Multislice Model: A Cross-Domain Synthesis

• Multislice model is a framework that decomposes high-dimensional data into slices and couples them via a shared operator across various domains.
• It underpins diverse applications such as dynamic MRI, CT reconstruction, electron scattering, and multilayer network analysis by linking local slice properties with global structure.
• This approach enables efficient joint inference and enhanced resolution in inverse problems, scattering simulations, and complex graph models.

Searching arXiv for recent and foundational uses of “multislice model” across domains.

“Multislice model” does not denote a single standardized construct. Across the cited literatures, it refers to several mathematically distinct models that all organize an object, signal, or state space through slices or layers and then impose cross-slice coupling through a shared operator, prior, or symmetry. In inverse imaging, this includes a shared 3D generative manifold for dynamic cardiac MRI, a fused lower-dimensional prior for 4D CT, and a joint 3D Fourier forward model for simultaneous multi-slice MRI (Zou et al., 2021, Majee et al., 2019, Pak et al., 2023). In wave physics, it denotes slice-by-slice propagation through projected potentials or tensor scattering media (Bangun et al., 2024, Mu et al., 2022). In discrete mathematics, the multislice is a fixed-composition space of colorings that underlies transposition walks, log-Sobolev inequalities, concentration, and FKN-type theorems (Salez, 2020, Sambale et al., 2020, Filmus, 2018). In network science, a multislice network is a collection of networks on the same set of nodes, represented by a supra-adjacency matrix with intra-layer and inter-layer blocks (Ren et al., 1 Apr 2025). A plausible unifying description is therefore methodological rather than ontological: slice decomposition plus structured coupling.

1. Terminological scope and recurrent structure

The literature uses “multislice” in several non-equivalent ways. In some domains it names a forward model, in others a prior, a latent-variable model, a state space, or a multilayer graph. A common misconception is that “multislice” always means a stack of physical image slices. In fact, the multislice of combinatorics is the level set of a color histogram, and the multislice of network theory is a layered graph on a common node set (Salez, 2020, Ren et al., 1 Apr 2025).

Domain	Slice object	Representative formulation
Dynamic MRI	2D acquired slices of a moving 3D heart	Shared 3D decoder and latent space (Zou et al., 2021)
4D CT	Space-time hyperplane slices	MACE fusion of $H_{xy,t}, H_{yz,t}, H_{zx,t}$ (Majee et al., 2019)
SMS MRI	Slice-encoding dimension in k-space	$Y = F_{3D}(X)$ (Pak et al., 2023)
TEM	Thin specimen slices along beam direction	$A=\prod_{m=1}^{M} G_m O_m$ (Bangun et al., 2024)
Combinatorics	Fixed-composition colorings	$\Omega_\kappa$ (Salez, 2020)
Multilayer networks	Copies of nodes across layers	Supra-adjacency $\hat{\mathcal A}$ (Ren et al., 1 Apr 2025)

This cross-domain recurrence is precise at the level of construction. A multislice model typically specifies a local slice representation and an operator that transmits information across slices: a shared latent prior in MRI, consensus equilibrium in CT, Fresnel propagation in electron microscopy, a dyadic Green’s function in birefringent optics, coordinate permutations in the combinatorial multislice, or inter-layer edges in multislice networks (Zou et al., 2021, Majee et al., 2019, Bangun et al., 2024, Mu et al., 2022, Salez, 2020, Ren et al., 1 Apr 2025).

