Compressed Sensing of Compton Profiles for Fermi Surface Reconstruction: Concept and Implementation

arXiv:2210.07701 · cond-mat.mtrl-sci · Submitted 2022-10-14 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Compressed Sensing of Compton Profiles for Fermi Surface Reconstruction".

Title and authors: Kai: So we're looking at this paper now called "Compressed Sensing of Compton Profiles for Fermi Surface Reconstruction: Concept and Implementation," and it’s about using compressed sensing to figure out the three-dimensional momentum distribution from just a limited set of scattering data.

Mira: That’s right, Kai, and what they are tackling is the problem that traditional Compton profiles make hard because they only give you a projection along one direction at a time, which hides the actual features of the electron momentum density.

Lev: It sounds like they're trying to do something that’s essentially tomography, but extended into three dimensions, and it has some real experimental hurdles with how many angles you can actually measure.

Kai: Exactly, and what they propose here is a way to reconstruct that three-dimensional electron momentum density using as little as fourteen directions of scattering data while maintaining high accuracy.

Mira: The paper suggests that the direct Fourier transform method is better than older techniques because its transformation matrix corresponds to the Fourier basis, which is dense and fits the compressed sensing requirement much better than Cormack's method which uses a sparse matrix.

Lev: That density of the matrix is key because if you have a dense A, missing some input data doesn't cause a complete failure of the procedure for all of x, it just leads to minor errors over all elements.

Kai: They also use an optimization problem called generalized least absolute shrinkage and selection operator, or LASSO, which helps by penalizing the absolute value of each component in the solution.

Mira: The function they minimize has a second term that forces the solutions to have more zeros in that matrix Bx, which is what makes compressed sensing work by favoring sparse solutions.

Lev: They set this up with constraints like non-negativity for the momentum density and a sum rule, which means you have to make sure the reconstructed density adds up correctly over all momentum space.

Kai: To fine-tune this whole process, they use an alternating direction method of multipliers to solve that optimization problem under those constraints.

Mira: And they found an optimal regularization parameter lambda equal to one point zero times ten to the minus one, which is what minimizes the validation error when testing how well the results generalize.

Lev: It’s interesting that even with noise in the input data, like a sigma level of one point three times ten to the minus three, they can still track a ridge in that gradient term and determine the Fermi surface feature.

Kai: That stability under noise is important for real experiments because you never get perfectly clean data from your synchrotron source.

Mira: The paper shows that even when noise increases up to a sigma level of one point one times ten to the minus two, they can still capture the feature around the Fermi surface by tracking that ridge in the gradient term of rho p.

Lev: That's solid because if we were building real hardware, we’d need to see how much overhead that noise sensitivity puts on our measurement fidelity.

Kai: Well, so to wrap up this paper on "Compressed Sensing of Compton Profiles for Fermi Surface Reconstruction: Concept and Implementation," they successfully reconstructed the momentum density from just fourteen projections.

Mira: This method points toward a future where we can get better characterization of the electron momentum distribution using much less experimental data than before.

Lev: For us, it means that if we want to use this on real hardware, we have to be careful about how the noise level impacts the final momentum resolution because they showed that it blurs the discontinuity in rho p.

Kai: The main implication for us is that this gives us a clear path forward for fermiology research, and it definitely suggests some exciting new experimental setups could be developed based on these results.

Mira: It really shows how compressed sensing, when applied to this specific problem, can allow you to use the characteristics of the electron momentum density to determine the shape of the Fermi surface with fewer scattering axes.

Lev: We'll see what it takes for us to translate this into a real measurement setup, but from what I see here, they’ve laid a pretty clear path on how to make that reconstruction stable even when things get noisy.

The paper's summary: Kai: So, to recap, they've taken Compton scattering data—which is basically how you measure electron momentum density projected along different directions—and used compressed sensing to reconstruct the full three-dimensional momentum distribution from a surprisingly small number of projections, specifically fourteen.

Mira: That’s right, and it’s not just about getting a picture; it's about overcoming the fundamental limitation in tomography where you need many more angles than you can practically measure. They argue that compressed sensing lets us use that sparsity assumption—that the final image is compressible—to get away with so few measurements.

Lev: From an engineering standpoint, that's huge because if we were trying to build a real experiment for this, needing only fourteen directions instead of what we previously thought might be necessary changes the entire experimental design and acquisition time.

