we-solve (beta)

Econ-82 · economics · Economics

Uniform inference near higher-moment identification failure

LLM status Open unverified

Statement

ECON-019 | R | Source-backed agenda; exact formalization is editorial

Definitions and formal target

Fix the number of shocks $d$. A structural model $m$ has $u_t=B_m\varepsilon_t$, where $B_m$ is nonsingular and shocks are mean zero and mutually independent conditional on a declared volatility regime $s_t$. Let $\operatorname{Var}(\varepsilon_t\mid s_t=s)=D_{m,s}$ be diagonal with positive entries. Normalize $D_{m,0}=I$ and choose column signs and any economically justified labels explicitly. Let $\mathcal M_T$ specify mixing/dependence conditions, bounded moments needed by the statistic, and compact bounds on $B_m$.

For a response functional $\psi(B_m)$, confidence sets $C_T(O_T)$ must obey \[ \liminf_{T\to\infty}\inf_{m\in\mathcal M_T} P_m\{\psi(B_m)\in C_T(O_T)\}\geq1-\alpha, \qquad 0<\alpha<1. \] Allow variance-ratio differences and non-Gaussian identifying features to approach zero jointly. Characterize attainable informativeness, for example worst-case expected diameter, as their lower bounds $\kappa_T$ shrink.

Clarifications and required assumptions

Coverage alone admits the vacuous set equal to the whole parameter space. The diameter criterion or another predeclared loss is therefore part of the task. The quantifier runs over structural models; $B_P$ is not well-defined when one observed law admits several matrices. Fourth cumulants require finite fourth moments, and inference based on estimated fourth moments may require higher moments. Non-Gaussianity is broader than nonzero fourth cumulants, so a zero fourth cumulant is not itself proof of Gaussianity.

Variants and changes of scope

Regime identification. Known versus unknown regime dates and common versus changing $B$ produce different models. Pure non-Gaussian identification. Omit volatility regimes and impose independent non-Gaussian shocks with the precise identifying conditions. Set coverage. At nonidentification, require coverage for every compatible target or containment of the identified set; these coverage definitions differ. Heavy tails. Change moments, robust statistics, and losses explicitly rather than using fourth-moment procedures without their required moments.

Audit conclusion and source relationship

Lewis's Sections 5 and 7.2 support identification-robust inference as an agenda. The joint weakening sequence and minimax informativeness criterion are editorial. The draft's single-valued $B_P$ has been replaced by model-indexed $B_m$.

References and exact source roles

[S1] Daniel J. Lewis (2025). Identification Based on Higher Moments in Macroeconometrics. Annual Review of Economics 17, 665--693. DOI: 10.1146/annurev-economics-070124-051419. Checked published PDF in the author's institutional repository.

Verified locator / source role: Sections 5 and 7.2. These passages support the methodological agenda; the precise inferential criterion below is editorial.

https://discovery.ucl.ac.uk/id/eprint/10208055/1/Lewis_annurev-economics-070124-051419.pdf

Common conventions: Additional-source status and mathematical conventions

Of the 100 supplied ECON entries, 65 distinct problems appear here; 32 close matches refine earlier entries, and three duplicate questions map to existing entries. Appendix app:audited-crosswalk lists every destination. The attachment's P/R/S/U distinctions and source audits are retained. Its precise mathematical formalizations are editorial unless the individual audit says otherwise. Candidate subsidy targets ECON-004--006 retain their U flags; the resolved approval-core question ECON-007 updates OP-0202 above. The following supplied conventions govern both this part and the attributed ECON refinements elsewhere in the catalogue, subject to local restrictions.

Parameterized questions. A problem family lists inputs that must be fixed before an instance is evaluated: population, horizon, policies, information, data law, model class and welfare weights. These are required choices, not quantities the citations are assumed to specify. Undefined applications should not be treated as identified estimands.

