Geodesic Synthetic Control

• Geodesic synthetic control is defined through geodesic paths, Fréchet means, and geometric projections instead of linear averaging, making it suitable for complex non-Euclidean data.
• The method extends traditional synthetic control by preserving linear aggregation at the expectation level while enabling geometric optimization over donor weights via divergence-based techniques.
• Its applications span causal inference and quantum sensing, with demonstrated success in recovering counterfactual trajectories and synthesizing precise frequency responses.

Searching arXiv for papers on geodesic synthetic control and related synthetic control geometry. Geodesic synthetic control denotes a family of geometric constructions in which a synthetic object is defined through geodesics, Fréchet means, or related geometric projections rather than ordinary Euclidean linear averaging. In causal inference, the term refers to extensions of synthetic control methods to outcomes that lie in a unique geodesic metric space, such as distributions, compositions, networks, trees, symmetric positive-definite matrices, and functional data; in that setting, the synthetic control is a weighted Fréchet mean of donor outcomes, and the treatment effect is represented by the geodesic from the synthetic counterfactual to the observed treated outcome (Kurisu et al., 1 May 2025). A complementary line of work on the assumptions of synthetic control shows that identification remains tied to linear aggregation at the level of expectations and convex combination of donor compositions, even when geometric or manifold-based estimation devices are introduced (Shi et al., 2021). A further related development formulates weight selection on the simplex through information-theoretic divergences and relaxed balance constraints, yielding an explicitly geometric optimization problem on donor weights (Liao et al., 3 Aug 2025). In a distinct quantum-sensing usage, “geodesic synthetic control” denotes the use of geodesic trajectories of the Bloch vector as a synthetic control tool to synthesize a desired, highly selective frequency response of a quantum sensor (Chen et al., 27 Jan 2026).

1. Classical synthetic control structure and its geometric constraints

The classical synthetic control method approximates a treated unit’s counterfactual outcome by a convex combination of donor outcomes. In the fine-grained reformulation developed in "On the Assumptions of Synthetic Control Methods" (Shi et al., 2021), the target estimand is

$\tau_T = E\big[Y_{1T}(1) - Y_{1T}(0)\big],$

and the identification problem reduces to identifying $E[Y_{1T}(0)]$. The paper introduces an individual-level model in which groups are distributions of individuals rather than indivisible “large units,” and group-level observed outcomes are expectations over those individual distributions. Under the independent causal mechanism assumption,

$P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$

and the stable-distribution factorization

$P_{jt}(X) = P_j(S)\,P_t(U\mid S),$

the expected untreated outcome admits the representation

$E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$

with $\lambda_t(s) := E_t[Y(0)\mid S=s]$ and $\gamma_j(s) := P_j(S=s)$ (Shi et al., 2021).

This reformulation is important because it clarifies that the linearity used by synthetic control is not imposed as a linear structural model for individuals. Rather, linearity is a consequence of aggregation. The paper’s identification theorem states that under A1 (ICM), A2 (stable distributions), A3 (sufficiently similar donors), and A4 (overlap), there exist time-invariant weights $\{\beta_j\}_{j\in D}$ such that

$$E[Y_{1t}(0)] = \sum_{j\in D} \beta_j E[Y_{jt}(0)], \qquad t \le T.$$

The same paper emphasizes that any geometric or manifold-based reinterpretation of synthetic control must preserve this underlying linear aggregation at the level of expectations (Shi et al., 2021).

A central consequence is that geometry can be introduced in donor selection, similarity measurement, or weight regularization without changing the estimand, but a nonlinear “synthetic outcome” that no longer corresponds to a convex combination at the expectation level generally targets something else. The paper makes this point sharply by noting that synthetic control fails if one aggregates via a nonlinear statistic such as the median rather than the mean (Shi et al., 2021).

2. Geodesic synthetic control for random objects and functional data

"Geodesic Synthetic Control Methods for Random Objects and Functional Data" formalizes geodesic synthetic control for settings where outcomes are elements of a unique geodesic metric space $(\mathcal{M}, d)$ rather than scalars or Euclidean vectors (Kurisu et al., 1 May 2025). A metric space is geodesic if any two points $E[Y_{1T}(0)]$0 can be connected by a constant-speed geodesic $E[Y_{1T}(0)]$1 satisfying

$E[Y_{1T}(0)]$2

The weighted Fréchet mean replaces the Euclidean weighted average: $E[Y_{1T}(0)]$3

With one treated unit and $E[Y_{1T}(0)]$4 controls over $E[Y_{1T}(0)]$5 periods, the donor barycenter at time $E[Y_{1T}(0)]$6 under unit weights $E[Y_{1T}(0)]$7 is

$E[Y_{1T}(0)]$8

and the pre-treatment fit criterion is

$E[Y_{1T}(0)]$9

The geodesic synthetic control weights are any minimizer

$P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$0

and the post-treatment synthetic counterfactual is

$P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$1

The treatment effect is defined not as subtraction but as the geodesic

$P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$2

In Euclidean spaces this construction collapses to ordinary synthetic control (Kurisu et al., 1 May 2025).

