PaperlayerView exact revision on arXiv

Source-linked AI summary

Information Transmission using the Nonlinear Fourier Transform, Part II: Numerical Methods

Mansoor I. Yousefi, Frank R. Kschischang

arXiv:1204.0830v2cs.IT

TL;DR

The paper addresses how to compute continuous and discrete nonlinear spectra numerically for the Zakharov-Shabat system, where practical signal studies were limited. It proposes and compares numerical schemes, then applies them to communication waveforms and parameter variations. Layer-peeling and spectral methods estimate nonlinear spectra accurately across a wide class of examples, while the experiments reveal eigenvalue trajectories and real-axis emergence or absorption.

  • Problem

    Numerical NFT computation requires robust, reliable methods because analytic solutions are uncommon, while numerical studies across practical signals were lacking.

  • Method

    The paper proposes and compares integration-based, layer-peeling, Ablowitz-Ladik, and spectral matrix eigenvalue methods for the Zakharov-Shabat system.

  • Results

    Layer-peeling and spectral methods gave accurate continuous- and discrete-spectrum estimates over a wide class of examples.

  • Takeaways & Limitations

    Amplitude changes eigenvalue locations and continuous-spectrum shape, whereas phase changes only spectral phase; bandlimited spectra appear confined to a bandwidth-related vertical strip.

Abstract

from arXiv · show

In this paper, numerical methods are suggested to compute the discrete and the continuous spectrum of a signal with respect to the Zakharov-Shabat system, a Lax operator underlying numerous integrable communication channels including the nonlinear Schrödinger channel, modeling pulse propagation in optical fibers. These methods are subsequently tested and their ability to estimate the spectrum are compared against each other. These methods are used to compute the spectrum of various signals commonly used in the optical fiber communications. It is found that the layer-peeling and the spectral methods are suitable schemes to estimate the nonlinear spectra with good accuracy. To illustrate the structure of the spectrum, the locus of the eigenvalues is determined under amplitude and phase modulation in a number of examples. It is observed that in some cases, as signal parameters vary, eigenvalues collide and change their course of motion. The real axis is typically the place from which new eigenvalues originate or are absorbed into after traveling a trajectory in the complex plane.

I. INTRODUCTION

The paper motivates numerical computation of the nonlinear Fourier transform for integrable communication channels and introduces the Zakharov-Shabat setting used throughout. It surveys the spectrum, computational problem, and proposed algorithmic classes.

  • I. INTRODUCTION: The nonlinear Fourier transform comprises continuous spectral values on the real axis and discrete spectral values at upper-half-plane eigenvalues.These spectral components propagate simply under integrable channel evolution, unlike the time-domain signal.
  • I. INTRODUCTION: NFDM encodes information in the nonlinear spectrum for communication over integrable channels such as the optical-fiber NLS channel.At reception, the NFT of the received signal provides spectra used to recover transmitted information.
  • I. INTRODUCTION: Numerical methods are needed because the forward NFT is analytically computable only in a few cases and must be robust, reliable, and sufficiently fast for real-time receivers.The paper addresses this need through algorithm evaluation and simulations over practical pulse shapes and parameters.
  • I. INTRODUCTION: The paper specializes general numerical schemes for the AKNS system to the Zakharov-Shabat system and compares them against analytically solvable cases.It then uses selected layer-peeling, Ablowitz-Ladik, and spectral matrix eigenvalue methods to compute spectra of practical signals.
  • I. INTRODUCTION: The Zakharov-Shabat operator is the Lax operator underlying the NLS equation, and its spectrum is obtained by solving Lv = λv under finite-support and integrability assumptions.The signal is assumed to belong to L1(R) and to be supported on [T1, T2].

III. NUMERICAL METHODS FOR COMPUTING THE CONTINUOUS SPECTRUM

The paper evaluates direct integration and layer-peeling approaches for computing the continuous spectrum. Basic Euler integration is unsatisfactory at affordable step sizes, while layer-peeling is reported as remarkably accurate.