Kai: It really does. The paper shows that even when the input data is noisy—and they test noise levels up to one point one times ten to the minus two—the method can still track that specific ridge in the gradient of the momentum density, which is what defines the Fermi surface feature.

Mira: That's a strong point because it means we don't need perfect input data to get a meaningful result; we just need enough structure left in the signal to follow. They use an optimization problem called generalized least absolute shrinkage and selection operator, or LASSO, with constraints like non-negativity and a sum rule to keep the reconstruction physically sensible.

Lev: And that LASSO setup is what lets them choose that optimal regularization parameter lambda—they found it at one point zero times ten to the minus one—which minimizes validation error when they test how general their results are across different datasets.

Kai: So, if we look at what this means for us working on superconducting qubits and cryogenics, it suggests a new way to characterize the electronic structure of materials like bcc lithium using much less experimental effort than traditional methods.

Mira: Exactly. It shifts the focus from needing perfect data acquisition to needing an algorithm smart enough to extract information from limited, noisy measurements, which is a very realistic scenario in condensed matter physics.

Lev: And this moves us toward developing more robust algorithms for error correction because if the reconstruction is stable under noise like this, it gives us a better benchmark for what kind of signal fidelity we actually need to achieve on real hardware.

Kai: So the paper suggests that compressed sensing combined with Fourier transforms offers a viable path to mapping out these complex electron momentum distributions efficiently.

Mira: It’s definitely a way forward in fermiology research, and it opens up exciting avenues for testing this reconstruction technique on more complicated materials than bcc lithium.

The paper's improvements: Kai: So, looking at the improvements section of this paper, they aren't just happy with their initial result; they're talking about how to make the reconstruction even better and more reliable.

Mira: They are focusing on two main things: first, using the direct Fourier transform method over Cormack’s method because that Fourier basis is dense, which helps satisfy the compressed sensing assumptions better.

Lev: That makes sense from a hardware standpoint; if you use a sparser matrix like in Cormack's, you have to measure way more input data points than if you use the dense Fourier basis.

Kai: They also emphasize using the generalized LASSO optimization problem with specific constraints—non-negativity and that sum rule we talked about earlier—to keep the density map physically correct during reconstruction.

Mira: And they are tuning that regularization parameter lambda even smarter, using cross validation to find the optimal value instead of just picking one number from a guess.

Lev: That dynamic tuning is what’s critical for real hardware; you don't want to bake in a fixed lambda that works perfectly everywhere, because experimental conditions always change.

Kai: The paper suggests this approach leads to better feature capture, even when the input data has some noise, because they showed tracking the ridge in the gradient term of the momentum density is still possible.

Mira: It means we can be more confident in using this technique for characterizing materials where we don't have perfect measurements from day one.

Lev: If we can handle that noise level and get a stable reconstruction with only fourteen directions, it means the overhead for error correction algorithms running on real hardware could actually be reduced because the input data quality is less of a bottleneck.

Kai: It’s about making the whole process more robust to those inevitable experimental perturbations we always have when dealing with electron scattering experiments.

Mira: So, they are pushing for an AI-driven approach that uses these constraints and dynamic tuning to ensure that the reconstructed Fermi surface features are captured cleanly without introducing artificial oscillations.

Lev: That stability is what we need if this method is going to be translated into a system you can actually run on a real machine, because instability in the reconstruction would just be noise amplified.

Kai: This work really sets a high bar for how we approach complex inverse problems in materials science, and it points toward needing more sophisticated AI tools to handle these types of reconstructions reliably.

Conclusion: Kai: So, to wrap up "Compressed Sensing of Compton Profiles for Fermi Surface Reconstruction: Concept and Implementation," they’ve successfully shown that you can reconstruct the three-dimensional momentum distribution from just fourteen scattering directions using compressed sensing.

Mira: It’s a solid demonstration that this approach works even when you have limited input data, provided you use the right mathematical framework to handle the sparsity assumption.

Lev: From an engineering viewpoint, it’s promising because if we can get this fidelity on real hardware without needing an impossible number of measurements, it changes how we plan out our experimental setups for fermiology.

Kai: Exactly. The implication is that we can characterize the shape of a Fermi surface using much less experimental effort than previously thought possible.

Mira: It really shows how these reconstruction techniques can provide a path toward more detailed study of electronic structure in materials, even when traditional methods struggle with the projection limitations.