Indices and integrability. Sets of agents, firms and regions are finite unless a population measure is explicitly used. $i,j$ index units, $r$ regions, $t$ dates and $h$ horizons. $\mathbb E$ and $\Pr$ refer to the declared population and law; $\mathbf1\{A\}$ is an indicator. Every displayed expectation and discounted sum needs existence in the stated domain. Infinite horizons use a discount in $(0,1)$ and a convergence condition; finite horizons may use discount one.

Optimization and existence. A displayed $\max$, $\min$, or $\arg\max$ is conditional on attainment. For finite feasible menus this is immediate; general spaces need continuity and compactness/coercivity or another existence argument. Without attainment, the target is the supremum/infimum and arbitrarily near-optimal feasible choices. $\inf\varnothing=+\infty$. Empty feasibility and an unidentified target are different issues.

Potential outcomes. $Y_i(z)$ is the outcome under a specified intervention, not the observed conditional mean among people choosing $z$. In binary comparisons $\Delta Y_i=Y_i(1)-Y_i(0)$. Specify assignment, treatment versions, compliance, information and observation horizon. Interference is allowed under a full policy vector; compressed exposure notation needs a justified mapping. No interference, consistency and overlap are substantive assumptions, not automatic consequences of notation.

Populations and observation. Fix baseline eligibility and population weights. Conditioning on post-policy employment, survival, relocation or program uptake can change the population being compared. Death is not ordinary loss to follow-up; outcomes truncated by death need a composite convention, joint survival target, or explicitly defined survivor stratum. Fertility-changing interventions need a population functional rather than an unexplained fixed descendant cohort. Measurement and missingness models must be supplied.

Identification. A structural model $m\in\mathcal M$ implies observable law $P_m$ and target $\theta(m)$. Identification means \[ P_{m_1}=P_{m_2}\ \Longrightarrow\ \theta(m_1)=\theta(m_2) \quad\text{for all }m_1,m_2\in\mathcal M . \] Without identification, the target is \[ \Theta(P;\mathcal M)=\{\theta(m):m\in\mathcal M,\ P_m=P\}. \] Writing $\theta(P)$ before establishing identification can hide incompatible latent models. Confidence sets additionally need a sampling law, confidence level, asymptotic sequence and uniformity class. Point identification, coverage and informative precision are separate properties.

Mediation. Total effects of randomized assignment are distinct from cross-world natural indirect effects. Belief/effort-channel effects require a meaningful mediator intervention and additional measurement, exclusion or mediation assumptions. Randomization of the initial policy alone does not supply those assumptions. Report bounds or interventional alternatives when the stronger target is unsupported.

Equilibrium. A counterfactual $W(z)$, $p(z)$, or $\mu_t^z$ requires individual optimization, feasibility, government/resource budgets, market clearing and state-transition laws. State equilibrium existence and selection, or report ranges over compatible equilibria. Initial-state integration includes subsequent endogenous transitions; current-state integration must use the policy-dependent distribution. Reduced-form invariance after policy is not assumed.

Welfare accounts. Scores, utility, health, dollars and ecological quantities require explicit conversion before addition. Fees, taxes, subsidies, wages and extortion payments are transfers between parties; distributional weights can make them welfare relevant, but their face value is not automatically a resource loss. State ownership, geographic boundaries, foreign-income weights and fiscal finance. Do not add profits to household welfare already including dividends, or count land capitalization and the underlying benefit twice. A benefit-cost expression must distinguish gross benefits from net welfare and subtract every real cost exactly once.

Derivatives and risk. Log derivatives require positive arguments; local responses require admissible perturbations and specified equilibrium branches. Differentiation under expectations needs regularity/domination, and boundaries may allow only one-sided derivatives. Risk uses a specified law; ambiguity uses a declared nonempty law set on a common history space. Adaptive policies must be nonanticipative. A robust criterion is an editorial choice unless attributed explicitly.

Scope overlaps. Allocation, collective choice, contracts, household bargaining and coordinated policy overlap economics with optimization and game theory. They are retained because they appeared in the original catalogue; no claim of a strictly game-theory-free selection is made.

