Angle Domain Guidance (ADG) Overview

• Angle Domain Guidance (ADG) is a paradigm that elevates angular variables from auxiliary roles to primary organizational coordinates across diverse applications.
• In nuclear matter, ADG retains full angular dependence, enhancing pairing probabilities and reducing free-energy by up to 35% compared to angle-averaged approaches.
• In latent diffusion and control, ADG employs bounded angular corrections to improve image quality and optimize terminal guidance, yielding measurable performance gains.

Searching arXiv for papers using “Angle Domain Guidance” and related ADG expansions to ground the article. Searching arXiv for papers using “Angle Domain Guidance” and related ADG expansions to ground the article. Angle Domain Guidance (ADG) is a field-dependent term rather than a single standardized method. In the cited literature it denotes, at minimum, an angle-dependent gap treatment in asymmetric nuclear matter, an angular guidance rule for latent diffusion sampling, and an adversarial domain generalization framework for time-series regression; related angle-domain formulations also appear in interception, impact-angle control, angle-only guidance, and angle-Doppler signal processing (Shang et al., 2013, Shang et al., 2013, Duan et al., 13 Feb 2026, Jin et al., 21 May 2025, Yahia et al., 6 Jan 2026, Lee et al., 2012, Wang et al., 19 Jun 2026, Bajpai et al., 1 Jun 2026, Gaudet et al., 2019, Maity et al., 13 Jun 2025). A common misconception is that ADG names a single transferable algorithm. The literature instead shows a recurring design principle: angle, orientation, or angular structure is elevated from an auxiliary variable to a primary organizing coordinate.

1. Terminological scope and disambiguation

Within the cited papers, ADG has multiple non-equivalent meanings. This suggests that the acronym is best interpreted locally, with the surrounding field providing the decisive definition.

Field	Expansion / usage	Central role
Asymmetric nuclear matter	angle-dependent gap	Retains full angular dependence of the $^3SD_1$ pairing gap
Latent diffusion	Angle Domain Guidance	Rotates toward the conditional direction while constraining magnitude growth
Drilling time-series regression	Adversarial Domain Generalization	Learns well-invariant representations for SSI prediction
Guidance and control	ADG-style / angle-domain guidance	Shapes terminal geometry through look angle, LOS angle, lead angle, or impact-angle variables
Radar and orbit determination	angle-domain / angle-Doppler formulations	Uses angular sparsity or angle-only observability as the structural basis of estimation

The terminological divergence is not superficial. In nuclear matter, ADG modifies the quasiparticle spectrum and superfluid stability landscape; in diffusion, it changes the geometry of classifier-free guidance; in drilling, it is a domain-invariant representation-learning objective; and in control and sensing, it often denotes a broader family of methods that parameterize guidance or estimation in angular coordinates rather than Cartesian force or state coordinates (Duan et al., 13 Feb 2026, Jin et al., 21 May 2025, Yahia et al., 6 Jan 2026, Wang et al., 19 Jun 2026, Maity et al., 13 Jun 2025).

2. ADG as an angle-dependent gap in asymmetric nuclear matter

In the nuclear-matter literature, ADG denotes an anisotropic neutron-proton pairing state in the isospin-singlet $^3SD_1$ channel. The basic contrast is between the angle-averaged gap (AAG), which replaces the angular structure by a spherical average, and the angle-dependent gap (ADG), which retains the full angular dependence of the gap. Because the tensor force mixes $S$- and $D$-wave components, the gap is intrinsically anisotropic, and the ADG treatment keeps that anisotropy explicit rather than replacing it by an isotropic proxy (Shang et al., 2013, Duan et al., 13 Feb 2026).

For the axisymmetric ansatz, the gap magnitude takes the form

$$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$

and the corresponding quasiparticle branches are

$$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$

This axisymmetric structure breaks full $O(3)$ rotational symmetry down to $O(2)$, so the superfluid keeps rotational symmetry around one axis but selects a symmetry axis. The physical interpretation given in the literature is that the neutron Fermi sphere becomes effectively oblate and the proton one prolate, increasing the pairing probability in selected angular sectors near the average Fermi surface (Shang et al., 2013, Shang et al., 2013).

