A High Precision Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity

arXiv:2611.07223 · gr-qc · Submitted 2026-11-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: General Relativity and Quantum Cosmology. The day in general relativity and quantum cosmology.

Kai: We're looking at the paper: "A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity".

Mira: Here's the gist of the paper. A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity presents a network that predicts quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on

−1100, 80: M Model Architecture and Representation The model utilizes a Transformer decoder that maps binary parameters to the coefficients of a quintic B-spline expansion for all 21 signed inertial strain modes through l = 4 The architecture involves a shared coefficient network fθ with trainable parameters θ mapping scaled input features through an MLP conditioner followed by an attention decoder This decoder applies pre-LayerNorm four-head cross-attention to conditioning tokens and a self-attention layer over all 384 coefficient tokens to couple temporal locations and mode channels The final output involves a shared linear readout mapping each token to the 42 real and imaginary components of the 21 modes, followed by an additive mode-attention branch that couples the 21 modes at each spline index Training Methodology The training proceeds in two stages, combining NR simulations with synthetic teacher waveforms The joint pretraining objective is defined as Lpre(s) = a⟨wiri(s)

U(1)i + µAi: ⟩NR + b r(s)⟨Vi⟩teacher + 10−6∥θ∥2, where a ≃ 0.092 and b ≃ 0.908 The model is pretrained on numerical-relativity (NR) waveforms from the SXS and RIT catalogues together with synthetic NRSur7dq4, SEOBNRv5PHM, and perturbative waveforms Subsequent fine-tuning on SXS strain yields the final model after training draws minibatches of 32 from the 2035 SXS training waveforms alone Performance and Accuracy On the held-out SXS evaluation set of 176 waveforms in 165 physical-system groups, the model achieves a mean balanced strain error of 0.0227 and a mean detector mismatch at 120 M⊙ of 0.0594% This performance is compared against SEOBNRv5PHM, which has a group-mean balanced strain error of 0.0844 and a mean mismatch of 0.7231% The Transformer's group-mean mismatch increases from 0.0245% to 0.0837% for the Transformer and SEOBNRv5PHM across total masses from 60 to 180 M⊙ Symmetry and Regularity The model enforces binary-exchange and equatorial-reflection equivariance through a four-element group G = G = I, E, P, EP The Reynolds average (Eq.

Kai: 11) guarantees exact equivariance c(gx) = Dgc(x) for any network weights Furthermore, the smooth network layers and the symmetric-mass-ratio factor s(ν) make the spline coefficients clmk(x) real analytic in the seven binary parameters on q > 0 The quintic B-spline basis provides C4 continuity for hlm(τ; x) and C2 continuity for Ψ4,lm(τ; x) on (−1100, 80)M Physical Scope and Diagnostics The spline domain τ ∈

−1100, 80: M covers the late-inspiral, merger, and ringdown signal The model's group-mean analytic curvature error at the strain-fitted alignment is 0.1128 on SXS targets Reference-frame diagnostics show a median fixed-azimuth offset of 3.87◦ and a median peak-time shift of 0.64M The model's groupmean analytic curvature error at the strain-fitted alignment is 0.1128 on SXS targets The comparison with native BAM Ψ4 extractions shows all-mode curvature errors growing from 0.085 at q ≃ 1 to 0.425 at q ≃ 8 This suggests the model accurately tracks the precessional amplitude modulation and cross-polarization phase through merger and ringdown We have presented a Transformer model that maps the seven intrinsic parameters of a precessing binary black hole directly to quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on

