Mirage Effect: Waves, Quantum, and AI

Updated 29 March 2026

Mirage Effect is a phenomenon where gradients in media cause electromagnetic, quantum, or AI inferences to produce illusionary displacements or inversions.
It spans classical optics to engineered photonic structures and quantum systems, with measurable effects such as ray bending, OAM reversal, and mirage gaps.
Applications include cloaking, super-resolution imaging, and AI model debugging, highlighting the need for mitigation strategies in high-stakes evaluations.

A mirage effect is a generic term for a class of phenomena in which electromagnetic waves, quantum excitations, or AI system inferences produce an illusion—an apparent displacement, inversion, or presence of information that does not correspond to the actual input or environment. The archetypal example is optical mirages arising from refractive-index gradients in atmospheric or engineered media. Contemporary usage spans physical wave propagation, nanophotonics, computational imaging, quantum materials, and vision-LLMs in artificial intelligence. This entry provides an exhaustive analysis across domains grounded in primary experimental and theoretical literature.

1. Mirage Formation in Classical and Structured Optical Media

Refractive-Index Gradients and Atmospheric Mirages

Mirages originate from ray curvature in vertically stratified refractive-index profiles $n(z)$ . The two canonical atmospheric cases are:

Inferior mirage: Overheated ground creates air layers with decreasing $n(z)$ toward the surface (e.g., deserts, hot roads). Incident rays from the sky bend upward as $n(z)$ decreases, reaching the observer’s eye from below the horizon. The observer perceives an apparent water-like layer due to the "virtual" image of the sky beneath the horizon.
Superior mirage: Temperature inversion over cold surfaces (e.g., sea ice) produces increasing $n(z)$ near the ground. Rays from distant objects curve downward, yielding multiple, stack-inverted images above the object—the Fata Morgana being the most elaborate manifestation.

The complete ray path $z(x)$ is governed by the differential Snell law,

$n(z)\sin\theta(x)=\text{const},$

together with

$\frac{d^2 z}{dx^2} + \frac{-k}{n_0-kz}\bigg(1+\left(\frac{dz}{dx}\right)^2\bigg)=0,$

for linear gradients $n(z)=n_0-kz$ (Polachini et al., 4 Oct 2025, Guenneau et al., 2017). Solutions may be analytic for linear $n(z)$ or require numerical integration otherwise. Laboratory-scale analogs using sugar-solution or thermal gradients quantitatively reproduce these classical mirage trajectories, and reconstructions of distorted images show explicit agreement between ray-tracing computation and optical measurements (Polachini et al., 4 Oct 2025).

Engineered Media and Twisted Light

In structured media, engineered vertical gradients (water-ethanol columns, graded-index slabs) recreate the mirage effect for beams with orbital angular momentum (OAM). A smoothly varying index profile,

$n(z)=n_\mathrm{H_2O}+a\mu(z)+b[\mu(z)]^2+c[\mu(z)]^3,\quad \mu(z)=\frac{1+\mathrm{erf}((z-z_0)/d)}{2},$

induces continuous U-turn trajectories for OAM-carrying beams. The paraxial Helmholtz equation

$2ik_0 n_0 \frac{\partial \psi}{\partial x}+\frac{\partial^2\psi}{\partial x^2}+\frac{\partial^2\psi}{\partial z^2}+2k_0^2 n_0\delta n(z)\psi=0$

controls the propagation-envelope (Cisowski et al., 2019). The mirage effect here encompasses not only spatial inversion but also a reversal of OAM handedness ( $\ell\to-\ell$ ), with astigmatic mode mixing emerging at the plane of maximal shear. This continuous transition is distinct from abrupt inversion at sharp interfaces, resulting in complex, spatially varying OAM fluxes on the output face.

Metastructures and Photonic Carpets

Mirage effects are engineered at the micron scale using quasi-conformal mapping to design photonic-crystal “carpets” that scatter incident waves as if from a flat boundary even when geometric perturbations (bumps) are present. The conformal grid prescribes the local refractive index as

$n'(x',y') = n_0/\sqrt{u_x^2 + u_y^2},$

mapped onto a dielectric pillar lattice. Near-field measurements confirm the recovery of straight interference fringes despite the underlying bump, supporting the device's operation as an optical mirage based on transformation optics principles (Scherrer et al., 2013).