  • III. NUMERICAL METHODS FOR COMPUTING THE CONTINUOUS SPECTRUM: The continuous spectrum is computed by solving the Zakharov-Shabat system numerically for real λ and obtaining the nonlinear Fourier coefficients.The calculation can be repeated over any desired finite set of continuous frequencies.
  • III. NUMERICAL METHODS FOR COMPUTING THE CONTINUOUS SPECTRUM: The forward Euler method iterates a one-step matrix update across a uniform time mesh and then uses the terminal vector to obtain a(λ) and b(λ).The mesh spans the finite signal-support interval [T1, T2].
  • III. NUMERICAL METHODS FOR COMPUTING THE CONTINUOUS SPECTRUM: The one-step Euler method does not produce satisfactory results for affordable small step sizes ǫ.The paper also considers central-difference and higher-order integration schemes to improve accuracy.
  • III. NUMERICAL METHODS FOR COMPUTING THE CONTINUOUS SPECTRUM: Fourth-order Runge-Kutta improves results but is slow for real-time spectrum estimation and does not outperform some later schemes.It is retained for comparison in the numerical simulations.
  • III. NUMERICAL METHODS FOR COMPUTING THE CONTINUOUS SPECTRUM: Layer-peeling applies the NFT’s causality and layer-peeling property to piecewise-constant signal sections through recursive coefficient updates.The implementation can be accelerated by dropping telescoping exponential factors, although this reduces accuracy because of large-small number products.
  • III. NUMERICAL METHODS FOR COMPUTING THE CONTINUOUS SPECTRUM: The layer-peeling method gives remarkably accurate results in estimating the nonlinear Fourier transform.The paper distinguishes this forward-NFT method from unrelated layer-peeling methods used in geophysics and fiber Bragg design.

D. Crank-Nicolson Method

The paper develops Crank-Nicolson and Ablowitz-Ladik discretizations for Zakharov-Shabat computation. Integrable discretization preserves structure and improves handling of oscillatory spectral factors compared with one-step finite differences.

  • D. Crank-Nicolson Method: Crank-Nicolson averages the system matrix across interval endpoints, producing an implicit update that can be rewritten as an explicit matrix iteration.The resulting scheme is reported to give good nonlinear-spectrum estimates.
  • E. The Ablowitz-Ladik Discretization: Ablowitz-Ladik discretization uses the Lax pair of an integrable discrete NLS equation to solve the continuous-time Zakharov-Shabat problem in the spectral domain.The discretization remains integrable for any step size ǫ, not only in the continuum limit.
  • E. The Ablowitz-Ladik Discretization: Under the z transformation, the continuous spectrum maps to the unit circle and the discrete spectrum maps outside it.This transforms the upper-half λ-plane into the exterior of the unit circle in the z domain.
  • E. The Ablowitz-Ladik Discretization: The Ablowitz-Ladik construction discretizes the signal in time while leaving the space derivative intact and averages the nonlinearity across three time samples.Its continuum limit approaches the continuous NLS equation.
  • E. The Ablowitz-Ladik Discretization: Ablowitz-Ladik is a non-finite-difference discretization that handles exp(±jλt) oscillations and thereby enhances accuracy over one-step finite-difference methods.The paper presents it as useful both for time-domain NLS discretization and spectral-domain Zakharov-Shabat computation.

IV. METHODS FOR CALCULATING THE DISCRETE SPECTRUM

The paper computes discrete eigenvalues either by searching for zeros of a(λ) or by solving a discretized matrix eigenproblem. Derivative propagation, Newton-Raphson refinement, and trace-formula checks support eigenvalue estimation and completeness.