−1100, 80: M Pretraining on NR waveforms together with synthetic NRSur7dq4, SEOBNRv5PHM (extending coverage to q = 20), and perturbative teachers precedes SXS strain fine-tuning The symmetry-projected spline representation guarantees binary-exchange and equatorial-reflection equivariance, C4 time continuity, and exact derivatives with respect to time and binary parameters On the held-out SXS evaluation set of 176 waveforms in 165 physical-system groups, the model achieves a mean balanced strain error of 0.0227 and a mean detector mismatch at 120 M⊙ of 0.0594% Reference-frame and curvature diagnostics on SXS targets show a median fixed-azimuth offset of 3.87◦, a median peak-time shift of 0.64M, and a mean strain-derived curvature error of 0.1128 Tightening the unaligned reference-frame anchor (where the peak-aligned strain error is 0.2306 at fixed azimuth versus 0.1213 after a rigid rotation) and hybridizing the

−1100, 80: M merger modes with earlier inspiral waveforms will be required before the model can be used in parameter estimation The waveform-generation and automatic differentiation interface, trained network weights, and scripts for reproducing the reported evaluations and figures will be made publicly available under the MIT license upon publication This paper is 2611.07223.

Mira: Page 1 --- A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity presents a network that predicts quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on

−1100, 80: M We have presented a Transformer model that maps the seven intrinsic parameters of a precessing binary black hole directly to quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on

−1100, 80: M Pretraining on NR waveforms together with synthetic NRSur7dq4, SEOBNRv5PHM (extending coverage to q = 20), and perturbative teachers precedes SXS strain fine-tuning The symmetry-projected spline representation guarantees binary-exchange and equatorial-reflection equivariance, C4 time continuity, and exact derivatives with respect to time and binary parameters On the held-out SXS evaluation set of 176 waveforms in 165 physical-system groups, the model achieves a mean balanced strain error of 0.0227 and a mean detector mismatch at 120 M⊙ of 0.0594% Reference-frame and curvature diagnostics on SXS targets show a median fixed-azimuth offset of 3.87◦, a median peak-time shift of 0.64M, and a mean strain-derived curvature error of 0.1128 Tightening the unaligned reference-frame anchor (where the peak-aligned strain error is 0.2306 at fixed azimuth versus 0.1213 after a rigid rotation) and hybridizing the

−1100, 80: M merger modes with earlier inspiral waveforms will be required before the model can be used in parameter estimation The waveform-generation and automatic differentiation interface, trained network weights, and scripts for reproducing the reported evaluations and figures will be made publicly available under the MIT license upon publication Page 2 --- A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity presents a network that predicts quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on

−1100, 80: M We have presented a Transformer model that maps the seven intrinsic parameters of a precessing binary black hole directly to quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on

−1100, 80: M Pretraining on NR waveforms together with synthetic NRSur7dq4, SEOBNRv5PHM (extending coverage to q = 20), and perturbative teachers precedes SXS strain fine-tuning The symmetry-projected spline representation guarantees binary-exchange and equatorial-reflection equivariance, C4 time continuity, and exact derivatives with respect to time and binary parameters On the held-out SXS evaluation set of 176 waveforms in 165 physical-system groups, the model achieves a mean balanced strain error of 0.0227 and a mean detector mismatch at 120 M⊙ of 0.0594% Reference-frame and curvature diagnostics on SXS targets show a median fixed-azimuth offset of 3.87◦, a median peak-time shift of 0.64M, and a mean strain-derived curvature error of 0.1128 Tightening the unaligned reference-frame anchor (where the peak-aligned strain error is 0.2306 at fixed azimuth versus 0.

Title and authors: Kai: So we're looking at this paper titled "A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity". It’s about using a Transformer to predict the coefficients of these merger waveforms.

Mira: That sounds like it’s tackling the complexity of precessing binaries, which is where things get really messy because they aren't just spinning straight. The authors are using this Transformer structure to map those binary parameters directly onto quintic B-spline coefficients for all twenty-one spin-weighted spherical harmonic strain modes through l equals four on a mass range from negative one thousand one hundred M to eighty M.

Lev: Mapping them directly is a big claim, because it implies you can get exact automatic differentiation and C four continuity with respect to time and binary parameters, which is pretty strong for waveform generation. We need to see how that maps onto actual hardware constraints later.