2. Mirage Effects in Diffusive and Quantum Regimes

Diffusive Photonics and Transformation Optics

Steady-state and time-dependent diffusion equations can be recast via coordinate transformations that produce effective material parameters (anisotropic, sign-alternating diffusivity tensors). This framework enables “mirage” cloaks that mask objects or create illusions analogous to ray-based atmospheric mirages. Specifically, space-folding maps generate domains in which an object plus its cloak reproduces the flux signature of a different target, effectively creating “anamorphic mirages” (e.g., transforming a circle into a square, or camouflaging external scatterers) (Guenneau et al., 2017). These concepts extend the mirage phenomenon to diffusive, non-wave regimes.

Quantum Mirage Gaps in Altermagnets

In altermagnet–superconductor heterostructures, intrinsic anisotropic exchange fields $T_J(k)=t_J(k_x^2-k_y^2)$ lead to the formation of energy bands with gapless nodal arcs and finite-energy “mirage gaps” in the quasiparticle spectrum,

$E_{\rm mirage}^\pm(\theta)=\pm\sqrt{[t_J k_F^2\cos2\theta]^2+\Delta^2},$

where $\Delta$ is the superconducting gap and $k_F$ the Fermi wavevector. The opening and splitting of these mirage gaps are reflected in quantum transport spectra as sharp conductance peaks, which quantify the underlying gapless superconductivity. Spin-singlet and triplet correlators (with $d$ -wave angular dependence) coexist, and the phenomenon is resolved via angle-resolved differential conductance (Wei et al., 2023).

3. Optical Mirage and Spinless Beams

Optical mirages have been widely attributed to spin–orbit coupling of light, leading to apparent far-field position shifts of scattering targets (spin-Hall and spin-dependent mirages). However, even spinless (linearly polarized) beams produce substantial spatial displacements (“mirages”) if the illuminated particle supports both electric and magnetic dipolar resonances with a nontrivial phase relationship. The shift,

$\Delta/\lambda = \frac{1}{\pi} \frac{\eta \sin\theta( \cos2\phi\,\hat{\theta} - \sin2\phi\cos\theta\,\hat{\phi})}{1-\gamma\sin^2\theta + 2g\cos\theta},\quad \eta = \mathrm{Im}(\alpha_E\alpha_M^*)/(|\alpha_E|^2+|\alpha_M|^2),$

where $\alpha_E$ , $\alpha_M$ are the electric and magnetic polarizabilities, originates from interference between the associated dipolar fields. The effect is maximized near points of strong electric-magnetic phase difference (off-Kerker conditions), attaining shifts exceeding several incident wavelengths (Olmos-Trigo et al., 2022).

4. Computational Mirage Effects: Imaging and AI

Emitter Localization in Resonant Environments

In super-resolution microscopy, a point emitter near a resonant (plasmonic) nanoparticle produces a composite far-field pattern comprising both direct and re-scattered fields. The observed point-spread function is shifted—a mirage effect—by tens of nanometers. By expanding the scattered field in the modal basis of the resonant object and correcting for the contribution of dominant modes, position errors far below the diffraction limit are achieved, restoring true localization. The modal correction involves solving a finite-dimensional nonlinear optimization without repeated PDE solutions, enabling rapid, noise-robust per-emitter position recovery (Baldassari et al., 2022).

Mirage Effect in Multimodal AI

In contemporary vision-LLMs, the mirage effect denotes the model’s production of detailed visual inferences and explanations for images that have not been provided. This is operationalized via “mirage mode,” where the image input is absent ( $X=\emptyset$ ) but the model continues to output appearances of visual reasoning, often with high decisiveness and without uncertainty hedges. The key metrics are:

Mirage Rate: Proportion of responses that do not acknowledge the missing image.
Mirage Score: Ratio of mirage-mode accuracy to standard accuracy,

$\text{MirageScore} = \frac{\text{Accuracy}^{\text{mirage}}}{\text{Accuracy}^{\text{original}}} \times 100\%.$

Frontier models (e.g., GPT-5.1, Gemini-3-Pro) exhibit mirage rates near 90–100% given evaluation prompts, and mirage scores up to 87% on standard visual question answering tasks—retaining much of their performance even when the image is withheld (Asadi et al., 23 Mar 2026).

