Understanding Path Angular Freedom (PAF)

• PAF is a framework coupling path information with angular degrees of freedom, enabling precise control in diverse domains.
• It underlies high-dimensional photonic Fourier transforms, programmable state engineering, and deterministic Bell-state analysis via path–OAM interplay.
• PAF implementations face challenges including optical calibration, interferometric stability, and managing trade-offs in design parameters.

Searching arXiv for the cited PAF-related papers and terminology. Path Angular Freedom (PAF) denotes a family of technical constructions in which path information is coupled to angular structure and used as an explicit resource. In the available literature, the term is not uniform across subfields. In image-based path planning, PAF is introduced explicitly as a smoothness-aware angular term inside A$^$-style search. In several photonic settings, the label is better understood as an operational description of the controlled interplay between transverse path and orbital angular momentum (OAM) degrees of freedom, including high-dimensional quantum Fourier transforms, programmable path–OAM state preparation, and hyperentanglement-assisted Bell-state analysis. A further path-distribution reading on the sphere defines PAF as a measure of the angular spread of path families contributing to OAM eigenstates (Xu, 12 Jul 2025, Kysela et al., 2020, Pabón et al., 2019, Yang et al., 21 Nov 2025, Feenstra, 2023).

1. Terminological scope and principal usages

The literature supports several distinct meanings of PAF rather than a single canonical definition. In "DAA*: Deep Angular A Star for Image-based Path Planning" (Xu, 12 Jul 2025), PAF is the paper’s own term for an adaptive angular regularizer during node expansion. In the high-dimensional QFT architecture of "Fourier Transform of the Orbital Angular Momentum of a Single Photon" (Kysela et al., 2020), the authors do not introduce the term explicitly; the label describes the capability to redistribute a photon’s state between OAM and path, and to use either degree of freedom as the computational register while the other serves as an auxiliary. The Bell-state analysis scheme similarly does not explicitly use the term, but the description fits the joint use of path and OAM ancillae. In the SLM-based state-generation work, PAF is operationalized as independent, programmable control over path index, per-path OAM charge, and per-path complex weight. In the path-distribution treatment of OAM eigenstates, PAF is a derived measure of angular spread rather than an original term of the paper (Pabón et al., 2019, Yang et al., 21 Nov 2025, Feenstra, 2023).

Context	Operational content	Representative paper
Image-based planning	Trade-off between minimizing $\theta_{jik}$ and minimizing $\pi-\theta_{jik}$ during A$^$ expansion	(Xu, 12 Jul 2025)
Single-photon QFT	Redistribution between path and OAM to implement OAM-only or path-only QFT efficiently and deterministically	(Kysela et al., 2020)
SLM state engineering	Independent control of path channel, per-path OAM charge, and per-path amplitude/phase	(Pabón et al., 2019)
Bell-state analysis	Use of path and OAM as auxiliary DOFs for deterministic polarization BSA	(Yang et al., 21 Nov 2025)
Path-distribution framework	Angular-spread functional over the distribution of contributing path families	(Feenstra, 2023)

A common thread is the use of a path variable together with an angular variable to enlarge representational freedom. This suggests a broader interpretation of PAF as a resource for reallocating information, constraints, or measurement signatures between path and angular structure, but the exact mathematical object depends on the application.

2. Path–angular freedom in high-dimensional photonic Fourier transforms

In the single-photon QFT construction, the basic system model uses a qudit encoded across OAM and path. The OAM register is $|\ell\rangle$ with $\ell \in \{0,\dots,d-1\}$ spanning $\mathcal{H}_{\text{OAM}}$, the path register is $|p\rangle$ with $p \in \{0,\dots,d_P-1\}$ spanning $\mathcal{H}_{\text{path}}$, and the composite space is $\theta_{jik}$0 with total dimension $\theta_{jik}$1. A single index $\theta_{jik}$2 is relabeled as $\theta_{jik}$3, with $\theta_{jik}$4. Within this relabeling, the core decomposition is

$\theta_{jik}$5

where

$\theta_{jik}$6

and

$\theta_{jik}$7

This decomposition underlies the paper’s two operational modes: an OAM-only QFT that uses path internally as an auxiliary, and a path-only QFT that uses OAM internally as an auxiliary (Kysela et al., 2020).