Kai: Right, so the core idea here is that this Transformer takes the basic inputs like the mass ratio and two three dimensional spin vectors and outputs these coefficients for all twenty-one modes across a huge time domain, from negative one thousand one hundred M up to eighty M.

Mira: And they enforce binary exchange and equatorial reflection equivariance through a four element group G equals I, E, P, EP symmetry transformations applied to the network inputs and outputs in section three. That's a big deal for physical consistency.

Lev: Equivariance is good for theory, but when you try to run this on real hardware or even simulate it on supercomputers, that structure has to be efficient. I’m wondering how much overhead that symmetry projection actually adds to the training time and inference speed.

The paper's summary: Kai: The summary says they pretrain this Transformer using a joint objective involving numerical relativity waveforms from catalogues like SXS and RIT, plus synthetic models like NRSur7dq4, SEOBNRv5PHM, and perturbative waveforms. Then they fine-tune it specifically on SXS strain data alone to get the final model.

Mira: That pretraining strategy is clever because it uses a mix of simulation data for the hard parts and synthetic models to cover parameter spaces where NR simulations are sparse, especially when the mass ratio q is greater than six, up to twenty.

Lev: Using those synthetic teachers like SEOBNRv5PHM, which extends multipolar pretraining to six q equals up to twenty, suggests they are trying to build a robust model that handles a wider range of physical configurations than just the ones covered by standard NR simulations. That’s where you start seeing if it’s actually going to run smoothly on real hardware.

Kai: Then they claim this final fine-tuning on SXS strain data yields results, and they report specific accuracy numbers on a held out set of one hundred seventy six waveforms across one hundred sixty five physical systems.

Mira: The performance metrics are pretty solid: a mean balanced root mean square strain error of zero point zero two three seven, and at twelve hundred M the detector mismatch is zero point zero five nine four percent. That’s quite low compared to other models they mentioned, like SEOBNRv5PHM, which had a group mean balanced strain error of zero point eight four four percent.

Lev: zero point zero two three seven for the strain error sounds good on paper, but we have to remember that this is a prediction based on B-spline coefficients over that specific domain from negative one thousand one hundred M to eighty M. That’s a huge range, and we need to see if it holds up when you push it outside those bounds.

The paper's improvements: Kai: One of the main improvements they highlight is the use of a quintic B-spline basis for the waveform representation, which gives C four continuity in time with respect to time and binary parameters. This is coupled with exact automatic differentiation, which I think makes parameter estimation much cleaner if it works out.

Mira: And they also emphasize that by making the spline coefficients c l m k x real analytic on q greater than zero, they get smooth interpolation between precessing and aligned spin paths without needing pairwise time or phase alignment for every single transition. That’s a big step in modeling precessional behavior smoothly.

Lev: Smoothness is nice, but analyticity is a very strict mathematical requirement. It means the function doesn't just have continuous derivatives; it has all of them infinitely many times in that parameter space, which is tough to enforce with standard neural network layers unless you are very careful with the layer design, like they did here.

Kai: They also focused on reference frame errors by aligning predicted SXS evaluation waveforms to their numerical relativity targets using a single rigid azimuthal rotation. This resulted in reducing the fixed azimuth strain error from zero point two three zero six at fixed azimuth down to zero point one two one three after a rigid rotation for physical system means.

Mira: That reduction in the reference frame error is telling, because it shows that aligning things with just one simple rotation is actually quite effective at fixing a significant part of the misalignment issue compared to what you might expect from a naive alignment.

Lev: So, while they achieved this specific reduction of zero point two three zero six to zero point one two one three, we need to be careful because that improvement only applies when you use that rigid rotation; if you try something more complex, the error could jump back up quickly.

Conclusion: Kai: So to wrap this up, the paper presents this Transformer model for precessing binary black hole merger waveforms calibrated to numerical relativity. It maps seven intrinsic parameters directly to quintic B-spline coefficients for all twenty-one signed spin-weighted spherical harmonic strain modes through l equals four on a massive time range from negative one thousand one hundred M to eighty M.

