**A complete instrument generates and analyzes a spanning set of polarization states.** A polarization-state generator placed before the sample creates known incident Stokes states, and a polarization-state analyzer after the sample measures output projections. In matrix form, a collection of detected intensities can be represented schematically as $$ \mathbf B=\mathbf A\mathbf M\mathbf W $$ where $\mathbf W$ characterizes generated states and $\mathbf A$ characterizes analyzer response. If both are invertible and well conditioned, the sample matrix can be reconstructed. Real systems use rotating compensators, photoelastic modulators, liquid-crystal retarders, division-of-amplitude channels, or other architectures whose modulation and demodulation models vary with wavelength. State diversity matters as much as count. Nearly identical states make inversion noise-sensitive; generator and analyzer condition numbers quantify that amplification. Useful states remain well distributed on the Poincaré sphere across the spectrum. Calibration must estimate the behavior of actual polarizers, retarders, modulators, mirrors, windows, detector channels, and azimuth offsets. Retardance is wavelength- and temperature-dependent; diattenuation and detector response can vary spectrally; rotation stages have zero and eccentricity errors. Eigenvalue calibration and related self-consistent procedures use reference elements to solve generator and analyzer matrices without assuming ideal components. Dark offsets, drift, nonlinearity, timing, stray light, and beam motion create correlated matrix errors. Air, isotropic mirrors, polarizers, and retarders test different functions; validation should use standards excluded from calibration. |Measurement scope|Observable set|Best suited sample|What it adds|Principal failure mode| |---|---|---|---|---| |Conventional $\Psi,\Delta$ ellipsometry|Amplitude ratio and phase difference in p/s basis|Isotropic nondepolarizing planar stack|Efficient thickness and scalar optical constants|Cross-polarization or depolarization forced into a wrong stack| |Selected generalized elements|Jones-like co- and cross-polarization terms|Deterministic anisotropic or patterned sample|Tensor axes and polarization conversion|Assuming nondepolarization when incoherent mixing exists| |Full Mueller matrix ellipsometry|Sixteen absolute or fifteen normalized transfer elements|Anisotropic and/or depolarizing sample|Diattenuation, polarizance, retardance, and depolarization constraints|Calibration error or model ambiguity across many correlated elements| |Spectroscopic Mueller mapping|Matrix versus wavelength and position|Spatially heterogeneous films or patterns|Domains, gradients, and validity masks|Pixel/footprint mixing and drift masquerading as depolarization| |Angle- and azimuth-resolved Mueller data|Matrix versus energy, incidence, and rotation|Crystals, gratings, metamaterials, complex stacks|Higher identifiability of dielectric tensors and geometry|Registration and convention errors across configurations| **Physical realizability must be checked before decomposition or fitting.** Not every arbitrary real 4×4 matrix maps physically allowable input Stokes vectors to allowable outputs. Noise and calibration error can yield negative intensities for some input state, polarization degree above unity, or a non-positive covariance/coherency representation. A physical projection may be appropriate, but it changes the data and uncertainty and must not conceal systematic instrument error. Checks include nonnegative output intensity, bounded diattenuation and polarizance, and positive coherency construction. Validate software against known matrices; element-wise clipping does not guarantee physicality and distorts correlations. Reciprocity and symmetry constrain specific sample classes. An isotropic planar reflector has a sparse matrix; anisotropic reciprocity relations require transformed forward and reverse frames. Deviations can also arise from alignment, azimuth, depolarization, or calibration. Uncertainty is matrix-valued because elements share intensity and calibration errors. Equal independent weighting can bias a fit; estimate covariance from propagation or repeats and use it in the residual metric. Normalization by noisy $M_{00}$ correlates every element and magnifies low-throughput noise. Save absolute data and distinguish polarization change from falling-reflectance normalization. **Depolarization usually means unresolved statistical mixing, not destruction at one ideal interface.** A deterministic homogeneous sample transforms fully polarized input into fully polarized output, even when it rotates polarization or couples p and s. Partial depolarization appears when the measurement averages mutually incoherent or fluctuating responses over space, angle, wavelength, time, depth, or multiple paths. Common causes include thickness or orientation variation inside the footprint, surface or volume scattering, mixed domains, finite source bandwidth, angular spread, backside reflection, patterned regions, and temporal change during modulation. The measured Mueller matrix describes the ensemble under that instrument’s resolution. A different footprint, numerical aperture, bandwidth, or integration time can produce a different depolarization index from the same specimen. Instrument imperfections can imitate sample depolarization. Unmodeled retardance dispersion, beam walk during rotating-element modulation, focus differences between states, detector integration mismatch, stray unpolarized light, or source instability reduces modulation contrast. Reference measurements across wavelength, angle, focus, and spot position must establish the instrument depolarization floor. Scalar depolarization metrics use different definitions and cannot identify whether variation arises from thickness mixture, roughness, domains, or multiple paths. Inspect the full matrix, spectrum, footprint dependence, and a mixture model. When the sample is a mixture of deterministic responses $\mathbf M_k$ with incoherent weights $w_k$, an ensemble representation is $$ \mathbf M_{mix}=\sum_k w_k\mathbf M_k,\qquad w_k\ge0,\quad\sum_k w_k=1 $$ This simple form illustrates why depolarization can encode unresolved heterogeneity, but the components and weights are generally not unique. A fitted two-domain mixture is a hypothesis requiring imaging, azimuth, footprint, or process evidence. **Mueller decomposition provides descriptors whose meaning depends on assumptions and order.