{"id":"4cea386e-3eeb-4c7a-a44c-8d8f5a07c514","arxiv_id":"1908.04243","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The estimated weights of many optimal portfolios have an exact stochastic representation under normality, and in high dimensions their asymptotic distribution is multivariate normal with explicit bias-corrected estimators.","lead":"This paper derives exact probability distributions for estimated optimal portfolio weights and characteristics, assuming asset returns are multivariate normal. The formulas enable fast simulation and lead to high-dimensional normal approximations, confidence regions, and tests for portfolio weights.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1(i) gives the wrong asymptotic variance for the estimated GMV variance: the exact chi-square law implies coefficient sqrt(2)(1-c)V_GMV, not sqrt(2(1-c))V_GMV, contradicting Corollary 4.1 and Theorem 4.4.","rationale":"The reader's conditional verdict is reasonable: the exact finite-sample derivations in Section 7 are largely coherent and rest on standard Wishart and t-distribution results, and the high-dimensional covariance formulas in Corollary 4.1 and Theorem 4.2 appear to be the intended correct ones. My stress-test found a more specific and more load-bearing mathematical issue than the normality assumption: Theorem 4.1(i) as printed is inconsistent with the exact chi-square law for hat(V)_GMV and with the paper's own Corollary 4.1 and Theorem 4.4. Since Theorem 4.1 is the source of every high-dimensional limit in the paper, an uncorrected reading overstates the asymptotic variance of the estimated GMV variance by a factor of 1/(1-c), with a 10-fold error already at c=0.9. This would directly affect confidence regions for V-dependent portfolios. The correction is only a coefficient factor, so the conceptual content of the paper survives, which is why the verdict remains conditional rather than reject. The normality limitation identified by the reader is real and acknowledged in Section 6, but the internal coefficient inconsistency is the place where the central high-dimensional claim is least secure under the paper's own assumptions.","tokens_in":27315,"tokens_out":24461,"duration_ms":240432,"concrete_test":"Derive the variance of sqrt(n-p)(hat(V)_GMV - (1-p/n)/(1-1/n)V_GMV) directly from (n-1)hat(V)_GMV/V_GMV ~ chi^2_{n-p} and compare it with the square of the coefficient in Theorem 4.1(i) and with the (1,1) entry of Xi in Corollary 4.1. For numerical confirmation, draw B=10^5 realizations of chi^2_m with n=1000, p=900 (m=100) and compute the empirical variance of sqrt(m) * (m/(n-1)) * (chi^2_m/m - 1). The result should be about 2(1-c)^2 = 0.02, not 2(1-c) = 0.2. If the simulation matches 0.02, Theorem 4.1(i) and the proof of Theorem 4.2 need a factor (1-c) correction; if it matches 0.2, then Corollary 4.1 and the Omega formulas in Theorem 4.2/4.4 are wrong.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Under Theorem 2.1(i), (n-1)hat(V)_GMV/V_GMV is exactly chi-squared with n-p degrees of freedom. Consequently, the variance of sqrt(n-p)(hat(V)_GMV - (1-p/n)/(1-1/n)V_GMV) is (n-p) * Var(V_GMV/(n-1) chi^2_{n-p}) = 2 V_GMV^2 (n-p)^2/(n-1)^2, which tends to 2(1-c)^2 V_GMV^2. The correct stochastic limit is therefore sqrt(2)(1-c) V_GMV u1. As printed, Theorem 4.1(i) states sqrt(2(1-c)) V_GMV u1, whose variance is 2(1-c)V_GMV^2, a factor 1/(1-c) too large. Corollary 4.1's (1,1) entry is 2V_GMV^2(1-c)^2 and Theorem 4.4(b)'s 2V_GMV^2 term both use the corrected coefficient, and the proof of Theorem 4.2 as written uses the uncorrected sqrt(2(1-c)) factor when combining the u1 term. The high-dimensional CLT statement is thus internally inconsistent with the paper's own exact finite-sample law and with its later covariance formulas. A reader relying on the printed Theorem 4.1(i) will overstate the asymptotic variance of the estimated GMV variance by 1/(1-c); at c=0.9 this is a 10-fold error, which would miscalibrate confidence intervals for V_GMV and for any portfolio whose asymptotic distribution enters through g2(lambda), such as tangency, Sharpe-ratio, and VaR/CVaR optimal portfolios. This concern is not about heavy-tailed robustness; it is a checkable internal consistency issue in the central asymptotic theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper characterizes the exact finite-sample