we-solve (beta)

Econ-433 · economics · Economics

Sparse synthetic controls

LLM status Solved unverified

Statement

Original subject group: Econometrics, causal inference and measurement

Audit classification: Published research agenda. The source research agenda is verified at p.423. Modern inference results exist; this finite-donor robust specification is editorial, and narrow coverage cannot follow solely from pretreatment fit and a rank bound.

Definitions and precise research target

Consider one treated region $j=0$, $J$ donor regions, $T_0$ pretreatment periods and a fixed post-treatment horizon $H$. Untreated outcomes satisfy $Y_{jt}(0)=\lambda_j^\top f_t+\eta_{jt}+\xi_{jt}$, where $\lambda_j,f_t\in\mathbb R^r$, the known rank bound is $r\leq r_{\max}$, and their Euclidean norms are bounded by stated constants $\Lambda,F$. Approximation errors obey $|\eta_{jt}|\leq a_{jt}$; noises $\xi_{jt}$ are mean-zero and independent across regions and periods, with a stated sub-Gaussian scale bound $\sigma$. Donor weights $w_j\geq0$ sum to one and are selected using pretreatment data. Post-treatment donor spillovers satisfy $|s_{jt}|\leq b_{jt}$, for prespecified nonnegative bounds $a_{jt},b_{jt}$. The target is $\tau=H^{-1}\sum_{t=T_0+1}^{T_0+H}\{Y_{0t}(1)-Y_{0t}(0)\}$, the treated region's average effect. For fixed small $J$, construct intervals with coverage at least $1-\alpha$, for specified $\alpha\in(0,1)$, uniformly over this model class, including weight selection. Determine when pretreatment fit and validation periods make intervals informative, and how removing contaminated donors trades spillover protection against worse fit. No inference guarantee is asserted without the declared factor, approximation and noise restrictions. The source's broad agenda supports inference and sensitivity concerns; this finite-donor specification is an adaptation whose contemporary unresolved status needs separate validation.

Definitions, assumptions and mathematical qualifications. The treated region is observed under treatment post-policy; donors satisfy $Y_{jt}=Y_{jt}(0)+s_{jt}$ after treatment, and all regions are untreated before it. Weights are measurable functions of pretreatment observations only. Conditional on deterministic factor primitives, $E[e^{u\xi_{jt}}]\leq e^{u^2\sigma^2/2}$ for every real $u$; clarify any dependence of treatment outcomes on untreated noise. Uniform coverage is $\inf_{M\in\mathcal M}\Pr_M(\tau_M\in C)\geq1-\alpha$, including weight randomness and counterfactual prediction noise. Pretreatment fit alone does not identify post-treatment factors: a factor direction absent before treatment can emerge afterward. Informative narrow intervals require a stated extrapolation, span or factor-excitation condition, not only bounded rank.

Variants and assumption changes

  1. Editorial, exact span. Assume $\lambda_0=\sum_jw_j\lambda_j$ for admissible nonnegative weights and no donor spillovers; account for estimated weights, approximation error and post-treatment noise.
  2. Editorial, bounded contamination. Permit $|s_{jt}|\leq b_{jt}$ and loading mismatch $\|\lambda_0-\sum_jw_j\lambda_j\|\leq d(w)$. Compare intervals using the worst-case factor bias $F d(w)$, approximation bounds and weighted spillover costs.

References and source audit

Checked source. Journal of Economic Literature 59(2) (2021), pp. 391--425, Section 9 p. 423; Section 6 pp. 410--411. The conclusion identifies inference and sensitivity as research areas. Full primary text retrieved. https://economics.mit.edu/sites/default/files/publications/jel.20191450.pdf.
  1. Alberto Abadie. 2021. Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects. Journal of Economic Literature 59(2) (2021), pp. 391--425, Section 9 p. 423; Section 6 pp. 410--411. https://economics.mit.edu/sites/default/files/publications/jel.20191450.pdf. DOI: https://doi.org/10.1257/jel.20191450.

Common conventions: Source mathematical conventions

These conventions apply to each entry unless explicitly replaced.

Models and identification. A primitive model $m$ specifies all probability laws, feasible choices, information, equilibrium and measurement rules used by its target. The admissible class $\mathcal M$ is fixed independently of outcomes. If $P_O(m)$ is its observable probability law and $\tau(m)$ the target, its identified set at observable law $P$ is \[ \mathcal I_\tau(P)=\{\tau(m):m\in\mathcal M,\ P_O(m)=P\}. \] Point identification means that this set is a singleton. An empty set signals incompatibility of data and assumptions. Coordinate bounds need not describe the joint set. Bounds are sharp only if every retained target is attainable or arbitrarily approachable by compatible models. Finite bounds require explicit restrictions on unobserved quantities; unrestricted confounding, hidden wealth, extrapolation or arbitrary equilibrium selection can yield uninformative sets.

