Smith Shadowing Assumption: Definition and Applications

• The Smith Shadowing Assumption has varied definitions, referring to four different concepts in physics, statistics, dynamical systems, and computer graphics, each with unique characteristics and contexts.
• In radiation physics, it's an assumption for exploiting the equivalence between real and pseudo-photon electron-field components for modeling the transverse field's suppression and recovery downstream after an absorber.
• In dynamical systems and computer graphics, it involves verifying functional stability and distinct visibility conditions such as path-level geographical correlated visibility

The Smith Shadowing Assumption is not a single standardized hypothesis across the literature. The name may refer to distinct concepts associated with relativistic-electron field shadowing, statistical trajectory consistency, ordinary dynamical shadowing, or Smith microfacet visibility. Among the supplied research papers, only “Smith Shadowing Assumption” (Young et al., 2019) explicitly attributes a criterion to Smith et al.; the other papers use “shadowing” in dynamical-systems or microfacet-rendering contexts without invoking a Smith-named assumption. The most precise treatment therefore distinguishes these meanings rather than treating them as one theorem or universal principle.

1. Terminological scope and attribution

The expression has at least four technically distinct uses. In relativistic-electron radiation, it denotes the assumption that an absorbing screen removes, or “shadows,” the electromagnetic field carried by a relativistic electron over a downstream region (Naumenko et al., 2011). In high-dimensional data assimilation and forecasting, the Smith criterion states that a model trajectory is consistent with observations when the distribution of forecast residuals is equivalent to the observational-error distribution (Young et al., 2019). In topological dynamics, shadowing is the property that sufficiently accurate pseudo-orbits are uniformly followed by genuine orbits; the papers on genericity, chaos, and entropy do not attribute this property to Smith (Meddaugh, 2018, Kawaguchi, 2021, Kawaguchi, 2023). In computer graphics, the Smith assumption is a spatial-statistical homogeneity assumption for microfacet normal distributions and associated visibility statistics (Cui et al., 2023).

Consequently, the phrase Smith Shadowing Assumption should not be used without specifying its domain. The dynamical-systems papers explicitly state that they neither formulate nor cite a result under that name (Meddaugh, 2018, Kawaguchi, 2021, Kawaguchi, 2023). The microfacet paper likewise uses “Smith assumption” in connection with stationary normal-distribution statistics rather than pseudo-orbit tracing (Cui et al., 2023).

2. Relativistic-electron field shadowing

In the radiation context, the assumption concerns the electromagnetic field of a relativistic electron passing through or near an absorbing or conducting screen. The electron’s field is decomposed into transverse and longitudinal components,

$$\mathbf E_e(\rho,z,\omega)= \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix},$$

where $\mathbf E_\perp$ is transverse to the electron velocity and $E_\parallel$ is parallel to it. For relativistic motion, the paper gives the approximate intensity ratio

$$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$

The conventional shadowing picture treats the transverse field as a collection of equivalent or “pseudo-photon” components that can be intercepted by a screen. Immediately downstream, the electron is regarded as semi-bare: its transverse field is reduced and subsequently recovers through propagation and diffraction over a characteristic recovery or formation scale.

The central experimental result is that this picture is component-dependent. The absorbing screen suppresses the transverse component downstream, but no corresponding distance-dependent suppression of the longitudinal component is observed over the experimentally accessible region (Naumenko et al., 2011). The paper models this behavior as

$$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$

rather than as proportional attenuation of both components,

$$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$

Here $\alpha(L)$ describes transverse-field suppression and recovery as a function of absorber-to-target distance $L$. The alternative proportional model is inconsistent with the measured backward diffraction-radiation asymmetry.