joint sampling distribution of the estimated five frontier parameters (GMV variance, GMV weights, GMV expected return, efficient-frontier slope, and self-financing portfolio weights) under i.i.d. multivariate normality, via a stochastic representation (Theorem 2.1). This representation is then used to derive exact stochastic representations for estimated optimal portfolio weights and characteristics (Theorems 3.1 and 3.2), and to obtain high-dimensional central limit theorems under p/n -> c (Theorems 4.1-4.4), together with consistent estimators and confidence regions in Section 4.3. A simulation study in Section 5 assesses the finite-sample quality of the asymptotic approximations and their robustness to t-distributed returns.","tokens_in":27695,"tokens_out":27691,"duration_ms":258855,"significance":"If the stated results were correct, the paper would be a valuable unified treatment: exact finite-sample distributions for a large class of optimal portfolios, an efficient simulation algorithm that avoids inverting the sample covariance matrix in each draw, explicit high-dimensional CLTs, and consistent estimators with confidence regions. The exact finite-sample representation in Theorem 2.1 is a substantial contribution and is derived from standard Wishart, inverse Wishart, and quadratic-form results rather than assumed. The simulation study is extensive and includes a robustness check against heavy-tailed returns. However, I find several load-bearing internal inconsistencies in the displayed formulas: a missing square root in Theorem 3.2, an incorrect asymptotic variance factor in Theorem 4.1(i), an omission of the term mu^T A mu in the asymptotic covariance of the slope estimator, and a structural problem with the claimed covariance-matrix form in Theorem 4.2 for general k. These issues must be resolved before the paper can be accepted.","major_comments":[{"comment":"The stochastic representation for hat R_GMV is missing a square root. The proof of Theorem 3.2 shows that, conditional on hat s, hat R_GMV has variance (V_GMV/n)(1 + n/(n-1) hat s), which equals (V_GMV/n)(1 + (p-1)/(n-p+1) psi). Therefore the correct representation is hat R_GMV = R_GMV + sqrt(V_GMV/n) * sqrt(1 + (p-1)/(n-p+1) psi) * z. As printed, Theorem 3.2 states hat R_GMV = R_GMV + sqrt(V_GMV/n)(1 + (p-1)/(n-p+1) psi) z, i.e. the parenthetical factor is not square-rooted. This changes the marginal and joint distributions of all six estimated optimal-portfolio characteristics in Section 3 and would produce incorrect QQ-plots and confidence statements if used directly.","section":"Section 3, Theorem 3.2"},{"comment":"The asymptotic coefficient in Theorem 4.1(i) is incorrect. From Theorem 2.1(i), (n-1) hat V_GMV / V_GMV is exactly chi-squared with n-p degrees of freedom, so Var(sqrt(n-p) hat V_GMV) = 2 V_GMV^2 (n-p)^2/(n-1)^2 -> 2(1-c)^2 V_GMV^2. The correct limit is sqrt(n-p)(hat V_GMV - (1-p/n)/(1-1/n) V_GMV) -> sqrt(2)(1-c) V_GMV u1. The printed expression sqrt(2(1-c)) V_GMV u1 has variance 2(1-c) V_GMV^2, a factor 1/(1-c) too large. This is not a typo isolated to one line: the same wrong factor appears in the proof of Theorem 4.1(i) and in the proof of Theorem 4.2 when the u1 term is combined, while Corollary 4.1's (1,1) entry 2 V_GMV^2(1-c)^2 and Theorem 4.4(b)'s 2 V_GMV^2 term both use the corrected coefficient. The printed Theorem 4.1(i) is therefore internally inconsistent with the paper's own exact finite-sample law and with its later covariance formulas. At c=0.9 this overstates the asymptotic variance of hat V_GMV by a factor of 10, which would badly miscalibrate confidence intervals for V_GMV and for any portfolio whose asymptotic distribution enters through g2.","section":"Theorem 4.1(i) and proof of Theorem 4.1"},{"comment":"The displayed variance Xi_{s,s} = 2(c+2s)/(1-c) + 2(s+c)^2/(1-c)^2 does not follow from Theorem 4.1(iv). Since u2, u3, u7 are independent, Theorem 4.1(iv) gives Xi_{s,s} = [2(c+2q)^2 + 4(s-q)]/(1-c) + 2(s+c)^2/(1-c)^2, where q = mu^T A mu and using eta^T (LQL^T)^-1 eta = (s-q)/s^2. The printed expression agrees with this only for a nongeneric value of q, and Assumption (A1) does not restrict q. Consequently the claimed 'direct application' in Corollary 4.1 is not correct as it stands, and the same omission propagates into the Xi_{RVs} matrix used by Theorem 4.3 and hence into the confidence regions