Causal interventions. A potential outcome $Y_i(p)$ is the outcome under a complete policy regime $p$, including timing, financing, information and market spillovers. The target population and units are fixed before treatment. With interference, use the complete assignment vector or a justified exposure mapping; individual no-interference notation alone cannot identify an economy-wide intervention. A difference of mean potential outcomes does not require the cross-world joint distribution. Conditioning on post-treatment migration, survival, enrollment or employment changes the estimand and needs separate justification.

Statistical statements. An estimator, confidence set, forecast or test specifies a sampling law, model class and loss/coverage criterion. Uniform $(1-\alpha)$-coverage means $\inf_{m\in\mathcal M}\Pr_m\{\tau(m)\in C_n\}\geq1-\alpha$, or an explicitly asymptotic version. The whole parameter space trivially has coverage, so informativeness requires a stated width, risk or power objective. A forecasting improvement is relative to a named benchmark, information set and evaluation population.

Equilibrium and optimization. An equilibrium comprises feasible plans, optimality, beliefs where needed, budget constraints and all market/resource clearing. The primitives or a stated benchmark define each correspondence $\mathcal E(p;m)$. Nonemptiness is assumed or proved, never inferred just from compact policy sets. A selected-equilibrium target uses a declared selection rule; a robust target quantifies over all permitted equilibria. Use suprema or infima unless attainment is established. An $\varepsilon$-optimal policy is feasible and within $\varepsilon$ of the finite optimum value. Report an empty feasible set separately.

Welfare and accounting. Normative utility, interpersonal normalization, social weights, numeraire, discounting and the use of public receipts are primitives. Behavioral utility need not coincide with the welfare criterion. Household welfare includes ownership income and rents once; adding profits again to compensating variation generally double-counts them. A monetary equivalent is defined by an explicit transfer and price convention, with monotonicity and range conditions. Average effects, mechanism contributions, transfers and welfare losses are different targets. Infinite sums require convergence and dynamic feasible plans require terminal or no-Ponzi conditions.

Source versions. A working-paper date, revision date and journal publication year are distinguished. A DOI for a journal version is not automatically the DOI of an earlier manuscript. Locators refer to the stated version and printed pages where specified. A metadata-only check does not verify a claim about a full-text conclusion. Every entry's source subsection narrows the provenance claim accordingly.

Statement licence: CC-BY-4.0 · Source file

Contributions 2

Econ-433-A2 · Full solution claimed · LLM-generated · gpt 6 astra pro ·

GPT 6 Astra Pro attempt on Econ-433

Finite-sample coverage certificates for sparse synthetic controls

Problem ID: C23-429 (JEL C23). Status: Solved for the stated finite-donor construction and comparison target. The results provide uniform coverage and explicit conditions for informative certificates; they do not assert that the proposed intervals are shortest or minimax optimal.

A uniform interval that permits estimated donor weights

The target is one treated region, $J\ge1$ donors, $T_0\ge0$ pretreatment periods, and $H\ge1$ post-treatment periods. Let ${\mathcal W}=\{w\in{\mathbb{R}}^J:w_j\ge0,\ \sum_j w_j=1\}$. All weights are selected from pretreatment observations (and possibly auxiliary randomization). The untreated outcomes obey \[ Y_{jt}(0)=\lambda_j^\top f_t+\eta_{jt}+\xi_{jt},\qquad r\le r_{\max},\quad\norm{\lambda_j}_2\le\Lambda,\quad\norm{f_t}_2\le F, \] where $|\eta_{jt}|\le a_{jt}$, and the mean-zero noises are mutually independent across regions and dates with ${\mathbb{E}} e^{u\xi_{jt}}\le e^{u^2\sigma^2/2}$ for all real $u$. The loadings and factors are deterministic model primitives (equivalently, the noise bound holds conditionally on fixed primitives). All regions are untreated before policy; afterward the observed donors satisfy $Y_{jt}=Y_{jt}(0)+s_{jt}$ with $|s_{jt}|\le b_{jt}$. The nonnegative envelopes $a_{jt},b_{jt}$ and constants $\Lambda,F,\sigma$ are declared in advance. Approximation errors and spillovers may otherwise be dependent on the data and noises.

For a post-treatment average, write a bar, and put \[ \begin{gathered} \Delta(w)=\lambda_0-\sum_{j=1}^Jw_j\lambda_j,\qquad Z_j=H^{-1}\sum_{t>T_0}\xi_{jt},\\ A_+(w)=\bar a_0+\sum_jw_j\bar a_j,\qquad S_+(w)=\sum_jw_j\bar b_j,\\ \tau=\bar Y_0(1)-\bar Y_0(0),\qquad \widehat\tau(w)=\bar Y_0(1)-\sum_jw_j\bar Y_j. \end{gathered} \] The exact decomposition is \begin{equation}\label{wsref-0-65-71-3a-64-65-63-6f-6d-70} \widehat\tau(w)-\tau =\Delta(w)^\top\bar f +\bar\eta_0-\sum_jw_j(\bar\eta_j+\bar s_j) +Z_0-\sum_jw_jZ_j. \end{equation} The treated outcome under treatment cancels, so no restriction on its dependence with untreated noise is necessary. Coverage concerns the realized average effect, including the treated unit's post-treatment counterfactual noise.