The architecture implements the QFT in arbitrarily large power-of-two dimensions using only conventional elements: beam splitters, mirrors, phase shifters, Dove prisms, interferometers, OAM parity sorters, and simple holograms. The OAM-only realization uses an OAM sorter $\theta_{jik}$8, local $\theta_{jik}$9-dimensional OAM-only QFT blocks on each path, an OAM–path controlled-$\pi-\theta_{jik}$0, a $\pi-\theta_{jik}$1-port path Fourier transform, an OAM–path SWAP, and a reverse OAM sorter. The sorter redistributes the input superposition from a single path into $\pi-\theta_{jik}$2 paths according to

$\pi-\theta_{jik}$3

The controlled phase is realized through rotated Dove prisms, since a Dove prism rotated by angle $\pi-\theta_{jik}$4 imparts

$\pi-\theta_{jik}$5

Choosing path-dependent rotations with $\pi-\theta_{jik}$6 yields the required $\pi-\theta_{jik}$7 phase factor. The path-only DFT is realized by a multiport network of beam splitters and phase shifters, with each beam splitter complemented by two mirrors so that OAM sign flips from interface reflections are canceled (Kysela et al., 2020).

The resource scaling is linear in the dimension for power-of-two dimensions. The paper reports empirical counts of approximately $\pi-\theta_{jik}$8 beam splitters, Dove prisms $\pi-\theta_{jik}$9, holograms $^$0, and phase shifters $^$1. The SWAP itself requires approximately $^$2 beam splitters, but the overall architecture still preserves $^$3 scaling. The scheme is passive and, in principle, 100% efficient. The path-only QFT obtained by leveraging OAM as an auxiliary improves on purely path-encoded multiports that scale as $^$4 in beam-splitter count, with the crossing dimension stated to be around $^$5. The paper presents an explicit scheme and scaling analysis, but not a full experimental demonstration of the complete high-dimensional network. Practical limits are interferometric stability, precise Dove-prism rotation, mode purity, beam-splitter ratio errors, hologram calibration, and residual path-dependent loss (Kysela et al., 2020).

3. Programmable path–OAM state engineering

In the SLM-based state-generation scheme, PAF is the capability to independently and programmably control the transverse path index $^$6, the OAM topological charge $^$7 assigned to each path, and the complex weight $^$8 carried by each path in a coherent superposition. The implementation uses a single phase-only SLM divided into $^$9 non-overlapping circular regions. In region $|\ell\rangle$0, a binary $|\ell\rangle$1–$|\ell\rangle$2 fork hologram sets the OAM charge $|\ell\rangle$3, while a blazed grating controls diffraction efficiency and phase. The prepared state has the form

$|\ell\rangle$4

with coefficients restricted to

$|\ell\rangle$5

Accordingly, the prepared state occupies a $|\ell\rangle$6 Hilbert space with one OAM value per path and no intra-path OAM superposition (Pabón et al., 2019).

Amplitude control is achieved through the first-order diffraction efficiency of the blazed grating,

$|\ell\rangle$7

where the phase depth $|\ell\rangle$8 is chosen independently per region. Relative phase control is obtained through a lateral shift $|\ell\rangle$9 of the blazed grating, which adds

$\ell \in \{0,\dots,d-1\}$0

to the $\ell \in \{0,\dots,d-1\}$1-th diffraction order. In the first order, this sets $\ell \in \{0,\dots,d-1\}$2. The field model is described by an aperture transmission

$\ell \in \{0,\dots,d-1\}$3

and an output field

$\ell \in \{0,\dots,d-1\}$4

The construction therefore realizes path–OAM correlations by directly programming the SLM’s spatial regions (Pabón et al., 2019).

The reported demonstrations reach dimensions up to $\ell \in \{0,\dots,d-1\}$5, with fidelities above $\ell \in \{0,\dots,d-1\}$6 in all cases and above $\ell \in \{0,\dots,d-1\}$7 for the $\ell \in \{0,\dots,d-1\}$8 and $\ell \in \{0,\dots,d-1\}$9 examples. The paper gives explicit examples of two-path, five-path, and nine-path states, including equal-weight and unequal-weight superpositions, and verifies independent amplitude, phase, and per-path OAM control by phase-shifting interferometry. The authors define an operational PAF tuple $\mathcal{H}_{\text{OAM}}$0, where $\mathcal{H}_{\text{OAM}}$1 is the number of path channels, $\mathcal{H}_{\text{OAM}}$2 is the number of distinct OAM modes used, and $\mathcal{H}_{\text{OAM}}$3 is the reachable coefficient set; they also propose the scalar metric $\mathcal{H}_{\text{OAM}}$4, which reaches $\mathcal{H}_{\text{OAM}}$5 in the reported experiments. At the same time, the reachable coefficient set is restricted: full arbitrary $\mathcal{H}_{\text{OAM}}$6 matrices $\mathcal{H}_{\text{OAM}}$7 are not targeted, and the architecture realizes a programmable one-OAM-per-path correlation. The main limits are SLM resolution, pixelation, diffraction efficiency, aperture effects, and path cross-talk controlled by the non-overlap condition $\mathcal{H}_{\text{OAM}}$8 (Pabón et al., 2019).