Lev: I just think it’s important that we keep pushing for algorithms that are stable under noise, because the paper did show that tracking the gradient ridge helps you survive experimental perturbations.

Kai: That stability is key; we need to see if this reconstruction method holds up when we start applying it to more complex materials, maybe things with higher electronic correlations.

Mira: It suggests that the theoretical framework of compressed sensing applied to this problem provides a powerful tool for interpreting experimental scattering data in ways that were previously inaccessible.

Lev: That's what I want to see next; if this reconstruction works reliably, we need to start thinking about how error correction overhead can be managed with such efficient input data requirements.

Research Institute for Interdisciplinary Science, Okayama University · Institute for Solid State Physics, University of Tokyo · Faculty of Culture and Information Science, Doshisha University · Institut f¨ur Physik, Martin-Luther Universit¨at Halle-Wittenberg · Theoretical Physics III, Center for Electronic Correlations and Magnetism, Institute of Physics, University of Augsburg · Center for Condensed Matter Theory and Quantum Computations, Ilia State University · Augsburg Center for Innovative Technologies, University of Augsburg

cond-mat.mtrl-sci

Submitted: 2022-10-14

Updated: 2022-11-09

Comments: 12 pages, 7 figures

Journal ref: J. Phys. Soc. Jpn. 95, 074707 (2026)

DOI: 10.7566/JPSJ.95.074707

License: http://creativecommons.org/licenses/by/4.0/

The gist: Compressed sensing of Compton profiles for Fermi surface reconstruction addresses the difficulty in determining the three-dimensional momentum distribution by utilizing compressed sensing to…

Key concepts

Compton Profile
When X-rays scatter off electrons in a material, the resulting pattern of scattered radiation is called the Compton profile. This profile contains information about how the electron momentum density is distributed along a specific scattering direction.
Compressed Sensing
This technique reconstructs an image from far fewer measurements than traditionally required by assuming the underlying data (the electron momentum density) is sparse. It allows for accurate reconstruction even with limited, incomplete data sets.
Generalized LASSO Optimization
This is a mathematical tool used to find the best solution for reconstructing the momentum density. It minimizes an objective function while enforcing constraints like non-negativity and a total sum rule, ensuring the resulting distribution is physically realistic.

Terminology

Summary

Compressed sensing of Compton profiles for Fermi surface reconstruction addresses the difficulty in determining the three-dimensional momentum distribution by utilizing compressed sensing to reconstruct it from a limited set of scattering data The authors propose a method that allows for the reconstruction of the three-dimensional electron momentum density using as little as 14 directions of scattering data with unprecedented accuracy

The Challenge of Traditional Reconstruction

Compton scattering provides information on the electron momentum density (EMD) projected along a scattering direction, but the resulting Compton profile obscures characteristics in the EMD because discontinuities in the EMD show up only as cusps in the profile, making Fermi-surface features difficult to identify The inverse problem of Eq. (1), which relates the Compton profile to the EMD, is a three-dimensional extension of computed tomography (CT) Compared with original CT, reconstructing the three-dimensional EMD requires recovering two axes eliminated by the double integral in Eq. p. 1

Where the paper says it

Compared with the original CT, the reconstruction of ρ(p) involves practical difficulties for following two reasons. Firstly, the reconstruction of ρ(p) requires recovering of two axes eliminated by the double integral in Eq.

(1), while original CT recovers only one axis Furthermore, experimental reasons limit the number of measurement axes to about 10, which is insufficient for full reconstruction of a three-dimensional Fermi surface

The Compressed Sensing Framework

Compressed sensing solves the underdetermined system y = Ax for x by assuming sparsity in the solution x This is achieved by solving the optimization problem called generalized least absolute shrinkage and selection operator (LASSO) The function to minimize is given by F(x) = 1/2 ky − Axk squared + λkBxk1, where the second term imposes a penalty for the absolute value of each component of Bx, which forces solutions to have more zeros in Bx The choice of the matrix B is essential in this generalized LASSO approach

The Reconstruction Method

The paper reviews two main approaches for reconstruction: the Radon transform method and the direct Fourier transform method The direct Fourier transform method is considered more suitable than Cormack's method because the transformation matrix A corresponds to the Fourier basis e(ip·r), which is dense and satisfies the requirement of compressed sensing, unlike Cormack’s polar coordinate representation where the matrix A is sparse The reconstruction involves solving a linear equation Bi = XAj, where Bi is given only on lines computed from several Compton profiles in Eq. (5) and ρj is defined on a dense grid