Theorem (Simultaneous finite-sample coverage): Fix $0\le\delta_0<1$ and $0<\delta_1<1$ with $\delta_0+\delta_1\le\alpha<1$. Suppose a pretreatment-computable loading certificate $\widehat d(w)$ satisfies, uniformly over the declared model class, \begin{equation}\label{wsref-0-65-71-3a-64-2d-63-65-72-74} {\mathbb{P}}\left\{\norm{\Delta(w)}_2\le\widehat d(w) \text{ for every }w\in{\mathcal W}\right\}\ge1-\delta_0. \end{equation} Define \begin{equation}\label{wsref-0-65-71-3a-72-61-64} R(w)=F\widehat d(w)+A_+(w)+S_+(w) +2\sigma\sqrt{\frac{\log(2J/\delta_1)}H}. \end{equation} Then \begin{equation}\label{wsref-0-65-71-3a-73-69-6d-75-6c-74-61-6e-65-6f-75-73} {\mathbb{P}}\left\{\tau\in[\widehat\tau(w)-R(w),\widehat\tau(w)+R(w)] \text{ for all }w\in{\mathcal W}\right\} \ge1-\delta_0-\delta_1. \end{equation} In particular, evaluating the interval at any pretreatment-selected $\widehat w$ yields coverage at least $1-\alpha$ uniformly over the model class.
Proof: Independence and the sub-Gaussian moment bound give ${\mathbb{E}} e^{u(Z_0-Z_j)}\le e^{u^2\sigma^2/H}$. Thus \[ {\mathbb{P}}\{|Z_0-Z_j|>x\}\le2\exp\{-Hx^2/(4\sigma^2)\}, \] with the zero-noise case interpreted exactly. A union bound gives, with probability at least $1-\delta_1$, $\max_j|Z_0-Z_j|\le2\sigma\sqrt{\log(2J/\delta_1)/H}$. On this same event for every $w\in{\mathcal W}$, \[ \left|Z_0-\sum_jw_jZ_j\right| =\left|\sum_jw_j(Z_0-Z_j)\right| \le\sum_jw_j|Z_0-Z_j| \le2\sigma\sqrt{\frac{\log(2J/\delta_1)}H}. \] Intersect with the event in (eq:d-cert). The factor term in (eq:decomp) is at most $F\widehat d(w)$, and the other terms have the respective pathwise envelopes. A union bound on these two events, without assuming that they are independent, proves (eq:simultaneous). ∎

The universal loading certificate is $\widehat d(w)=2\Lambda$ by the triangle inequality. Inserting $\delta_0=0$, $\delta_1=\alpha$ gives the explicit answer requiring no additional assumptions: \begin{equation}\label{wsref-0-65-71-3a-75-6e-69-76-65-72-73-61-6c} \boxed{\quad C(\widehat w)=\widehat\tau(\widehat w)\ \pm\ \left[2\Lambda F+\bar a_0+\sum_j\widehat w_j(\bar a_j+\bar b_j) +2\sigma\sqrt{\frac{\log(2J/\alpha)}H}\right].\quad} \end{equation} If $r_{\max}=0$, the factor-mismatch term vanishes. For comparison, the donor-free interval \[ \bar Y_0(1)\ \pm\ \left[\Lambda F+\bar a_0+ \sigma\sqrt{\frac{2\log(2/\alpha)}H}\right] \] is also uniformly valid. Choosing the shorter of separately calibrated intervals requires joint error control; individual validity alone does not justify adaptive selection.

Loading-mismatch certificates from pretreatment observations

The universal radius is generally conservative. There are two ways to improve it while preserving (eq:d-cert) for all candidate weights.

Certificate without excitation. Suppose $T_0\ge1$ and $\delta_0>0$. Put \[ u_0=\sigma\sqrt{2\log\!\left(\frac{2(J+1)T_0}{\delta_0}\right)}. \] Pad all possible ranks to $r_{\max}$ coordinates. Define the observable feasible set $\mathcal K_0$ of pretreatment loadings and factors by \[ \norm{\lambda_j}_2\le\Lambda,\quad \norm{f_t}_2\le F, \quad |Y_{jt}-\lambda_j^\top f_t|\le a_{jt}+u_0 \quad (0\le j\le J,\ 1\le t\le T_0). \] For $\mathcal K_0\ne\varnothing$ let \begin{equation}\label{wsref-0-65-71-3a-70-72-6f-66-69-6c-65} \widehat d_{\mathcal K}(w) =\max_{(\lambda,f)\in\mathcal K_0} \left\|\lambda_0-\sum_jw_j\lambda_j\right\|_2, \end{equation} otherwise set $\widehat d_{\mathcal K}=2\Lambda$ and record that the calibrated feasibility set was empty. The set is compact, so the displayed maximum exists. The true factorization lies in $\mathcal K_0$ whenever all pretreatment noises satisfy $|\xi_{jt}|\le u_0$, an event of probability at least $1-\delta_0$ by a union bound. Therefore (eq:profile) satisfies (eq:d-cert), without known factors, known exact rank, or factor excitation. Computing a global maximum (or a certified upper bound) is essential: a local maximum is not a coverage certificate. Empty feasibility does not itself refute the stochastic model.