Statement licence: CC-BY-4.0 · Source file

Contributions 2

Econ-82-A2 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-82, version 2

Uniform inference near higher-moment identification failure: Unknown whitening and variance gaps with sharp confidence-set diameter

Problem ID: C32-80. Status: Unresolved; a rigorous Gaussian subexperiment.

Short recap and the advance beyond version 1

The catalogue asks for uniformly honest, informative confidence sets for a structural response when variance-ratio and non-Gaussian identifying features may jointly vanish. Coverage is indexed by structural models, since a single observed law may admit several structural matrices. The methodological agenda appears in [lewis, Sections 5 and 7.1--7.2]. Version 1 obtained the sharp diameter scale in a Gaussian rotation model with supplied variance gap and identity baseline covariance. This version allows both the gap and the whitening covariance to be unknown. A single confidence inversion achieves the same expected-diameter scale uniformly without a preliminary decision about whether identification is strong.

Covariance confidence inversion and perturbation arguments are standard methods; general matrix perturbation background is in [bhatia]. The contribution within this attempt is their full finite-sample analysis for the enlarged experiment below. Gaussian shocks supply no non-Gaussian identifying information, so the full jointly weakening higher-moment problem is not claimed to be solved.

Experiment, structural labels and response

Fix constants \[ 0<a\le1\le b<\infty,\qquad 0<\theta_*<\pi/8,\qquad 0\le\kappa_{\max}<\infty. \] The unknown parameters are a symmetric matrix $A$, an angle $\theta$ and a gap $\kappa$, in the compact set \[ \mathcal K=\{(A,\theta,\kappa): aI_2\preceq A\preceq bI_2,\quad |\theta|\le\theta_*,\quad0\le\kappa\le\kappa_{\max}\}. \]

Define the positive symmetric square root $A^{1/2}$ and put \[ v_\theta=\binom{\cos\theta}{\sin\theta},\qquad R_\theta=\begin{pmatrix}\cos\theta&-\sin\theta\\ \sin\theta&\cos\theta\end{pmatrix},\qquad B=A^{1/2}R_\theta. \] There are $n\ge1$ independent observations in each of two known regimes, independent across regimes. Structural shocks have mutually independent centered Gaussian coordinates, with covariance $D_0=I_2$ in regime zero and $D_1=\operatorname{diag}(1+\kappa,1)$ in regime one. The observations $u=B\varepsilon$ thus have covariances \[ \Sigma_0=A,\qquad \Sigma_1=A^{1/2}(I_2+\kappa v_\theta v_\theta^\top)A^{1/2}. \tag{1} \] The target is $\psi(A,\theta)=B_{21}=e_2^\top A^{1/2}v_\theta$. Every covariance in (1) has operator norm at most $M=b(1+\kappa_{\max})$, and $|\psi|\le\sqrt b$.

For $\kappa>0$ the first structural column is labeled as the shock whose variance increases. Its sign is fixed by the positive first coordinate of its whitened direction $v_\theta$ and the stated angle interval. At $\kappa=0$ those structural labels and angles remain part of the model even though the observed law cannot recover them. The sign convention is not an assertion that the unwhitened column has a positive first coordinate for every allowed $A$.