2. Generative and inverse-problem formulations in medical imaging

In dynamic cardiac MRI, the multislice model in “Variational manifold learning from incomplete data: application to multislice dynamic MRI” is the multislice variational manifold model, multislice V-SToRM (Zou et al., 2021). The object is a time series of 3D cardiac volumes $\mathbf{x}(\mathbf r,t_z)$, observed through 2D multislice k–t measurements

$$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$

The central modeling step is

$$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$

where $\mathcal D_\theta$ is a shared 3D CNN decoder and $\mathbf c(t_z)$ are low-dimensional latent variables. Because fully sampled images are unavailable, the paper replaces a global encoder by per-frame Gaussian latent distributions $Y = F_{3D}(X)$0, learned directly from undersampled data with a Gaussian prior $Y = F_{3D}(X)$1 and MRI forward-model back-propagation. The same generator is shared across all slices and time points, and the KL term forces latent variables toward a common coordinate system, so latent coordinates become a global motion-state index for simultaneous temporal and inter-slice alignment. This directly addresses the limitation that existing slice-independent cardiac methods cannot exploit inter-slice redundancies and require post-processing for phase alignment (Zou et al., 2021).

In 4D X-ray CT, “Multi-Slice Fusion” defines the multislice model as an implicit 4D prior built from multiple lower-dimensional denoisers (Majee et al., 2019). The reconstruction variable is a sequence of 3D volumes over time,

$Y = F_{3D}(X)$2

and the data term is combined with a data-fidelity agent $Y = F_{3D}(X)$3 and denoiser agents $Y = F_{3D}(X)$4 inside Multi-Agent Consensus Equilibrium. For the 4D prior, the paper uses $Y = F_{3D}(X)$5, $Y = F_{3D}(X)$6, and $Y = F_{3D}(X)$7, each a 2.5D CNN denoiser operating slice by slice in a different orientation. The consensus equilibrium condition

$Y = F_{3D}(X)$8

makes the final reconstruction consistent with cone-beam CT physics and with all three slice-oriented priors simultaneously. The later sparse-view and limited-angle study preserves the same multislice construction and reports PSNR/SSIM gains over FBP, MBIR+TV, and MBIR+4D-MRF, including $Y = F_{3D}(X)$9 dB and $A=\prod_{m=1}^{M} G_m O_m$0 in the simulated $A=\prod_{m=1}^{M} G_m O_m$1 case and $A=\prod_{m=1}^{M} G_m O_m$2 dB and $A=\prod_{m=1}^{M} G_m O_m$3 in the $A=\prod_{m=1}^{M} G_m O_m$4 limited-angle case (Majee et al., 2020).

In dynamic simultaneous multi-slice MRI, the “Holistic Multi-Slice Framework” reinterprets SMS acquisition as a 3D Fourier transform over space and slice index: $A=\prod_{m=1}^{M} G_m O_m$5 with undersampling handled through

$A=\prod_{m=1}^{M} G_m O_m$6

Here the multislice model is not a per-slice reconstruction followed by slice separation; slice separation is inherent in the forward model and solved jointly by an unrolled CRNN+DC architecture (Pak et al., 2023). The paper is explicit that the effective multislice model is not just “use 3D convolutions”; it is the combination of a correct joint forward model $A=\prod_{m=1}^{M} G_m O_m$7, a data-consistency step in that joint k-space, and slice-aware network design. On cine SMS, the proposed network reports NMSE $A=\prod_{m=1}^{M} G_m O_m$8, PSNR $A=\prod_{m=1}^{M} G_m O_m$9 dB, and SSIM $\Omega_\kappa$0, outperforming both independent slices and a 3D-convolution baseline (Pak et al., 2023).

3. Electron-scattering multislice formulations

In transmission electron microscopy, the multislice model is the classical slice-by-slice propagation of an electron wave through a crystal potential. In “Eigenstructure Analysis of Bloch Wave and Multislice Matrix Formulations for Dynamical Scattering in Transmission Electron Microscopy”, the traditional multislice method is reformulated as a global matrix operator

$\Omega_\kappa$1

where $\Omega_\kappa$2 is the real-space phase-grating operator of slice $\Omega_\kappa$3 and $\Omega_\kappa$4 is the Fresnel propagator (Bangun et al., 2024). This transmission matrix $\Omega_\kappa$5 enables direct comparison with the Bloch-wave scattering matrix $\Omega_\kappa$6. The paper shows that $\Omega_\kappa$7 when eigenvalue phases differ only by integer multiples of $\Omega_\kappa$8, and that the eigenvectors are related by

$\Omega_\kappa$9

It further shows that the determinant of $\hat{\mathcal A}$0 can be used to estimate the mean inner potential, thereby turning the multislice operator into a directly interpretable physical object (Bangun et al., 2024).