  • IV. METHODS FOR CALCULATING THE DISCRETE SPECTRUM: Discrete eigenvalues are zeros of a(λ) in the upper half complex plane and can be visualized as wells in |a(λ)| over a complex mesh.The wells may be deep and narrow, making eigenvalue localization sensitive.
  • IV. METHODS FOR CALCULATING THE DISCRETE SPECTRUM: The paper uses two discrete-spectrum approaches: integration-based coefficient calculation with root finding, and operator discretization as a matrix eigenvalue problem.The integration-based approach requires good initial points and careful convergence handling.
  • IV. METHODS FOR CALCULATING THE DISCRETE SPECTRUM: Derivative information for a(λ) is updated recursively alongside a, avoiding approximate numerical differentiation when computing discrete spectral amplitudes.The same augmented propagation idea updates v and v′ through method-dependent matrices.
  • IV. METHODS FOR CALCULATING THE DISCRETE SPECTRUM: Newton-Raphson uses the available derivative information to search for eigenvalues and often converges quadratically in only a few iterations.For noisy communication signals, transmitted constellation points can repeatedly provide initial conditions; unknown signals use random initial conditions.
  • IV. METHODS FOR CALCULATING THE DISCRETE SPECTRUM: Trace formulas for n = 1, 2, 3 check whether the computed discrete and continuous spectral components account for the signal’s conserved quantities.For n = 1, the identity is analogous to Parseval’s theorem and checks total energy accounting.
  • IV. METHODS FOR CALCULATING THE DISCRETE SPECTRUM: The search process estimates remaining eigenvalues from a continuous-spectrum error vector and repeats until its norm falls below a prescribed tolerance.The error is updated whenever a new eigenvalue is found.

B. Discrete Spectrum as a Matrix Eigenvalue Problem

Matrix eigenvalue formulations compute the discrete spectrum directly, but their computational trade-offs depend strongly on problem size and matrix properties.

  • Matrix-based schemes obtain all discrete eigenvalues at once by solving a regular matrix eigenvalue problem.They do not provide the continuous spectrum.
  • Small matrix eigenproblems can require less effort and do not require prior knowledge of the continuous spectrum or discrete-spectrum size.
  • Large non-Hermitian matrix problems can be slow, generate spurious eigenvalues, and resist restriction to a chosen complex-plane region.
  • The formulation starts from the Zakharov-Shabat operator equation Lv = λv.

1) Central-Difference Eigenproblem:

The central-difference eigenproblem discretizes the Zakharov-Shabat system in time and identifies the point spectrum through a matrix eigenvalue problem.

  • Central-Difference Eigenproblem:: The algorithmic workflow samples the signal, scans a real-axis mesh for continuous spectra, and searches a prescribed upper-half-plane region for eigenvalues.
  • Central-Difference Eigenproblem:: The time-domain formulation replaces the time derivative with a central finite-difference matrix.
  • Central-Difference Eigenproblem:: The point spectrum is contained in the eigenvalues of the discretized matrix.
  • Central-Difference Eigenproblem:: The method can also be expressed through a second-order recursive equation and polynomial root computation in O(N^2) operations.
  • Central-Difference Eigenproblem:: A general-purpose eigenvalue routine may cost O(N^3) because the discretized operator is non-Hermitian.

2) Ablowitz-Ladik Eigenproblem:

The Ablowitz-Ladik eigenproblem rewrites discrete integrable iterations as matrix problems, with time- and frequency-domain formulations for spectral computation.

  • Ablowitz-Ladik Eigenproblem:: The discretized Ablowitz-Ladik equations form coupled recurrences involving shifted signal samples and cyclic vector indices.
  • Ablowitz-Ladik Eigenproblem:: Simplified, normalized, and unnormalized Ablowitz-Ladik variants modify operator entries or recurrence terms while preserving the matrix-eigenproblem framework.
  • Ablowitz-Ladik Eigenproblem:: In the frequency domain, Fourier transforms approximate derivatives and produce a Zakharov-Shabat system involving a Toeplitz matrix.
  • Ablowitz-Ladik Eigenproblem:: The point spectrum is found from the eigenvalues of the resulting matrix.

V. RUNNING TIME, CONVERGENCE AND STABILITY OF THE NUMERICAL METHODS

The methods have first-order convergence but differ in efficiency and stability: continuous-spectrum computation is O(N^2), while non-Hermitian and unstable discretizations impose practical constraints.