The experiment used a $6.1\ \mathrm{MeV}$ electron beam with $\gamma=12$, wavelengths of approximately $\mathbf E_\perp$0–$\mathbf E_\perp$1, and absorber-to-target distances up to approximately $\mathbf E_\perp$2. A conducting mirror inclined at $\mathbf E_\perp$3 generated backward diffraction radiation. The longitudinal field was inferred indirectly because it produces an asymmetry in the angular distribution. If the maxima have intensities $\mathbf E_\perp$4 and $\mathbf E_\perp$5, the asymmetry is

$\mathbf E_\perp$6

The average intensity is dominated by the transverse field, whereas $\mathbf E_\perp$7 is sensitive to interference between transverse and longitudinal contributions. The measured intensity dependence was used to infer $\mathbf E_\perp$8, and the resulting asymmetry agreed with the model in which only the transverse field is distance-dependent (Naumenko et al., 2011).

The relevant transverse-field recovery scale is of order $\mathbf E_\perp$9, whereas the far-field criterion involves $E_\parallel$0. For $E_\parallel$1 and $E_\parallel$2–$E_\parallel$3, $E_\parallel$4 is approximately $E_\parallel$5–$E_\parallel$6, comparable with the scanned distance range. The experiment therefore probes transverse-field recovery rather than the full conventional radiation formation length of order $E_\parallel$7.

The result has implications for Smith–Purcell radiation, grazing-incidence diffraction radiation, backward diffraction radiation, and related near-field interactions with material boundaries. An upstream absorber need not reduce the radiation signal according to a law derived from uniform shadowing of the entire electron field. The longitudinal contribution may remain available for interaction with a grating or conducting surface. The conclusion is qualified by the absence of a direct measurement of $E_\parallel$8, the phenomenological nature of $E_\parallel$9, the finite distance range, detector uncertainty, beam divergence, and the specific absorber–mirror geometry (Naumenko et al., 2011).

3. Smith’s statistical trajectory-consistency criterion

In high-dimensional forecasting, Smith et al.’s assumption is statistical rather than electromagnetic. Let $$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$0 denote forecast residuals and $$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$1 realizations of observational noise. The consistency criterion is

$$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$2

where $$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$3 is the distribution of forecast residuals and $$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$4 is the distribution specified by the observational-error model (Young et al., 2019). A candidate trajectory shadows observations while this equivalence holds.

For a model–observation pair, the ideal shadowing time is

$$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$5

where the maximum is over all possible initial conditions or candidate trajectories. A finite experiment can test only finitely many candidates, so the largest measured candidate shadowing time is a lower bound on $$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$6.

The original sequence method uses the Euclidean residual length $$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$7. Over progressively longer trajectory segments, its median and 90th percentile are compared with the corresponding statistics generated by the observational-noise model. The first failed statistical test determines the candidate shadowing time. This approach becomes computationally expensive and statistically fragile when the state dimension is large: generating high-dimensional noise samples is costly, isolated outliers can distort short segments, and short segments provide poor estimates of the 90th percentile.

A modified state method instead evaluates the distribution of the $$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$8 residual components at each observation time. It tests the residual median and 90th percentile against analytical order-statistic distributions implied by the observational-noise model. If $$\frac{E_\parallel^2}{E_\perp^2}\sim\frac{1}{\gamma^2}.$$9 are independent draws with cumulative distribution function $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$0, the cumulative distribution function of the $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$1-th order statistic is

$$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$2

The method is intended for high-dimensional states, approximately $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$3 or larger, and is not appropriate for a three-dimensional system such as Lorenz ’63 when estimating a 90th percentile from the state components.

The method was evaluated with the MORALS rotating-annulus model, whose state dimension is $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$4. The perfect-model scenario was used, so the model and system shared the same numerical evolution operator. Observations were generated every $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$5 with independent Gaussian errors scaled to $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$6 of local natural variability. Candidate initial conditions were randomly oriented perturbations on a scaled-space hypersphere of radius $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$7.