The 2013 analysis of the pure ADG state reports that, at $\rho=\rho_{0}=0.17\,\mathrm{fm}^{-3}$ with the Argonne $V_{18}$ interaction and $^3SD_1$0 MeV, the ADG gap $^3SD_1$1 exceeds the AAG result for $^3SD_1$2, with a maximum enhancement of about 22% at $^3SD_1$3. The free-energy reduction is larger: the ADG $^3SD_1$4 is lower than the AAG value for $^3SD_1$5, and about 35% lower for $^3SD_1$6. The transition to the normal state is reported as continuous and second order, and the critical asymmetry coincides with the AAG case; at $^3SD_1$7, the quoted values are $^3SD_1$8 for $^3SD_1$9 MeV, respectively (Shang et al., 2013).

When ADG is combined with FFLO pairing, the pair momentum $S$0 introduces an additional directional degree of freedom through

$S$1

so the angle $S$2 between $S$3 and the ADG symmetry axis becomes a variational parameter. The 2013 FFLO-ADG study finds only two locally stable orientations, $S$4 and $S$5, with the former located at small asymmetry and the latter favored at large asymmetry. It also reports that the pure ADG state disappears at about $S$6 at $S$7 MeV, whereas the FFLO-ADG-parallel state survives up to about $S$8 (Shang et al., 2013).

The 2026 phase-diagram study refines that picture. It states that ADG by itself does not extend the asymmetry window for homogeneous superfluidity relative to AAG, but it does suppress normal-superfluid phase separation in the weak-coupling BCS regime, and in combination with FFLO it enlarges the asymmetry range over which superfluidity survives while significantly reducing phase separation. At high density these combined effects can nearly eliminate phase separation; at low density the ADG effect weakens because the destructive effect of asymmetry decreases and the $S$9-wave fraction in the $D$0 channel decreases monotonically. In the BEC regime, the paper states that both ADG and FFLO vanish, while phase separation persists and the superfluid component forms a BEC of deuterons (Duan et al., 13 Feb 2026).

The orientation labels are not uniform across the literature. The 2013 FFLO-ADG paper names $D$1 the FFLO-ADG-orthogonal state and $D$2 the FFLO-ADG-parallel state, whereas the 2026 phase-diagram paper denotes $D$3 by FFLO-ADG-O and $D$4 by FFLO-ADG-P. This suggests that the physically invariant content is the angle $D$5 itself, not the letter label attached to it (Shang et al., 2013, Duan et al., 13 Feb 2026).

3. ADG as angular guidance in latent diffusion

In diffusion modeling, ADG refers to the method introduced in "Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation" (Jin et al., 21 May 2025). Its point of departure is a failure mode of classifier-free guidance (CFG): large guidance weights improve text-image alignment but also produce oversaturation, color distortion, loss of detail, and overall visual degradation. The central empirical claim is that higher CFG weights are associated with larger latent sample norms, and that these larger norms correlate strongly with higher image saturation.

The paper formulates CFG as

$D$6

and argues that linear extrapolation in score space amplifies latent norms. Within its Gaussian-mixture analysis, Theorem 3.2 states that for the same initial point $D$7, the CFG sample is pushed farther in an outer direction than the ordinary ODE sample, and Theorem 3.3 defines an anomalous region

$D$8

whose size increases with the guidance weight $D$9. The paper summarizes the mechanism as a linear enhancement of $$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$0, which improves semantic alignment but increases norms and induces oversaturation and distortion.

ADG replaces extrapolation by a bounded angular correction. Given conditional and unconditional clean estimates, it computes the angle

$$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$1

caps the turning angle by

$$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$2

and then rotates the unconditional estimate toward the conditional one. Proposition 4.1 gives the norm bound

$$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$3

The intended effect is geometric: preserve the beneficial directional or text-alignment effect of CFG while constraining magnitude variation.

The reported experiments are on COCO10k using Stable Diffusion v3.5, with additional compatibility tests on SD v2.1 + DPM-Solver. The metrics are CLIP score, ImageReward (IR), and FID. Relative to CFG, CFG++, and APG, ADG is reported to achieve higher CLIP, much better IR, and competitive or better FID, with reported COCO10k 10-NFE values roughly in the ranges CLIP 0.319–0.324, IR 0.928–0.970, and FID 16.6–17.4. The paper also states that ADG without the angle constraint becomes unstable, whereas ADG with normalization performs similarly to full ADG, supporting the claim that angular adjustment is the primary improvement. The method is explicitly described as a heuristic, like CFG, and as primarily designed for latent diffusion rather than image-domain diffusion (Jin et al., 21 May 2025).

