, uniqueness as information set under the intersection property, gauge invariance) and the forced two-window architecture that makes screening-off and the decision-node covariance simultaneously true. The causal half (A4, soundness, identification, interventional distinction of proxies) is carefully delimited rather than smuggled in; the intervention experiment is the right test. Experiments E1–E7 are confirmatory on a DGP built to satisfy the normal form, identities hold at machine precision, the two-stage solver scales, and the package is reproducible. The paper itself states the classical-factor-model limit when the space is known, so that is not a hidden soft spot.\n\nThe real perimeter is A4 (returns as sinks, declared universe covers common causes) and the open empirical question whether small certified ε* exist in liquid markets. The paper does not pretend otherwise; it ships a pre-specified field protocol and leaves performance claims to the earlier empirical line. Free parameters (dep functional, κ, geometry Ut) are ordinary modeling choices, not free lunch.\n\nThis is for people who care about conditioning-set selection as a primitive in mean-variance, or who want a clean static foundation under a screening-off condition. The math is solid enough for a serious referee. I would engage: read the uniqueness and projection theorems, check the two-window claim against your own factor setups, and treat A4 and real-market ε* as empirical inputs. Worth peer review.","headline":"Clean static theory: one CI condition yields diagonal-plus-low-rank risk, projected Markowitz, uniqueness/invariance, and first-order ε-bounds; novelty is real but bounded by the classical factor limit the paper itself states.","tokens_in":27511,"tokens_out":588,"would_cite":true,"duration_ms":6656,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91G10","62P05","90C20"],"pacs":[],"model":"grok-4.5","headline":"One screening-off condition on drivers yields a closed-form projected Markowitz portfolio and a diagonal-plus-low-rank risk structure.","keywords":["causal separation","projected Markowitz","conditional risk","screening-off","diagonal-plus-low-rank covariance","Hansen-Jagannathan bound","portfolio selection","approximate separation"],"falsifier":"On a controlled panel where the true common drivers are known, inject residual cross-correlation of size epsilon into the residual block (or intervene by destroying a proxy's link to a true cause while preserving its marginal law) and check whether the first-order displacement formulas, the dependence floor, and the invariance of the causal separator hold at the predicted magnitudes while a correlational alternative collapses.","tokens_in":27383,"feed_emoji":"📊","tokens_out":842,"duration_ms":6224,"temperature":0.7,"pith_summary":"Mean-variance portfolio choice is solvable but fragile because estimation error in the mean and the inverse covariance is amplified. This paper answers a prior question: conditional on what information should those moments be computed? It formalizes causal separation: if a declared set of drivers screens off the assets so that, given the drivers' realized path through the investment horizon, asset returns are mutually independent, then the entire static mean-variance theory follows. The condition forces a diagonal-plus-low-rank conditional covariance via an exact tower decomposition, with the low-rank block identified as the assets' response to driver innovations. The constrained mean-variance problem then admits a closed-form projected Markowitz solution in which the classical information matrix is replaced by its projection onto the constraint-compatible subspace. The paper also proves uniqueness of the minimal sufficient separator as an information set, invariance under reparametrization, a conditional efficient frontier, an exact Hansen-Jagannathan gap priced by shadow prices of the constraints, regularization by the idiosyncratic variance floor, and first-order sensitivity bounds under approximate separation. The causal content is proved under an explicit structural margin rather than assumed: common causes form a separator, observational data identify their realized information only, and interventions distinguish true causes from equally fitting correlational proxies. Seven reproducible experiments check the identities at machine precision and measure estimation risk, robustness, intervention invariance, and scaling to thousands of assets.","feed_headline":"One screening condition yields projected Markowitz portfolios","feed_subtitle":"Independence given drivers forces diagonal-plus-low-rank risk and a closed-form solution","key_machinery":"Causal separation (screening-off on the horizon-closed driver path together with asset-past screening S0), which forces the tower decomposition of the decision-node covariance into a diagonal idiosyncratic block plus a low-rank response-to-innovations block, and thereby the projected Markowitz solution in which the information matrix is replaced by its orthogonal projection onto the constraint-compatible subspace.","core_discovery":"A single structural condition—causal separation, mutual independence of asset returns given the drivers' realized closed-horizon path, plus screening of the asset past—induces the complete static portfolio theory that follows from it: a diagonal-plus-low-rank decision-node covariance via tower decomposition, a unique closed-form projected Markowitz solution, uniqueness of the minimal sufficient separator as a sigma-algebra, invariance of all derived objects, a conditional efficient frontier, an exact Hansen-Jagannathan gap, regularization by the idiosyncratic floor, and first-order sensitivity under approximate separation at a certified tolerance.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Causal separation yields closed-form projected Markowitz portfolios","Asset independence given drivers forces projected Markowitz solution","Single causal condition delivers diagonal-plus-low-rank portfolio theory","Causal separation induces projected Markowitz and conditional frontier","Drivers-path independence gives exact projected Markowitz portfolios"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That every common cause of the assets' horizon returns is already among the declared drivers and that returns do not cause each other or feed back into the drivers inside the investment horizon; if fire-sale contagion or latent common causes operate inside the window, separation and its certificates fail.","fun_headline_variants_meta":{"raw":{"variants":["Causal separation yields closed-form projected Markowitz portfolios","Asset independence given drivers forces projected Markowitz solution","Single causal condition delivers diagonal-plus-low-rank portfolio theory","Causal separation induces projected Markowitz and conditional frontier","Drivers-path independence gives exact projected Markowitz portfolios"]},"model":"grok-4.5","effort":"low","cost_usd":0.004718,"raw_usage":{"total_tokens":1397,"prompt_tokens":868,"num_sources_used":0,"completion_tokens":60,"cost_in_usd_ticks":47180000,"prompt_tokens_details":{"text_tokens":868,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":469,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":868,"tokens_out":60,"duration_ms":3378,"temperature":1.0,"reasoning_tokens":469,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T07:18:11.939564+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a controlled panel where the true common drivers are known, inject residual cross-correlation of size epsilon into the residual block (or intervene by destroying a proxy's link to a true cause while preserving its marginal law) and check whether the first-order displacement formulas, the dependence floor, and the invariance of the causal separator hold at the predicted magnitudes while a correlational alternative collapses.","supporting_citations":[],"review_version":2}