of Section 4.3. The authors should re-derive the full covariance matrix from Theorem 4.1 and correct all entries involving mu^T A mu.","section":"Corollary 4.1, entry (4,4)"},{"comment":"The limiting covariance matrix of L hat w_g is not generally of the form a LQL^T + b eta eta^T for k>1. The proof's own expansion contains the independent u3 contribution with coefficient matrix M3 = sqrt(1-c)/(s+c)(g(lam) LQL^T + 2s^2 (g3(lam)/(1-c) - g(lam)/(s+c)) eta eta^T)(LQL^T)^-1/2. The cross term in M3 M3^T is proportional to (LQL^T)^{1/2} eta eta^T (LQL^T)^-1/2 + (LQL^T)^-1/2 eta eta^T (LQL^T)^{1/2}, which for k>1 is not in the linear span of LQL^T and eta eta^T in general; for example, with D = diag(2,1) and eta = (1,1), the cross term cannot be written as aD + b eta eta^T. Since u3 is independent of the other normal variables, this term cannot cancel. Thus Eq. (4.7) is not the correct covariance matrix for general k, and the confidence region in (4.19) is invalid in that setting. The k=1 simulation study would not reveal the problem because every scalar covariance can be represented in the printed form.","section":"Theorem 4.2, Eq. (4.7)"}],"minor_comments":[{"comment":"There are numerous typos: 'quantites' should be 'quantities', 'On of the issues' should be 'One of the issues', 'pure performance' should be 'poor performance', 'cannot be longer used' should be 'can no longer be used', 'the later property' should be 'the latter property', and 'consequent paper' should be 'subsequent paper'.","section":"Throughout"},{"comment":"The proof uses the notation xi2 for the chi-squared variable driving hat V_GMV, but in Theorem 2.1 xi1 is the chi-squared variable corresponding to hat V_GMV and xi2 corresponds to hat s. Please rename to avoid confusion.","section":"Proof of Theorem 4.1(i)"},{"comment":"The sentence describing the covariance between hat theta and hat R_GMV says it is 'partly determined by the estimated self-financing portfolio hat eta du to the deterministic expression close to u5'; this is hard to parse and should be rewritten.","section":"Section 4.1, text after Theorem 4.1"},{"comment":"The discussion says the asymptotic approximation works well for the t(10) scenario, but for hat s and hat V_GMV at c=0.9 the QQ-plots show visible deviations and positive bias. Please state more carefully which quantities are 'well approximated' and which are only approximately so, and note that the bias corrections are deferred to future work.","section":"Section 5"},{"comment":"Reference [16] is cited as an arXiv preprint; if a published version exists, it should be updated. Also, the reference to 'Bodnar et al. [16, Lemma 5.3]' in Section 4.3 should give the lemma's exact statement for the consistency of (1-p/n) l^T hat Sigma^{-1} l.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The central idea of the paper is strong and the exact finite-sample representation in Theorem 2.1 appears sound, but the number of algebra errors in the displayed distributional results is concerning: a missing square root in Theorem 3.2, a wrong variance factor in Theorem 4.1(i), an omitted mu^T A mu term in Corollary 4.1, and a structural problem with Theorem 4.2's covariance-matrix form for general k. These are not merely editorial: they affect the confidence intervals and tests that are a main deliverable. I recommend that the authors re-derive the asymptotic covariance matrices from the stochastic representations, verify the k>1 case numerically, and correct all downstream formulas before the paper is reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful part of this paper is the exact finite-sample stochastic representation of the five frontier quantities and the clean way it generates sampling distributions for GMV, mean-variance, tangency, Sharpe-ratio, VaR/CVaR and other optimal portfolios. That part is new and largely sound. The high-dimensional asymptotics are also a real step beyond earlier fixed-dimension results, and the consistent estimators in Section 4.3 are a practical addition. I give credit for the simulation study being honest about the normality assumption rather than hiding the t(10) biases.