Repeated testing requires stringent control of Type-I errors. Because two statistical tests are performed at each verification time, the false-rejection probability accumulates over the sequence of tests. For the reported ensembles, significance levels of $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$8 and $$\mathbf E(L)= \begin{pmatrix} \alpha(L)\mathbf E_\perp\ E_\parallel \end{pmatrix},$$9 were used. At the recommended value $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$0, candidate shadowing times for the $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$1 ensemble ranged from approximately $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$2 to $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$3, broadly agreeing with visual and distance-based divergence.

The expected candidate shadowing time decreased approximately logarithmically with initial scaled distance:

$$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$4

with $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$5 in seconds. The relation is empirical and applies to randomly oriented off-attractor perturbations in the chosen parameter regime. It is not a universal law and does not establish the existence of an optimal or mathematically guaranteed shadowing trajectory. Finite candidate ensembles provide lower bounds; candidates on an attractor or stable manifold could shadow for longer. Conversely, perfect-model results exclude model error and may be optimistic for real applications (Young et al., 2019).

4. Ordinary shadowing in topological dynamics

In dynamical systems, ordinary shadowing is a quantified pseudo-orbit tracing property. For a compact metric space $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$6 and continuous map $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$7, a sequence $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$8 is a $$\mathbf E(L)=\alpha(L) \begin{pmatrix} \mathbf E_\perp\ E_\parallel \end{pmatrix}.$$9-pseudo-orbit if

$\alpha(L)$0

It is $\alpha(L)$1-shadowed by $\alpha(L)$2 if

$\alpha(L)$3

The map has the shadowing property when

$\alpha(L)$4

This is a one-sided, uniform, all-time property. It does not require surjectivity, expansivity, invertibility, or uniqueness of the tracing point in the principal results of the cited papers (Kawaguchi, 2021, Kawaguchi, 2023).

The paper “On genericity of shadowing in one dimension” proves that shadowing is Baire-generic among continuous maps and continuous surjections on graphites (Meddaugh, 2018). A graphite is a locally connected graphoid: a one-dimensional graph-like continuum approximable by finite graphs through retractions with arbitrarily small fibers. The class includes graphs, dendrites, the Menger curve, the Sierpiński carpet, the Sierpiński gasket, and certain locally connected Julia sets.

The proof uses arbitrarily fine finite open covers whose members have small diameter, nonempty core, a taut structure, and a nerve that is a graph without three-cycles. Maps are approximated by finite graph maps, and admissible cover-transition patterns are realized by genuine orbits. The construction yields dense open sets $\alpha(L)$5 of maps for which sufficiently accurate pseudo-orbits are $\alpha(L)$6-shadowed. Their intersection is a dense $\alpha(L)$7 set of shadowing maps.

The paper does not prove genericity for homeomorphisms on graphites and does not establish that every locally connected one-dimensional continuum is a graphite. It therefore supports a genericity theorem for ordinary shadowing, not an assumption attributed to Smith (Meddaugh, 2018).

5. Shadowing, chaos, and local entropy

Under ordinary shadowing, chain-theoretic properties can be transferred to genuine-orbit properties. For a continuous map on a compact metric space, a finite $\alpha(L)$8-chain satisfies

$\alpha(L)$9

Chain recurrence, chain components, chain proximality, and chain sensitivity describe approximate orbit structure. The paper “Generic and dense distributional chaos with shadowing” proves equivalences between these chain properties and several forms of generic or dense chaos under the shadowing hypothesis (Kawaguchi, 2021).

For generic uniform chaos, the following are equivalent under shadowing: chain sensitivity together with chain proximality; chain proximality together with a unique nontrivial chain-stable chain component; generic uniform chaos; dense uniform chaos; and generic chaos. Thus, shadowing acts as the bridge that converts approximate chain behavior into actual orbit separation and recurrence.

For DC1 chaos, the corresponding equivalences involve a pseudo-orbit separation property and the existence of a distal pair in the unique chain-stable component. The analogous result for DC2 is weaker: even with shadowing, dense DC2 does not imply generic DC2. The cited examples also show that dense DC1 need not imply generic chaos, and that the chain conditions can fail to produce distributional chaos when shadowing is absent (Kawaguchi, 2021).

