PaperlayerView exact revision on arXiv

Source-linked AI summary

Information Transmission using the Nonlinear Fourier Transform, Part I: Mathematical Tools

Mansoor I. Yousefi, Frank R. Kschischang

arXiv:1202.3653v3cs.IT

TL;DR

The paper addresses how to transmit information over nonlinear dispersive channels whose signal degrees-of-freedom are coupled by dispersion and nonlinearity. It develops the nonlinear Fourier transform and uses its spectral evolution to formulate nonlinear frequency-division multiplexing. The resulting framework represents signals in discrete and continuous nonlinear spectra, with independent linear channel evolution in the spectral domain, and this first paper develops the underlying mathematical tools.

  • Problem

    Nonlinear dispersive channels such as optical fibers have complicated coupled signal dynamics, making their deterministic input-output map difficult to establish.

  • Method

    The paper develops the nonlinear Fourier transform for integrable channels and uses it to represent signals spectrally and formulate nonlinear frequency-division multiplexing.

  • Results

    NFT spectral components evolve according to simple independent linear equations, while the signal is represented by discrete and continuous nonlinear spectra.

  • Takeaways & Limitations

    The mathematical tools support OFDM-like information encoding in nonlinear spectra for channels generated by integrable Lax-pair systems.

Abstract

from arXiv · show

The nonlinear Fourier transform (NFT), a powerful tool in soliton theory and exactly solvable models, is a method for solving integrable partial differential equations governing wave propagation in certain nonlinear media. The NFT decorrelates signal degrees-of-freedom in such models, in much the same way that the Fourier transform does for linear systems. In this three-part series of papers, this observation is exploited for data transmission over integrable channels such as optical fibers, where pulse propagation is governed by the nonlinear Schrödinger equation. In this transmission scheme, which can be viewed as a nonlinear analogue of orthogonal frequency-division multiplexing commonly used in linear channels, information is encoded in the nonlinear frequencies and their spectral amplitudes. Unlike most other fiber-optic transmission schemes, this technique deals with both dispersion and nonlinearity directly and unconditionally without the need for dispersion or nonlinearity compensation methods. This first paper explains the mathematical tools that underlie the method.

I. INTRODUCTION

The paper introduces the nonlinear Fourier transform as a mathematical framework for transmission over nonlinear dispersive channels. It represents signals through nonlinear spectra so propagation becomes independently evolving spectral channels, supporting nonlinear frequency-division multiplexing.

  • Channel model: The stochastic NLS equation models optical-fiber propagation through the competing effects of chromatic dispersion and Kerr nonlinearity.These effects produce temporal and spectral broadening; their balance can yield solitonic propagation.
  • Motivation: Nonlinear dispersive channels couple signal degrees-of-freedom, making their deterministic input-output relationship difficult to establish.Existing approaches often assume linear dominance or treat nonlinearity as a perturbation or nuisance.
  • Nonlinear Fourier transform: The nonlinear Fourier transform diagonalizes the nonlinear Schrödinger channel by exposing linear structure hidden in an integrable nonlinear PDE.It generalizes the ordinary Fourier transform to certain nonlinear systems.
  • Nonlinear spectra: NFT representations contain discrete and continuous nonlinear spectra, corresponding respectively to solitons and non-solitonic radiation.Solitons retain their shape or return periodically to their initial shape during propagation.
  • Transmission principle: In the spectral domain, the channel acts through simple independent linear equations despite complicated time-domain NLS propagation.This motivates encoding information in the nonlinear spectra, analogously to OFDM in linear channels.
  • Scope: This first article develops the mathematical tools underlying NFT-based transmission and presents them as applicable to a broader class of integrable nonlinear dispersive PDEs.Later parts address numerical NFT methods and NFDM communication algorithms and examples.
  • Motivation: The authors distinguish their NFT motivation from conserved-quantity communication: conventional linear multiplexing in nonlinear channels leads to interference-limited multiuser systems.They instead pursue NFDM by exploiting NLS integrability to seek transmission without deterministic distortions.

III. CANONICAL LAX FORM FOR EXACTLY SOLVABLE MODELS

The canonical Lax framework describes nonlinear evolution through operators whose spectra remain invariant. Its commutator equation connects isospectral operator flows to nonlinear PDEs such as KdV.