\n\nThe stress-test note is right, and it is not a minor cosmetic thing. From Theorem 2.1(i), (n-1)hat(V)_GMV/V_GMV is exactly chi-square with n-p degrees of freedom. That pins the asymptotic coefficient: sqrt(n-p)(hat(V)_GMV - (1-p/n)/(1-1/n)V_GMV) converges to sqrt(2)(1-c)V_GMV u1, not sqrt(2(1-c))V_GMV u1 as printed in Theorem 4.1(i). The paper itself contradicts the printed coefficient: Corollary 4.1 and Theorem 4.4(b) use the corrected factor, and the covariance formula in Theorem 4.2 is consistent with the corrected factor. So the statement of Theorem 4.1(i) and the proof line that repeats the wrong coefficient are typos, but they are load-bearing typos. At c=0.9, the printed theorem overstates the asymptotic variance of hat(V)_GMV by a factor of 10, and any portfolio whose asymptotic distribution enters through g2(lambda) inherits the error. A referee needs to require this fix before publication.\n\nOther soft spots are proportionately smaller. The exact results rely on multivariate normality; the paper is upfront about that, but the title's general promise is broader than what the assumptions deliver. The high-dimensional statements rely on Assumption (A1), which is reasonable but should be stated more carefully in the abstract and introduction. The simulation evidence is QQ-plots without coverage tables or code, so it is suggestive rather than conclusive. Notation is reused in ways that make the long covariance formulas harder to check than they need to be.\n\nWho is this for? People working on estimation risk in portfolio choice, especially those who need exact sampling distributions or high-dimensional confidence regions for optimal portfolios. The exact representation alone justifies the paper. I would send this to peer review with a specific request to re-derive every constant in Theorem 4.1 and trace it through Theorems 4.2-4.4. The core machinery is worth the referee time.","headline":"Sharp exact stochastic representations for optimal portfolios, but the high-dimensional CLT for the GMV variance has a normalization error that must be fixed before anyone builds confidence intervals on it.","tokens_in":28220,"tokens_out":4569,"would_cite":true,"duration_ms":43609,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H10","62H12","62E20","91G10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that the exact finite-sample joint distribution of estimated optimal portfolio weights and characteristics is captured by a stochastic representation built from five efficient-frontier parameters, and that in high…","keywords":["optimal portfolio weights","sampling distribution","stochastic representation","efficient frontier","high-dimensional asymptotics","parameter uncertainty","global minimum variance portfolio","confidence regions"],"falsifier":"Simulate returns from a multivariate t distribution with five degrees of freedom, set $n=1000$ and $p/n=0.9$, construct the paper's nominal 95% confidence sets for a non-GMV portfolio such as the expected-utility portfolio, and measure empirical coverage; if coverage falls substantially below 95% while Gaussian-return simulations match, the normality assumption is the load-bearing condition.","tokens_in":27112,"feed_emoji":"📊","tokens_out":4745,"duration_ms":49971,"temperature":0.7,"pith_summary":"This paper establishes that the sampling distributions of essentially all classical optimal portfolio estimators can be obtained from one joint stochastic representation of five estimated efficient-frontier quantities. Under independent multivariate normal returns with more observations than assets, the plug-in estimators of the global minimum-variance portfolio variance, mean, slope parameter, and related portfolio weights are represented using independent chi-square, normal, and t variables. The same representation is then used to show that, when dimension and sample size grow together, the estimated optimal portfolio weights converge to a multivariate normal distribution whose covariance matrix has an explicit closed form. The practical payoff is exact and fast simulation of estimation risk, plus new confidence regions and tests for portfolio weights and characteristics.","feed_headline":"One stochastic representation captures portfolio-weight estimation risk","feed_subtitle":"Exact and asymptotic distributions for GMV, tangency, and VaR-optimal portfolios follow from five frontier parameters.","key_machinery":"The load-bearing mechanism is the efficient-frontier parametrization: every Markowitz-type optimal portfolio is a linear combination of the global