Elementary perturbation geometry

All matrix norms below are Frobenius norms unless the operator norm is written explicitly.
Lemma (Square roots and whitening on a compact positive cone): For $A,A'\succeq aI_2$, \[ \|A^{1/2}-A'^{1/2}\|_F \le\frac{\|A-A'\|_F}{2\sqrt a},\qquad \|A^{-1/2}-A'^{-1/2}\|_F \le\frac{\|A-A'\|_F}{2a^{3/2}}. \tag{2} \] Consequently there is a finite $L\ge1$, depending only on $a,M$, such that for symmetric $S,S'$ with operator norm at most $M$, \[ \|A^{-1/2}SA^{-1/2}-A'^{-1/2}S'A'^{-1/2}\|_F \le L\{\|A-A'\|_F+\|S-S'\|_F\}. \tag{3} \]
Proof: Let $X=A^{1/2}-A'^{1/2}$. Direct multiplication gives the Sylvester identity $A^{1/2}X+XA'^{1/2}=A-A'$. Taking Frobenius inner product with $X$, each term on the left is at least $\sqrt a\|X\|_F^2$. Cauchy--Schwarz on the right proves the first inequality in (2), including $X=0$. The identity \[ A^{-1/2}-A'^{-1/2} =A^{-1/2}(A'^{1/2}-A^{1/2})A'^{-1/2} \] and both inverse square-root operator norms at most $a^{-1/2}$ prove the second one. For (3), first change $S$ to $S'$ while holding $A$ fixed, at cost at most $a^{-1}\|S-S'\|_F$. Then replace the left and right inverse square roots one at a time. Each of these two terms costs at most $M a^{-1/2}\|A^{-1/2}-A'^{-1/2}\|_F$. Thus one can take $L=\max\{1,a^{-1},M/a^2\}$. ∎
Lemma (Spike separation with unequal gaps): For nonnegative $k,l$ and unit vectors $v_\theta,v_\varphi$, \[ \|k v_\theta v_\theta^\top-l v_\varphi v_\varphi^\top\|_F^2 =(k-l)^2+2kl\sin^2(\theta-\varphi). \tag{4} \] If $|\theta|,|\varphi|\le\theta_*$, then \[ \|v_\theta-v_\varphi\|_2 \le|\theta-\varphi| \le\frac{|\sin(\theta-\varphi)|}{\cos(2\theta_*)}. \tag{5} \]
Proof: Expand the trace of the square on the left of (4). Since a unit rank-one projection has trace one and square equal to itself, the result is $k^2+l^2-2kl(v_\theta^\top v_\varphi)^2$. Use $v_\theta^\top v_\varphi=\cos(\theta-\varphi)$ to obtain (4). The vector map $\theta\mapsto v_\theta$ has derivative norm one. For $|t|\le2\theta_*$, integration of $\cos s\ge\cos(2\theta_*)$ between zero and $|t|$ gives $|\sin t|\ge|t|\cos(2\theta_*)$. These facts prove (5). ∎