  • A. Lax Pairs and Evolution Equations: The framework is presented using operator intuition while warning that infinite-dimensional operator results require care and are reviewed separately.The paper notes that finite-dimensional matrix properties do not automatically carry over to function spaces.
  • A. Lax Pairs and Evolution Equations: An isospectral operator family has eigenvalues independent of the evolution parameter even while its operator entries or defining waveform change.For diagonalizable operators, the family is similar to a fixed multiplication operator.
  • A. Lax Pairs and Evolution Equations: The Lax equation expresses evolution as L_z = [M,L], where the commutator is ML − LM.Every diagonalizable isospectral operator satisfies this differential equation.
  • A. Lax Pairs and Evolution Equations: Conversely, evolving L according to the Lax equation preserves its initial spectrum, producing an isospectral family.The construction uses an invertible solution G to G_z = MG and similarity evolution from the initial operator.
  • A. Lax Pairs and Evolution Equations: For a diagonalizable operator, isospectrality is equivalent to satisfying the Lax commutator equation for some M.If L is self-adjoint, M must be skew-Hermitian.
  • A. Lax Pairs and Evolution Equations: L and M may share dependence on a waveform q, allowing their commutator to generate nonlinear evolution equations.The resulting equations can involve q and its time derivatives; KdV is an example.
  • A. Lax Pairs and Evolution Equations: For KdV, choosing L = D^2 + q and a corresponding M reduces the Lax equation to q_z = q_ttt + q q_t.The zero-order operator term yields the nonlinear PDE, while operator adjoint properties preserve L's eigenvalues.
  • A. Lax Pairs and Evolution Equations: A Lax pair is defined as operators L and M satisfying the Lax equation, with L's eigenvalues independent of z.The operators need not be uniquely chosen for a given nonlinear equation.

B. The Zero-Curvature Condition

The zero-curvature formulation recasts the Lax compatibility condition as an equation involving P, M, and their commutator. This hidden linear structure generates integrable nonlinear PDEs, including NLS, sine-Gordon, and KdV examples.

  • B. The Zero-Curvature Condition: The eigenvalues of L remain constant during an isospectral flow and serve as the spectral quantities underlying the construction.The associated eigenvector evolution is obtained from the Lax equation.
  • B. The Zero-Curvature Condition: Rewriting the Lax system with an operator P and enforcing equality of mixed derivatives yields the zero-curvature condition P_z − M_t + [P,M] = 0.This condition is equivalent to the Lax equation in the formulation used here.
  • B. The Zero-Curvature Condition: The nonlinear equation derived from zero curvature is a compatibility condition between two linear equations, revealing hidden linearity in the nonlinear system.The formulation separates the linear auxiliary equations from the resulting nonlinear evolution.
  • B. The Zero-Curvature Condition: The AKNS system fixes P in a two-component formulation, with the Zakharov-Shabat system arising when r = q and s = −q*.These systems are central to the nonlinear Fourier transform.
  • B. The Zero-Curvature Condition: For the sine-Gordon choices of r and s, the zero-curvature equation produces q_tz = sin(q), while an alternative choice gives q_tz = sinh(q).The two outcomes follow from distinct coefficient selections in the auxiliary operator.
  • B. The Zero-Curvature Condition: For the NLS choice r = q and s = −q*, zero curvature reduces to j q_z = q_tt + 2|q|^2q.This is the normalized nonlinear Schrödinger evolution used as the principal example.
  • B. The Zero-Curvature Condition: A KdV auxiliary-operator choice similarly produces q_z = q q_t + q_ttt.The examples demonstrate that different operator choices generate different integrable equations.
  • B. The Zero-Curvature Condition: The choice of Lax pair for a given nonlinear equation is not unique because scaling, constant shifts, and orthogonal transformations can preserve the equations.Different pairs may therefore represent the same evolution while providing different formulations.

C. Lax Convolution and Integrable Communication Channels

The paper defines integrable channels through Lax pairs and Lax convolution, then develops the spectral machinery underlying the nonlinear Fourier transform. Canonical eigenvectors and time-independent nonlinear Fourier coefficients provide the key representation.