Mira: They achieved this by pretraining on NR waveforms and synthetic teachers like NRSur7dq4 and SEOBNRv5PHM, then fine-tuning on SXS strain data, resulting in a mean balanced root mean square strain error of zero point zero two three seven.

Lev: From my side, it’s important to remember that the model enforces binary exchange and equatorial reflection equivariance through a four element group G equals I, E, P, EP symmetry transformations applied to the network inputs and outputs in section three. That structural guarantee is what makes the output physically sound.

Kai: And they showed that reference frame diagnostics on SXS targets show a median fixed azimuth offset of three point eight seven degrees and a median peak time shift of zero point six four M when comparing to NR targets.

Mira: The comparison with native BAM extractions shows all mode curvature errors growing from zero point zero eight five at q equals one to zero point four two five at q equals eight, which suggests the model accurately tracks the precessional amplitude modulation and cross polarization phase through merger and ringdown.

Lev: They also flag that before this model can be used in parameter estimation, we need to tighten that unaligned reference frame anchor where the peak aligned strain error is zero point two three zero six at fixed azimuth versus zero point one two one three after a rigid rotation, and we’ll need to hybridize those merger modes with earlier inspiral waveforms.

Kai: That means the next step isn't just deploying this model; it’s about refining the anchor before we can actually use it for parameter estimation on real data.

Mira: It sounds like a very detailed map, but that final requirement to tighten the anchor makes it clear that we’re still in the development phase of applying this to actual discovery science.

Lev: Yeah, yeah. The full title is "A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity", and that's what we have today.

William H. Miller III Department of Physics and Astronomy Johns Hopkins University

gr-qc

Submitted: 2026-11-04

Updated: 2026-11-08

Comments: 23 pages, 11 figures, 9 tables; ancillary files include the NR dataset membership and exposure tables; inference code and weights at https://github.com/jdong745/merger-transformer

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity presents a network that predicts quintic B-spline coefficients for all 21 signed spin-weighted…

Key concepts

Quintic B-spline coefficients
These are mathematical coefficients used to represent the waveform shape over time. By using a quintic basis, the model can capture very smooth and continuous changes in the signal, which is crucial for accurately describing how black holes merge and interact during their inspiral, merger, and ringdown phases.
Transformer Decoder
This is a type of neural network architecture designed to process sequential data. In this paper, it takes the binary parameters as input and uses attention mechanisms to map them onto the specific B-spline coefficients needed for all 21 strain modes, effectively learning the complex relationship between physical parameters and waveform features.
Equivariance (Binary-Exchange & Equatorial-Reflection)
Equivariance means the model respects certain symmetries inherent in physics. This paper enforces binary-exchange and equatorial-reflection equivariance, ensuring that swapping the two black holes or reflecting them across the equator results in a predictable change in the predicted strain, which validates physical consistency.
Strain Modes (l=4)
These are specific mathematical components used to decompose gravitational waves into different types of patterns. The model predicts coefficients for all 21 signed spin-weighted spherical-harmonic modes up to l=4, allowing it to capture a wide range of physical aspects of the merger signal.

Terminology

Summary

A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity presents a network that predicts quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on [−1100, 80]M

Model Architecture and Representation

The model utilizes a Transformer decoder that maps binary parameters to the coefficients of a quintic B-spline expansion for all 21 signed inertial strain modes through l = 4 The architecture involves a shared coefficient network fθ with trainable parameters θ mapping scaled input features through an MLP conditioner followed by an attention decoder This decoder applies pre-LayerNorm four-head cross-attention to conditioning tokens and a self-attention layer over all 384 coefficient tokens to couple temporal locations and mode channels The final output involves a shared linear readout mapping each token to the 42 real and imaginary components of the 21 modes, followed by an additive mode-attention branch that couples the 21 modes at each spline index