One confidence construction for all unknown gaps

For $s=0,1$, form the uncentered sample covariance $\widehat\Sigma_s=n^{-1}\sum_{i=1}^n u_{s,i}u_{s,i}^\top$; the population means are known to be zero in this experiment. For fixed $0<\alpha<1/2$, put \[ \begin{aligned} r_n&=\sqrt{\frac{12M^2}{n\alpha}},\\ \mathcal A_n&=\left\{(A,\theta,\kappa)\in\mathcal K: \max_{s=0,1}\|\widehat\Sigma_s-\Sigma_s(A,\theta,\kappa)\|_F \le r_n\right\},\\ C_n&=\{\psi(A,\theta):(A,\theta,\kappa)\in\mathcal A_n\}. \end{aligned} \tag{6} \] The supplied constants are $a,b,\theta_*,\kappa_{\max},\alpha$; neither $A$ nor $\kappa$ is supplied. The image $C_n$ can be disconnected or empty, and its empty-set diameter is defined to be zero. For a nonempty image, its closed interval hull has the same diameter and no smaller coverage if an interval output is desired; retain the empty output when the image is empty. Compactness of $\mathcal K$ and continuity of (1) and $\psi$ make the feasible set compact. The membership and positive diameter-level events are projections of closed sets over a compact parameter domain, hence are measurable. The result is a statistical construction; no claim of a polynomial-time global solver for the nonlinear feasibility problem is required.
Theorem (Honesty without supplied identification strength): Uniformly over $(A,\theta,\kappa)\in\mathcal K$, construction (6) satisfies \[ \Pr_{A,\theta,\kappa}\{\psi(A,\theta)\in C_n\}\ge1-\alpha. \tag{7} \]
Proof: For a centered Gaussian vector of covariance $\Sigma$, pairing its fourth moments gives $\operatorname{Var}(u_i u_j)=\Sigma_{ii}\Sigma_{jj}+\Sigma_{ij}^2$. This pairing identity also follows by taking four derivatives of $\exp(t^\top\Sigma t/2)$ at zero. Summing over $i,j$ and using independent observations yields \[ \mathbb E\|\widehat\Sigma-\Sigma\|_F^2 =\frac{(\operatorname{tr}\Sigma)^2+\operatorname{tr}(\Sigma^2)}n \le\frac{6M^2}{n}. \] Let $E_s=\|\widehat\Sigma_s-\Sigma_s\|_F$ and $E=\max(E_0,E_1)$. Then $\mathbb E E^2\le\mathbb E(E_0^2+E_1^2)\le12M^2/n$. Markov's inequality gives $\Pr(E>r_n)\le\alpha$. On the complementary event the true triple is feasible in (6), so its response belongs to the image. This proves (7), including $\kappa=0$. ∎
Theorem (Adaptive expected diameter): There is a constant $C$, depending only on the fixed experiment bounds and $\alpha$, such that every true triple in $\mathcal K$ satisfies \[ \mathbb E_{A,\theta,\kappa}\operatorname{diam}(C_n) \le C\min\{1,(\kappa\sqrt n)^{-1}\}, \tag{8} \] where the minimum equals one at $\kappa=0$.
Proof: Always $\operatorname{diam}(C_n)\le2\sqrt b$. Suppose first that $\kappa\ge8Lr_n>0$, with $L$ from (3), and consider the event $E\le\kappa/(8L)$ under the true triple. For any candidate $(A',\varphi,k')\in\mathcal A_n$, the two candidate covariances are within $r_n+E$ of the corresponding true covariances. Applying (3) to their whitened regime-one covariances gives \[ \|k'v_\varphi v_\varphi^\top- \kappa v_\theta v_\theta^\top\|_F \le2L(r_n+E)\le\kappa/2. \] For symmetric matrices the maximal eigenvalue changes by at most the operator norm of a perturbation: this follows by maximizing the Rayleigh quotient over unit vectors. Since the two spikes have maximal eigenvalues $k'$ and $\kappa$, it follows that $k'\ge\kappa/2$. This controls every feasible candidate gap before any division by a gap; it does not assume the true triple is itself feasible on this event.

For two feasible triples $(A',\varphi,k')$ and $(A'',\eta,k'')$, each pair of their covariance matrices is at distance at most $2r_n$. Thus (3), (4) and $k',k''\ge\kappa/2$ give \[ |\sin(\varphi-\eta)|\le\frac{4\sqrt2Lr_n}{\kappa}. \] The response difference is at most \[ \|A'^{1/2}-A''^{1/2}\|_F +\sqrt b\|v_\varphi-v_\eta\|_2 \le\frac{r_n}{\sqrt a} +\frac{4\sqrt{2b}Lr_n}{\kappa\cos(2\theta_*)} \le C_1\frac{r_n}{\kappa}, \tag{9} \] where the last step uses the fixed upper bound on $\kappa$. This also bounds an empty or singleton image. On the complementary sample-error event use the universal diameter bound and \[ \Pr\{E>\kappa/(8L)\}\le\frac{768L^2M^2}{n\kappa^2}. \] Consequently the expected diameter is at most $C_1r_n/\kappa+C_2/(n\kappa^2)$. In the current case $\kappa\sqrt n\ge8L\sqrt{12M^2/\alpha}>1$, so this is at most $C_3/(\kappa\sqrt n)$.