Identification is stated through the structural model

$P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$3

where $P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$4 is a time-invariant random object and $P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$5 is measurable. For each $P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$6, define the optimal-weight set

$P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$7

Assumption 1 requires existence and uniqueness of the relevant Fréchet means and a stable nonempty intersection of these sets across time. Under this assumption, Theorem 3.1 states that

$P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$8

so the GSC synthetic path recovers the untreated counterfactual path exactly (Kurisu et al., 1 May 2025).

The framework is explicitly designed for non-Euclidean outcomes. The paper lists univariate distributions in Wasserstein space, compositions on the positive orthant of the unit sphere, networks represented by graph Laplacians, SPD matrices under several metrics, and functional data in $P_{jt}(X, Y(0)) = P_{jt}(X)\,P_t(Y(0)\mid X),$9 as canonical examples (Kurisu et al., 1 May 2025).

3. Geodesic synthetic difference-in-differences and augmentation

The same framework extends synthetic difference-in-differences to geodesic metric spaces. The average post-treatment outcome for unit $P_{jt}(X) = P_j(S)\,P_t(U\mid S),$0 is defined as the Fréchet mean

$P_{jt}(X) = P_j(S)\,P_t(U\mid S),$1

and a time-weighted pre-treatment mean is

$P_{jt}(X) = P_j(S)\,P_t(U\mid S),$2

Time weights are chosen by minimizing

$P_{jt}(X) = P_j(S)\,P_t(U\mid S),$3

while unit weights $P_{jt}(X) = P_j(S)\,P_t(U\mid S),$4 are estimated as in GSC (Kurisu et al., 1 May 2025).

A distinctive ingredient is the geodesic transport map $P_{jt}(X) = P_j(S)\,P_t(U\mid S),$5, defined so that $P_{jt}(X) = P_j(S)\,P_t(U\mid S),$6 and, for any $P_{jt}(X) = P_j(S)\,P_t(U\mid S),$7, $P_{jt}(X) = P_j(S)\,P_t(U\mid S),$8 is well-defined and unique. Using this map, the GSDID synthetic post-treatment mean is

$P_{jt}(X) = P_j(S)\,P_t(U\mid S),$9

and the treatment effect is the geodesic

$E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$0

The paper states that GSDID generalizes both geodesic synthetic control and geodesic difference-in-differences and exhibits a double robustness property. Theorem 6.1 shows that, under the geodesic transport assumption and basic existence conditions, either a GSC-type condition or a GDID-type parallel trend condition suffices for

$E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$1

(Kurisu et al., 1 May 2025).

An augmented version, AGSC, is also proposed for cases where exact pre-treatment fit is unattainable. It uses a regression prediction $E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$2, forms a barycentric prediction over controls, and then transports the GSC synthetic control through the map

$E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$3

The paper states that in Euclidean settings this reduces to the usual augmented synthetic control formula (Kurisu et al., 1 May 2025).

4. Weight geometry, simplex divergences, and relaxed balance

A separate but directly relevant development is "A Relaxation Approach to Synthetic Control," which keeps the donor-weight simplex but replaces least-squares fitting with an information-theoretic objective under relaxed balance conditions (Liao et al., 3 Aug 2025). Classical outcomes-only SCM solves

$E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$4

where

$E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$5

SCM-relaxation instead minimizes a divergence on $E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$6 subject to a relaxed first-order condition. In the baseline $E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$7 case,

$E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$8

with

$E[Y_{jt}(0)] = \sum_s \lambda_t(s)\,\gamma_j(s),$9

The paper generalizes this to

$\lambda_t(s) := E_t[Y(0)\mid S=s]$0

where $\lambda_t(s) := E_t[Y(0)\mid S=s]$1 may be quadratic, empirical likelihood, or entropy balancing (Liao et al., 3 Aug 2025).

The geometric content is explicit. The paper interprets these objectives as Bregman divergences from the equal-weight vector $\lambda_t(s) := E_t[Y(0)\mid S=s]$2. For the quadratic case,

$\lambda_t(s) := E_t[Y(0)\mid S=s]$3

while the entropy case corresponds to

$\lambda_t(s) := E_t[Y(0)\mid S=s]$4

which is $\lambda_t(s) := E_t[Y(0)\mid S=s]$5 up to a constant (Liao et al., 3 Aug 2025). The paper further states that the donor pool may exhibit a group structure, in which case the oracle weights are within-group equal: $\lambda_t(s) := E_t[Y(0)\mid S=s]$6 Asymptotically, the proposed estimator achieves oracle performance in terms of out-of-sample prediction accuracy (Liao et al., 3 Aug 2025).