A sharper certificate with factor excitation. Let $V$ be a prespecified pretreatment fitting or validation block of size $m\ge1$. Add the substantive assumption \begin{equation}\label{wsref-0-65-71-3a-65-78-63-69-74-65} m^{-1}\sum_{t\in V}f_tf_t^\top\succeq\kappa I_r, \qquad\kappa>0\text{ known}, \end{equation} on the actual factor dimension (not on artificial padded coordinates). Put \[ \begin{split} e_t(w)&=Y_{0t}-\sum_jw_jY_{jt},\\ R_V(w)&=\left[m^{-1}\sum_{t\in V}e_t(w)^2\right]^{1/2},\\ A_V(w)&=\left[m^{-1}\sum_{t\in V} \left(a_{0t}+\sum_jw_ja_{jt}\right)^2\right]^{1/2}. \end{split} \] For $d=\min(r_{\max},m)>0$, define \[ K(d,x)=d+2\sqrt{dx}+2x,\qquad q_V=\sigma\sqrt{\frac{2K(d,\log(J/\delta_0))}m}. \]

Lemma (Projected sub-Gaussian noise): If a random vector $X\in{\mathbb{R}}^n$ obeys ${\mathbb{E}} e^{u^\top X}\le e^{s^2\norm u_2^2/2}$ for every $u\in{\mathbb{R}}^n$, and $P$ is a deterministic orthogonal projector of rank $d$, then \[ {\mathbb{P}}\{\norm{PX}_2^2>s^2K(d,x)\}\le e^{-x},\qquad x>0. \]
Proof: Let $G\sim N(0,I_n)$ be independent of $X$. Gaussian integration and the assumed moment-generating bound give, for $0<t<1/(2s^2)$, \[ {\mathbb{E}} e^{t\norm{PX}_2^2} ={\mathbb{E}}_G{\mathbb{E}}_X e^{\sqrt{2t}(PG)^\top X} \le{\mathbb{E}}_Ge^{s^2t\norm{PG}_2^2} =(1-2s^2t)^{-d/2}. \] Consequently, for $U=\norm{PX}_2^2/s^2$ and $0<t<1/2$, \[ \log{\mathbb{E}} e^{t(U-d)} \le-dt-\frac d2\log(1-2t) \le\frac{dt^2}{1-2t}. \] Chernoff's inequality with $t=\sqrt{x}/(\sqrt d+2\sqrt{x})$ yields ${\mathbb{P}}\{U>d+2\sqrt{dx}+2x\}\le e^{-x}$. The cases $s=0$ and $d=0$ are immediate. ∎
Proposition (Simultaneous excitation certificate): Under (eq:excite), with probability at least $1-\delta_0$, simultaneously for every $w\in{\mathcal W}$, \begin{equation}\label{wsref-0-65-71-3a-64-2d-65-78-63-69-74-65} \norm{\Delta(w)}_2\le \widehat d_V(w):= \min\left\{2\Lambda,\ \frac{R_V(w)+A_V(w)+q_V}{\sqrt\kappa}\right\}. \end{equation} If $r_{\max}=0$, the certificate is zero without needing (eq:excite).
Proof: Let $B$ be the $m\times r$ matrix with rows $f_t^\top$, and let $P_V$ project onto its column space, of dimension at most $d$. Each vector $\xi_{0,V}-\xi_{j,V}$ has sub-Gaussian variance proxy $2\sigma^2I_m$. Lemma Lemma (Projected sub-Gaussian noise), followed by a union bound over donors, gives \[ \max_j\frac{\norm{P_V(\xi_{0,V}-\xi_{j,V})}_2}{\sqrt m} \le q_V \] with probability at least $1-\delta_0$. Every weighted contrast is a convex combination of these donor contrasts. The latent residual $g(w)=B\Delta(w)$ belongs to the projector's range, hence projection and the triangle inequality imply simultaneously \[ \frac{\norm{g(w)}_2}{\sqrt m} \le R_V(w)+A_V(w)+q_V. \] Excitation gives $\norm{g(w)}_2/\sqrt m\ge\sqrt\kappa\norm{\Delta(w)}_2$. Combine the two inequalities with $\norm{\Delta(w)}_2\le2\Lambda$. ∎

The projection $P_V$ is used only to prove the bound; it is not estimated to evaluate (eq:d-excite). Its stochastic allowance is of order $\sigma\sqrt{(r_{\max}+\log(J/\delta_0))/m}$. The actual fitting residual $R_V$ may still reflect noise. As the guarantee is simultaneous, $V$ may be used in the selection of the final weights. Multiple data-selected validation blocks require a joint event or an error-budget adjustment. An empirically guessed value of $\kappa$ does not replace the structural assumption (eq:excite).

