Ghost in Science: Concepts Across Disciplines

• Ghost is a polysemous term used in various fields, characterized by phenomena such as dynamical bottlenecks, negative-residue modes in quantum fields, and spurious LiDAR returns.
• In dynamical systems, ghost phenomena describe slow flow near vanished equilibria, with quantifiable delays (e.g., t_delay ∝ μ^(-1/2)), while in optics ghost imaging reconstructs images using correlation techniques.
• In technical applications, ghosts also serve as algebraic bookkeeping devices and acronyms in machine learning, offering insights into phase slowing, non-unitary behaviors, and computational optimizations.

In contemporary technical literature, ghost is a polysemous term rather than a single concept. It can denote a remnant slow-flow structure left by a saddle-node bifurcation, a propagating degree of freedom with a wrong-sign kinetic term or negative-residue pole, a false LiDAR return produced by multi-path reflections, a correlation-based imaging or projection protocol, a bookkeeping decoration in diagrammatic algebra, or simply a method title or acronym in machine learning and computer vision (Zhang et al., 12 Feb 2026, Gumrukcuoglu et al., 2016, Ikeda et al., 30 Mar 2026, Bornman et al., 2019, Nurcombe, 2023, Chen et al., 15 May 2026).

1. Technical scope and principal meanings

The shared label masks strong domain dependence. In some fields, a ghost is a dynamical remnant of vanished equilibria; in others it is a pathology of the spectrum; in still others it is an operational artifact or a constructive mathematical device.

Domain	Meaning of “ghost”	Diagnostic feature
Dynamical systems	Remnant slow region after a saddle-node bifurcation	Long bottleneck without a true fixed point
QFT and gravity	Degree of freedom with wrong-sign kinetic term or negative residue	Negative norm, non-unitarity, or instability
LiDAR sensing	False 3D return from multi-path reflections	Spurious points behind glass or reflective surfaces
Imaging and projection	Correlation-based image formation or synthesis using random masks	Bucket detection or weighted mask projection
Diagrammatic algebra	Boundary bookkeeping dots in a two-boundary TL generalization	Odd boundary connectivities made associative

A central source of confusion is that these meanings are not interchangeable. The saddle-node ghost in nonlinear dynamics is explicitly described as classical and topological, and is not a quantum-field ghost such as a Faddeev–Popov ghost or a ghost condensate (Zhang et al., 12 Feb 2026). Conversely, field-theoretic ghosts are defined by wrong-sign kinetic structure or negative-residue poles, not by slow passage through phase space (Gumrukcuoglu et al., 2016, Buoninfante, 27 May 2026).

2. Dynamical-systems ghosts and bottleneck dynamics

In dynamical systems, the canonical ghost is the saddle-node ghost. After a stable and an unstable fixed point collide and annihilate, the flow remains very slow near their former location, producing a bottleneck. In the overdamped normal form,

$\dot{x}=\mu+x^2,$

the delay scales as $t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$ (Zhang et al., 12 Feb 2026). This basic picture is generalized in the formalism of ghost attracting sets, ghost channels, and ghost cycles, where slow directed flow is organized without reliance on invariant sets such as fixed points or invariant manifolds. The diagnostic $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$ identifies slow regions, and the framework emphasizes that ghost-based scaffolds remain robust under additive white noise in regimes where heteroclinic channels and cycles become fragile (Koch et al., 2023).

This dynamical notion has been extended to strongly gravitating systems. In Einstein–Maxwell–scalar theory, highly excited black-hole remnants can be trapped near the remnant of a saddle-node in solution space. The resulting nonlinear bottleneck delays the onset of linear ringdown and yields a predicted quiescence–burst–delayed ringdown pattern. Because the relevant reduced dynamics is inertial rather than overdamped, the paper derives

$$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$

so the universal exponent is $1/4$, not $1/2$ (Zhang et al., 12 Feb 2026).

A closely related mechanism appears in nonlinear optics. “Ghost State of Light” reports a long-lived non-stationary state in a single-mode optical cavity with a nonlinear response with memory. The cavity transmission exhibits a plateau interpreted as a ghost bottleneck. The observed state persists for lifetimes exceeding the cavity photon lifetime by more than ten orders of magnitude, and the paper identifies minimal conditions for realizing parametrically long-lived non-stationary states: proximity to a saddle-node boundary and a memory time satisfying $\tau\Gamma \gtrsim 1$ (Boer et al., 13 May 2026). This suggests that ghost-mediated slowing is not restricted to low-dimensional normal forms but can be engineered in physical platforms with an additional slow variable.

3. Ghosts in quantum field theory and gravity

In field theory and gravity, a ghost is a propagating degree of freedom whose quadratic action has a wrong-sign kinetic term. A representative criterion is $K<0$ for a mode $q$ with Lagrangian contribution $L \supset -(1/2)K\dot q^2$, or, equivalently, a propagator pole with negative residue (Gumrukcuoglu et al., 2016, Lambiase et al., 20 Oct 2025). These ghosts are conceptually distinct from Faddeev–Popov ghosts, which are auxiliary gauge-fixing fields and do not appear as physical asymptotic states (Buoninfante, 27 May 2026, Holdom, 2024).