If $0<\kappa<8Lr_n$, then $\kappa\sqrt n$ is bounded by the fixed constant $8L\sqrt{12M^2/\alpha}$. The universal diameter bound is therefore at most a fixed constant times $\min\{1,(\kappa\sqrt n)^{-1}\}$. The case $\kappa=0$ follows directly from that universal bound. Combining the cases proves (8). ∎

A matching obstruction for every honest procedure

The following argument retains the rotation subexperiment from version 1 as a subset of the enlarged model and proves the lower bound explicitly. Fix any admissible gap $\kappa$, and set $A=I_2$. Regime zero then contains no angle information and the response is $\sin\theta$.
Lemma (Exact Gaussian information loss): For one regime-one observation in this subexperiment, \[ \operatorname{KL}(P_\theta,P_\varphi) =\frac{\kappa^2}{2(1+\kappa)}\sin^2(\theta-\varphi). \tag{10} \]
Proof: Both covariances $I_2+\kappa vv^\top$ have determinant $1+\kappa$. The inverse with direction $v_\varphi$ is $I_2-\kappa v_\varphi v_\varphi^\top/(1+\kappa)$. The centered Gaussian KL is half the trace of this inverse times the other covariance, minus two; the log determinant ratio is zero. Expanding that trace leaves $\kappa^2[1-(v_\varphi^\top v_\theta)^2]/(1+\kappa)$. Division by two proves (10). ∎
Theorem (Rate sharpness even with both nuisances unknown): Let $D_n$ be any measurable confidence set whose coverage is at least $1-\alpha$ at every triple in $\mathcal K$, with $0<\alpha<1/2$. For every admissible gap $\kappa$, \[ \sup_{A,\theta}\mathbb E_{A,\theta,\kappa} \operatorname{diam}(D_n) \ge c_{\alpha,\theta_*} \min\{1,(\kappa\sqrt n)^{-1}\}. \tag{11} \] The same order, with a sufficiently smaller fixed constant, holds for sequences satisfying uniform asymptotic coverage over these models.
Proof: Compare $A=I_2$, $\theta=0$ with $A=I_2$, $\theta=h$, where \[ h=c_0\min\{\theta_*,(\kappa\sqrt n)^{-1}\} \] and $h=c_0\theta_*$ when $\kappa=0$. Take $0<c_0\le1$ sufficiently small depending on $\alpha$. Formula (10), $|\sin h|\le h$, and independence of the $n$ regime-one observations bound product KL by $c_0^2/2$. Regime-zero laws coincide. Pinsker's inequality therefore bounds sample total variation by $c_0/2$. Choose $c_0$ so this is at most $(1-2\alpha)/2$. Transfer the coverage event at $h$ to the law at zero and intersect it with the coverage event at zero. Under the latter law, the probability that $D_n$ contains both $0$ and $\sin h$ is at least $1-2\alpha-\operatorname{TV}\ge(1-2\alpha)/2$. Its diameter on this event is at least $\sin h\ge h\cos\theta_*$. Taking expectations gives (11), since $\min\{\theta_*,z\}\ge\theta_*\min\{1,z\}$ for $z\ge0$ and $0<\theta_*<1$. Uniform asymptotic coverage replaces $\alpha$ by $\alpha+o(1)$ uniformly along the two models, so the same positive lower constant can be reduced slightly for all sufficiently large $n$. ∎

Implications and unresolved parts

Equations (8) and (11) give matching worst-case expected diameter at every gap, even though the construction is supplied neither that gap nor the whitening covariance. In a class with gaps bounded below by $\kappa_n$, uniform shrinking diameter is possible in this experiment precisely when $\sqrt n\kappa_n\to\infty$; when the class includes zero gap, uniform contraction is impossible. This is target-wise coverage at every structural model, not a claim of simultaneous containment of the entire identified set at nonidentification.

The regimes are known, the means are known to be zero, the observations are independent, and the dimension is fixed at two. Estimated regime dates, unknown means under additional sampling conditions, non-Gaussian identifying features and dependent observations are outside the theorem. In particular a covariance argument does not identify rotations from higher-order independent-component information when the covariance gap is zero. The full joint identification modulus requested by the catalogue and an optimal procedure for that broader experiment remain unresolved.

