Transverse Phase Space Tomography

• Transverse phase space tomography is a diagnostic method that reconstructs the full phase-space density from a series of lower-dimensional projections, revealing non-Gaussian features and beam coupling.
• The approach uses controlled phase-space rotations and linear transport to generate diverse projections, which are then inverted using techniques like MLEM, SVD, and kernel methods.
• This method is essential for accurately diagnosing weak nonlinear distortions, slice emittance variations, and multi-dimensional couplings in accelerator and cold-atom experiments.

Searching arXiv for the specified paper and closely related transverse phase-space tomography work. Transverse phase space tomographic reconstruction is a measurement and inversion framework for recovering a phase-space distribution from a set of lower-dimensional projections acquired under controlled phase-space rotations or equivalent linear transforms. In accelerator and cold-atom settings, the reconstructed object is typically a distribution such as $f(x,x')$, $f(y,y')$, $f(x,p)$, or, in coupled problems, a higher-dimensional density over $(x,x',y,y')$ or beyond. The method is used because rms-based emittance or Twiss analyses can miss non-Gaussian structure, coupling, halo, and weak nonlinear distortions, whereas the reconstructed density can reveal the actual geometry of the occupied phase-space volume, including dephasing, crescent formation, X-shaped superpositions, and slice-dependent structure (Zhou et al., 2014).

1. Definition, scope, and relation to conventional beam characterization

Transverse phase space describes the distribution of particle coordinates in a transverse plane, for example $(x,x')$ with $x$ the transverse position and $x' = dx/ds$ the angle with respect to the beam axis, and similarly for $(y,y')$ (Ye et al., 8 Aug 2025). In cold-atom work, the analogous pair is $(x,p)$, where $x$ is position and $f(y,y')$0 momentum; the same tomographic logic applies because the measured density is a projection of a two-dimensional phase-space distribution after a controlled rotation (Zhou et al., 2014).

The basic objective is to reconstruct the full distribution rather than only its second moments. This distinction is central in systems where the beam or cloud “does not produce a strictly elliptical phase space,” where profiles show non-Gaussian traits, or where coupling introduces interplane correlations that are not represented by uncoupled rms fits (Garcia et al., 2013). At FACET-II, this is stated explicitly: emittance measurement and Twiss characterization traditionally rely on fitting rms beam sizes versus quadrupole strength under a Gaussian-beam assumption, and this can be inaccurate for non-Gaussian beams; tomography instead recovers the actual phase-space density and reveals non-Gaussian features such as X-shaped superpositions (Ye et al., 8 Aug 2025).

The same formalism extends beyond uncoupled 2D reconstructions. FAST reported a fully coupled 4D transverse reconstruction using mixed projections generated by seven quadrupoles and 2D screen images (Romanov, 2018). SwissFEL ACHIP used nanofabricated wire scanners to reconstruct $f(y,y')$1 and $f(y,y')$2 for micrometer-scale beams (Hermann et al., 2021). FLASHForward experimentally demonstrated a 5D reconstruction of $f(y,y')$3 by combining quadrupole-based phase-space rotation with a polarizable X-band transverse deflection structure (Jaster-Merz et al., 19 May 2025). PAL-XFEL implemented a full 6D generative reconstruction with standard accelerator elements, within which the transverse tomograms appear as marginals of the reconstructed six-dimensional density (Kim et al., 28 Feb 2025).

A common misconception is that tomography is merely a more elaborate quadrupole scan. The published work indicates a stricter claim: tomography is an inverse problem whose fidelity depends on how projections are generated, how the forward operator is modeled, and how incomplete or noisy measurements are regularized. This is especially evident when weak nonlinearities, dispersion, space charge, or limited-angle coverage alter the effective projection geometry (Zhou et al., 2014).

2. Phase-space rotation, transport, and projection formation

The operational core of transverse tomography is the generation of a family of distinct projections. In the ideal harmonic case, phase-space points rotate uniformly. For cold atoms in a nearly harmonic trap, the exact harmonic mapping is

$f(y,y')$4

with $f(y,y')$5, which permits tomography by interpreting time evolution as a rotation of the phase-space frame (Zhou et al., 2014). In accelerator beamlines, the corresponding mechanism is linear transport through drifts, quadrupoles, solenoids, or TDS streaking sections. For a single-quadrupole scan at FACET-II, the linear optics between the quadrupole and the screen is represented by

$f(y,y')$6

and each setting produces a distinct projection labeled by

$f(y,y')$7

The measured line profiles are then rescaled and labeled by $f(y,y')$8 and $f(y,y')$9 to form a sinogram (Ye et al., 8 Aug 2025).

