A tomographic analysis of reflectometry data I: Component factorization

Many signals in Nature, technology and experiment have a multi-component structure. By spectral decomposition and projection on the eigenvectors of a family of unitary operators, a robust method is developed to decompose a signals in its components. Different signal traits may be emphasized by different choices of the unitary family. The method is illustrated in simulated data and on data obtained from plasma reflectometry experiments in the tore Supra.

💡 Research Summary

The paper introduces a novel tomographic framework for decomposing multi‑component signals, with a particular focus on plasma reflectometry data. Traditional time‑frequency representations (TFRs) such as those belonging to Cohen’s class suffer from cross‑terms, negative values, and marginal inconsistencies, especially when dealing with non‑stationary signals that simultaneously exhibit time, frequency, and scale characteristics. To overcome these limitations, the authors construct families of unitary operators derived from linear combinations of the non‑commuting time ( t̂ ) and frequency ( ω̂ = i d/dt) operators, as well as the dilation operator D. Three principal operators are defined:

1. B_S(μ, ν) = μ t̂ + ν ω̂ (time‑frequency),
2. B_A1(μ, ν) = μ ω̂ + ν D (frequency‑scale),
3. B_A2(μ, ν) = μ t̂ + ν D (time‑scale).

For each operator a unitary evolution U(μ, ν) = exp(i B) is introduced. The eigenfunctions of U, denoted Ψ_{θ,T,x}(t) with θ = arctan(ν/μ) and eigenvalue x, form an orthonormal basis for any chosen θ. Projecting a signal s(t) onto this basis yields coefficients c_{θ,x}(s) = ⟨s, Ψ_{θ,T,x}⟩, and the corresponding “tomogram” M_s(x, θ) = |c_{θ,x}(s)|². For a fixed θ, M_s is a genuine probability distribution over x; varying θ continuously interpolates between pure time and pure frequency (or scale) representations, thereby providing a family of joint‑feature descriptions without the artifacts typical of conventional TFRs.

The decomposition algorithm proceeds as follows:

1. Choose a θ (or a set of θ values) that balances time and frequency resolution for the signal of interest.
2. Construct a discrete set of eigenvalues x_n = x₀ + 2π n T sin θ, guaranteeing orthonormality.
3. Compute the projections c_{θ,x_n}(s) for all n.
4. Apply an energy threshold ε to discard coefficients whose magnitude squared falls below ε, thereby denoising the signal.
5. Group the remaining coefficients into subsets F_k (spectral intervals) and reconstruct partial signals s_k(t) = ∑{x_n∈F_k} c{θ,x_n}(s) Ψ_{θ,T,x_n}(t). Each s_k is defined as a component of the original signal.

The authors illustrate the method with two simulated examples. In the first, a signal composed of three sinusoidal components (frequencies 25 rad/s and 75 rad/s, the latter active only in two separate time windows) plus white noise is analyzed. By selecting θ = π/5, the tomogram reveals three distinct spectral ridges corresponding to the three components. Using appropriate x‑intervals (45–65, 135–145, 145–155) the authors reconstruct each component with reconstruction errors ranging from –14 dB to –18 dB, while the overall reconstruction error reaches –27 dB. A conventional Fourier projection (θ = π/2) fails to separate the two 75 rad/s bursts, demonstrating the superiority of the tomographic approach.

The second simulated case involves an “incident” chirp whose instantaneous frequency sweeps linearly from 75 Hz to 50 Hz over 20 s, and a delayed “reflected” chirp with a non‑linear phase term and a 3 s delay. After adding noise (SNR ≈ 15 dB), the tomogram again shows two well‑separated ridges when θ = π/5. The incident component is reconstructed from x ≈ 45–50, achieving an error of –9.5 dB; the reflected component is reconstructed from x ≈ 47.5–50.5 with an error of –10 dB. These results confirm that the method can disentangle overlapping time‑frequency structures even when their instantaneous frequencies are nearly identical.

Finally, the technique is applied to real plasma reflectometry data obtained from the Tore Supra tokamak. Reflectometry measures the phase shift of a microwave beam reflected from the plasma cutoff layer, producing signals that contain both the direct (incident) wave and the reflected wave, the latter being delayed and phase‑modulated by plasma density fluctuations. Using the same tomographic framework (θ = π/5), the authors separate the incident and reflected contributions, reconstructing each with errors of roughly –9 dB and –10 dB respectively. This demonstrates that the method is robust against experimental noise and can provide clean component extraction for diagnostic purposes.

Key advantages highlighted include:

• Flexibility: By choosing different operator pairs (time‑frequency, frequency‑scale, time‑scale) the analyst can emphasize the most relevant physical attributes of the signal.
• Exact probabilistic interpretation: The tomogram’s normalization guarantees that the sum of squared projections equals the signal energy, eliminating cross‑terms and negative values.
• Noise resilience: Thresholding in the coefficient domain effectively denoises the signal without sacrificing component fidelity.
• Applicability to real‑world data: Successful decomposition of noisy reflectometry measurements validates the method for plasma diagnostics and potentially other fields involving non‑stationary, multi‑component signals.

The paper also notes practical considerations: the choice of θ and the discretization step for x critically affect resolution and must be guided either by prior knowledge of the signal structure or by automated optimization strategies. Future work may explore adaptive selection of (μ, ν) pairs and extensions to higher‑dimensional data.

In summary, the authors present a mathematically rigorous, experimentally validated tomographic decomposition technique that overcomes the inherent limitations of traditional time‑frequency analysis, offering a powerful tool for extracting physically meaningful components from complex, noisy signals.