Training Methodology

The training proceeds in two stages, combining NR simulations with synthetic teacher waveforms The joint pretraining objective is defined as Lpre(s) = a⟨wiri(s)[U(1)i + µAi]⟩NR + b r(s)⟨Vi⟩teacher + 10−6∥θ∥2, where a ≃ 0.092 and b ≃ 0. p. 4908 The model is pretrained on numerical-relativity (NR) waveforms from the SXS and RIT catalogues together with synthetic NRSur7dq4, SEOBNRv5PHM, and perturbative waveforms Subsequent fine-tuning on SXS strain yields the final model after training draws minibatches of 32 from the 2035 SXS training waveforms alone

Performance and Accuracy

On the held-out SXS evaluation set of 176 waveforms in 165 physical-system groups, the model achieves a mean balanced strain error of 0.0227 and a mean detector mismatch at 120 M⊙ of 0. p. 0 p. 0 p. 0 p. 00594% This performance is compared against SEOBNRv5PHM, which has a group-mean balanced strain error of 0. p. 60844 and a mean mismatch of 0. p. 67231% The Transformer's group-mean mismatch increases from 0.0245% to 0. p. 60837% for the Transformer and SEOBNRv5PHM across total masses from 60 to 180 M⊙

Symmetry and Regularity

The model enforces binary-exchange and equatorial-reflection equivariance through a four-element group G = G = I, E, P, EP The Reynolds average (Eq. 11) guarantees exact equivariance c(gx) = Dgc(x) for any network weights Furthermore, the smooth network layers and the symmetric-mass-ratio factor s(ν) make the spline coefficients clmk(x) real analytic in the seven binary parameters on q > 0 The quintic B-spline basis provides C4 continuity for hlm(τ; x) and C2 continuity for Ψ4,lm(τ; x) on (−1100, 80)M

Physical Scope and Diagnostics

The spline domain τ ∈ [−1100, 80]M covers the late-inspiral, merger, and ringdown signal The model's group-mean analytic curvature error at the strain-fitted alignment is 0.1128 on SXS targets Reference-frame diagnostics show a median fixed-azimuth offset of 3. p. 187◦ and a median peak-time shift of 0. p. 564M The model's groupmean analytic curvature error at the strain-fitted alignment is 0.1128 on SXS targets The comparison with native BAM Ψ4 extractions shows all-mode curvature errors growing from 0. p. 5085 at q ≃ 1 to 0.425 at q ≃ 8 This suggests the model accurately tracks the precessional amplitude modulation and cross-polarization phase through merger and ringdown

Conclusion

We have presented a Transformer model that maps the seven intrinsic parameters of a precessing binary black hole directly to quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on [−1100, 80]M Pretraining on NR waveforms together with synthetic NRSur7dq4, SEOBNRv5PHM (extending coverage to q = 20), and perturbative teachers precedes SXS strain fine-tuning The symmetry-projected spline representation guarantees binary-exchange and equatorial-reflection equivariance, C4 time continuity, and exact derivatives with respect to time and binary parameters On the held-out SXS evaluation set of 176 waveforms in 165 physical-system groups, the model achieves a mean balanced strain error of 0.0227 and a mean detector mismatch at 120 M⊙ of 0.0594% Reference-frame and curvature diagnostics on SXS targets show a median fixed-azimuth offset of 3. p. 1 p. 1 p. 187◦, a median peak-time shift of 0. p. 5 p. 5 p. 564M, and a mean strain-derived curvature error of 0. p. 6 p. 6 p. 61128 Tightening the unaligned reference-frame anchor (where the peak-aligned strain error is 0. p. 1 p. 1 p. 12306 at fixed azimuth versus 0. p. 1 p. 1 p. 11213 after a rigid rotation) and hybridizing the [−1100, 80]M merger modes with earlier inspiral waveforms will be required before the model can be used in parameter estimation The waveform-generation and automatic differentiation interface, trained network weights, and scripts for reproducing the reported evaluations and figures will be made publicly available under the MIT license upon publication This paper is 2611. p. 007223.