4. Hyperentanglement-assisted Bell-state analysis

In the linear-optical Bell-state analyzer, PAF refers to the joint use of path and OAM as auxiliary degrees of freedom for distinguishing polarization Bell states. The logical Bell states are

$\mathcal{H}_{\text{OAM}}$9

The auxiliary OAM qubit is the subspace $|p\rangle$0, implemented with $|p\rangle$1, and the path qubit is the pair of spatial modes $|p\rangle$2. The analyzed input is hyperentangled across polarization and path, with trivial initial OAM,

$|p\rangle$3

This enlarges the accessible Hilbert space beyond polarization alone and enables deterministic Bell-state analysis in an idealized linear-optical model (Yang et al., 21 Nov 2025).

The optical stack consists of a polarization-controlled OAM shift (P-COS), an OAM-controlled path shift (O-CPS), an OAM Hadamard gate, a polarization rotation by HWP$|p\rangle$4, and final single-photon projective measurement. P-COS is implemented with QWPs and a q-plate and maps

$|p\rangle$5

O-CPS, realized by a Mach–Zehnder interferometer with beam splitters, Dove prisms, phase plates, and spiral phase plates, routes $|p\rangle$6 to remain on the same path and $|p\rangle$7 to flip paths:

$|p\rangle$8

$|p\rangle$9

After P-COS and O-CPS, the $p \in \{0,\dots,d_P-1\}$0 states acquire same-path correlations $p \in \{0,\dots,d_P-1\}$1, whereas the $p \in \{0,\dots,d_P-1\}$2 states acquire cross-path correlations $p \in \{0,\dots,d_P-1\}$3. A subsequent Dove prism on selected arms of photon $p \in \{0,\dots,d_P-1\}$4, an OAM Hadamard in the $p \in \{0,\dots,d_P-1\}$5 subspace, and the polarization rotation convert the $p \in \{0,\dots,d_P-1\}$6 phase into distinct OAM–polarization interference signatures (Yang et al., 21 Nov 2025).

The final single-photon projective measurement uses a PBS followed by an OAM $p \in \{0,\dots,d_P-1\}$7 sorter, so that each photon is measured in the orthonormal product basis $p \in \{0,\dots,d_P-1\}$8. The paper’s central claim is that the four polarization Bell states are mapped to disjoint sets of detector coincidences. In the ideal model, the success probability is therefore 100%, circumventing the 50% linear-optics limit that applies when only polarization and passive linear optics are used without ancillae. The proposal is theoretical rather than experimental and assumes hyperentangled input, ideal q-plates, Dove prisms, spiral phase plates, lossless and phase-stable interferometers, perfect OAM orthogonality and sorting, and ideal detectors. The main practical vulnerabilities are OAM crosstalk, interferometer phase noise, alignment errors in Dove prisms and q-plates, polarization drift, and unequal detector efficiencies (Yang et al., 21 Nov 2025).

5. Smoothness-aware PAF in image-based path planning

In DAA$p \in \{0,\dots,d_P-1\}$9, PAF is introduced as a higher-order angular term that makes A$\mathcal{H}_{\text{path}}$0 adaptive to path smoothness in imitation learning. The motivation is that classic A$\mathcal{H}_{\text{path}}$1 prioritizes path shortening via admissible heuristics but does not explicitly model path smoothness, whereas reference trajectories in imitation-learning datasets may be smooth, human-labeled, or not strictly shortest. The move angle at step $\mathcal{H}_{\text{path}}$2 for the triplet $\mathcal{H}_{\text{path}}$3 is

$\mathcal{H}_{\text{path}}$4

with edges defined by Cartesian coordinates. PAF is then

$\mathcal{H}_{\text{path}}$5

where $\mathcal{H}_{\text{path}}$6 trades off between minimizing $\mathcal{H}_{\text{path}}$7 and maximizing $\mathcal{H}_{\text{path}}$8 through minimizing $\mathcal{H}_{\text{path}}$9. The paper evaluates DAA$\theta_{jik}$00-min with $\theta_{jik}$01, DAA$\theta_{jik}$02-max with $\theta_{jik}$03, and DAA$\theta_{jik}$04-mix with learned $\theta_{jik}$05 (Xu, 12 Jul 2025).