For geodesic synthetic control in the causal-inference sense, this literature is significant because it locates one natural geometric layer not in the outcome space but in the weight simplex. The assumptions paper (Shi et al., 2021) indicates that such geometry is compatible with identification so long as the synthetic outcome remains a convex combination at the expectation level. This suggests that divergence-based or manifold-based optimization over $\lambda_t(s) := E_t[Y(0)\mid S=s]$7 can be used for estimation and regularization without abandoning the causal estimand, whereas replacing the outcome-side convex combination by an unrelated nonlinear construction would generally require a new identification theory (Shi et al., 2021).

5. Applications, metric choice, and empirical practice

The geodesic synthetic control framework has been demonstrated in both simulations and empirical applications. In simulations with graph Laplacians of weighted stochastic block model networks and with SPD matrices under the log-Euclidean metric, GSC exactly recovers the true counterfactual object at the treated time under the paper’s structural model (Kurisu et al., 1 May 2025). In applications, employment-share compositions after the 2011 Great East Japan Earthquake were mapped by square-root transform to $\lambda_t(s) := E_t[Y(0)\mid S=s]$8 and analyzed with the geodesic arc-length metric; age-specific fertility rate curves in East Germany were treated as $\lambda_t(s) := E_t[Y(0)\mid S=s]$9 functions; and age-at-death distributions for Russia after the Soviet collapse were analyzed in Wasserstein space with GSDID (Kurisu et al., 1 May 2025).

These applications illustrate that the choice of metric is substantive rather than merely computational. The paper notes that distributions may be studied under Wasserstein, Fisher–Rao, or sliced Wasserstein metrics; compositions under spherical or Aitchison-type geometries; SPD matrices under Frobenius, log-Euclidean, affine-invariant, power, or Log-Cholesky metrics; and functional data in $\gamma_j(s) := P_j(S=s)$0 (Kurisu et al., 1 May 2025). It also states that different metrics can yield different causal conclusions and identifies metric selection as an open problem (Kurisu et al., 1 May 2025).

Inference in the current framework is primarily permutation-based. The paper recommends placebo tests in which each donor is treated as pseudo-treated, synthetic paths are recomputed, and post-treatment distances are compared. It simultaneously lists several limitations: the need for unique geodesics and well-behaved Fréchet means, computational cost of barycenter calculation in some spaces, the finite-sample structural nature of the identification results, and the absence of a general asymptotic inference theory (Kurisu et al., 1 May 2025).

From the perspective of synthetic control theory, the assumptions paper provides an additional diagnostic principle: suitable auxiliary covariates are those that are linear in the same $\gamma_j(s) := P_j(S=s)$1 probabilities as the outcome or are different measurements of the same outcome, whereas unsuitable covariates can bias the weights and harm counterfactual estimation (Shi et al., 2021). This is especially relevant when geometric distances in covariate or trajectory space are used to define donor neighborhoods.

6. Distinct usage in quantum sensing

In a separate literature on quantum metrology, “geodesic synthetic control” refers not to causal inference but to geometric control of a quantum sensor’s frequency response. "Experimental High-Accuracy and Broadband Quantum Frequency Sensing via Geodesic Control" studies the electron spin of a single nitrogen-vacancy center in diamond with bare Hamiltonian

$\gamma_j(s) := P_j(S=s)$2

and introduces a control Hamiltonian chosen so that the Bloch vector follows piecewise geodesic arcs on the Bloch sphere with a continuously rotating rotation axis (Chen et al., 27 Jan 2026). In the paper’s own summary, thinking of “geodesic synthetic control” in this context means using geodesic trajectories of the Bloch vector as a synthetic control tool to synthesize a desired, highly selective frequency response of a quantum sensor, with strong suppression of higher harmonics and associated systematic errors (Chen et al., 27 Jan 2026).

For the MHz protocol, the effective rotating-frame control Hamiltonian is

$\gamma_j(s) := P_j(S=s)$3

and in the large-$\gamma_j(s) := P_j(S=s)$4 limit the induced toggling-frame modulation function satisfies

$\gamma_j(s) := P_j(S=s)$5

The accumulated phase is

$\gamma_j(s) := P_j(S=s)$6

so the filter function is sharply peaked at the scan frequency. By contrast, for XY and CPMG dynamical decoupling, the modulation functions contain infinite odd harmonics,

$\gamma_j(s) := P_j(S=s)$7

which produce spurious responses at $\gamma_j(s) := P_j(S=s)$8 (Chen et al., 27 Jan 2026).

The paper reports experimentally reconstructed filter spectra with a single dominant peak at $\gamma_j(s) := P_j(S=s)$9 for geodesic protocols and multi-harmonic structure for XY/CPMG, robustness against injected harmonic noise, bias-free frequency estimation in multi-frequency environments, and synchronized readout achieving $\{\beta_j\}_{j\in D}$0 resolution for GD$\{\beta_j\}_{j\in D}$1 and $\{\beta_j\}_{j\in D}$2 for GD$\{\beta_j\}_{j\in D}$3 (Chen et al., 27 Jan 2026). The shared structural motif with causal geodesic synthetic control is not identity of method but use of geometry to synthesize an effective response: in one case a counterfactual object in a metric space, in the other a single-frequency filter in the toggling frame.