Shadow Variable: Theory & Applications

• Shadow variable is an auxiliary construct that makes latent structures accessible, evident in fields from MNAR inference to quantum state reconstruction.
• In missing-data analysis, it restores identifiability by decoupling outcome dependence from nonresponse through observed proxies.
• In graphics and celestial theory, shadow variables serve as explicit geometric controls or basis operators that ensure model consistency and enhance inference.

“Shadow variable” is a context-dependent technical term rather than a single standardized concept. In missing-data theory it denotes a fully observed variable associated with an outcome but conditionally independent of the missingness mechanism; in practical continuous-variable quantum shadow estimation it denotes a random operator-valued snapshot reconstructed from a single randomized homodyne measurement; in image compositing it denotes explicit geometric or appearance controls such as pixel height maps and softness parameters; in celestial conformal field theory it is realized as a shadow-basis operator; and in shadow-aware satellite reconstruction it appears as explicit solar-visibility and shadow-map variables (Miao et al., 2015, Miao et al., 2015, Li et al., 7 Jun 2026, Yang et al., 15 Dec 2025, Sheng et al., 2022, Liu et al., 18 Jun 2026, Luo et al., 4 Jan 2026). A plausible unifying interpretation is that a shadow variable is an auxiliary object that makes latent structure accessible to inference, reconstruction, or operator algebra.

1. Terminological scope and taxonomy

Across the cited literature, the term names objects with sharply different ontological status. Some are ordinary observed covariates, some are random estimators, some are geometry-encoding image fields, and some are conformal primaries. What they share is not substance but role: each is introduced to mediate access to an otherwise inaccessible quantity.

Domain	Shadow variable	Immediate role
MNAR statistics	Fully observed $Z$	Identifies full-data law
DID with MNAR	Covariate component $Z$	Identifies ATT under MNAR
CV quantum estimation	$\hat{\rho}_{i,k}$	Unbiased single-shot snapshot
Image compositing	Pixel height map, softness	Controls shadow geometry and penumbra
Celestial OPEs	Shadow-basis operator	Completes OPE closure
Satellite 3DGS	$S_{\text{sun},i}$, $S$	Encodes solar visibility and shadows

This dispersion of meanings makes local definition essential. A common misconception is that “shadow variable” always refers to a proxy variable in statistics or always refers to an optical shadow in graphics. The cited work shows instead that the term ranges from semiparametric identification devices to operator-valued estimators and boundary conformal primaries. This suggests that the phrase should always be interpreted relative to the model class in which it is introduced.

2. Shadow variables in missing-not-at-random inference

In the missing-data literature, a shadow variable is a fully observed outcome proxy used to recover identification under missing not at random (MNAR) missingness. The foundational assumption is that $Z$ is associated with the outcome $Y$ but independent of the missingness indicator $R$ conditional on $Y$ and fully observed covariates $X$. One formulation is

$Z$0

while a closely related formulation is

$Z$1

Under these assumptions, the joint law can be parameterized through a pattern-mixture decomposition and an odds-ratio function $Z$2, and identification is obtained by a completeness condition imposed on the complete-case distribution $Z$3 (Miao et al., 2015).

The main identification logic is that the observed distribution of $Z$4 among respondents and nonrespondents constrains the unobserved dependence of $Z$5 on $Z$6. In the pattern-mixture framework, the odds ratio

$Z$7

reduces, under the shadow-variable assumption, to a function of $Z$8 alone. The resulting Fredholm integral equation of the first kind links $Z$9 to the observable ratio $\hat{\rho}_{i,k}$0. When the completeness condition holds, this equation has a unique solution, so the full joint distribution $\hat{\rho}_{i,k}$1 is nonparametrically identified (Miao et al., 2015).

This framework supports semiparametric estimation as well as efficiency theory. The literature develops regression-based, inverse-probability-weighted, and doubly robust estimators for generic full-data functionals $\hat{\rho}_{i,k}$2. It also derives the semiparametric efficiency bound, the efficient score for the odds-ratio parameter $\hat{\rho}_{i,k}$3, and a closed-form projection formula for the efficient influence function. In a related development, three doubly robust estimators for $\hat{\rho}_{i,k}$4 are proposed: a regression estimator with residual bias correction $\hat{\rho}_{i,k}$5, a Horvitz–Thompson estimator with extended weights $\hat{\rho}_{i,k}$6, and a regression estimator with an extended outcome model $\hat{\rho}_{i,k}$7. Their double robustness is with respect to the baseline outcome model and baseline propensity model, conditional on a correctly specified log odds-ratio model $\hat{\rho}_{i,k}$8; the same framework also yields goodness-of-fit diagnostics through extension parameters $\hat{\rho}_{i,k}$9 and $S_{\text{sun},i}$0, which converge to zero if the corresponding baseline model is correct (Miao et al., 2015).