Inverse problems use the same slice-by-slice structure. In inverse multislice ptychography, the exit wave after $\hat{\mathcal A}$1 slices is

$\hat{\mathcal A}$2

and far-field intensities satisfy

$\hat{\mathcal A}$3

The paper introduces layer-wise optimisation and sparse matrix decomposition to recover slice transmission functions $\hat{\mathcal A}$4 and optionally the probe, and argues that unique separation of atomic layers is possible for simulated data when considering a low acceleration voltage (Bangun et al., 2022).

Multislice electron ptychography further turns slice-resolved phase retrieval into a tool for point-defect metrology. In the SiC case study, the object is reconstructed as $\hat{\mathcal A}$5 slices of thickness $\hat{\mathcal A}$6 nm, and defect depth is inferred from slice-wise phase contrast along atomic columns (Bhat et al., 2024). The reported depth precision is $\hat{\mathcal A}$7 nm and the depth resolution is $\hat{\mathcal A}$8 nm, even though the reconstruction slice thickness is $\hat{\mathcal A}$9 nm. The same study shows that isolated point defects can be located within a unit cell along the sample’s depth, and that electron energy, dose, defocus, and convergence semi-angle govern defect contrast (Bhat et al., 2024).

The multislice formalism also underlies vibrational EELS and quantum algorithms. The frequency-resolved frozen phonon multislice method defines

$\mathbf{x}(\mathbf r,t_z)$0

interpreting the variance of the exit wave over frequency-selective frozen-phonon ensembles as vibrational scattering intensity at energy $\mathbf{x}(\mathbf r,t_z)$1 (Zeiger et al., 2021). In quantum computing, the multislice update

$\mathbf{x}(\mathbf r,t_z)$2

is implemented with QFT and diagonal phase circuits; the improved algorithm reconstructs the phase-shifting quantum circuit without using the multi-controlled quantum gates and reports parameter settings that keep relative error within $\mathbf{x}(\mathbf r,t_z)$3 while reducing gate count (Wang et al., 2024).

4. Vectorial and birefringent generalizations in optics

In optics, the multislice model becomes a vectorial forward model for field propagation through a birefringent scattering medium. “A Multislice computational model for birefringent scattering” discretizes the medium along the propagation axis $\mathbf{x}(\mathbf r,t_z)$4 into thin slices and treats each slice as a single-scattering event driven by a 3D tensor scattering potential (Mu et al., 2022). The electric field is a full three-component vector

$\mathbf{x}(\mathbf r,t_z)$5

and the scattering potential tensor is

$\mathbf{x}(\mathbf r,t_z)$6

The model is built from a dyadic Green’s function and a polarization transfer function tensor

$\mathbf{x}(\mathbf r,t_z)$7

which enters both the propagation operator

$\mathbf{x}(\mathbf r,t_z)$8

and the slice-scattering update (Mu et al., 2022).

This formulation is explicitly presented as a fully vectorial alternative to scalar BPM-style models and as a computationally lighter alternative to full-wave FDTD or FEM in highly anisotropic media. The validation compares amplitude and phase of $\mathbf{x}(\mathbf r,t_z)$9 against FDTD for four birefringent spheres and reports runtime $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$0 s for the multislice model versus $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$1 s for FDTD in the same test case (Mu et al., 2022). Experimental validation with trapped vaterite particles shows good agreement between measured and simulated $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$2 and $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$3, especially for cross-polarized scattering, supporting the claim that the full tensor nature of birefringence and the longitudinal field $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$4 matter in strongly birefringent scattering (Mu et al., 2022).

A plausible implication is that the optical multislice model shares the same architectural pattern as the electron-scattering multislice method—local slice interaction plus inter-slice propagation—while replacing scalar phase gratings by tensor scattering operators.