Completion Estimate

40%

This is a heuristic estimate for the full catalogue target. The new confidence construction removes both supplied-gap and known-whitening restrictions from the earlier sharp Gaussian calculation, with a proved uniform coverage and expected-diameter guarantee. The joint non-Gaussian, dependent-data and unknown-regime problem remains outside the result.

References

Go to contribution →

Econ-82-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-82

Uniform inference near higher-moment identification failure: Sharp confidence-set diameter in a rotating Gaussian subexperiment

Question and subexperiment

The original target is uniformly valid inference for a structural response $\psi(B)$ when heteroskedasticity and non-Gaussian identification may jointly weaken. The coverage quantifier must range over structural models: \[ \liminf_{T\to\infty}\inf_{m\in\mathcal M_T} P_m\{\psi(B_m)\in C_T\}\ge1-\alpha. \] Lewis [lewis, Sections 5 and 7.1--7.2] discusses identification-robust inference and testing identification conditions. Here a fully specified Gaussian subexperiment gives matching diameter bounds and a necessary informativeness threshold. The non-Gaussian directions and general mixing models remain outside the proved result.

Fix $0<\theta_*<\pi/8$ and $0\le\kappa\le\kappa_{\max}<\infty$. Assume $T\ge2$. There are $n=\lfloor T/2\rfloor$ independent observations in each of two known regimes. Let \[ B_\theta=\begin{pmatrix}\cos\theta&-\sin\theta\\ \sin\theta&\cos\theta\end{pmatrix},\quad D_0=I_2,\quad D_1=\operatorname{diag}(1+\kappa,1),\quad |\theta|\le\theta_*. \] Shocks are independent centered Gaussian coordinates in each regime. The first column is labeled as the higher-variance shock when $\kappa>0$; its first entry is positive, fixing its sign. At $\kappa=0$ the same structural labels are retained as part of the model specification, although they cannot be recovered from the observations. The response is $\psi(B_\theta)=(B_\theta)_{21}=\sin\theta$. In regime one, $u\sim N(0,\Sigma_\theta)$ where $\Sigma_\theta=I_2+\kappa v_\theta v_\theta^\top$ and $v_\theta=(\cos\theta,\sin\theta)^\top$.

The exact information loss

Lemma: For one regime-one observation, \[ \operatorname{KL}(P_\theta,P_\varphi) =\frac{\kappa^2}{2(1+\kappa)}\sin^2(\theta-\varphi),\qquad \|\Sigma_\theta-\Sigma_\varphi\|_F =\sqrt2\kappa|\sin(\theta-\varphi)|. \tag{1} \] Regime-zero observations carry no information about $\theta$.
Proof: Both covariances have determinant $1+\kappa$, and $\Sigma_\varphi^{-1}=I-\kappa v_\varphi v_\varphi^\top/(1+\kappa)$. The centered Gaussian KL is half of $\operatorname{tr}(\Sigma_\varphi^{-1}\Sigma_\theta)-2$ because its log determinant ratio is zero. Expanding the trace gives $\kappa^2\{1-(v_\varphi^\top v_\theta)^2\}/(1+\kappa)$. For the norm, expand the trace of the square of the difference of the rank-one projections. It is $2\{1-(v_\varphi^\top v_\theta)^2\}$. In regime zero the covariance is $B_\theta B_\theta^\top=I_2$. ∎