Precisely when fitting and validation improve the interval

Given a desired half-width $\varepsilon>0$, a sufficient, observable condition for a radius at most $\varepsilon$ is \begin{equation}\label{wsref-0-65-71-3a-69-6e-66-6f-72-6d-61-74-69-76-65} F\widehat d_V(\widehat w)+A_+(\widehat w)+S_+(\widehat w) +2\sigma\sqrt{\frac{\log(2J/\delta_1)}H}\le\varepsilon. \end{equation} With $F>0$, the excitation-based loading allowance is strictly smaller than $2\Lambda F$ whenever \[ R_V(\widehat w)+A_V(\widehat w)+q_V <2\Lambda\sqrt\kappa. \] Thus informative fit requires excitation in the directions needed for extrapolation, sufficiently small residual and approximation envelopes, and non-dominant spillover and post-treatment-noise allowances. Enlarging a period in which the factors are unexcited does not accomplish this.

Full excitation is sufficient but not necessary for an average effect. Another valid additional restriction is a known average-factor span: \begin{equation}\label{wsref-0-65-71-3a-6b-6e-6f-77-6e-73-70-61-6e} \bar f=\sum_{t\in V}q_tf_t+v_f, \qquad\norm{v_f}_2\le\rho, \end{equation} where $q=(q_t)_{t\in V}$ is prespecified and deterministic. Form the validation-adjusted center $\widehat\tau_q(w)=\widehat\tau(w)-\sum_{t\in V}q_te_t(w)$. Substitution in (eq:decomp), and the same donor-uniform noise argument, give a simultaneous interval of radius \begin{align}\label{wsref-0-65-71-3a-73-70-61-6e-2d-72-61-64-69-75-73} R_q(w)={}&\rho\widehat d(w)+A_+(w)+S_+(w) +\sum_{t\in V}|q_t|\left(a_{0t}+\sum_jw_ja_{jt}\right)\notag\\ &+2\sigma\sqrt{\left(H^{-1}+\norm q_2^2\right) \log(2J/\delta_1)}. \end{align} To verify it, replace the factor term $\Delta(w)^\top\bar f$ by $\Delta(w)^\top v_f$, and note that each region's combined post-minus-pre noise $Z_j-\sum_{t\in V}q_t\xi_{jt}$ has variance proxy $\sigma^2(H^{-1}+\norm q_2^2)$, by independence across dates. The certificate event and this simultaneous noise event yield coverage $1-\delta_0-\delta_1$. If $\rho=0$, the loading certificate is unnecessary, and one may take $\delta_0=0$.

Exact span, estimated weights, and the limits of pretreatment fit

The exact-span variant needs care about what is assumed to balance. If the selected weights themselves obey $\lambda_0=\sum_j\widehat w_j\lambda_j$ almost surely, and spillovers are absent, then $\widehat d=0$, $S_+=0$ and Theorem Theorem (Simultaneous finite-sample coverage) gives \[ C(\widehat w)=\widehat\tau(\widehat w)\ \pm\ \left[\bar a_0+\sum_j\widehat w_j\bar a_j +2\sigma\sqrt{\frac{\log(2J/\alpha)}H}\right]. \] By contrast, the existence of $w^\star\in{\mathcal W}$ satisfying exact balance does not imply that an estimated $\widehat w$ equals $w^\star$. The selected mismatch obeys \begin{equation}\label{wsref-0-65-71-3a-65-73-74-69-6d-61-74-65-64} \norm{\Delta(\widehat w)}_2 =\left\|\sum_j(w_j^\star-\widehat w_j)\lambda_j\right\|_2 \le\Lambda\norm{\widehat w-w^\star}_1, \end{equation} but the unknown right side needs a certified bound. One direct implementation is to impose existence of a balancing vector $v\in{\mathcal W}$, $\lambda_0=\sum_jv_j\lambda_j$, inside the feasible set $\mathcal K_0$ of (eq:profile), and then maximize the mismatch of the actual chosen weights. Proposition Proposition (Simultaneous excitation certificate) is an alternative under excitation. Neither method treats the fitted weights as oracle weights.
Proposition (Perfect pretreatment fit need not imply small uncertainty): If $r_{\max}\ge1$ and $\Lambda F>0$, the model includes an observationally indistinguishable noiseless family with perfect pretreatment fit for any $T_0$, but with possible effects filling $[-\Lambda F,\Lambda F]$.
Proof: Take one donor with $\lambda_1=0$, a scalar factor $f_t=0$ on every pretreatment date and $f_t=F$ afterward, and set all noises, approximations, and spillovers to zero. Make the observed post-treatment treated outcomes zero. Allow $\lambda_0\in[-\Lambda,\Lambda]$. All observed data stay zero; the pretreatment donor residual vanishes identically; yet $\tau=-\lambda_0F$ fills the asserted interval. ∎

For deterministic interval rules and $1-\alpha>0$, coverage for all the indistinguishable models forces an interval to contain the entire set. With independent auxiliary randomization, endpoint coverage implies \[ {\mathbb{P}}\{\operatorname{length}(C)\ge2\Lambda F\}\ge1-2\alpha \quad(\alpha<\tfrac12), \] and integration of coverage over that compatible interval gives ${\mathbb{E}}\operatorname{length}(C)\ge2\Lambda F(1-\alpha)$. Even existence of exact balance does not eliminate this example's logic: use two donors with loadings $-\Lambda$ and $+\Lambda$; any treated loading in between is exactly representable, but all pretreatment observations still vanish and the effect remains undetermined.