The projection itself is a Radon-type marginal. In the cold-atom formulation,

$f(x,p)$0

while in FACET-II the idealized projection is written as

$f(x,p)$1

FAST expresses the same idea in 4D: for optics setting $f(x,p)$2, the 2D screen image is a projection of the upstream 4D density along the unmeasured angles $f(x,p)$3 under a transport matrix $f(x,p)$4 [(Zhou et al., 2014); (Ye et al., 8 Aug 2025); (Romanov, 2018)].

Several measurement schemes instantiate this principle differently. Fermilab’s 400 MeV Linac used an upstream quadrupole focusing triplet and a 10 m, dispersion-free, magnet-free straight with three multiwire profile monitors; different triplet settings altered the projection direction through the effective transport matrix $f(x,p)$5, with the measured coordinate

$f(x,p)$6

and the corresponding normal angle

$f(x,p)$7

Tomography was described as “rotating the phase space distribution using different waist focusing conditions of the upstream triplet and performing a de-convolution of the profile data” (Garcia et al., 2013).

SwissFEL ACHIP replaced screens with nanofabricated wire projections. Nine free-standing gold stripes, each $f(x,p)$8 wide and arranged radially and uniformly in angle, were scanned across the beam at six longitudinal positions around the waist. The measured beam loss monitor signal is proportional to the line integral of the beam density along the wire, yielding 54 projections in total. Near the waist, small changes in longitudinal position produce comparatively large changes in phase advance, which makes that region especially informative for tomography (Hermann et al., 2021).

In higher-dimensional tomography, projection generation combines more than one rotation mechanism. The 5D method proposed for ARES and later demonstrated at FLASHForward used transverse phase advances $f(x,p)$9 from quadrupole scans together with adjustable streaking angle $(x,x',y,y')$0 or $(x,x',y,y')$1 from a polarizable X-band TDS. Near RF zero-crossing, the TDS maps time into an angular kick, and after downstream drift this becomes a screen shear proportional to time, enabling reconstruction of 3D $(x,x',y,y')$2 charge density before slice-wise 4D inversion (Jaster-Merz et al., 2023, Jaster-Merz et al., 19 May 2025).

3. Inversion algorithms and reconstruction representations

The inversion stage converts a set of measured projections into a phase-space density. Different experiments use different representations and regularization strategies, but all solve an ill-conditioned inverse problem.

In the cold-atom work, the reconstruction uses the Leonhardt kernel inversion rather than standard filtered back-projection:

$(x,x',y,y')$3

with a regularized, nondivergent kernel $(x,x',y,y')$4 and cutoff $(x,x',y,y')$5, which sets the phase-space resolution and stabilizes inversion against noise and finite sampling (Zhou et al., 2014). The same paper notes that a maximum-likelihood approach was also tested and yielded similar reconstructions.

FACET-II uses Maximum Likelihood Expectation Maximization (MLEM) in a discretized non-negative grid model,

$(x,x',y,y')$6

with the measured projections stacked into a sinogram. The paper states that MLEM is used rather than filtered back-projection and that regularization such as early stopping or smoothing can be applied, though the details are not specified there (Ye et al., 8 Aug 2025).

FAST formulates the coupled 4D reconstruction as a large linear system $(x,x',y,y')$7 over a 6×6×6×6 grid, giving $(x,x',y,y')$8 voxel amplitudes as unknowns. To suppress ringing and keep projections analytically tractable, each voxel basis function is a multivariate normal in a normalized 4D phase-space basis constructed via Williamson normalization from an initial second-moment estimate. The inversion uses truncated singular value decomposition, retaining 200 singular values, with a 1296×40659 linear system whose SVD took approximately one hour on a single Intel Xeon E5-1650 (3.5 GHz) core (Romanov, 2018).

SwissFEL ACHIP departs from grid-based inversion and represents the distribution by an ensemble of macro-particles. Each particle carries $(x,x',y,y')$9, and a smooth density is generated by separable Gaussian kernels with typical choices $(x,x')$0 and $(x,x')$1. The reconstruction iterates a stochastic birth–death update based on the normalized projection residuals,

$(x,x')$2

and stops when

$(x,x')$3

which produced stable convergence in approximately 100–110 iterations for the reported data (Hermann et al., 2021).