The best-known nonlinear gravitational example is the Boulware–Deser ghost. In multimetric gravity, its presence or absence is determined by the Hamiltonian constraint structure. Bimetric gravity is ghost-free, trimetric gravity generically contains a ghost, and cutting one interaction in the trimetric graph restores ghost-freedom; more generally, loop-type interactions in multimetric gravity never become ghost-free (Nomura et al., 2012). In massive gravity, decoupling-limit tuning can eliminate the ghost at $t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$0, yet “Massive Gravity: Exorcising the Ghost” shows that the ghost reappears at fourth order away from the decoupling limit, so decoupling-limit ghost-freedom is not sufficient for full nonlinear ghost-freedom (Alberte et al., 2010).

Quadratic and fourth-order gravities supply another standard setting. “Cosmology with a light ghost” studies a Weyl-squared theory with a massive spin-two ghost state and concludes that, for all ghost masses allowed by laboratory constraints, ghosts would have overclosed the Universe at temperatures higher than that of primordial nucleosynthesis (Ivanov et al., 2016). “Exorcising ghosts with gravitational waves” contrasts ghostful and ghost-free fourth-order gravity: in the ghostful theory, the massless spin-$t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$1 GW flux cancels the combined GW fluxes of the massive spin-$t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$2 ghost and massive spin-$t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$3 scalar in the vanishing-mass limit, so the GR quadrupole formula is not recovered at the leading order, leading to the bound $t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$4; by contrast, the torsion-enabled ghost-free theory smoothly reproduces the Newtonian potential and GR quadrupole formulae when $t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$5 and $t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$6 vanish (Lambiase et al., 20 Oct 2025).

Several papers re-evaluate whether every ghost implies the same pathology. “Jeans’ Ghost” shows that a low-energy tachyonic ghost around a spacelike scalar-gradient background can be recast, via a canonical transformation, as a fluid variable with positive kinetic terms and a Jeans-like infrared instability; the instability is then classical and long-wavelength rather than a catastrophic UV vacuum decay (Gumrukcuoglu et al., 2016). “Ghosts without runaway” constructs integrable classical mechanical models with a negative-kinetic-term oscillator and proves complete classical stability for all initial conditions despite an unbounded Hamiltonian (Ye et al., 2021). “Making sense of ghosts” argues in $t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$7d QFT that an appropriate positive-definite metric $t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$8 can define a sensible Born rule, while “Asymptotic Quantum Dynamics of Ghost Fields” shows that a dressed propagating ghost coupled to ordinary fields develops first-sheet complex poles and no free asymptotic one-particle ghost state exists because the negative-norm state becomes non-orthogonal to a superposition of positive-norm multiparticle states on a timescale set by the inverse imaginary part of the complex mass (Holdom, 2024, Buoninfante, 27 May 2026). In a different strong-coupling proposal, “A ghost and a naked singularity; facing our demons” argues that in classically scale-invariant quadratic gravity the full graviton propagator yields slight acausal behavior rather than a propagating ghost (Holdom, 2019).

4. Ghosts in sensing, imaging, and projection

In mobile LiDAR, a ghost point is a false 3D return that does not correspond to a physical surface. Ghosts arise from multi-path propagation through or off glass and other reflective or transparent surfaces, generating extra peaks in the time-of-flight signal that conventional peak-only LiDAR converts into spurious 3D points (Ikeda et al., 30 Mar 2026). Because geometry-only heuristics assume dense, static scans, they fail in sparse, dynamic mobile settings. “Ghost-FWL” addresses this by using full-waveform LiDAR, which preserves the complete temporal intensity histogram of each beam. The dataset comprises 24,412 annotated frames across 10 scenes with 7.5 billion peak-level annotations, and the baseline with FWL-MAE pretraining reaches recall $t_{\mathrm{delay}} \sim \pi \mu^{-1/2}$9 and removal rate $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$0. Ghost removal reduces SLAM trajectory errors by $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$1 relative to factory peak-selection strategies and reduces ghost-induced pedestrian false positives from $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$2 to $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$3, about a $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$4 reduction (Ikeda et al., 30 Mar 2026).

In optics, ghost imaging uses correlations rather than direct image formation. One photon or beam interacts with the object and is detected by a non-imaging bucket detector, while a correlated partner is measured with spatial resolution; the image is reconstructed from second-order correlations (Bornman et al., 2019). In computational ghost imaging, the standard relation is

$q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$5

with known illumination patterns $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$6, single-pixel measurements $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$7, and object $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$8 (Ye et al., 2021, Ceddia et al., 2018). “Ghost imaging with engineered quantum states by Hong–Ou–Mandel interference” shows that HOM filtering can select anti-symmetric two-photon states, producing a coherent superposition of counter-rotated images rather than the single rotated image obtained from the symmetric SPDC state (Bornman et al., 2019).

