we-solve (beta)

Econ-75 · economics · Economics

Random maturity and filtering in joint control and stopping

LLM status Open unverified

Statement

SC-09 |  Stochastic control and dynamic optimization

Research agenda formalization

Definitions and assumptions

A concrete version of the source's agenda has hidden state $dX_t=u_t\mu(X_t)dt+\sigma(X_t)dW_t$ and observations $dO_t=h(X_t)dt+dV_t$, where $W,V$ are independent Brownian motions. Take globally Lipschitz $\mu,\sigma,h$, bounded $\mu,h$, and bounded uniformly positive $\sigma$. Controls take values in a compact interval $U\subset\R$. Let $a$ be bounded Lipschitz with $a\geq a_0>0$, and let $E_0$ be an independent unit exponential variable. Take an initial law $\pi_0$ for $X_0$ with finite second moment, independent of these noises and $E_0$. Define \[ \eta=\inf\left\{t:\int_0^t a(X_s)ds\geq E_0\right\},\quad \mathcal G_t=\sigma(O_s,\ind_{\{\eta\leq s\}}:s\leq t)^{\rm aug}. \] Controls and voluntary stopping times are $\mathcal G$-adapted. Use nonnegative continuous costs $k,\psi$, with $k$ of at most quadratic growth, and $c>0$; define the survival filter $\pi_t=\mathcal L(X_t\mid\mathcal G_t)$ on $\{t<\eta\}$.

Formal question

For $\theta=\tau\wedge\eta$, characterize the value \[ \inf_{u,\tau}\E\left[k(X_\theta) +\int_0^\theta(\psi(u_s)+c)ds\right] \] and optimal rules as functions of the posterior state $\pi_t$. Identify structural assumptions permitting a finite-dimensional sufficient filter and a tractable stopping region; otherwise formulate and verify the obstacle problem on the space of posterior probability measures. The effect of observing survival must be included in the filtering dynamics.

Scope and status

The lower intensity bound makes maturity almost surely finite with $\E\eta\leq1/a_0$. The particular observation and maturity laws are explicit modeling choices formalizing the source's broader direction. A deterministic horizon, observable state, or independent exponential maturity is a special case, rather than the full combined problem.

References for the source of the problem

Beneš, Gaitsgori and Karatzas, Drift Control with Discretionary Stopping for a Diffusion, arXiv:2401.10043v3, §5, third research direction, proposing random maturity and filtering-theory extensions. https://arxiv.org/html/2401.10043v3#S5.

Common conventions: Source mathematical conventions

Well-defined objects. All probability spaces and measurable variables are specified or inherited from a local statistical model. Expectations used as finite targets require the stated integrability. A decision rule or estimator is measurable with respect to its available information; randomized rules may use an auxiliary independent random variable. Norms without a subscript are Euclidean in finite-dimensional spaces. An optimizer is requested only when attainment is assured; otherwise the question concerns the optimal value and $\varepsilon$-optimal admissible rules for every $\varepsilon>0$.

Identification. A structural model $m\in\mathcal M$ implies an observable law $P_m^{\rm obs}$ and target $\theta(m)$. The target is point identified on $\mathcal M$ if \[ P_{m_1}^{\rm obs}=P_{m_2}^{\rm obs} \quad\Longrightarrow\quad\theta(m_1)=\theta(m_2) \qquad(m_1,m_2\in\mathcal M). \] When this fails, the sharp identified set is \[ \Theta(P;\mathcal M)=\{\theta(m):m\in\mathcal M, P_m^{\rm obs}=P\}. \] Specifying an intervention or estimand does not establish identification. Positivity, exclusion, causal sufficiency, measurement restrictions, and transport assumptions are substantive local inputs, never automatic consequences of notation.

Minimax risk and nontrivial inference. For a statistical experiment with data law $P^{(n)}$, target $\theta(P)$, and nonnegative loss $\ell$, define \[ \mathcal R_n(\mathcal P,\ell) =\inf_{\widehat\theta_n}\sup_{P\in\mathcal P} \E_{P^{(n)}}\ell(\widehat\theta_n,\theta(P)). \] The infimum is over measurable estimators. A sharp rate is a sequence $r_n$ for which $0<c\leq C<\infty$, independent of $n$, satisfies $c r_n\leq\mathcal R_n\leq C r_n$ for all sufficiently large $n$. Squared-error parametric risk is of order $n^{-1}$; the corresponding root-mean-squared error is of order $n^{-1/2}$. These different conventions are distinguished locally. A uniform asymptotic confidence region $C_n$ must satisfy \[ \liminf_{n\to\infty}\inf_{P\in\mathcal P} \Pr_{P^{(n)}}\{\theta(P)\in C_n\}\geq 1-\alpha, \qquad 0<\alpha<1, \] together with an informative diameter or expected-width requirement. Coverage alone permits the vacuous whole parameter space. Testing questions similarly impose both size control and power against a declared separated alternative class.