  • V. RUNNING TIME, CONVERGENCE AND STABILITY OF THE NUMERICAL METHODS: All iterative methods require O(N) operations per eigenvalue and have the same asymptotic running time with different coefficients.
  • V. RUNNING TIME, CONVERGENCE AND STABILITY OF THE NUMERICAL METHODS: O(N^2) operations are required to compute the continuous spectrum on an N-point eigenvalue mesh, versus O(N log2 N) for an FFT.
  • V. RUNNING TIME, CONVERGENCE AND STABILITY OF THE NUMERICAL METHODS: The global error of all described methods is at least O(ε), so the methods are convergent as N increases for fixed time bounds.
  • V. RUNNING TIME, CONVERGENCE AND STABILITY OF THE NUMERICAL METHODS: Some methods are unstable because unbounded Zakharov-Shabat solutions can amplify discretization errors.
  • V. RUNNING TIME, CONVERGENCE AND STABILITY OF THE NUMERICAL METHODS: Ablowitz-Ladik discretization is reported immune to chaos and numerical instability among the studied discretizations, including in amplifier-noise conditions.
  • VI. TESTING AND COMPARING THE NUMERICAL METHODS: Testing compares matrix-based and search-based schemes on uniformly sampled signals against analytical formulas where available.

A. Satsuma-Yajima Signals

The Satsuma-Yajima signal provides an analytically tractable test case for comparing numerical estimates of its discrete and continuous spectra. Across these tests, spectral and layer-peeling methods are accurate, while Ablowitz-Ladik accuracy depends strongly on sampling and eigenvalue size.

  • Discrete-spectrum estimation: The spectral and central-difference methods accurately estimate the discrete eigenvalues, whereas Ablowitz-Ladik methods add a vertical line of spurious eigenvalues.The spurious eigenvalues from all methods generally lie near the real axis and can be filtered by their negligible imaginary parts.
  • Matrix-based methods: As sample count decreases, spectral and central-difference methods retain reasonable precision, while Ablowitz-Ladik accuracy rapidly deteriorates.This deterioration is associated with the approximation e^(jλε) ≈ 1 + jλε becoming inaccurate when ε is large.
  • Matrix-based methods: The spectral method is generally the most accurate matrix-based scheme; Ablowitz-Ladik performs well when λε ≪ 1 or N ≥ 200 but breaks down near N = 50.Its relative performance also depends on the eigenvalue being estimated.
  • Search-based methods: For the largest eigenvalue, Runge-Kutta, layer-peeling, and Crank-Nicolson have similar accuracy, followed by forward discretization, while Ablowitz-Ladik is less accurate.The comparison concerns search-based estimation as a function of the number of sample points.
  • Computational trade-offs: Runge-Kutta is too slow for practical implementation; search-based methods take an order of magnitude longer than matrix-based methods at small N, while QR costs O(N^3) for large N.Search-based methods also fail below N < 200 because continuous-spectrum energy errors degrade Newton stopping criteria and increments.

B. Rectangular Pulse

The rectangular-pulse example derives the continuous and discrete spectra, then compares numerical methods against the known eigenvalue and spectral behavior. Results show accurate eigenvalue estimation alongside spurious roots, slow discrete-amplitude convergence, and method-dependent behavior under amplitude and phase changes.

  • Analytical spectrum: The rectangular-pulse analysis derives the continuous spectrum and reduces the discrete-spectrum calculation to a constant-coefficient second-order ordinary differential equation.The resulting eigenvalue condition can be generalized to piece-wise constant pulses through NFT causality and layer peeling.
  • Numerical eigenvalues: The exact eigenvalue for q(t) = 2 rect(t) is λ = 1.5713j, and all methods recover it while also producing spurious eigenvalues across the real axis.The central-difference scheme visibly produces fewer spurious eigenvalues, whereas Ablowitz-Ladik schemes add two eigenvalues with large negative real part.
  • Method comparison: The modified Ablowitz-Ladik scheme performs the same as the basic scheme in the rectangular-pulse precision comparison.Consequently, the modified scheme is omitted from the corresponding graphs.
  • Convergence: Convergence of the rectangular pulse's discrete spectral amplitude is generally much slower than eigenvalue convergence.For this pulse, convergence does not occur until N > 1000; the continuous spectra are essentially identical across methods except for slight variations near zero frequency.
  • Parameter variation: As |A| increases, more eigenvalues appear on the imaginary axis, while their spacing decreases as |λ| increases.Ablowitz-Ladik methods reproduce purely imaginary eigenvalues at spurious locations with large real part, and phase addition has no influence on the methods.