4. ADG as adversarial domain generalization for drilling time series

In drilling analytics, ADG denotes Adversarial Domain Generalization rather than angle-domain geometry. The target problem is regression: predict a continuous Stick-Slip Index (SSI) from 60-second sequences sampled at 1 Hz using surface drilling measurements such as surface torque, surface weight on bit, rate of penetration, flow rate, and total rotation speed variations including surface rotation speed and downhole motor rotation speed. The setting is cross-well generalization: each well is treated as a domain, training occurs on source wells, and testing occurs on entirely unseen wells (Yahia et al., 6 Jan 2026).

The SSI is defined as

$$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$4

where $$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$5 is the downhole bit rotation speed over the 60-second window. ADG uses three modules: a generator or feature extractor $$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$6, implemented with LSTM + Layer Normalization layers; an SSI predictor $$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$7, implemented as a fully connected network; and a domain discriminator $$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$8, also implemented as a fully connected network. The discriminator is attached through a Gradient Reversal Layer (GRL) that multiplies gradients by $$D^{2}(\mathbf{k}) = \Delta_{0}^{2}(k) + \sqrt{2}\,\Delta_{0}(k)\Delta_{2}(k)\left(3\cos^{2}\theta-1\right) + \Delta_{2}^{2}(k)\left(\frac{3\cos^{2}\theta+1}{2}\right),$$9 during backpropagation. The practical objective is written as

$$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$0

where the regression loss is MSE and the domain-classification loss is cross-entropy.

The comparison baseline is a stacked LSTM + LayerNorm regressor trained with MSE only. The paper also evaluates IRM, which uses the same generator and SSI predictor architecture but replaces the domain discriminator by an invariance penalty controlled by $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$1. Hyperparameters are selected by grid search using validation wells distinct from the training wells. For ADG, the searched values are generator regularization $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$2, generator hidden layers $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$3, and trade-off $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$4. Each configuration is evaluated with three random initializations using validation MSE and normalized DTW, across three validation split cases. The selected ADG configuration is generator regularization $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$5, generator hidden layers 6, and $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$6; for IRM, the selected value is $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$7.

The reported cross-well results state that ADG improves performance by 10.86% over the baseline, whereas IRM improves performance by 8.42% over the baseline, with the percentages reported using normalized DTW. The paper also bins SSI into four severity classes—no stick-slip $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$8, low $$E_{\mathbf{k}}^{\pm}=\sqrt{\xi_{\mathbf{k}}^{2}+D^{2}(\mathbf{k})}\pm \delta\varepsilon_{\mathbf{k}}.$$9, moderate $O(3)$0, and severe $O(3)$1—and reports that severe events are detected 60% of the time by ADG, compared with 20% for the baseline. Transfer learning is then applied at test time using the first 10% of labeled sequences from the target well, with fine-tuning of the first two layers of the generator and the first two layers of the SSI predictor; this retraining takes about 1.5 minutes and improves performance for both models, while ADG remains better than the baseline even after TL (Yahia et al., 6 Jan 2026).

5. Angle-domain guidance in control and interception

A broader control literature uses what the papers describe as ADG-style or angle-domain formulations, even when the acronym ADG is not the paper’s primary title. The common structure is that terminal geometry is enforced through angular variables—impact angle, look angle, lead angle, or LOS angle—rather than through direct force-vector design alone (Lee et al., 2012, Wang et al., 19 Jun 2026, Bajpai et al., 1 Jun 2026, Gaudet et al., 2019, Park et al., 2023).

A classical example is weighted optimal impact-angle guidance. The 2012 formulation considers a planar homing engagement against a stationary target under the linearized dynamics

$O(3)$2

with terminal constraints $O(3)$3 and $O(3)$4 in a frame rotated by the desired impact angle. The cost is

$O(3)$5

and the central contribution is a characterization of the weighting functions whose inverse and first three indefinite integrals are analytically obtainable, yielding closed-form state-feedback laws. The special cases $O(3)$6 and $O(3)$7 recover the familiar minimum-energy impact-angle law and the time-to-go weighted law, respectively (Lee et al., 2012).

A more explicitly angle-domain construction appears in the 2026 trajectory-shaping method for simultaneous arrival time and arrival angle control. There the look angle $O(3)$8 is parameterized directly as

$O(3)$9

so that $O(2)$0 and $O(2)$1 are built into the profile. The guidance problem becomes a two-parameter nonlinear solve for $O(2)$2, and the paper develops a two-stage procedure: an analytical warm start based on small-angle and near-linear range approximations, followed by one-dimensional refinement and then the full two-dimensional solve. In the reported cases, the obtained solution is very close to the open-loop optimal solution and remains feasible in highly nonlinear scenarios where some existing methods fail (Wang et al., 19 Jun 2026).