Optimization and Validation

The goal is to minimize the function F in Eq. (8) with respect to ρj under constraints, including non-negativity (ρ(p) ≥ 0) and the sum rule (Σ ρ(p) dp = n), which are represented by discrete relations The alternating direction method of multipliers (ADMM) is used to solve this optimization problem under these constraints The regularization parameter λ is determined using the cross validation (CV) method, minimizing the validation error MSEV The optimal value found is λ = 1. p. 5

Where the paper says it

The reconstructed EMD ρ(p) along p k [100] (Left panel) and p k [110] (Right panel) compared with the exact one. The regularization parameter λ is varied among (a) λ = 101 , (b) λ = 10−1 [the optimal value λopt determined by the cross validation method (see Section IV D)], and (c) λ = 10−4 .

0 × 10−1 ≡ λopt

Noise Sensitivity Analysis

The noise level affects the momentum resolution in the reconstructed ρ(p) because as the noise level increases from σ = 10−3 to σ = 10−1, the discontinuity in ρ(p) gets blurred and the peak in ∇ρ(p) becomes broader Nevertheless, even with noise, the Fermi surface can still be determined by tracking the ridge in ∇ρ(p) The demonstration with bcc-Li shows that the method stably captures the feature around the Fermi surface despite experimental perturbations

Conclusion and Future Scope

The compressed sensing technique successfully reconstructs ρ(p) from 14 projections to characterize the shape of the Fermi surface This method provides a path toward further applications in fermiology and can stimulate development on the experiment side The results demonstrate that even if Compton profiles are perturbed by noise, the method stably captures the feature around the Fermi surface

Pruning redundant reasoning tokens

Compton scattering is a well-established technique that can provide detailed information about electronic states in solids Making use of the principle of tomography, it is possible to determine the Fermi surface from sets of Compton-scattering data with different scattering axes Practical applications, however, are limited due to long acquisition time required for measuring along enough number of scattering directions

How the training signal is built

The Compton profile Jζ (pz) is related to the EMD ρ(p) by a double integral Jζ (pz) = Z Z ρ(p) dpxdpy The reconstruction of ρ(p) involves practical difficulties for following two reasons, namely recovering two axes eliminated by the double integral and the limited number of measurement axes The direct Fourier transform method is more suitable because the matrix A in this approach corresponds to the Fourier basis e(ip·r), which is dense and satisfies compressed sensing requirements

Where the gains come from

The key idea of compressed sensing is that the final image of the density map is compressible, meaning the information, hence the number of measured Fourier signals, can be less than the number of pixels in the final image The success indicates that using characteristics of the EMD allows for Fermi-surface reconstruction with a much fewer number of scattering axes than previously required The L1-norm regularization in the generalized LASSO helps in choosing a sparse solution The optimal value λ = 1.0 × 10−1 is determined by minimizing the validation error MSEV

The Compton scattering comprises the collision events in which photons (usually X-rays) are inelastically scattered by electrons in materials Since these electrons are in motion, the scattered radiation is Doppler-broadened and its measurement provides information on the electron momentum density (EMD) projected along the scattering direction [1, 2] Compton scattering measurements play an important role in investigations of the finitetemperature electronic structure, and supplies complementary information to other experiments such as the angle-resolved photoemission spectroscopy (ARPES) and the de Haas–van Alphen measurement Experimental Compton scattering studies on elemental Li and Al revealed marked influences of electronic correlations in particular in Li [3–5]

The reconstruction of ρ(p) is performed by representing the integral in Eq. (4) with a discrete sum over pj on a uniformly spaced grid, ∆pcalc, within a cube of volume (2Pmax) cubed The linear equation Bi = XAj forms an underdetermined system that has fewer equations than the number of unknown variables The objective function F in Eq. p. 3

Where the paper says it

We represent the integral in Eq. (4) with a discrete sum over pj on a uniformly spaced grid, ∆pcalc, within a cube of volume (2Pmax) 3 .

(8) is minimized with respect to ρj under constraints, including the non-negativity and sum-rule relations

The computational cost for the ADMM algorithm grows as O(N 3), which limits the feasible maximum system size The matrix inversion in Eq. (C4) is the most expensive computation, with a cost of O(N 3), determining the upper limit of the system size The optimal λ = 1. p. 9

Where the paper says it

The cost for the LU decomposition is O(N3 ), while the cost for the updates is O(N2 ), where N is the dimension of x (N should be replaced with Nirr when the symmetry is applied as presented in Appendix B). Therefore, the computational cost and the memory storage required for the LU decomposition determine the upper limit of the system size.