There is also an unavoidable prediction-noise scale at fixed $H$. Set all loadings, approximation errors, and spillovers to zero and take independent $N(0,\sigma^2)$ untreated noises, with the treated observed post-treatment outcomes fixed at zero. Then $\tau=-Z_0\sim N(0,\sigma^2/H)$ independently of the observed data. An interval of random length $L$ has conditional coverage at most $g(L)=2\Phi(L\sqrt H/(2\sigma))-1$. Since $g$ is concave on $[0,\infty)$, Jensen's inequality yields \begin{equation}\label{wsref-0-65-71-3a-6e-6f-69-73-65-2d-6c-6f-77-65-72} {\mathbb{E}} L\ge\frac{2\sigma}{\sqrt H}\Phi^{-1}(1-\alpha/2). \end{equation} When $\sigma=0$ the lower bound is zero. Increasing $T_0$ alone does not remove either this uncertainty or the factor-direction obstruction.

Sparse selection and the donor-contamination tradeoff

It can be useful to make the noise radius sensitive to diversification. Besides the constant noise term in (eq:rad), the following alternative is also simultaneous: \begin{equation}\label{wsref-0-65-71-3a-65-6c-6c-69-70-73-6f-69-64} N_E(w)=\frac{\sigma}{\sqrt H} \sqrt{(1+\norm w_2^2)K(J,\log(1/\delta_1))}. \end{equation} Indeed, with $Z=(Z_0,\ldots,Z_J)$ and $P$ projecting onto the $J$-dimensional subspace perpendicular to $(1,\ldots,1)$, the contrast vector $(1,-w)$ lies in that subspace. Lemma Lemma (Projected sub-Gaussian noise) bounds $\norm{PZ}_2$, and Cauchy--Schwarz gives (eq:ellipsoid) simultaneously for every $w$. Either noise bound may be chosen in advance; taking their pointwise minimum requires joint calibration, such as a split of $\delta_1$.

For a prespecified available set $D$ and support limit $1\le k\le|D|$, a concrete sparse rule is to choose from pretreatment information \begin{equation}\label{wsref-0-65-71-3a-73-70-61-72-73-65-6c-61-77} \widehat w\in{\operatorname*{arg\,min}}_{\substack{w\in{\mathcal W},\ {\operatorname{supp}}(w)\subseteq D\\ \norm w_0\le k}} \left\{F\widehat d(w)+\sum_jw_j(\bar a_j+\bar b_j)+N(w)\right\}, \end{equation} with $N=q_+$ or $N_E$, and a deterministic tie-break. For the certificates above, the feasible set is a finite union of compact simplex faces and the objective is continuous, so a minimizer exists. For fixed small $J$, supports can be enumerated. Any feasible approximate minimizer also retains coverage; accuracy affects width, not the theorem. An $\ell_1$ penalty is constant on the simplex, $\norm w_1=1$, so it does not itself encourage sparse weights.

When exact balance exists in ${\mathbb{R}}^r$, there is a balancing representation using at most $r+1$ donors: if a representation has more than $r+1$ positive weights, the corresponding augmented vectors $(1,\lambda_j)$ are linearly dependent. A nonzero perturbation along that dependence preserves both the weight sum and the balanced loading. Move until one weight hits zero and iterate. This is an existence statement; it does not guarantee that noisy fitting attains that representation or that its contamination allowance is small.

For any original weights $w$ and refitted weights $v$, write $R(w)=F\widehat d(w)+\bar a_0+\sum_jw_j\bar a_j+S_+(w)+N(w)$. Exactly, \begin{align}\label{wsref-0-65-71-3a-64-65-6c-65-74-69-6f-6e} R(v)-R(w)={}&F\{\widehat d(v)-\widehat d(w)\} +\sum_j(v_j-w_j)\bar a_j\notag\\ &+\{S_+(v)-S_+(w)\}+\{N(v)-N(w)\}. \end{align} If the remaining donors used by $v$ are uncontaminated, deletion reduces the certified half-width if and only if \begin{equation}\label{wsref-0-65-71-3a-64-65-6c-65-74-65-2d-63-6f-6e-64-69-74-69-6f-6e} S_+(w)>F\{\widehat d(v)-\widehat d(w)\} +\sum_j(v_j-w_j)\bar a_j+N(v)-N(w). \end{equation} This states the spillover protection versus the loss of balance, approximation control, and noise diversification. For example, if $\lambda_0=\lambda_1=0$, $\lambda_2=\delta$ in a scalar model, donor 1 has spillover envelope $b$, and donor 2 is clean, the deterministic single-donor allowances are $b$ and $F|\delta|$. Their noise allowances agree, so switching to the clean donor reduces the radius precisely when $b>F|\delta|$.