The same correlation logic has been extended in two directions. “Ghost Panorama” transforms Hadamard-based circular patterns into omnidirectional structured illumination using a convex spherical mirror with curvature radius $q(\mathbf{x})=\tfrac12\|\mathbf{F}(\mathbf{x})\|^2$9 cm, producing a panoramic field of view of approximately $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$0 with a single, lens-free photodetector and a reconstructed panoramic belt of $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$1 pixels (Ye et al., 2021). “Ghost Projection” reverses the measurement paradigm: any desired spatial distribution of radiant exposure may be produced, up to an additive constant, by spatially-uniformly illuminating a set of random masks in succession, either by weighting each mask by its correlation with the target image, selecting a correlated subset, or numerically optimizing mask exposures (Ceddia et al., 2021). In x-ray phase-contrast computational ghost imaging, this random-mask framework is combined with analyzer-based differential phase contrast and a stable Fourier-domain inversion to recover projected thickness quantitatively (Ceddia et al., 2018).

5. Ghosts as diagrammatic bookkeeping in algebra

In algebraic and integrable-model contexts, ghosts are not particles or artifacts but bookkeeping decorations. “The ghost algebra and the dilute ghost algebra” introduces the ghost algebra as a two-boundary generalisation of the Temperley–Lieb algebra. The standard two-boundary TL algebra requires the number of strings connected to each boundary to be even. The ghost algebra lifts that restriction by decorating boundaries with bookkeeping dots called ghosts and imposing an evenness condition on “number of ghosts + number of strings attached to that boundary” (Nurcombe, 2023).

This modification permits parity-sensitive boundary weights without losing associativity. Boundary arcs on the top boundary carry $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$2, $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$3, or $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$4 according to endpoint parity; bottom arcs analogously carry $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$5, $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$6, or $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$7; top–bottom arcs carry $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$8 or $$\frac{d^2 \Lambda}{dt^2}=-\mu\,\varepsilon-\beta\,\Lambda^2, \qquad t_b \propto \varepsilon^{-1/4},$$9. The resulting algebra is associative and cellular, contains TL, one-boundary TL, and one-boundary ghost subalgebras, and admits a dilute extension dGh$1/4$0 with empty nodes (Nurcombe, 2023).

The algebraic payoff is integrability. The paper constructs loop models associated with Gh$1/4$1 and dGh$1/4$2, classifies solutions to their boundary Yang–Baxter equations given existing bulk solutions for TL and dilute TL, and thereby obtains one-parameter families of commuting transfer tangles. In this setting, “ghost” names an auxiliary parity-tracking device that enlarges the admissible diagrammatics while preserving the algebraic structures needed for Yang–Baxter integrability (Nurcombe, 2023).

6. “Ghost” as a method name and acronym

Recent machine-learning and computer-vision literature also uses Ghost or GHOST as a title or acronym rather than as a phenomenon. In these cases the semantic content is supplied by the expansion, not by any shared ontology with dynamical, field-theoretic, or optical ghosts.

“GHOST: Geometry-Hierarchical Online Streaming Token Eviction for Efficient 3D Reconstruction” is a training-free KV-cache management framework for long monocular video. It scores cached tokens using geometry outputs such as camera pose, dense depth, and per-patch confidence maps, adds a privilege mechanism for camera and register tokens, and allocates layer-wise budgets by cosine-similarity profiling. Experiments report that GHOST preserves excellent reconstruction quality while cutting the KV cache by nearly half and delivering $1/4$3 faster inference (Chen et al., 15 May 2026).

“Ghost: Plausible Yet Unlearnable Trajectories via On-Manifold Substitution for Next-POI Privacy” addresses privacy in released check-in trajectories. It generates perturbed sequences that remain geographically and semantically plausible while degrading next-POI predictors trained on the released data. The method uses a frozen trajectory LLM as a manifold prior, replaces entropy-floor randomization with on-manifold substitution, and achieves protection-gap competitive with PGD while attaining the lowest restored accuracy under the bigram adaptive purification adversary on both datasets (Yu et al., 2 Jun 2026). A plausible implication is that, in current AI literature, “Ghost” frequently signals hidden or difficult-to-invert structure, but that meaning is metaphorical and domain-specific rather than technical in the older physical sense.

Across these literatures, the decisive question is therefore not what a ghost is in general, but which operational definition is in force. In nonlinear dynamics it is a bottleneck in phase flow; in QFT it is a wrong-sign or negative-norm mode; in LiDAR it is a false return; in correlation optics it is an image or exposure synthesized without direct spatial detection; in algebra it is a boundary bookkeeping dot; and in contemporary ML it may be an acronymic method name (Koch et al., 2023, Gumrukcuoglu et al., 2016, Ikeda et al., 30 Mar 2026, Bornman et al., 2019, Nurcombe, 2023, Chen et al., 15 May 2026).