minimum-variance portfolio $w_{\\mathrm{GMV}} = \\Sigma^{-1}\\mathbf{1}/(\\mathbf{1}^\\top\\Sigma^{-1}\\mathbf{1})$ and the self-financing portfolio $v = Q\\mu/(\\mu^\\top Q\\mu)$, where $Q = \\Sigma^{-1} - \\Sigma^{-1}\\mathbf{1}\\mathbf{1}^\\top\\Sigma^{-1}/(\\mathbf{1}^\\top\\Sigma^{-1}\\mathbf{1})$, with the scalar function $g$ selecting the specific portfolio. The paper derives the joint stochastic representation of the plug-in estimators of the five determining quantities $V_{\\mathrm{GMV}}$, $R_{\\mathrm{GMV}}$, $s$, $\\theta = Lw_{\\mathrm{GMV}}$, and $\\eta = Lv$. This representation separates deterministic population matrices from independent standard normal, chi-square, and t components, so samples from the finite-sample distribution can be drawn without recomputing the inverse sample covariance matrix in each simulation, and the high-dimensional limit follows by applying standard central limit results to the independent components.","core_discovery":"The central claim is Theorem 2.1: for $n>p$ independent $p$-dimensional normal returns, the joint sampling distribution of $(\\hat{V}_{\\mathrm{GMV}}, \\hat{\\theta}, \\hat{R}_{\\mathrm{GMV}}, \\hat{s}, \\hat{\\eta})$ is exactly given by a stochastic representation in which all randomness comes from independent chi-square, normal, and t random variables with scale factors that depend only on the population quantities $V_{\\mathrm{GMV}}$, $R_{\\mathrm{GMV}}$, $s$, $\\theta$, $\\eta$, and $LQL^\\top$. The paper then shows that every optimal portfolio of the form $w_g = w_{\\mathrm{GMV}} + g(R_{\\mathrm{GMV}}, V_{\\mathrm{GMV}}, s)v$, including the mean-variance, expected-utility, tangency, Sharpe-ratio, minimum-VaR, and minimum-CVaR portfolios, inherits a complete exact sampling distribution through this representation. Under the high-dimensional regime $p/n \\to c \\in (0,1)$, the scaled estimated weights converge to a multivariate normal distribution with covariance matrix $\\Omega_{L,g}$ given in Theorem 4.2, and the paper provides consistent estimators and confidence sets built from these limits.","pith_inferences":["The paper's own robustness simulations show that heavy-tailed $t$-distributed returns mainly distort $\\hat{s}$ and $\\hat{V}_{\\mathrm{GMV}}$; an extension to elliptical distributions might replace the normal components by t components in the stochastic representation, but this is not claimed in the paper.","Because the stochastic representation uses only a small number of independent variables regardless of dimension $p$, it could accelerate Bayesian posterior sampling for portfolio weights in large asset universes, although the paper does not develop that connection.","For concentration ratios close to 1, the normal approximation for $\\hat{s}$ is visibly weaker in the simulations, suggesting that a finite-sample correction or a different limit regime for $p/n \\to 1$ could be a productive next step.","The consistent estimator of $LQL^\\top$ via $(1-p/n)L\\hat{Q}L^\\top$ and the resulting estimated covariance matrices open a direct route to power analysis of portfolio-weight tests, which the paper does not pursue."],"forward_implications":["For any portfolio in the paper's class, exact finite-sample draws of estimated weights and characteristics can be generated from low-dimensional standard distributions, making estimation-risk assessment substantially cheaper than direct Wishart simulation.","The plug-in estimator of a general optimal portfolio is inconsistent in high dimensions, but the paper's bias-corrected estimators in (4.10)-(4.14) are consistent and can serve as the basis for inference.","Under $p/n \\to c$, the frontier parameter estimators are asymptotically normally distributed with a block-diagonal covariance structure, implying asymptotic independence between certain components such as $\\hat{V}_{\\mathrm{GMV}}$ and the slope estimator $\\hat{s}$.","The explicit covariance formulas support chi-square-based simultaneous confidence regions and tests on linear combinations of optimal portfolio weights, including a closed-form covariance for the expected-utility portfolio.","The consistent estimators of the asymptotic covariance matrices allow construction of confidence intervals for optimal portfolio characteristics such as expected return, variance, VaR, and CVaR in the high-dimensional setting."],"supporting_citations":[{"why":"Supplies