C. N-Soliton Signals

The paper examines how numerical methods recover nonlinear spectra and how amplitude, phase, chirp, and bandwidth variations move sinc-signal eigenvalues. These studies reveal eigenvalue emergence, collisions, absorption, and numerical sensitivity.

  • Layer-peeling and spectral methods give accurate nonlinear-spectrum estimates and are selected for subsequent simulations.
  • Amplitude and phase modulation: For A sinc(2t), the spectrum is initially continuous; near A = π, new eigenvalues emerge, including symmetric pairs with nonzero real parts.
  • Amplitude and phase modulation: The truncated sinc function demonstrates that real symmetric signals can have eigenvalues that are not purely imaginary.
  • Amplitude and phase modulation: Eigenvalues near the real axis may be difficult to distinguish from numerical artifacts because their abundance and location depend on sampling and truncation.
  • Amplitude and phase modulation: Constant phase modulation leaves eigenvalue locations and continuous-spectrum magnitude unchanged, while shifting the continuous-spectrum phase.
  • Amplitude and phase modulation: Linear chirp shifts both discrete and continuous spectra, whereas quadratic chirp can move eigenvalues upward, downward, and into the real axis.
  • Amplitude and phase modulation: At about ω = 41.41, two eigenvalues collide, leave the jω axis, and are later absorbed into the real axis.
  • Amplitude and phase modulation: Increasing sinc bandwidth expands the continuous real-frequency range and can cause eigenvalues to collide, divert into quadrants, or reach the real axis.

2) Sinc Wavetrains:

The paper studies how amplitude, phase, pulse separation, and bandwidth constraints shape the discrete spectra of sinc wavetrains. Eigenvalue trajectories include emergence, collision, and absorption events.

  • For two sinc pulses, changing the phase of one amplitude moves an eigenvalue off the jω axis and produces another eigenvalue from the real axis.
  • Increasing pulse separation produces eigenvalue trajectories, real-axis absorption, new eigenvalue emergence, and collisions on the jω axis.
  • When eigenvalues appear to collide or unite, the continuous spectrum develops delta-like spikes and the eigenvalue problem becomes ill-conditioned.
  • For wavetrains containing 16 sinc signals, all 16 bandlimited degrees of freedom are modulated to examine the region occupied by the discrete spectrum.

B. Preservation of the Spectrum

The paper evaluates whether computed spectra are preserved during NLS propagation and compares numerical methods for practical nonlinear Fourier analysis. It reports accurate spectrum estimation but important sensitivity and scope limitations.

  • Spectra are evaluated by comparing numerical outputs with propagation behavior and, more generally, by testing confidence in the computed nonlinear transform.
  • Under NLS propagation, the discrete spectrum is preserved while continuous spectral amplitudes undergo the expected phase change.
  • Layer-peeling and spectral methods accurately estimate continuous and discrete spectra across a wide class of examples.
  • Matrix-based methods are recommended when eigenvalue locations are unknown, while search-based methods suit approximately known eigenvalues.
  • Discrete spectral amplitudes remain highly sensitive to eigenvalue-location errors and are unlikely to provide viable timing-centered information transmission.
  • For an isolated normalized sinc, A < π yields a purely continuous spectrum, while larger amplitudes produce imaginary-axis and nonzero-real-part eigenvalues.
  • Amplitude changes alter eigenvalue locations and continuous-spectrum shape, whereas phase changes affect only the spectrum phase.
  • The paper identifies development of a fast nonlinear Fourier transform analogous to the FFT as requiring significant additional effort.
Loading 1204.0830v2…