A frequent source of confusion is the relation between a shadow variable and an instrumental variable for nonresponse. The two are explicitly contrasted in this literature. An instrumental variable for missingness typically affects $S_{\text{sun},i}$1 but is independent of $S_{\text{sun},i}$2 conditional on $S_{\text{sun},i}$3; a shadow variable instead is associated with $S_{\text{sun},i}$4 and does not affect $S_{\text{sun},i}$5 beyond $S_{\text{sun},i}$6 and $S_{\text{sun},i}$7. The direction of exclusion is therefore reversed.

3. Semiparametric difference-in-differences with MNAR outcomes

The difference-in-differences extension treats the shadow variable as part of the covariate vector $S_{\text{sun},i}$8 in a two-period setting with outcome evolution $S_{\text{sun},i}$9, treatment indicator $S$0, and post-treatment response indicator $S$1. The parameter of interest is the treatment effect on the treated,

$S$2

The paper states the shadow-variable property as

$S$3

and describes $S$4 as fully observed, related to the outcome evolution, but independent of the missingness mechanism once conditioning variables are included. Under this structure, the odds ratio

$S$5

simplifies to $S$6, yielding an identification equation in which the observable density ratio $S$7 determines the MNAR response mechanism and the nonrespondent outcome distribution (Li et al., 7 Jun 2026).

Once the odds ratio is identified, the response probability

$S$8

and the nonrespondent mean

$S$9

become recoverable. This leads to an MNAR identification formula for the ATT,

$Z$0

The formulation nests the MAR case: if $Z$1 collapses to a function of $Z$2, then $Z$3 reduces to the standard MAR estimand (Li et al., 7 Jun 2026).

The proposed estimator is semiparametric. It specifies a treatment model $Z$4, a baseline response model $Z$5, and an odds-ratio model such as

$Z$6

Parameters $Z$7 are estimated by generalized method of moments using

$Z$8

and $Z$9 is estimated by a weighted calibration equation. The final ATT estimator is

$Y$0

Under correct specification of the missingness mechanism and propensity score, the estimator is consistent and asymptotically normal with sandwich variance $Y$1. The paper also proves a double-robust-type property conditional on correct $Y$2: consistency is preserved if either the propensity score or the control outcome model is correct (Li et al., 7 Jun 2026).

The simulation study contrasts a naive complete-case DID estimator, a MAR estimator, and the MNAR shadow-variable estimator under $Y$3 and $Y$4. The empirical application to China’s two-child policy uses $Y$5 hukou as shadow variable, reports $Y$6 observations, about $Y$7 missing post-treatment debt, about $Y$8 treated, and estimates $Y$9. The ATT on log debt is $R$0 under MAR DID and $R$1 under MNAR DID, illustrating that shadow-variable-based correction can materially alter inference when nonresponse depends on the unobserved outcome evolution (Li et al., 7 Jun 2026).

4. Shadow variables in continuous-variable quantum shadow estimation

In continuous-variable quantum information, a shadow variable is not a covariate but a random operator-valued estimator associated with a single randomized measurement outcome. “Practical Homodyne Shadow Estimation” develops this notion for discretized homodyne detection in a truncated Fock space $R$2. The local oscillator phase is restricted to

$R$3

and the quadrature axis is partitioned into finitely many bins $R$4. The corresponding POVM elements are

$R$5

with outcome probability $R$6. The measurement channel is defined by

$R$7

and the single-shot shadow variable is

$R$8

Its defining property is unbiasedness: $R$9 For any observable $Y$0, the scalar estimator $Y$1 is then unbiased for $Y$2 (Yang et al., 15 Dec 2025).

The existence of $Y$3 requires informational completeness of the discretized POVM. The paper proves a sufficient condition: $Y$4 under which there exists a choice of quadrature bins making the POVM informationally complete. It also proves necessary conditions on $Y$5: either $Y$6, or $Y$7 with odd $Y$8. If $Y$9, or if $X$0 with even $X$1, the paper constructs explicit pairs of states with identical measurement statistics. An explicit Algorithm 1 selects equal-spaced bins over a growing range until the measurement matrix reaches full rank $X$2 (Yang et al., 15 Dec 2025).

The variance analysis introduces a shadow norm

$X$3

with Bernstein’s inequality linking $X$4 to sample complexity. The main bound is

$X$5

Using $X$6, the shadow norm scales as $X$7, improving earlier $X$8-type bounds. The term “shadow variable” here therefore refers to a randomized classical representation of a quantum state, not to an auxiliary covariate or a literal optical shadow (Yang et al., 15 Dec 2025).