0 × 10−1 results in only 1. p. 1
Where the paper says it

Our reconstruction method and technical details in practical calculations are presented in Section III and IV, respectively. Demonstrative results are presented in Section V focusing on the noise of the input data.

8% (98.2%) of∇ρ(p) being finite (zero) in the final result

The electronic structure of alkali metals is calculated within density functional theory (DFT) [65, 66] using the spin-polarized relativistic Korringa-Kohn-Rostoker (SPR-KKR) method [67] The selfconsistent LSDA calculations are performed with a 62 × 62 × 62 mesh in the Brillouin zone [67] Jζ (pz) and ρ(p) are obtained by the energy integral in the complex plane on a semi-circular contour with 32 points and a rectangular grid in the momentum space with a cutoff pmax = 10 a.u. [69] The bin size of Jζ (pz) is ∆pexp = 0. p. 3

Where the paper says it

The results for the principal directions, [001], [110], and [111], have been published in Ref. The bin size of Jζ (pz) is ∆pexp = 0.

01 a.u., which is comparable to the experimental one ∆pexp = 0. p. 3
Where the paper says it

The bin size of Jζ (pz) is ∆pexp = 0. , which is comparable to the experimental one ∆pexp = 0.

02 in Ref. [4]

The discrete cosine transformation in Appendix A defines Bn by Bn = ∆pexp " J0 + (−1)nJN−1 + 2/N X-2 k=1 Jk cos (πkn / N − 1) This discrete cosine transformation is classified as Type I in SciPy python package Symmetry operations are applied to reduce the number of p-points from N to Nirr, where Nirr/N ∼ 1/48 for Oh pointgroup symmetry The size of the original matrix A is (M × N), while A˜ is (M × Nirr)

The computational cost for the ADMM update steps is O(N 2), and with symmetry applied, this cost scales as O(Nirr 3), which limits the feasible maximum system size The downfolding of the derivative term kDρk1 eliminates the N-scale in kDρk1, leading to a more manageable problem size The optimal λ = 1.0 × 10−1 results in only 2,103 components being finite out of 3Nirr = 119,133 in Dξρ

The paper is organized as follows: first review the concept of the compressed sensing in Section II, followed by reconstruction method and technical details in practical calculations in Sections III and IV, and demonstrative results focusing on the noise of the input data in Section V The paper is summarized in Section VI

The reconstruction of ρ(p) is obtained by solving Eq. (8) with the constraints, leading to the solution x∗ The minimization process involves minimizing F in Eq. p. 9

Where the paper says it

(2) plus the constraints in Eq. The advantage of the latter form, Fe(x, z, z 0 ), is that the minimization with respect to x, z, and z 0 can be done using analytical formulas.

p. 9
Where the paper says it

(2) plus the constraints in Eq. The advantage of the latter form, Fe(x, z, z 0 ), is that the minimization with respect to x, z, and z 0 can be done using analytical formulas.

(8) with respect to ρj under the constraints, Eq. p. 3
Where the paper says it

In the discrete representation, the above two relations are written as ρj ≥ 0 , X j ρj = n, (11) Our goal is to minimize the function F in Eq. (8) with respect to ρj under the constraints, Eq.

p. 3
Where the paper says it

In the discrete representation, the above two relations are written as ρj ≥ 0 , X j ρj = n, (11) Our goal is to minimize the function F in Eq. (8) with respect to ρj under the constraints, Eq.

(11) The optimization-validation process is done for K combinations of yT and yV to determine the optimal λ

The validation error MSEV represents to what extent the fitting result is general, meaning the ability that results infer different datasets The obtained optimal value λ = 1.0 × 10−1 is better understood by analyzing the effect of the L1-norm regularization At λ = λopt, 2,103 components are finite out of 3Nirr = 119,133 in Dξρ The noise level affects the momentum resolution, but the Fermi surface can still be determined by tracking the ridge in ∇ρ(p)

The obtained optimal value is better understood by analyzing the effect of the L1-norm regularization Figure 6(b) shows the number of non-zero components N>0 of the L1 term [the second term in Eq. (8)] as a function of the regularization parameter λ Below the optimal λ, N>0 rapidly increases as λ decreases, which means components that appear only for small-λ are used to fit minute structure of the input data, i. p. 6