For the same globally defined radius and bounds, restricting the donor set cannot decrease the optimized minimum: \[ \inf_{\substack{w\in{\mathcal W},\ {\operatorname{supp}}(w)\subseteq D}}R(w) \ \ge\ \inf_{w\in{\mathcal W}}R(w). \] A contamination-aware optimizer over the full pool can already assign zero weight to a contaminated donor. Removal can improve on a fit-only rule or hedge against larger uncertainty about contamination, but not beat the globally minimized common criterion. If donors are deleted adaptively, one cannot simply replace $J$ in the simultaneous bound by the selected subset size without recalibration.

Completion Estimate

COMPLETION: 100% The finite-donor problem as stated is resolved constructively: (eq:universal) has uniform coverage without additional assumptions, (eq:profile) and (eq:d-excite) permit valid data-dependent reductions, (eq:informative)--(eq:span-radius) give explicit informative-validation criteria, and (eq:deletion)--(eq:delete-condition) settle the tradeoff for the proposed certified radii. The exact-span variants and counterexamples are covered as well. The completion estimate refers only to the original specified target, not to dependent-shock extensions, necessary-and-sufficient minimax informativeness conditions, or shortest possible intervals. It is an LLM claim, not an independently verified mathematical verdict.

Sources and verification

The mathematical arguments above were generated and checked by OpenAI GPT-6 on 11 October 2026, using the user-provided problem and earlier attempt. LaTeX was compiled to check syntax and references. No independent human mathematical proof check or formal verification is reported; source checking does not establish novelty. The original problem is an editorial finite-donor adaptation of a methodological agenda and should not be read as a theorem already stated by its source. The prior attempt's stronger independence assumption was removed by using simultaneous noise events; no assumption that pretreatment errors are independent of future noises is required. No empirical effect is claimed.

References

Go to contribution →

Econ-433-A1 · Partial result · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-433

Uniform synthetic-control certificates with selected weights and contamination

Problem ID: C23-429. Status: Partial. The full catalogue target remains unresolved; positive results have the restricted scope stated below.

A coverage target that includes counterfactual prediction noise

The canonical model has a treated region, finitely many donors, bounded factor loadings and factors, approximation error, independent sub-Gaussian noise, and bounded donor spillovers. The requested interval covers the treated region's realized average effect, not merely its expectation over untreated noise. This distinction means that even perfectly balanced factors do not remove post-treatment prediction noise.

Write $t=1,\ldots,H$ for post-treatment periods and let $w$ be any nonnegative donor weights summing to one, chosen from pretreatment data only. Assume deterministic factor primitives, $\|f_t\|\le F$, $\|\lambda_j\|\le\Lambda$, $|\eta_{jt}|\le a_{jt}$, and $|s_{jt}|\le b_{jt}$. All untreated noises are independent over regions and dates, with moment-generating function bounded by $\exp(u^2\sigma^2/2)$. In addition, the full vector of post-treatment noises is independent of the sigma-field generated by all pretreatment observations and any auxiliary randomization used to choose the weights. This extra restriction excludes pretreatment approximation errors that encode future noise; independence of the noise coordinates alone would not exclude that leakage. Observed donors equal $Y_{jt}(0)+s_{jt}$. Define \[ \widehat\tau(w)=H^{-1}\sum_t\left(Y_{0t}(1)-\sum_jw_jY_{jt}\right),\qquad \tau=H^{-1}\sum_t\{Y_{0t}(1)-Y_{0t}(0)\}. \] No independence between treated potential outcomes and untreated outcomes is needed: the treated outcome cancels from their difference.

Theorem (A valid interval for arbitrary pretreatment selection): Suppose a deterministic or pretreatment-measurable certificate $d(w)$ satisfies $\|\lambda_0-\sum_jw_j\lambda_j\|\le d(w)$ almost surely. Put \[ B(w)=Fd(w)+H^{-1}\sum_t\left[a_{0t}+\sum_jw_j(a_{jt}+b_{jt})\right], \] \[ R_\alpha(w)=B(w)+\sigma\sqrt{\frac{2(1+\|w\|_2^2)}H\log\frac2\alpha}. \tag{1} \] Then $[\widehat\tau(w)-R_\alpha(w),\widehat\tau(w)+R_\alpha(w)]$ has coverage at least $1-\alpha$ uniformly over the declared class, including any pretreatment weight-selection rule.
Proof: Subtracting the target gives \[ \widehat\tau-\tau=H^{-1}\sum_t\left[ (\lambda_0-\sum_jw_j\lambda_j)^Tf_t+ \eta_{0t}-\sum_jw_j(\eta_{jt}+s_{jt})+ \xi_{0t}-\sum_jw_j\xi_{jt}\right]. \] The first two groups have absolute sum bounded by $B(w)$ pathwise. Conditional on all pretreatment observations, weights are fixed and post-treatment noises retain their independent sub-Gaussian laws. Their averaged weighted contrast has scale squared $\sigma^2(1+\|w\|_2^2)/H$. The two-sided exponential tail bound gives (1) conditionally and hence unconditionally. Bounded spillovers or approximation errors may be dependent on other variables; their pathwise envelopes suffice. ∎

The universal certificate $d(w)=2\Lambda$ follows from the triangle inequality. This supplies a valid, potentially wide interval without a factor-excitation assumption. Exact loading balance supplies $d=0$, but pretreatment fit by itself does not certify exact balance.