** Polar decomposition methods factor a measured matrix into idealized depolarizer, retarder, and diattenuator matrices. Because matrix multiplication is not commutative, changing factor order changes derived parameters. The factors summarize the chosen algebraic representation; they are not automatically literal layers arranged in the specimen. Cloude or covariance decompositions express a physical Mueller matrix as an incoherent sum of nondepolarizing components and can provide rank or entropy-like measures. Differential decomposition uses a logarithmic or differential-generator viewpoint suited to distributed anisotropy and depolarization under its assumptions. Each approach answers a different question, and singular matrices, noise, branch choices, or strong effects can create instability. Retardance is phase delay between eigenpolarizations, diattenuation is differential attenuation, polarizance describes generated polarization from unpolarized input, and depolarization describes reduced polarization degree for an ensemble. Optical rotation, circular retardance, linear retardance, and reference-frame rotation can share similar matrix structure. Sign and axis ambiguities require declared conventions and often sample-azimuth measurements. Decomposition is valuable for visualization, anomaly detection, and initializing a physical model. It is usually not a substitute for solving Maxwell’s equations for the actual layered, anisotropic, or patterned structure. A decomposition-derived “linear retardance” does not by itself yield birefringence or film thickness because the same retardance can arise from different products of optical anisotropy and path length. Derived maps should include decomposition stability and physicality flags. Near low reflectance, matrix elements and decompositions become noisy. Angle wrapping, eigenvalue ordering, and axis degeneracy can create discontinuous color maps even when the sample varies smoothly. Unwrap and regularize only with documented rules, and preserve the raw matrix. **Anisotropic films and periodic structures require a forward electromagnetic model.** For a homogeneous anisotropic layer, the dielectric response is a tensor whose principal values and Euler orientation enter the propagation problem. Berreman-type 4×4 transfer methods or equivalent formalisms handle coupled field components through stratified anisotropic media. Multiple wavelengths, angles, and sample azimuths help separate tensor elements, thickness, and orientation. Generalized ellipsometry often refers to deterministic p–s coupling described through Jones reflection or transmission matrices. Mueller matrix ellipsometry includes that information while also detecting depolarization. The names overlap in practice, so the reported observables—Jones terms, selected Mueller elements, or full matrix—should be stated instead of relying on the technique label. Periodic gratings and semiconductor structures require rigorous coupled-wave analysis, finite-element, finite-difference, or another validated electromagnetic solver. Pitch, height, linewidth, sidewall angle, corner rounding, overlay, material optical constants, roughness, and line-width variation can all influence the matrix. Mueller elements add polarization diversity, but geometric parameters remain correlated and must be constrained by design information or orthogonal metrology. For patterned structures, azimuth is especially powerful: rotating the grating relative to the plane of incidence changes cross-polarization and symmetry. An incorrect azimuth or sample tilt can resemble structural asymmetry. Fit or calibrate alignment parameters, and acquire symmetry-related azimuths to separate geometry from stage error. Circular terms can arise from chirality or magneto-optics, but also coordinate error, retarder offset, and off-axis linear anisotropy. Use azimuth and reversal tests plus a model that excludes these artifacts. ```flowchart Define whether anisotropy, cross-polarization, or depolarization drives the decision -> Fix Stokes handedness, reference frames, normalization, wavelength, angle, and azimuth -> Calibrate PSG and PSA matrices, detector response, timing, and instrument depolarization floor -> Validate with independent isotropic, polarizer, retarder, and depolarizing references -> Acquire complete intensity states with repeats and drift monitors -> Reconstruct Mueller matrices with covariance and physical-realizability tests -> Inspect raw elements, symmetry, absolute throughput, residuals, and footprint dependence -> Apply declared decomposition only for bounded descriptive questions -> Fit a physical anisotropic, mixture, or patterned-structure forward model -> Confirm material or geometry parameters using azimuths, angles, and orthogonal metrology ``` **A production-ready method preserves the matrix, its covariance, and its conventions.** The recipe should freeze source spectrum, incidence angle, spot and footprint, sample azimuth, focus, polarizer and compensator states, modulation frequencies, detector settings, wavelength grid, normalization, coordinate frame, handedness, calibration artifacts, reconstruction algorithm, physical projection, and exclusion rules. Store raw intensity harmonics or state measurements, calibrated PSG and PSA matrices, absolute and normalized Mueller elements, covariance, physicality metrics, decomposition outputs, model predictions, residuals, and acquisition timestamps. A table of derived retardance and depolarization without the original matrix cannot be reinterpreted when conventions or decomposition methods change. Monitor calibration condition numbers, reference-matrix residuals, $M_{00}$ throughput, repeatability, and the instrument’s apparent depolarization. Validate after source, detector, polarizer, compensator, objective, angle, or alignment changes. Spectral regions with weak modulation or poor state conditioning should be masked by rule rather than rescued by unconstrained inversion. Report only parameters identifiable within the measured wavelength, angle, azimuth, and footprint range. A full matrix can reveal that a scalar model is invalid; it does not guarantee that a unique complex model exists. The strongest result combines physically valid matrices, calibrated uncertainty, forward-model agreement, symmetry tests, and orthogonal structural evidence. The durable way to interpret Mueller matrix ellipsometry is through a Stokes-convention-state-generation-matrix-physicality-depolarization-decomposition-forward-model-and-uncertainty lens.