5. Image-space, rendering, and remote-sensing shadow variables

In image compositing, the central shadow variable is a 2.5D geometric representation called pixel height. For an object point $X$9 with ground footpoint $Z$00, the pixel height is

$Z$01

A pixel height map assigns this value to every pixel in the object mask. Given a light position $Z$02 with pixel height $Z$03, the shadow point $Z$04 is computed analytically by

$Z$05

In this framework, the pixel height map is the geometric shadow variable controlling direction and shape, while the softness parameter $Z$06 controls penumbra width and blur. The method learns pixel height from RGB, object mask, and Y-Coordinate Map using HENet with a MiT backbone, trained with per-pixel MSE and total variation regularization; soft shadows are generated by a U-Net-style decoder with AdaIN modulation from a log-binned softness embedding. On Real1500, the baseline YCM has Abs error $Z$07 and relative error $Z$08, whereas the best HENet configuration reports Abs error $Z$09 and relative error $Z$10. In soft-shadow evaluation, SSN has mean Abs $Z$11 and mean ZNCC $Z$12, while SSG reports mean Abs $Z$13 and mean ZNCC $Z$14; in a $Z$15AFC user study, the full system is preferred $Z$16 of the time (Sheng et al., 2022).

In shadow-aware satellite reconstruction, the term is attached to explicit shadow-state variables in a 3D Gaussian Splatting pipeline. ShadowGS assigns each Gaussian a solar visibility $Z$17 and renders a per-pixel shadow map $Z$18 alongside skylight radiance $Z$19, near-surface reflection $Z$20, and albedo $Z$21. These are blended as

$Z$22

with incident radiance

$Z$23

and final color

$Z$24

Solar visibility is computed by ray marching through Gaussians: $Z$25 The model further imposes a shadow consistency loss

$Z$26

an entropy regularizer

$Z$27

and a sparse-view shadow prior $Z$28 from FDRNet masks. In ablation, adding the shadow consistency constraint reduces MAE from $Z$29 m to $Z$30 m and increases PSNR from $Z$31 dB to $Z$32 dB. In sparse-view evaluation over $Z$33 JAX AOIs, the shadow prior improves mean MAE from $Z$34 m to $Z$35 m and PSNR from $Z$36 dB to $Z$37 dB (Luo et al., 4 Jan 2026).

These two graphics-oriented meanings differ sharply. In pixel-height compositing, the shadow variable is an explicit image-space surrogate for object–ground geometry; in ShadowGS, it is an illumination-geometry state variable embedded in a differentiable renderer. Both, however, externalize shadows as controllable model components rather than letting them be absorbed implicitly into texture or appearance.

6. Shadow-basis operators and shadow completion in celestial OPEs

In celestial conformal field theory, the relevant object is a shadow-basis operator. For a primary $Z$38 on the celestial sphere, the shadow transform is

$Z$39

with

$Z$40

and $Z$41, $Z$42. The paper argues that the ordinary celestial OPE does not close on Mellin-basis exchanges alone, because Mellin–Mellin two-point functions are contact-supported, whereas the OPE limit of regular celestial amplitudes requires non-contact behavior. The resolution is a shadow-completed OPE in which the same exchanged bulk particle appears both in Mellin basis and in shadow basis (Liu et al., 18 Jun 2026).

For scalar $Z$43 theory, the ordinary collinear coefficient is

$Z$44

and the shadow OPE coefficient is fixed, not independent: $Z$45 The two are related by the universal shadow factor $Z$46, obtained from the star–triangle relation. The resulting scalar OPE contains both the Mellin representative and the shadow representative of the exchanged particle. Analogous shadow-completed OPEs are derived for gluons and gravitons, with the shadow transform flipping helicity as $Z$47 (Liu et al., 18 Jun 2026).

A central clarification is that shadow-basis operators do not add new bulk degrees of freedom. Mellin and shadow bases are two conformal-primary bases for the same bulk one-particle representation, but they define distinct local primary states in the boundary theory. The paper shows that the shadow primary is not, in general, a convergent linear combination of descendants of the Mellin primary. In this sense, the “shadow variable” language here names a basis completion mechanism in boundary operator algebra, not a proxy variable, estimator, or geometric control. This is another common misconception dispelled by the literature.

The cross-disciplinary record therefore supports a strongly local reading of the term. In statistics, a shadow variable is an observed auxiliary variable that restores identifiability under MNAR. In quantum tomography, it is a random snapshot generated by inverting a measurement channel. In graphics and remote sensing, it is an explicit state variable governing shadow geometry or visibility. In celestial theory, it is a shadow transform of a conformal primary required for OPE consistency. The shared motif is structural indirection: each shadow variable encodes latent content through a representation that is experimentally observable, computationally reconstructible, or algebraically necessary.