The paper “Some results on shadowing and local entropy properties of dynamical systems” studies pointwise shadowing. A point $L$0 is shadowable if every sufficiently accurate pseudo-orbit beginning at $L$1 is uniformly shadowed. The chain-accessible set

$L$2

organizes the local consequences of shadowability. The paper proves that $L$3 is shadowable if and only if $L$4 consists of shadowable points, equivalently if $L$5 has shadowing on $L$6 (Kawaguchi, 2023).

For a shadowable point, positive local entropy at some fixed scale occurs precisely when $L$7 contains a chain component that is neither a periodic orbit nor an odometer. Uniform local entropy near $L$8 is equivalent to positive topological entropy of $L$9. Under global shadowing and chain transitivity, failure of $6.1\ \mathrm{MeV}$0-expansiveness is equivalent to the existence of arbitrarily small-scale separated chain structures and factor dynamics onto the full one-sided shift.

These results concern ordinary shadowing as a dynamical property. They neither define nor invoke a Smith-named assumption.

6. Smith microfacet visibility and multiple scattering

In computer graphics, the Smith assumption has a separate meaning. The microfacet normal-distribution function is assumed to be statistically invariant across positions in the microgeometry: “the NDFs are always the same regardless of different positions” (Cui et al., 2023). This is a stationarity or spatial-homogeneity assumption, not a claim that all microfacets are independent.

For a microfacet normal-distribution function $6.1\ \mathrm{MeV}$1, Fresnel factor $6.1\ \mathrm{MeV}$2, and Smith function $6.1\ \mathrm{MeV}$3, the conventional single-bounce geometric attenuation is represented by a masking–shadowing term. The height-correlated form is

$6.1\ \mathrm{MeV}$4

The paper’s single-bounce segment factor is

$6.1\ \mathrm{MeV}$5

with the sign determined by the convention that the incident direction points into the microgeometry. This reproduces the conventional height-correlated Smith shadowing–masking function.

For multiple scattering, the paper interprets the Smith microfacet model as equivalent to a semi-infinite homogeneous medium. A path is represented by a direction sequence

$6.1\ \mathrm{MeV}$6

Rather than multiplying independent per-bounce visibility factors, it writes the path contribution as

$6.1\ \mathrm{MeV}$7

where $6.1\ \mathrm{MeV}$8 is a local scattering or vertex term and $6.1\ \mathrm{MeV}$9 is a correlated path-level segment factor.

The invariance principle requires that adding an infinitesimal layer with the same optical properties as the existing semi-infinite medium leave total reflectance unchanged. The resulting recurrence is

$\gamma=12$0

subject to the validity conditions on the initial and final directions. The recurrence treats the two ends of the path symmetrically and captures correlations between multiple scattering segments.

The paper distinguishes this correlated formulation from independent-bounce approximations. It reports agreement with unbiased random-walk references, passage of the white-furnace test, reciprocity, and recovery of multiple-scattering energy lost by the classical single-bounce Smith BRDF. Its assumptions concern homogeneous microfacet statistics, Smith visibility functions, reflection-only transport, and the microfacet–medium equivalence; they do not concern pseudo-orbit tracing or statistical consistency with observations (Cui et al., 2023).

The phrase Smith Shadowing Assumption therefore has no single cross-disciplinary definition. In radiation physics it is challenged by evidence that transverse and longitudinal electron-field components shadow differently. In forecasting it denotes residual-distribution consistency between a model and observations. In topological dynamics, ordinary shadowing is an unattributed pseudo-orbit tracing property whose genericity and consequences are studied independently of Smith. In microfacet rendering, the Smith assumption is spatial statistical homogeneity leading to directional masking–shadowing factors and, in the multiple-bounce setting, correlated path-level visibility.