A transparent excitation-based loading certificate

There are $T_0$ pretreatment dates. Assume the factor Gram matrix obeys \[ T_0^{-1}\sum_{t\le T_0}f_tf_t^T\succeq\kappa I_r, \qquad\kappa>0, \tag{2} \] where the lower bound is substantively known. Define observed root-mean-square fit and its approximation envelope by \[ R_0(w)=\left[T_0^{-1}\sum_{t\le T_0}(Y_{0t}-\sum_jw_jY_{jt})^2\right]^{1/2}, \] \[ A_0(w)=\left[T_0^{-1}\sum_{t\le T_0}(a_{0t}+\sum_jw_ja_{jt})^2\right]^{1/2}. \] Let $u=\sigma\sqrt{2\log(2(J+1)T_0/\alpha_0)}$ and set \[ \widehat d(w)=\min\left\{2\Lambda, \frac{R_0(w)+A_0(w)+2u}{\sqrt\kappa}\right\}. \tag{3} \]
Proposition (Simultaneous validity for selected weights): With probability at least $1-\alpha_0$, (3) bounds the loading mismatch for every weight vector in the simplex simultaneously. Inserting it in (1), using post-treatment tail level $\alpha_1$, gives unconditional coverage at least $1-\alpha_0-\alpha_1$.
Proof: A union bound on all pretreatment noises gives $|\xi_{jt}|\le u$ simultaneously with probability $1-\alpha_0$. Every simplex contrast then has noise magnitude at most $2u$ at every date, including adaptively selected weights. The triangle inequality in the normalized Euclidean norm bounds the latent factor-fit norm by $R_0+A_0+2u$. By (2), that norm is at least $\sqrt\kappa\|\lambda_0-\sum_jw_j\lambda_j\|$. Intersecting with the universal bound proves (3). On this pretreatment event the bias certificate holds, and independent post-treatment tails add failure probability at most $\alpha_1$. ∎

This certificate is deliberately conservative: a maximum-noise bound does not exploit averaging efficiently. It is nevertheless finite-sample valid under arbitrary weight selection and clarifies what a tighter result must improve. A validation period can replace part of the fitting sample, but absent a lower excitation bound it cannot rule out factors invisible throughout both periods.

Perfect pre-fit can coexist with irreducible ambiguity

Take one factor, no noise, no approximation error, no spillovers, donor loadings zero, and pretreatment factors zero. In two models set the treated loading to $+\Lambda$ and $-\Lambda$ respectively, with every post-treatment factor equal to $F$. Set the observed treated outcome under treatment to zero in both models. All observed data are identical, and every synthetic control has perfect pretreatment fit. The realized effects are $-\Lambda F$ and $+\Lambda F$.

For any interval procedure with coverage $1-\alpha$ in both models, the common observed-data distribution assigns probability at least $1-2\alpha$ to covering both targets, by the union bound. On that event its length is at least $2\Lambda F$. This proves a quantitative obstruction to uniformly narrow intervals from pre-fit and a rank bound alone. The example satisfies rank one and arbitrarily many pretreatment periods; more dates with zero excitation do not resolve it.

Sparse donor choice and contamination tradeoffs

Suppose two donors have loadings $0$ and $\delta$ and the treated loading is zero in a scalar model. The first donor has a spillover bound $b$, while the second is uncontaminated; ignore approximation errors. Using only the first gives bias allowance $b$, using only the second gives $F|\delta|$. Equal noise scales imply equal variance allowances for these two single-donor choices. Removing the contaminated donor narrows the interval exactly when $F|\delta|<b$ under these bounds; the opposite inequality makes removal costly. A mixture adds both loading and weighted contamination terms but reduces $\|w\|_2^2$ relative to an extreme point. Therefore sparsity itself has no universal precision advantage.

One can minimize the explicit radius over a prespecified simplex or sparsity-constrained set using pretreatment information only. The theorem remains valid for that selection. For observed loading certificates, the simultaneous event in Proposition 2.1 is what protects the optimization; a separate pointwise guarantee for every fixed weight would not automatically do so. These intervals establish a restricted finite-sample solution, but sharpness, less conservative excitation estimation, unknown factor rank and realistic dependent shocks remain unresolved. No empirical synthetic-control effect is asserted.

Completion Estimate

80%

This is a post hoc, heuristic estimate for the full catalogue target. The attempt proves finite-sample synthetic-control intervals for selected weights, excitation-based loading certificates, and quantitative fit and contamination tradeoffs. Coverage uses an added prohibition on pretreatment information leaking future noise, while the full declared class and sharper practical excitation controls remain unresolved.

References

  • [abadie] A. Abadie (2021), Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects, Journal of Economic Literature 59(2), 391--425. https://economics.mit.edu/sites/default/files/publications/jel.20191450.pdf. The primary article provides the methodological context. The explicit radius and lower-bound example here follow directly from the stated factor class.

Go to contribution →