The 5D ARES and FLASHForward reconstructions use Simultaneous Algebraic Reconstruction Technique (SART), implemented slice-wise. In the 5D method, the discrete forward model is

$(x,x')$4

with positivity and charge-related constraints. The authors report that SART converges to good results in 1–2 iterations; two iterations were used in both the simulation study and the later FLASHForward experiment (Jaster-Merz et al., 2023, Jaster-Merz et al., 19 May 2025).

The 2026 resolution framework introduces a different level of analysis. It states that the experimentally accessible degrees of freedom are determined by the Gram or sampling operator $(x,x')$5, with eigenvalues that quantify which Hilbert-space modes are transmitted reliably and which are weakly supported. In that framework, reconstruction in the Gram eigenbasis is a measurement-adapted compression of the tomographic problem, and features dominated by small-$(x,x')$6 modes should be treated as artefacts or as hypotheses requiring additional measurements (Hradil et al., 28 May 2026). This suggests a unifying interpretation of regularization across kernel, SVD, MLEM, and algebraic methods: each, implicitly or explicitly, suppresses modes that are poorly supported by the measurement geometry.

4. Experimental realizations across cold atoms and accelerators

The published implementations span cold atoms, proton linacs, photoinjectors, FEL beamlines, and GeV-class electron transport. The measurement geometries differ, but the recurring structure is projection acquisition under known transport plus inversion under explicit regularization.

Platform	Projection generation	Reconstruction
Atom-chip cold atoms	Harmonic rotation with 13 images over $(x,x')$7	Leonhardt kernel (Zhou et al., 2014)
Fermilab FAST	Seven quadrupoles, 22 screen projections	Truncated SVD in 4D (Romanov, 2018)
SwissFEL ACHIP	Nine wire angles at six $(x,x')$8 positions, 54 projections	Particle-based birth–death inversion (Hermann et al., 2021)
FACET-II	Single-quadrupole scan upstream of a screen	MLEM sinogram inversion (Ye et al., 8 Aug 2025)
FLASHForward	Phase-advance scans plus ten TDS angles	3D + slice-wise 4D SART (Jaster-Merz et al., 19 May 2025)

The cold-atom demonstration remains notable because it isolates deformation by weak anharmonicity with unusually clear control. The experiment used approximately 3,000 $(x,x')$9 atoms at approximately $x$0 in a magnetic atom-chip trap with longitudinal harmonic frequency $x$1 and transverse frequencies approximately $x$2. For an initial oscillation amplitude of $x$3 and corrugation wire current of $x$4, slight deformation appears at $x$5, becomes distinct at $x$6, and becomes a pronounced crescent-shaped distribution at $x$7. Without corrugation, the distribution remains approximately isotropic Gaussian up to $x$8 (Zhou et al., 2014).

Fermilab’s 400 MeV proton Linac provided an earlier beamline realization motivated by the statement that the beam “does not produce a strictly elliptical phase space.” The 10 m straight used three multiwire profile monitors, with MW5 at the waist having $x$9 pitch, and preliminary rms analyses reported $x' = dx/ds$0 by the analytical three-screen method and $x' = dx/ds$1, with notably larger approximate 95% emittances of approximately $x' = dx/ds$2 in the vertical plane and approximately $x' = dx/ds$3 in the horizontal plane (Garcia et al., 2013). The tomography program there was explicitly intended to recover the non-elliptical distribution hidden behind those rms summaries.

FAST moved to a coupled 4D reconstruction at 32 MeV and 100 nC bunch charge. Seven of eight quadrupoles were varied so as to rotate the beam’s transverse phase space by $x' = dx/ds$4 in each plane while keeping the Twiss functions at the screen approximately constant, producing 22 distinct projections and 40,659 informative pixel values after image preparation. The reconstructed 4D density revealed small but nonzero interplane correlations consistent with solenoidal coupling and possible laser spot asymmetries or space-charge distortions (Romanov, 2018).

SwissFEL ACHIP addressed the opposite scale regime: micrometer beams requiring submicrometer diagnostics. The reported experiment used a beam energy of approximately 3.2 GeV and bunch charge of approximately 1 pC, with scans at six $x' = dx/ds$5-locations spanning approximately 8 cm around the waist. Typical SNR ranged from 25 to 45, charge fluctuations were approximately 1.3% rms, and PMT readout noise was less than 1%. The reconstructed waist beam sizes were $x' = dx/ds$6 and $x' = dx/ds$7, with $x' = dx/ds$8 in both planes; normalized emittances from Gaussian fits to the reconstructed phase spaces were $x' = dx/ds$9 and $(y,y')$0 (Hermann et al., 2021).