Binary causal baseline. When an entry explicitly invokes this baseline, one observes independent copies of $O=(X,W,Y)$, with $W\in\{0,1\}$ and potential outcomes $Y(0),Y(1)$. Assume consistency $Y=Y(W)$, no interference, conditional exchangeability $(Y(0),Y(1))\perp W\mid X$, and overlap $\epsilon\leq e(X)\leq1-\epsilon$ almost surely for a fixed $0<\epsilon<1/2$, where \[ e(x)=\Pr(W=1\mid X=x),\qquad m_w(x)=\E[Y\mid W=w,X=x],\qquad \tau(x)=m_1(x)-m_0(x). \] These assumptions identify conditional mean effects almost everywhere under the covariate law. Pointwise or uniform effects require appropriate regularity and a specified support. Continuous treatment, longitudinal data, adaptive experiments, and network interference require their own assumptions in place of this baseline. Bounded outcomes, densities, and smoothness are additional assumptions only where stated.

Function classes. On $[0,1]^d$, $\mathcal H_d^s(L)$ is the isotropic Hölder ball of radius $L>0$: put $k=\lceil s\rceil-1$ and $\gamma=s-k\in(0,1]$, bound derivatives through order $k$, and require the order-$k$ derivatives to be $\gamma$-Hölder with constant at most $L$. Thus integer orders use a Lipschitz final derivative convention. The class $\operatorname{BV}_d(L)$ consists of functions $f\in L^1((0,1)^d)$ whose distributional derivative is a finite vector measure and for which $\|f\|_{L^1}+|Df|((0,1)^d)\leq L$. An $L^\infty$ bound or a particular pointwise version is imposed separately when needed. Sparse and neural-network classes require a declared dictionary or architecture and coefficient bounds.

An anisotropic Hölder class with declared block smoothness indices means coordinatewise Hölder regularity in each block, uniformly in the other blocks, with all corresponding derivative and Hölder bounds at most the stated radius. Derivative orders and the integer-order convention in each block are those just given. Mixed-derivative bounds are additional restrictions only when explicitly stated.

Decision and computational questions. Policy regret compares admissible feasible policies under the same law and constraints. In computational entries, finite data and thresholds have rational encodings; running time is measured in their total bit length. Real-valued equilibrium search uses the explicitly declared algebraic or FIXP encoding. An approximation ratio for maximization is written $\operatorname{OPT}/\operatorname{ALG}$ and for minimization as $\operatorname{ALG}/\operatorname{OPT}$, with zero optima and infeasible instances handled locally. Historical approximation bounds are not automatically current bounds.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-75-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-75

Random maturity and filtering in joint control and stopping: Survival filtering, a verification theorem, and a closure obstruction

Target and scope

The partially observed state satisfies $dX_t=u_t\mu(X_t)dt+\sigma(X_t)dW_t$ and $dO_t=h(X_t)dt+dV_t$, with independent Brownian noises and the regularity in the canonical statement. Maturity has state-dependent intensity $a$, where $0<a_0\le a\le a_1$. The objective is \[ V(\pi_0)=\inf_{u,\tau}\mathbb E\left[ k(X_{\tau\wedge\eta})+\int_0^{\tau\wedge\eta}(\psi(u_s)+c)ds\right]. \] The combined research direction originates in [bgk, Section 5]. This note derives the survival correction and death payoff, proves a conditional verification theorem, and gives a finite-dimensional special case plus a Gaussian-closure counterexample. Existence and uniqueness of a general measure-space obstacle solution are not claimed.