A diagnostic regime distinction emerges between “mirage” (unacknowledged hallucination) and “explicit guessing” (where uncertainty is expressed and performance drops). These findings reveal deep susceptibility of public benchmarks to textual non-visual inference, including systematic biases such as pathology over-prediction in medical VQA. Benchmark cleaning protocols (e.g., B-Clean) remove test items that any candidate model can answer without visual input, substantially reducing model accuracy and altering rankings, especially in clinical datasets.

5. Unified Theoretical Principles and Mathematical Formalism

Key equations underpinning the mirage effect across physical and computational domains converge on the refractive (or effective potential) gradient-induced curvature of information-carrying trajectories—whether rays, quantum excitations, or probabilistic inferences. The common structure is a constraint or optimization over trajectories in an inhomogeneous medium or an information manifold, with the illusion manifest when the environment or input is incompletely specified or artificially mapped.

Ray optics / wave propagation: Eikonal equations, index mapping, or transformation optics encode the physical bending or redirection of wavefronts.
Quantum/transport: Mirage gaps emerge as finite-energy band edges displaced from zero in $k$ , $\theta$ .
Computational imaging: Modal expansions and linear/nonlinear inversion correct for far-field shifts engendered by composite scattering.
AI models: Mirage responses arise from the implicit default to knowledge-based or training-set prior inference, in the absence of true input, with diagnostic regimes governed by ablation of the input channel.

6. Experimental Realizations and Quantitative Benchmarks

Mirage phenomena have been realized and quantified in:

Sugar-gradient and water-ethanol tanks for classic optical mirages, with explicit measurement and theoretical reconstruction matching to sub-percent accuracy (Polachini et al., 4 Oct 2025, Cisowski et al., 2019).
Photonic crystal carpets, with near-field scanning optical microscopy confirming cloaked bump invisibility—RMS error reduction in wavefronts by nearly a factor of 5 (Scherrer et al., 2013).
Altermagnet/superconductor junctions, with differential conductance measurements resolving mirage gap features in the millikelvin regime, and clear parametric separation of quantum transport signatures by exchange field strength (Wei et al., 2023).
Nanophotonic scatterers exhibiting optical mirage shifts of up to 12 wavelengths, directly probed by far-field scattering and control of phase between resonant dipoles (Olmos-Trigo et al., 2022).
Image-based AI benchmarks, where mirage and cleaned benchmarks reveal striking drops (up to 70%) in accuracy and inversion of model rankings, with direct implication for safety-critical evaluation (Asadi et al., 23 Mar 2026).

7. Implications, Applications, and Mitigation Strategies

Mirage effects underscore fundamental limits and subtleties in both physical and information processing systems, including:

Wave control and invisibility: Mirage carpets and transformation optics enable cloaking, wavefront shaping, and spatial illusions at scales from nanophotonics to macroscopic diffusion (Guenneau et al., 2017, Scherrer et al., 2013).
OAM light and quantum device engineering: Handedness inversion and spatial mixing in twisted beams offer routes to spatially structured light manipulation and quantum resource encoding (Cisowski et al., 2019).
Metrology and super-resolution: Modal correction of mirage localization errors enables nanometric emitter recovery in plasmonic imaging (Baldassari et al., 2022).
AI robustness: Diagnostic regimes (mirage, guess, cleaned) and cleaning frameworks (B-Clean) are necessary for reliable, vision-grounded evaluation, revealing the high prevalence and quantitative impact of mirage reasoning in state-of-the-art AI (Asadi et al., 23 Mar 2026).

A plausible implication is that benchmarking and safety assessment in fields relying on coupled modalities (vision, language, quantum, transport) must adopt ablation, cross-modal consistency checks, and clean benchmark construction to avoid spurious mirage-induced artifacts. In high-stakes domains (e.g., medical AI), uncorrected mirages can lead to catastrophic over-confidence and erroneous decisions.

Summary Table: Select Mirage Effect Implementations

Domain	Physical/Algorithmic Mechanism	Signature/Metric
Atmospheric Optics	$dn/dz$ -driven ray curvature	Virtual/inverted image
Graded-index Photonics	Engineered $n(z),n(x,y)$ profiles	Wavefront restoration, OAM sign reversal
Diffusive Cloaks	Transform-optical tensor mapping	Undisturbed flux, shape illusions
Quantum Materials	Anisotropic exchange ( $T_J$ )	Finite-energy mirage gaps
Nanophotonic Scattering	E–M interference (spinless beams)	Apparent shift $\Delta/\lambda$
AI Vision–Language	Missing input + prior inference	Mirage Rate, Mirage Score