the Wishart and inverse Wishart distribution results for the sample covariance matrix that underlie the conditional distribution derivations.","marker":"[48]"},{"why":"Provides the partitioned Wishart and inverse Wishart properties used to obtain independence and conditional t-distributions in Lemma 7.1.","marker":"[15]"},{"why":"Gives the quadratic-form independence and distribution results used to derive the chi-square stochastic representation for $\\hat{s}$ and related quantities.","marker":"[46]"},{"why":"Establishes the independence of the sample mean vector and sample covariance matrix under normality, a key step in the conditional argument.","marker":"[51]"},{"why":"Earlier work on sampling error in mean-variance efficient portfolio weights that the paper extends to a unified representation for a large class of portfolios.","marker":"[22]"},{"why":"Prior derivation of the distribution of the sample minimum-variance frontier, which the present stochastic representation generalizes and reproves.","marker":"[41]"},{"why":"High-dimensional estimation of the global minimum variance portfolio, providing background and baselines for the large-dimensional asymptotic results.","marker":"[17]"}],"fun_headline_variants":["Exact sampling distributions for all optimal portfolio weights","Stochastic representation yields exact and asymptotic weight laws","High-dim limit and exact finite-sample laws for portfolio weights","One representation: exact and asymptotic portfolio-weight distributions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation is exact only if asset returns are independent multivariate normal with $n>p$, because that assumption makes the sample mean and sample covariance independent and permits the Wishart and inverse Wishart calculations.","fun_headline_variants_meta":{"raw":{"variants":["Exact sampling distributions for all optimal portfolio weights","Stochastic representation yields exact and asymptotic weight laws","High-dim limit and exact finite-sample laws for portfolio weights","One representation: exact and asymptotic portfolio-weight distributions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000317,"raw_usage":{"total_tokens":1841,"prompt_tokens":1040,"completion_tokens":801,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":740}},"tokens_in":656,"tokens_out":801,"duration_ms":8449,"temperature":1.0,"reasoning_tokens":740,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:47:24.146164+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate returns from a multivariate t distribution with five degrees of freedom, set $n=1000$ and $p/n=0.9$, construct the paper's nominal 95% confidence sets for a non-GMV portfolio such as the expected-utility portfolio, and measure empirical coverage; if coverage falls substantially below 95% while Gaussian-return simulations match, the normality assumption is the load-bearing condition.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Wishart and inverse Wishart distribution results for the sample covariance matrix that underlie the conditional distribution derivations."},{"cited_title":"and Okhrin, Y","cited_arxiv_id":null,"evidence_quote":"Provides the partitioned Wishart and inverse Wishart properties used to obtain independence and conditional t-distributions in Lemma 7.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the quadratic-form independence and distribution results used to derive the chi-square stochastic representation for $\\hat{s}$ and related quantities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the independence of the sample mean vector and sample covariance matrix under normality, a key step in the conditional argument."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Earlier work on sampling error in mean-variance efficient portfolio weights that the paper extends to a unified representation for a large class of portfolios."},{"cited_title":"and Smith, D","cited_arxiv_id":null,"evidence_quote":"Prior derivation of the distribution of the sample minimum-variance frontier, which the present stochastic representation generalizes and reproves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"High-dimensional estimation of the global minimum variance portfolio, providing background and baselines for the large-dimensional asymptotic results."}],"review_version":1}