  • Lax Convolution: A Lax-pair system maps q(t,0) to q(t,L) through the evolution equation induced by L_z = [M,L], defining a nonlinear channel operation.The corresponding input-output map has the form q_z = K(q).
  • Lax Convolution: Lax convolution names the action of an integrable system on an input waveform, with output q(t,L) = q(t,0) ˙ (L,M;L).This extends the system viewpoint from ordinary convolution to integrable evolution systems.
  • Integrable Communication Channels: An integrable communication channel is a waveform channel whose noise-free channel is an integrable system.The paper develops communication schemes by selecting different Lax pairs, including models motivated by fiber-optic transmission.
  • Integrable Communication Channels: The adopted noisy model adds distributed band-limited noise during evolution, so noise interacts with the signal according to the governing equation.The paper assumes sufficiently high signal-to-noise ratio to treat the stochastic system as a perturbation of the deterministic one.
  • Nonlinear Fourier Transform: The nonlinear Fourier transform is defined through spectral analysis of the L operator, whose solutions form eigenspaces for each spectral parameter.For the NLS case, the development uses the Zakharov-Shabat system and assumes q(t) belongs to L1(R) and vanishes as |t| approaches infinity.
  • Canonical Eigenvectors and Spectral Coefficients: Canonical eigenvectors are constructed from boundary-conditioned solutions, and their projections yield time-independent nonlinear Fourier coefficients a(λ) and b(λ).The time independence follows from the invariant bilinear structure of the eigenspace.

B. The Nonlinear Fourier Transform

The NFT represents a signal through continuous and discrete spectral functions derived from the Zakharov-Shabat problem. The continuous spectrum corresponds to radiation, while discrete eigenvalues correspond to solitons.

  • Spectral representation: Discrete eigenvalues are isolated zeros of the analytic function a(λ) in the upper half-plane.For the NLS Zakharov-Shabat operator, these satisfy a(λ_j)=0.
  • Spectral interpretation: The continuous spectrum represents the non-solitonic radiation component, whereas the discrete spectrum represents soliton pulses.The continuous component has an ordinary Fourier-transform analogue; the discrete component does not.
  • Transform construction: The NFT is formally defined by the continuous and discrete spectral functions obtained from canonical eigenvectors and spectral coefficients.The continuous function can be computed from a limiting differential-equation solution, while the discrete spectrum is found from zeros of a(λ).
  • Spectral representation: The NFT consists of continuous spectral values λ on the real line and discrete spectral values λ_j in the upper half-plane.Discrete values are zeros of a(λ), while continuous values are associated with the real axis.
  • Transform construction: Analytical computation of the NFT is generally unavailable except in a few special cases.The paper introduces a rectangular-pulse example as one such analytically tractable case.

C. Example: Nonlinear Fourier Transform of a Rectangular Pulse

For a rectangular pulse, the nonlinear spectrum changes with amplitude: weak pulses resemble the ordinary Fourier transform, while stronger pulses develop discrete spectral mass points.

  • Pulse setup: The example considers a rectangular pulse of amplitude A and duration T=t_2−t_1.The pulse is nonzero with amplitude A on the interval [t_1,t_2].
  • Spectral behavior: For small A, the rectangular pulse has no discrete spectrum and its continuous spectrum is essentially the ordinary Fourier transform.This behavior is also consistent with the weak-amplitude limit described for the example.
  • Spectral behavior: As A increases, the continuous spectrum deviates from the ordinary Fourier transform and one or more discrete mass points appear on the jω axis.Figure 4 compares these spectra for T=1 and A=1, 2, and 6.

D. Elementary Properties of the Nonlinear Fourier Transform

The NFT has linear-limit, scaling, shifting, convolution, energy, causality, and propagation properties that make nonlinear channel evolution tractable in spectral coordinates.