Path shortening is retained through a heuristic distance

$\theta_{jik}$06

with $\theta_{jik}$07, $\theta_{jik}$08, and $\theta_{jik}$09. The higher-order objective augments an A$\theta_{jik}$10-style unary term by adding the angular regularizer:

$\theta_{jik}$11

subject to $\theta_{jik}$12 for internal path nodes. The search is implemented through message passing. Messages are updated as

$\theta_{jik}$13

and the A$\theta_{jik}$14-like evaluation cost is

$\theta_{jik}$15

Differentiable successor selection uses

$\theta_{jik}$16

The relation $\theta_{jik}$17 aligns the message-passing implementation with the higher-order objective (Xu, 12 Jul 2025).

Training uses path loss and, when available, path probability map loss. The path loss is an $\theta_{jik}$18 discrepancy between the search-history map constructed from the softmax scores and the reference path map. The combined objective is

$\theta_{jik}$19

The encoder weights and the parameters $\theta_{jik}$20, $\theta_{jik}$21, and $\theta_{jik}$22 are learned by backpropagation. The paper reports training settings for UNet, CNN, and Transformer backbones across maze, video-game, drone-view, and Aug-TMPD datasets. Evaluation uses SPR, PSIM, ASIM, CD, Hist, and Ep. Over seven datasets, the abstract reports improvements over neural A$\theta_{jik}$23 of $\theta_{jik}$24 SPR, $\theta_{jik}$25 ASIM, and $\theta_{jik}$26 PSIM when the shortest path is plausible. In the supervised PPM setting, the abstract reports that DAA$\theta_{jik}$27 surpasses TransPath by $\theta_{jik}$28 SPR, $\theta_{jik}$29 PSIM, and $\theta_{jik}$30 ASIM. Learned $\theta_{jik}$31 values are dataset dependent: approximately $\theta_{jik}$32, $\theta_{jik}$33, and $\theta_{jik}$34 on MPD, TMPD, and CSM; approximately $\theta_{jik}$35 and $\theta_{jik}$36 on Warcraft and Pokémon; and approximately $\theta_{jik}$37 and $\theta_{jik}$38 on SDD-intra and SDD-inter. The trade-off figure shows that DAA$\theta_{jik}$39 may increase Hist modestly by $\theta_{jik}$40–$\theta_{jik}$41 on maze and SDD datasets, but can decrease Hist by $\theta_{jik}$42–$\theta_{jik}$43 on video-game datasets. The stated failure modes are poor learned $\theta_{jik}$44 maps, excessively small $\theta_{jik}$45 with weak heuristics, and reduced gradient flow due to discretized activation (Xu, 12 Jul 2025).

6. PAF as angular spread in path distributions for OAM eigenstates

The path-distribution framework for OAM eigenstates begins from motion on the sphere $\theta_{jik}$46 and an exact propagator containing both classical geodesics and nonclassical elastica. Introducing an extended angle $\theta_{jik}$47 yields a characteristic angular momentum

$\theta_{jik}$48

which labels how endpoints are related by geodesic transport. The generalized stationary-phase analysis identifies the dominant saddle

$\theta_{jik}$49

with width $\theta_{jik}$50 for large $\theta_{jik}$51. From the amplitude density $\theta_{jik}$52, the construction defines a one-dimensional distribution $\theta_{jik}$53 and then a bivariate angular distribution $\theta_{jik}$54, which is folded to a purely angular distribution $\theta_{jik}$55 satisfying

$\theta_{jik}$56

Numerically, the imaginary part of $\theta_{jik}$57 is suppressed as $\theta_{jik}$58 grows, the real part is non-negative, and the $\theta_{jik}$59 distribution shows peaks at $\theta_{jik}$60 with sidebands that diminish with $\theta_{jik}$61 (Feenstra, 2023).