Where the paper says it

Below the optimal λ, N>0 rapidly increases as λ decreases. Such components that appear only for small-λ are used to fit minute structure of the input data, that is, noise, and thus increase validation errors.

e., noise At λ = λopt, 2,103 components are finite out of 3Nirr = 119,133 in Dξρ The validation error MSEV is minimized at λ = 1. p. 6
Where the paper says it

Such components that appear only for small-λ are used to fit minute structure of the input data, that is, noise, and thus increase validation errors. At λ = λopt, 2,103 components are finite out of 3Nirr = 119,133 in Dξρ, meaning that only 1.

0 × 10−1 The demonstration with bcc-Li will lead to further applications to more complicated materials This work will contribute to accelerate the research into fermiology and stimulate development on the experiment side

The paper is organized as follows: we first review the concept of compressed sensing in Section II, our reconstruction method and technical details in practical calculations are presented in Section III and IV, demonstrative results are presented in Section V focusing on the noise of the input data The paper is summarized in Section VI The authors acknowledge support from JSPS KAKENHI grants and JST CREST, and MM acknowledges financial support from the Deutsche Forschungsgemeinschaft

The input data y is split into K groups randomly, with (K − 1)-groups of data being used for training The summation of the index i in Eq. (8) is taken over the subset yT, which represents the training data The converged solution x∗ is validated with yV, where MV = M/K and PV = 1 − PT The validation error MSEV represents to what extent the fitting result is general, meaning the ability that results infer different datasets

The obtained optimal value is better understood by analyzing the effect of the L1-norm regularization, as shown in Figure 6(b) The noise level affects the momentum resolution in reconstructed ρ(p), but tracking the ridge in ∇ρ(p) allows determination of the Fermi surface The demonstration with bcc-Li will lead to further applications to more complicated materials This work will contribute to accelerate the research into fermiology and stimulate development on the experiment side

The paper is organized as follows: we first review the concept of compressed sensing in Section II, our reconstruction method and technical details in practical calculations are presented in Section III and IV, demonstrative results are presented in Section V focusing on the noise of the input data The paper is summarized in Section VI

The validation error MSEV represents to what extent the fitting result is general, meaning the ability that results infer different datasets The obtained optimal value is better understood by analyzing the effect of the L1-norm regularization, as shown in Figure 6(b) The noise level affects the momentum resolution, but tracking the ridge in ∇ρ(p) allows determination of the Fermi surface The demonstration with bcc-Li will lead to further applications to more complicated materials This work will contribute to accelerate the research into fermiology and stimulate development on the experiment side

The reconstruction of ρ(p) is obtained by solving Eq. (8) with the constraints, leading to the solution x∗ The minimization process involves minimizing F in Eq. (8) with respect to ρj under the constraints, Eq.

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 AI system can reconstruct three-dimensional momentum distribution of bcc-Li from as little as 14 directions of scattering data with unprecedented accuracy by leveraging compressed sensing to overcome limitations in traditional tomography.

  2. The improved system can perform Fermi surface reconstruction using Fourier transform based methods, which is more suitable than Cormack's method because the transformation matrix A is dense and satisfies the requirements of compressed sensing, thereby avoiding artificial oscillations.

  3. The AI system can utilize a generalized LASSO optimization problem with constraints to minimize a function F that enforces non-negativity and a sum rule on the reconstructed momentum density rho p, leading to stable reconstruction even in the presence of noise.

  4. The improved system can dynamically determine an optimal regularization parameter lambda using cross validation, ensuring that the reconstructed Fermi surface features are captured without introducing unphysical features like artificial oscillations observed in direct Fourier transform methods.

Abstract

Compton scattering is a well-established technique that can provide detailed information about electronic states in solids. Making use of the principle of tomography, it is possible to determine the Fermi surface from sets of Compton-scattering data with different scattering axes. Practical applications, however, are limited due to long acquisition time required for measuring along enough number of scattering directions. Here, we propose to overcome this difficulty using compressed sensing. Taking advantage of a hidden sparsity in the momentum distribution, we are able to reconstruct the three-dimensional momentum distribution of bcc-Li, and identify the Fermi surface with as little as 14 directions of scattering data with unprecedented accuracy. This compressed-sensing approach will permit further wider applications of the Compton scattering experiments.

Related papers