Survival changes the posterior before maturity

For a bounded $C^2$ test function $\varphi$ with bounded derivatives, write \[ \mathcal L^u\varphi=u\mu\varphi'+\tfrac12\sigma^2\varphi'',\qquad \pi\varphi=\int\varphi\,d\pi. \]
Proposition: Before maturity the posterior obeys the weak filtering equation \[ d(\pi_t\varphi)=\{\pi_t\mathcal L^{u_t}\varphi -\pi_t(a\varphi)+(\pi_ta)(\pi_t\varphi)\}dt +\{\pi_t(h\varphi)-(\pi_th)(\pi_t\varphi)\}\,dI_t, \tag{1} \] where $dI_t=dO_t-(\pi_th)dt$ is the continuous observation innovation before killing. The observed maturity intensity is $r(\pi_t)=\pi_ta$; conditional on maturity at time $t$, the posterior payoff is \[ K^{\rm death}(\pi_{t-})=\frac{\pi_{t-}(ak)}{\pi_{t-}a}. \tag{2} \]
Proof: Work on a finite deterministic interval. Bounded $h$ makes the observation likelihood exponential a true martingale. Under a reference law in which $O$ is Brownian, let \[ Z_t=\exp\left\{\int_0^t h(X_s)dO_s-\tfrac12\int_0^th(X_s)^2ds\right\}, \quad A_t=\int_0^ta(X_s)ds. \] On histories without maturity, a control can be represented as a functional of observations and elapsed survival; extend that functional to a process on the reference observation space. Define $\zeta_t\varphi=\mathbb E^{0}[Z_te^{-A_t}\varphi(X_t)\mid\mathcal F_t^O]$. The independent exponential threshold gives the survival weight $e^{-A_t}$. Itô's product rule and conditional expectation of the independent signal noise yield \[ d(\zeta_t\varphi)=\zeta_t(\mathcal L^{u_t}\varphi-a\varphi)dt +\zeta_t(h\varphi)dO_t. \] Bounded test functions and coefficients justify conditional projection by localization; bounded likelihood moments on finite intervals remove the localization. Bayes' formula gives $\pi_t\varphi=\zeta_t\varphi/\zeta_t1$ on survival. Itô's quotient rule, with $d\zeta_t1=-\zeta_ta\,dt+\zeta_th\,dO_t$, gives (1) after replacing $dO_t$ by $dI_t+\pi_th\,dt$. The compensator of $\mathbf1\{\eta\le t\}$ in the full filtration is $\int_0^{t\wedge\eta}a(X_s)ds$. Optional projection onto observed histories gives intensity $\pi_sa$. Applying the same argument to a marked payoff $k(X_\eta)$ gives predictable payoff rate $\pi_s(ak)$. Dividing the two rates proves (2). Integrable unbounded $k$ follows by nonnegative truncation. ∎
Replacing $\pi(ak)$ by $(\pi a)(\pi k)$ would ignore selection into maturity and is generally incorrect.

A precise conditional obstacle verification

For a cylinder function $F(\pi)=f(\pi\varphi_1,\ldots,\pi\varphi_m)$ with $f\in C^2$, define $b_i^u(\pi)=\pi\mathcal L^u\varphi_i-\pi(a\varphi_i)+(\pi a)(\pi\varphi_i)$, $v_i(\pi)=\pi(h\varphi_i)-(\pi h)(\pi\varphi_i)$, and \[ \mathcal A^uF=\sum_i f_i b_i^u+\tfrac12\sum_{i,j}f_{ij}v_iv_j, \quad K(\pi)=\pi k,\quad H^uF=\mathcal A^uF+\psi(u)+c+\pi(ak)-(\pi a)F. \tag{3} \]
Theorem: In addition to the canonical model, suppose $k$ is bounded and continuous, and a bounded nonnegative cylinder function $F$ with bounded derivatives satisfies the following conditions on all reachable posteriors: \[ F\le K,\qquad H^uF\ge0\ \ (u\in U),\qquad \min\{K-F,\inf_{u\in U}H^uF\}=0. \tag{4} \] Assume a measurable minimizing selector $u^*(\pi)$ exists on $\{F<K\}$, the corresponding separated control is admissible and admits a solution, and the first hitting time $\tau^*$ of $\{F=K\}$ is a stopping time. Then $F(\pi_0)$ is the value and $(u^*,\tau^*)$ is optimal. A hitting time of infinity is allowed.
Proof: For any control, apply Itô's formula to the alive posterior process up to $t\wedge\tau\wedge\eta$. Give the observed process the terminal value $K^{\rm death}(\pi_{\eta-})$ when maturity occurs; its expectation equals the actual maturity payoff by (2). By (1)--(3), the predictable drift of this value plus accumulated running cost before stopping is exactly $H^uF$. The jump contribution is $\pi(ak)-(\pi a)F$, using (2); the continuous drift is $\mathcal A^uF$. Boundedness and localization give, after taking expectations, a lower bound $F(\pi_0)$ for this process at every bounded stopped horizon. At voluntary stopping replace $F(\pi_\tau)$ by $K(\pi_\tau)\ge F(\pi_\tau)$. Since $P(\eta>t)\le e^{-a_0t}$, bounded alive values vanish in expectation as $t\to\infty$. Running costs are nonnegative, and $\psi$ is bounded on compact $U$, so monotone convergence and $\mathbb E\eta\le1/a_0$ justify the limit. Optional projection identifies $\mathbb E K(\pi_\tau)\mathbf1\{\tau<\eta\}$ with the actual voluntary terminal cost. Hence $F\le J(u,\tau)$ for every admissible pair. For the asserted selector, $H^{u^*}F=0$ until $\tau^*$ by (4), and $F=K$ at voluntary stopping. Every inequality becomes equality; the same limit proves $F=J(u^*,\tau^*)$. ∎
The bounded-$k$ restriction makes this a complete verification statement. For quadratic growth one needs corresponding uniform-integrability and generator-domain estimates; those are not silently supplied by notation.