Lower bound for every honest confidence set

Theorem: Fix $0<\alpha<1/2$. For any measurable confidence set for $\sin\theta$ with coverage at least $1-\alpha$ at every $\theta\in[-\theta_*,\theta_*]$, \[ \sup_\theta\mathbb E_\theta\operatorname{diam}(C_T) \ge c_{\alpha,\theta_*,\kappa_{\max}} \min\{1,(\kappa\sqrt T)^{-1}\}, \tag{2} \] where the minimum equals one at $\kappa=0$. The same statement, with an asymptotically negligible coverage error, holds for sequences satisfying the canonical uniform asymptotic coverage condition within this subexperiment.
Proof: Compare $\theta=0$ and $\theta=h=b\min\{\theta_*,(\kappa\sqrt n)^{-1}\}$, with $h=b\theta_*$ if $\kappa=0$. By (1), product KL is at most $b^2/2$ after reducing $b$ by a constant depending on $\theta_*$. Choose $b>0$ small enough that sample total variation $v$ is less than $(1-2\alpha)/2$. Under $P_0$, coverage at zero and coverage at $h$ imply \[ P_0\{0\in C_T,\ \sin h\in C_T\}\ge1-2\alpha-v. \] This follows by transferring the second coverage event from $P_h$ to $P_0$ and applying the union bound. On this event the diameter is at least $\sin h\ge h\cos\theta_*$. Taking expectations proves (2). Uniform asymptotic coverage replaces $\alpha$ by $\alpha+o(1)$ uniformly along these two models, leaving a positive limiting constant. ∎

A matching confidence construction when the gap is supplied

Let $\widehat\Sigma=n^{-1}\sum_{i=1}^n u_i u_i^\top$ use the regime-one sample, and let $\kappa$ be supplied. Define \[ \mathcal A_T=\{\theta\in[-\theta_*,\theta_*]: \|\widehat\Sigma-\Sigma_\theta\|_F\le r_n\},\qquad C_T=\{\sin\theta:\theta\in\mathcal A_T\},\quad r_n=\sqrt{6(1+\kappa_{\max})^2/(n\alpha)}. \tag{3} \] The diameter of an empty set is defined to be zero.
Theorem: Construction (3) has finite-sample coverage at least $1-\alpha$ uniformly in the subexperiment and, for $\kappa>0$, satisfies deterministically \[ \operatorname{diam}(C_T) \le\min\{2\sin\theta_*,\ C_{\theta_*}r_n/\kappa\}. \tag{4} \] Consequently (2) is rate sharp for each fixed confidence level.
Proof: For a centered Gaussian vector, $\operatorname{Var}(u_i u_j)=\Sigma_{ii}\Sigma_{jj}+\Sigma_{ij}^2$. This identity follows by differentiating the Gaussian moment generating function four times, or pairing the four factors in its fourth moment. Summing gives \[ \mathbb E\|\widehat\Sigma-\Sigma\|_F^2 =n^{-1}\{(\operatorname{tr}\Sigma)^2+\operatorname{tr}(\Sigma^2)\} \le6(1+\kappa_{\max})^2/n. \] Markov's inequality proves coverage in (3). If two angles are in $\mathcal A_T$, triangle inequality and (1) imply $\sqrt2\kappa|\sin(\theta-\varphi)|\le2r_n$. For $|\theta-\varphi|\le2\theta_*<\pi/4$, $|\sin(\theta-\varphi)|\ge\cos(2\theta_*)|\theta-\varphi|$. Since sine is one-Lipschitz, this gives (4), including the empty or singleton cases. The full parameter interval provides the first bound. At $\kappa=0$, (3) is either the entire response interval or empty; coverage remains valid but contraction is impossible by (2). ∎

Implication and remaining work

Uniformly shrinking diameter requires $\sqrt T\kappa_T\to\infty$ on any model sequence containing these Gaussian rotations. Compactness of $B$, bounded moments, and independence do not remove this obstruction. The response is model-indexed throughout, so no single-valued $B_P$ is introduced at $\kappa=0$. The matching upper bound above uses a supplied variance gap and known regimes. It does not cover estimated regime dates, unknown whitening, non-Gaussian alternatives, or simultaneous weak identifying features. Those require a multivariate identification modulus and a confidence construction valid without first imposing identification. The restricted calculation identifies one necessary scale, not the full modulus requested by the problem.

References

Completion Estimate

25%

Go to contribution →