Within a path-distribution interpretation, PAF can be defined as a functional of the angular marginal

$\theta_{jik}$62

where $\theta_{jik}$63 is a momentum–angle distribution obtained by re-expressing the $\theta_{jik}$64 dependence in terms of linear momentum $\theta_{jik}$65 through $\theta_{jik}$66. Two natural measures are then

$\theta_{jik}$67

and the angular entropy

$\theta_{jik}$68

These quantities are normalized and rotationally invariant about the $\theta_{jik}$69 axis. For a uniform distribution on the sphere, $\theta_{jik}$70, one obtains $\theta_{jik}$71 and $\theta_{jik}$72 (Feenstra, 2023).

The framework provides a replacement for the vector model of orbital angular momentum by assigning a distribution over geodesic-plane tilts rather than a single polar angle. Representative features are explicit. For $\theta_{jik}$73, the angular distribution is uniform, so $\theta_{jik}$74 and $\theta_{jik}$75, which naturally includes the $\theta_{jik}$76 state. For $\theta_{jik}$77 and $\theta_{jik}$78, the distribution peaks near $\theta_{jik}$79, so $\theta_{jik}$80 approaches $\theta_{jik}$81 at large $\theta_{jik}$82. For $\theta_{jik}$83, the distribution sharpens along the axis, so $\theta_{jik}$84 at large $\theta_{jik}$85. A key sum rule is that $\theta_{jik}$86 is constant in $\theta_{jik}$87, so averaging over $\theta_{jik}$88 recovers spherical symmetry. In this setting, PAF is therefore a measure of the angular freedom of the path families whose amplitudes build the OAM eigenstate (Feenstra, 2023).

7. Cross-cutting themes, misconceptions, and technical limits

Across these uses, PAF does not denote a single standardized formalism. In planning, the angular variable is the turn angle between successive path segments. In the photonic works, the angular variable is OAM or a closely related angular-path distribution. In the QFT and Bell-state-analysis papers, the term itself is not introduced explicitly; it is a descriptive label for the controlled interplay between path and OAM. This distinction matters because otherwise one might incorrectly read the literature as describing a single transferable method rather than several structurally related ideas (Kysela et al., 2020, Yang et al., 21 Nov 2025, Xu, 12 Jul 2025).

A second misconception is that PAF always means “minimize angles.” The DAA$\theta_{jik}$89 formulation rejects that simplification directly: purely minimizing $\theta_{jik}$90 can conflict with roundabouts, obstacle-dense regions, and path similarity, while maximizing angle everywhere can over-curve or be inefficient. The photonic uses show an analogous point in a different language: the relevant freedom is not a preference for one angular value, but the ability to redistribute information between registers, engineer orthogonal signatures, or shape an angular distribution. This suggests that PAF is best understood as controlled angular adaptability rather than a single extremal criterion (Xu, 12 Jul 2025, Feenstra, 2023).

The engineering limits are likewise context dependent. In the QFT network, the dominant issues are nested-interferometer stability, precise Dove-prism rotation, beam-splitter ratio errors, hologram calibration, mode purity, and insertion loss. In the single-SLM state generator, the principal bounds come from pixelation, finite diffraction efficiency, aperture effects, and the fact that the reachable coefficients are restricted to $\theta_{jik}$91. In the Bell-state analyzer, determinism is idealized and presumes hyperentangled inputs together with high-fidelity OAM sorting and phase-stable interferometry. In DAA$\theta_{jik}$92, performance depends on the quality of the learned node-prior map, the balance among $\theta_{jik}$93, $\theta_{jik}$94, and $\theta_{jik}$95, and the modest but real trade-off between path optimality and search efficiency. These limits show that the practical value of PAF rests not only on added degrees of freedom, but also on how robustly those freedoms can be calibrated, stabilized, and learned (Kysela et al., 2020, Pabón et al., 2019, Yang et al., 21 Nov 2025, Xu, 12 Jul 2025).

The broader significance of PAF is therefore not terminological uniformity but recurring structure. In high-dimensional photonics it enables efficient and deterministic single-photon transformations, programmable hybrid encodings, and ancillary DOFs for measurement. In planning it provides a differentiable mechanism for jointly optimizing shortening and smoothing. In path-integral descriptions of OAM it furnishes an invariant way to quantify how broadly the contributing path families are distributed in angle. Taken together, these uses identify PAF as a general resource for coupling path variables to angular structure in order to gain expressivity, controllability, or discriminatory power, while the precise mathematical realization remains domain specific (Kysela et al., 2020, Pabón et al., 2019, Yang et al., 21 Nov 2025, Xu, 12 Jul 2025, Feenstra, 2023).