--- Page 1 ---

A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity presents a network that predicts quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on [−1100, 80]M

We have presented a Transformer model that maps the seven intrinsic parameters of a precessing binary black hole directly to quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on [−1100, 80]M Pretraining on NR waveforms together with synthetic NRSur7dq4, SEOBNRv5PHM (extending coverage to q = 20), and perturbative teachers precedes SXS strain fine-tuning The symmetry-projected spline representation guarantees binary-exchange and equatorial-reflection equivariance, C4 time continuity, and exact derivatives with respect to time and binary parameters On the held-out SXS evaluation set of 176 waveforms in 165 physical-system groups, the model achieves a mean balanced strain error of 0.0227 and a mean detector mismatch at 120 M⊙ of 0.0594% Reference-frame and curvature diagnostics on SXS targets show a median fixed-azimuth offset of 3.87◦, a median peak-time shift of 0.64M, and a mean strain-derived curvature error of 0.1128 Tightening the unaligned reference-frame anchor (where the peak-aligned strain error is 0.2306 at fixed azimuth versus 0.1213 after a rigid rotation) and hybridizing the [−1100, 80]M merger modes with earlier inspiral waveforms will be required before the model can be used in parameter estimation The waveform-generation and automatic differentiation interface, trained network weights, and scripts for reproducing the reported evaluations and figures will be made publicly available under the MIT license upon publication

--- Page 2 ---

A Transformer model for precessing binary black-hole merger waveforms calibrated to numerical relativity presents a network that predicts quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on [−1100, 80]M

We have presented a Transformer model that maps the seven intrinsic parameters of a precessing binary black hole directly to quintic B-spline coefficients for all 21 signed spin-weighted spherical-harmonic strain modes through l = 4 on [−1100, 80]M Pretraining on NR waveforms together with synthetic NRSur7dq4, SEOBNRv5PHM (extending coverage to q = 20), and perturbative teachers precedes SXS strain fine-tuning The symmetry-projected spline representation guarantees binary-exchange and equatorial-reflection equivariance, C4 time continuity, and exact derivatives with respect to time and binary parameters On the held-out SXS evaluation set of 176 waveforms in 165 physical-system groups, the model achieves a mean balanced strain error of 0.0227 and a mean detector mismatch at 120 M⊙ of 0.0594% Reference-frame and curvature diagnostics on SXS targets show a median fixed-azimuth offset of 3.87◦, a median peak-time shift of 0.64M, and a mean strain-derived curvature error of 0.1128 Tightening the unaligned reference-frame anchor (where the peak-aligned strain error is 0.2306 at fixed azimuth versus 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved prediction of all 21 signed inertial strain modes through l=4 on [−1100, 80]M by mapping binary parameters to quintic B-spline coefficients. This allows for exact automatic differentiation and C4 time continuity with respect to time and binary parameters.

  2. Exact enforcement of binary-exchange and equatorial-reflection equivariance in the model through a four element group G = I, E, P, EP symmetry transformations applied to the network output.

  3. Training of the Transformer model using a joint pretraining objective that combines NR simulations with synthetic teachers (NRSur7dq4, SEOBNRv5PHM) and subsequent fine-tuning on SXS strain data to achieve a mean balanced root-mean-square strain error of 0.023.

  4. Ability to evaluate performance across seven observer inclinations by minimizing a shape loss over the time translation delta and rigid azimuthal rotation alpha, allowing for the calculation of detector mismatch M that is sensitive to inclination and mass.

  5. Enhanced parameter-space continuity by ensuring the resulting quintic B-spline coefficients clmk(x) are real analytic on q > 0, enabling smooth interpolation between neighboring waveforms along precessing and aligned-spin paths without pairwise time or phase alignment.

  6. Accurate identification of reference frame errors by aligning predicted SXS evaluation waveforms to their NR targets using a single rigid azimuthal rotation, reducing the fixed azimuth strain error from 0.2306 to 0.1213 for physical-system means.

Sources

Related papers