FACET-II demonstrated that a single quadrupole and a downstream screen can suffice for useful transverse tomography. In nominal two-bunch operation, with a 1.6 nC driver and 0.5 nC witness at about 7 ps separation, injector tomography revealed a distinctly non-Gaussian X-shaped structure in the horizontal phase space when the two bunches were combined. The interpretation was verified by characterizing each bunch individually. With zero pulse separation and charge varied from approximately 0.5 nC to 2 nC, both planes showed increasingly stretched distributions with higher charge, and in the horizontal plane the phase space was observed to “flip” when the bunch charge was halved (Ye et al., 8 Aug 2025).

FLASHForward demonstrated a 5D experimental reconstruction at 1.09 GeV and 297 pC using a PolariX TDS and a GAGG:Ce screen. Ten steps over approximately $(y,y')$1 in each of $(y,y')$2 and $(y,y')$3 and ten TDS angles over approximately $(y,y')$4 yielded 960 successfully recorded projections over 28 hours. The tomography gave an rms bunch duration of $(y,y')$5, in excellent agreement with $(y,y')$6 from independent streaked-screen measurements, and projected normalized emittances of $(y,y')$7 and $(y,y')$8 (Jaster-Merz et al., 19 May 2025).

5. Nonlinearity, coupling, multi-bunch structure, and higher-dimensional generalizations

One of the clearest lessons from the literature is that tomography is unusually sensitive to weak distortions that are difficult to detect in single projections. In the atom-chip experiment, weak static corrugation $(y,y')$9 produced energy-dependent angular velocity,

$(x,p)$0

with $(x,p)$1, making the effective rotation frequency $(x,p)$2 energy dependent and causing angular-velocity dispersion across the ensemble. For $(x,p)$3 and $(x,p)$4, the ensemble spans $(x,p)$5 to $(x,p)$6, giving an angular dispersion after $(x,p)$7 of $(x,p)$8, which explains the observed crescent-shaped reconstructions (Zhou et al., 2014).

The accelerator analogue is amplitude- or energy-dependent distortion from field errors, fringe fields, misalignments, dispersion, or space charge. FAST frames the coupled 4D problem in terms of linear transport and symplectic invariants such as eigenemittances, showing that mixed projections can reveal correlations that uncoupled fits miss (Romanov, 2018). FACET-II shows a different class of departure from rms optics: superimposed bunches with different transverse phase spaces can form an X-shaped composite distribution, and charge tuning can reveal space-charge-dominated dynamics that are not evident in standard Gaussian-based emittance extraction (Ye et al., 8 Aug 2025).

The higher-dimensional extensions make these couplings explicit rather than treating them as nuisances. The 5D method reconstructs

$(x,p)$9

by combining quadrupole-based transverse rotations with a TDS whose polarization angle $x$0 is adjustable over approximately $x$1. In normalized coordinates, the transport in each plane is a pure rotation,

$x$2

and similarly in $x$3, which is why slice-wise 4D inversion becomes possible (Jaster-Merz et al., 2023). FLASHForward then used the reconstructed 5D density to generate a particle distribution for simulations and to extract the transverse 4D slice emittance $x$4, finding an average relative discrepancy of approximately 5% between $x$5 and $x$6, with a minimum of approximately 1% near $x$7, indicating minor residual cross-plane correlations in the core (Jaster-Merz et al., 19 May 2025).

The 6D extensions pursue a still broader goal: using transverse images under multiple machine settings to infer the full source distribution over $x$8. The two-stage CNN trained on KEK-ATF ASTRA data reconstructs 15 pairwise 2D histograms from only sixteen $x$9–$f(y,y')$00 screen images measured at a dispersive location under different RF phases and solenoid fields, with inference in under a minute on an NVIDIA RTX A400 GPU (Mukherjee et al., 3 Mar 2026). PAL-XFEL’s generative phase space reconstruction instead uses a transformer-based generative neural network combined with differentiable Bmad-X tracking and standard accelerator elements, including a quadrupole, bunch compressor, X-band linearizer, and screens, to reconstruct a near-unique 6D state whose downstream predictions match withheld measurements (Kim et al., 28 Feb 2025).

These higher-dimensional works do not replace transverse tomography so much as subsume it. The published descriptions state directly that within a full 6D reconstruction, transverse tomography corresponds to the marginal distributions over $f(y,y')$01 and $f(y,y')$02 extracted from the reconstructed 6D density (Kim et al., 28 Feb 2025). This suggests that the boundary between “transverse tomography” and “full phase-space diagnostics” is increasingly methodological rather than conceptual.