The 2026 hierarchical sliding-mode framework for simultaneous impact-time and impact-angle control uses the lead angle $O(2)$3 as the central geometric variable. The design introduces an impact-time submanifold $O(2)$4, an impact-angle submanifold

$O(2)$5

and a composite manifold

$O(2)$6

with a variable-gain adaptive law acting through the interceptor’s lateral acceleration as the sole control input. The paper proves asymptotic convergence of the composite manifold by a Lyapunov argument and extends the method from stationary targets to constant-velocity targets through the predicted interception point concept (Bajpai et al., 1 Jun 2026).

Angle-only terminal guidance provides a different but related trajectory. The 2019 reinforcement-learning study learns a policy that maps stabilized seeker LOS-angle errors and LOS-angle-rate increments,

$O(2)$7

directly to four binary divert-thruster commands, without range estimation. The policy is trained with PPO and recurrent networks, and is compared with augmented zero-effort miss guidance with perfect target-acceleration knowledge. In the reported randomized and worst-case scenarios, the learned angle-only policy attains better hit statistics and lower or similar fuel use, with reported runtime less than 1 ms on a 2.3 GHz CPU and memory use around 64 KB (Gaudet et al., 2019).

The optimization-based impact-angle literature also contains a nonconvex angle-domain formulation in which the maneuver vector is constrained to remain perpendicular to the LOS vector,

$O(2)$8

The proposed method alternates least-squares control updates with Euclidean projections onto convex and nonconvex feasible sets using an ADMM-like scheme, including a closed-form projection onto the angular constraint set. This direct handling of angular feasibility is explicitly contrasted with linearized optimal guidance laws and convexification-based approaches (Park et al., 2023).

6. Angle-domain representations in sensing and estimation

Angle-domain structure also organizes recent sensing and estimation methods. In angle-only initial relative orbit determination, the measurements are LOS directions without direct range information, so observability is weak in coplanar or nearly stationary configurations. The 2025 active-learning framework models relative motion in the LVLH frame with discrete-time Clohessy–Wiltshire dynamics and uses the normalized relative position as the output,

$O(2)$9

Its contribution is a dual-control design in which thrust inputs both maintain tracking near a reference trajectory and deliberately excite the system to improve observability. In the reported simulations, the closed-loop AL-IROD variant is more stable than open-loop design, the relative MAE remains generally below 2.5% as initial distances vary from 1 km to 4 km, the velocity estimation error norm decreases to about $\rho=\rho_{0}=0.17\,\mathrm{fm}^{-3}$0 m/s by the end, and the reported CPU times are approximately 37.46 s for PD-only offline design, 36.58 s for PD with dithering, 54.10 s for PD with AL offline design, and 58.59 s for AL-IROD closed loop (Xie et al., 27 May 2025).

In bistatic mmWave MIMO radar, the organizing structure is the angle-Doppler (AD) domain rather than a terminal guidance law. The target-plus-clutter scene is modeled as sparse in the joint AoD-AoA-Doppler space, with the vectorized observation written as

$\rho=\rho_{0}=0.17\,\mathrm{fm}^{-3}$1

and then refined by an off-grid first-order expansion in AoD, AoA, and Doppler offsets. The method uses sparse Bayesian learning with Gamma hyperpriors on row precisions, a block majorization-minimization update of the posterior mean, covariance, and off-grid offsets, and a practical clutter rule that identifies near-zero-Doppler components as clutter. The paper also derives CRB and Bayesian CRB benchmarks for joint AoD, AoA, and velocity estimation (Maity et al., 13 Jun 2025).

Taken together, these literatures suggest that ADG is best understood not as a single algorithmic family but as an angular-structure paradigm. In nuclear matter, the decisive structure is anisotropic gap geometry; in latent diffusion, it is rotation in latent direction space; in drilling, it is domain-invariant representation learning despite the acronymal coincidence; in interception, it is the use of angle-referenced terminal manifolds; and in sensing, it is the exploitation of angular observability or angular sparsity. The shared theme is methodological rather than disciplinary: angle or orientation is treated as a primary state, symmetry axis, latent direction, or sparse coordinate, and guidance or inference is built around that choice (Duan et al., 13 Feb 2026, Jin et al., 21 May 2025, Yahia et al., 6 Jan 2026, Wang et al., 19 Jun 2026, Maity et al., 13 Jun 2025).