  • Basic properties: As ||q||_L1→0, the discrete spectrum vanishes and the continuous NFT approaches the ordinary Fourier transform.The limiting transform is the ordinary Fourier transform of −q*(t).
  • Basic properties: The NFT also has stated weak-nonlinearity, phase, dilation, time-shift, and frequency-shift transformation rules.These properties relate time-domain operations to corresponding spectral-coordinate transformations.
  • Basic properties: Parseval’s identity partitions signal energy into continuous-spectrum energy and discrete-spectrum energy.The quantities ĤE and Ẽ represent the energy in the continuous and discrete spectra, respectively.
  • Causality: For non-overlapping signal portions, nonlinear Fourier coefficients satisfy a causality and layer-peeling property that yields a Markov structure.The property applies when the portions have separated time supports, such as segments of a pulse train.
  • Propagation: Under NLS propagation, eigenvalues remain fixed while nonlinear spectral coefficients acquire the phase factor exp(−4jλ^2z).For discrete eigenvalues, λ_j(z)=λ_j(0), and the same propagation law applies at the corresponding spectral locations.
  • Propagation: Lax convolution becomes multiplication by the channel filter H(λ,z)=exp(−4jλ^2z) in the nonlinear Fourier domain.This is the spectral-domain diagonalization underlying independent component evolution.

VI. AN APPROACH TO COMMUNICATION OVER INTEGRABLE CHANNELS

The paper proposes NFDM by encoding information in nonlinear spectra, applying the inverse NFT for transmission, and recovering spectra with the forward NFT at reception.

  • NFDM architecture: NFDM uses the NFT to convert an integrable nonlinear channel into parallel spectral channels with independent noiseless component propagation.This motivates a nonlinear analogue of orthogonal frequency-division multiplexing.
  • NFDM channel model: The channel model represents output spectral components as input components multiplied by H plus effective spectral noise.It applies to both continuous components and discrete eigenvalue components.
  • NFDM channel model: Spectral-domain noise is generally non-Gaussian, correlated across distinct eigenvalues, and dependent on the entire signal spectrum.A small-noise perturbation approach is suggested for its analysis.
  • NFDM architecture: The transmitter encodes information in nonlinear spectral degrees of freedom and applies the inverse NFT to generate the time-domain signal.The resulting waveform is then sent through the channel.
  • NFDM architecture: At the receiver, the forward NFT is applied and the recovered spectra are compared with the transmitted spectra using a metric d.Constellation design, coding, and modulation are formulated in the spectral domain.
  • Implementation boundary: In practice, the paper’s continuous-time real-line NFT should be replaced by a discrete-time NFT with periodic boundary conditions.The practical implementation is deferred to Part III and related references.

B. The Inverse Transform

The inverse NFT reconstructs the signal from its discrete and continuous spectral functions by solving a matrix Riemann-Hilbert system for canonical eigenvectors. The resulting representation supports transmission over integrable channels.

  • The inverse transform maps the discrete and continuous spectral functions to the signal q(t).
  • The Riemann-Hilbert factorization yields a linear system with 2N + 2 equations for discrete and continuous canonical eigenvectors.
  • Canonical eigenvectors are related to the signal through the Zakharov-Shabat system and to the nonlinear Fourier transform through the Riemann-Hilbert formulation.
  • Solving the Riemann-Hilbert system for V^1 and substituting it with the spectral functions into the reconstruction formula produces q(t).
  • The inverse transform is performed once at the transmitter, while only the forward transform is required in real time at the receiver.
  • The NFT maps a Lax convolution to multiplication and enables nonlinear frequency-division multiplexing for integrable channels.

APPENDIX A SPECTRUM OF BOUNDED LINEAR OPERATORS

This appendix reviews bounded linear-operator concepts needed to distinguish discrete and continuous spectra and to understand spectral diagonalization in Hilbert spaces.

  • Moving from finite-dimensional matrices to infinite-dimensional operators requires care because finite-dimensional results may not carry over.
  • An operator is invertible when it is one-to-one, onto, and has a bounded inverse; in finite dimensions, only one-to-one is additionally required.
  • The discrete spectrum contains eigenvalues with nonzero eigenvectors, whereas the continuous spectrum contains values for which the operator is not surjective.
  • In infinite-dimensional spaces, the spectrum can include both discrete and continuous parts, unlike finite-dimensional Hilbert spaces where it is entirely discrete.
  • Multiplication operators act by sample-wise multiplication and are analogous to diagonal matrices.
  • Every bounded self-adjoint operator in a separable Hilbert space is unitarily equivalent to a multiplication operator.
  • The appendix also sketches NFT properties, including its linear-limit behavior, trace formulas, and propagator composition for concatenated signals.