6. Resolution limits, artefacts, validation, and methodological outlook

Tomography is limited not only by detector resolution and shot noise but by the measurement geometry itself. The most explicit formulation of this point is given by the Gram-operator framework, which states that the measurement induces a sampling operator $f(y,y')$03 whose eigenvalues determine the experimentally accessible modes and the reconstruction bandwidth. Large eigenvalues correspond to reliably resolved modes; small eigenvalues correspond to weakly accessible modes that are highly sensitive to noise. Features supported predominantly by modes below a threshold $f(y,y')$04 should be treated as artefacts or as hypotheses requiring additional measurements (Hradil et al., 28 May 2026).

In practical accelerator tomography, the same limitation appears as finite-angle conditioning, sparse projection sets, or insufficiently diverse optics. FAST notes that 22 images, corresponding to approximately $f(y,y')$05 coverage in each plane, were sufficient for a stable 4D inversion on a $f(y,y')$06 grid, but also remarks that more angles or mixed-coupling optics can improve conditioning, especially for finer grids (Romanov, 2018). The 5D ARES study quantified this dependence: to reach less than or equal to 5% discrepancies in all reconstructed transverse planes, approximately 50 transverse rotation angles and 50 streaking angles were needed; for less than or equal to 10% discrepancies, approximately 30 rotation angles and approximately 25 streaking angles sufficed (Jaster-Merz et al., 2023).

Detector and instrument models matter comparably. Fermilab’s multiwire data require deconvolution of a point-spread function set primarily by wire pitch, wire diameter, mechanical alignment, and electronic response (Garcia et al., 2013). ACHIP explicitly convolves the forward model with the measured wire profile of width $f(y,y')$07 and reports that spatial resolution is ultimately limited by wire roughness at approximately 100 nm (Hermann et al., 2021). FLASHForward propagates uncertainty from the measured shear parameters by performing 100 reconstructions with $f(y,y')$08 drawn from Gaussian distributions and notes that a 1% beam-energy error leads to approximately 5% emittance error (Jaster-Merz et al., 19 May 2025).

Validation strategies across the literature are correspondingly stringent. The atom-chip work compares reconstructions against simulations using both energy-dependent phase angles and full numerical integration of Newton’s equations including rare collisions, finding quantitative agreement and phase-space area conservation when dimensions are separable and collisions are infrequent (Zhou et al., 2014). FAST compares measured profiles for the $f(y,y')$09-scan and $f(y,y')$10-scan with corresponding projections of the reconstructed model and propagates the reconstructed second moments to the screen using $f(y,y')$11 (Romanov, 2018). FLASHForward tracks a five-million-particle distribution sampled from the reconstructed 5D density to the screen for all measured settings and reports a centroid-based discrepancy of $f(y,y')$12 across projections, with qualitative agreement of the main features (Jaster-Merz et al., 19 May 2025). PAL-XFEL validates its 6D reconstruction by predicting downstream longitudinal phase space and $f(y,y')$13 correlations that were withheld from training, supporting the claim of near-unique reconstruction (Kim et al., 28 Feb 2025).

The field also contains an objective tension between model-based and learned reconstructions. Kernel inversions, SVD, MLEM, and SART make the forward model explicit and their regularization comparatively transparent [(Zhou et al., 2014); (Romanov, 2018); (Ye et al., 8 Aug 2025); (Jaster-Merz et al., 19 May 2025)]. CNN and generative methods reduce measurement time and can incorporate rich coupled physics, but they depend on the representativeness of simulation data, the fidelity of the transport model, and calibration stability (Mukherjee et al., 3 Mar 2026, Kim et al., 28 Feb 2025). A plausible implication is that future transverse phase-space tomography will increasingly combine both approaches: measurement-adapted subspaces or Gram-eigenmode criteria for what is genuinely resolved, together with flexible learned priors for interpolation within that accessible bandwidth (Hradil et al., 28 May 2026).

Across cold atoms, linacs, FEL injectors, and GeV-class beamlines, the topic is now defined less by a single inversion formula than by a shared architecture: generate controlled phase-space projections, model the transport operator faithfully, regularize according to measurement support, and validate by forward prediction. Within that architecture, transverse phase space tomographic reconstruction has become a diagnostic for weak anharmonicity, 4D coupling, multi-bunch superposition, slice emittance, and full 5D or 6D phase-space structure (Zhou et al., 2014).