5. The multislice as a combinatorial and probabilistic state space

In combinatorics, probability, and Boolean function analysis, the multislice is not a propagation model at all. It is the finite set

$$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$5

equivalently the level set of the color histogram (Salez, 2020). Its cardinality is the multinomial coefficient $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$6, and the canonical measure is uniform. The Boolean slice / Johnson graph is the two-color case $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$7, and the symmetric group arises when $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$8 and $$\mathbf b(t_z) = \mathcal{A}_{t_z}\Big(\mathbf{x}(\mathbf r,t_z)\Big) + \mathbf n_{t_z}.$$9 (Salez, 2020, Sambale et al., 2020).

The associated dynamics is the transposition walk, or multi-urn Bernoulli–Laplace diffusion model, in which a pair of coordinates is chosen and their colors are swapped. The generator is

$$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$0

and the Dirichlet form is

$$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$1

Within this model, the sharp dependence of the log-Sobolev constant on the rarest color is

$$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$2

confirming a conjecture of Filmus, O’Donnell and Wu (Salez, 2020). The same paper derives small-set expansion bounds and extends the analysis to colored exclusion processes on general graphs (Salez, 2020).

Concentration theory on the multislice uses modified log-Sobolev inequalities and yields analogues of product-space inequalities for this non-product domain (Sambale et al., 2020). The paper proves a bounded-difference inequality on the multislice, Talagrand’s convex distance inequality, multilevel concentration for multilinear polynomials, and a triangle-count concentration bound for $$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$3 resembling the $$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$4 case. Interpreting $$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$5 as sampling without replacement also yields a finite-sampling-corrected bounded-difference inequality and a Serfling-type bound (Sambale et al., 2020).

The FKN theorem on the multislice gives the structural counterpart. For a $$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$6-balanced multislice, if a Boolean function $$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$7 satisfies $$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$8, then there exists a Boolean dictator $$\mathbf{x}(\mathbf r,t_z) = \mathcal D_\theta\big(\mathbf c(t_z)\big),$$9 such that

$\mathcal D_\theta$0

The same framework yields a stability version of the edge-isoperimetric inequality in parameter regimes where the optimal set is a dictator (Filmus, 2018). Here, “multislice model” means a highly symmetric fixed-composition domain, and “degree 1” is defined representation-theoretically by the components $\mathcal D_\theta$1 (Filmus, 2018).

6. Multislice networks and cross-domain synthesis

In network science, a multislice network is a collection of networks on the same set of nodes, with copies of each basis node in each layer and with connections between the different layers to be determined (Ren et al., 1 Apr 2025). The structure is encoded by the supra-adjacency matrix

$\mathcal D_\theta$2

decomposed as $\mathcal D_\theta$3 into intra-layer and inter-layer parts. The paper defines 1D, 2D, and 3D triangles through triadic paths in the supra-graph and expresses their counts as

$\mathcal D_\theta$4

$\mathcal D_\theta$5

$\mathcal D_\theta$6

For multislice Erdős–Rényi networks, the joint law of $\mathcal D_\theta$7 is shown to be close in total variation to a product of Poisson distributions with means $\mathcal D_\theta$8, extending classical triangle-count asymptotics from sparse $\mathcal D_\theta$9 to multislice networks (Ren et al., 1 Apr 2025).

A recurring misconception is that “multislice” names a single model family with a stable semantics. The cited literature shows the opposite. In MRI and CT, multislice typically means joint inference across acquired slices or slice-oriented priors (Zou et al., 2021, Majee et al., 2019, Pak et al., 2023). In TEM and optics, it is a forward scattering model generated by alternating local slice interaction and inter-slice propagation (Bangun et al., 2024, Mu et al., 2022). In combinatorics, it is a state space with fixed composition (Salez, 2020). In network theory, it is a supra-graph over multiple layers (Ren et al., 1 Apr 2025). A plausible synthesis is that the term is unified less by one equation than by one design move: decompose a high-dimensional object into slices, then recover or analyze global structure by the operator that ties those slices together.