APPENDIX C RIEMANN-HILBERT FACTORIZATION PROBLEM

The scalar Riemann-Hilbert factorization problem seeks analytic functions on opposite sides of a contour whose boundary values satisfy a prescribed jump relation. Projection operators and Plemelj formulae provide its solution under regularity conditions.

  • Analytic functions satisfy the Cauchy-Riemann conditions, which constrain their real and imaginary parts.
  • A scalar Riemann-Hilbert problem finds functions analytic inside and outside a closed contour with prescribed boundary relation.
  • The projection operator creates sectionally analytic functions whose boundary limits obey a jump condition across the contour.
  • In the homogeneous case, logarithms of the unknown functions convert the multiplicative boundary relation into an additive jump relation.
  • When log g does not satisfy a Hölder condition, multiplying g by a suitable decaying factor can restore the required regularity.
  • The nonhomogeneous problem factors g into g+ and g− and is then solved using Plemelj formulae.

B. The Matrix Riemann-Hilbert Problem

The inverse NFT leads to a matrix Riemann-Hilbert problem whose canonical eigenvectors are constructed using integral representations and analytic continuation. For this particular problem, a projection operator suffices for the solution.

  • Matrix Riemann-Hilbert problems are generally more involved and may not have closed-form solutions.
  • For the matrix problem arising in the inverse NFT, the projection operator is sufficient to solve the problem.
  • The canonical eigenvectors satisfy analyticity properties inherited from the Zakharov-Shabat system.
  • Duhamel’s formula converts the eigenvector differential equation into an integral representation, with transient terms omitted because the boundary condition begins at t = −∞.
  • The impulse-response factors determine analyticity regions: V^1 and V^2 are analytic in C+, while their tilde counterparts are analytic in C−.
  • The integral representation can be treated as a fixed-point map and expanded as a uniformly convergent series under q(t) ∈ L1(R).

APPENDIX E ASYMPTOTIC FORM OF CANONICAL EIGENVECTORS AND

The appendix derives large-|λ| asymptotics for canonical eigenvectors and develops a contour-integration solution of the inverse nonlinear Fourier transform’s Riemann–Hilbert factorization problem.

  • APPENDIX E ASYMPTOTIC FORM OF CANONICAL EIGENVECTORS AND: The appendix uses inverse Fourier transformation and substitutions into equations (41)–(43) to obtain large-|λ| approximations.
  • APPENDIX E ASYMPTOTIC FORM OF CANONICAL EIGENVECTORS AND: An analogous asymptotic expression is obtained for V 1 as λ becomes large.
  • APPENDIX E ASYMPTOTIC FORM OF CANONICAL EIGENVECTORS AND: For |λ| → ∞, q is negligible relative to jλ, so the canonical eigenfunction v(t, λ) approaches its boundary conditions at t = ±∞.
  • APPENDIX E ASYMPTOTIC FORM OF CANONICAL EIGENVECTORS AND: The resulting asymptotic relations include the limiting inner product shown in equation (46).
  • APPENDIX F SOLUTION OF THE RIEMANN-HILBERT PROBLEM: The inverse nonlinear Fourier transform is formulated as a Riemann–Hilbert factorization problem and simplified through appropriate contour integration.
  • APPENDIX F SOLUTION OF THE RIEMANN-HILBERT PROBLEM: Projection equations are divided by a(λ)(λ − ζ) and integrated along the real axis, with the path passing the singularity λ = ζ from above.
  • APPENDIX F SOLUTION OF THE RIEMANN-HILBERT PROBLEM: Cauchy integrals are evaluated using the residue theorem by closing the real-axis path in the upper or lower half-plane.
  • APPENDIX F SOLUTION OF THE RIEMANN-HILBERT PROBLEM: The resulting integral equation relates canonical eigenvectors V 1 and Ṽ 1 to q̂(λ) and q̃(λj), while subsequent evaluations recover equations of the Riemann–Hilbert system.
Loading 1202.3653v3…