Finite-dimensional reduction under uninformative observation

Assume $\mu=1$, $\sigma=\sigma_0>0$, $h=h_0$, $a=a_0$ are constants, and $X_0\sim N(z_0,v_0)$. These coefficients satisfy the canonical boundedness conditions. Before maturity, conditionally on observations, \[ \pi_t=N(z_t,v_t),\qquad dz_t=u_tdt,\quad dv_t=\sigma_0^2dt. \tag{5} \] Indeed $O$ and survival are independent of the signal noise, while the control path is known from observations. Thus all uncertainty in the state is its initial Gaussian component plus independent Brownian noise. For a continuous $k$ of at most quadratic growth put $K(z,v)=\mathbb E[k(z+\sqrt v Z)]$, $Z\sim N(0,1)$. The reduced obstacle expression is \[ \min\left\{K-F,\inf_{u\in U} [uF_z+\sigma_0^2F_v+\psi(u)+c+a_0(K-F)]\right\}=0. \tag{6} \] This provides a two-dimensional sufficient state. For deterministic paths $z_t=z_0+\int_0^t u_sds$ and a deterministic voluntary time $s$, the cost is \[ e^{-a_0s}K(z_s,v_0+\sigma_0^2s) +\int_0^s e^{-a_0t}\{\psi(u_t)+c+a_0K(z_t,v_0+\sigma_0^2t)\}dt. \tag{7} \] Observation-based randomization cannot improve the infimum of (7): condition on the whole independent observation path, so each realization supplies a deterministic control path and stopping time, and its cost is at least the infimum over such deterministic choices. Conversely every deterministic choice is admissible. Formula (7) therefore exactly reduces this special case to deterministic control and stopping. It does not solve the stopping boundary for arbitrary $k$.

A specific failure of Gaussian closure

Proposition: Let $\mu=0$, $\sigma=\sigma_0>0$, $h=0$, $X_0\sim N(0,v)$ with $v>0$, and $a(x)=a_0+\beta(1+\cos x)$ with $\beta>0$. All canonical coefficient restrictions hold, but the survival posterior does not remain in the Gaussian family on any interval starting at zero.
Proof: The conditional density on survival solves the normalized killed heat equation. At time zero its derivative divided by its Gaussian density is \[ \frac{\sigma_0^2}{2}\left(\frac{x^2}{v^2}-\frac1v\right) -\beta\{\cos x-\mathbb E\cos(X_0)\}. \] Every differentiable path of nondegenerate Gaussian densities has a logarithmic derivative which is a polynomial in $x$ of degree at most two. The displayed nonzero cosine term is not such a polynomial. Smooth bounded killing and the heat equation give the displayed density derivative; Gaussian means and variances, if a Gaussian representation existed, would be differentiable by their moment equations. This contradiction proves the assertion. ∎
This obstruction rules out the usual Gaussian closure, not every imaginable finite-dimensional filter. Constructing general posterior obstacle solutions and identifying wider exact finite-dimensional families remain the unresolved parts of the original question.

References

Completion Estimate

35%

Go to contribution →