Integrable Systems and Inverse Scattering

Integrable systems are nonlinear evolution equations (ODEs/PDEs) exactly solvable via the inverse scattering transform (IST) — a nonlinear analogue of the Fourier transform. The key result: equations like KdV and NLS possess infinitely many conserved quantities and can be linearized by mapping the solution onto scattering data of an associated linear spectral problem, which evolves trivially. Applies to shallow-water waves, nonlinear optics, plasma physics, and Bose–Einstein condensates. Pioneered by Gardner, Greene, Kruskal & Miura (1967) and Lax (1968).

An integrable system in the soliton sense is a nonlinear PDE that can be written as the compatibility condition of an overdetermined linear system (a Lax pair), endowing it with an infinite hierarchy of conservation laws and a method of exact solution by inverse scattering. The inverse scattering transform proceeds in three stages: solve the direct scattering problem to obtain scattering data from initial data, evolve the scattering data linearly in time, then reconstruct the solution by solving a linear integral equation (Gelfand–Levitan–Marchenko). This matters because it provides exact, multi-parameter nonlinear solutions — solitons — that emerge unscathed from collisions, defying the usual expectation that nonlinearity destroys structure. The surprising fact: an infinite-dimensional, nonlinear Hamiltonian system can possess action-angle coordinates and be completely integrable in the Liouville sense, with solitons behaving like elementary particles.

The Core Insight

A nonlinear PDE $u_{t} = K (u)$ is integrable if it is equivalent to the statement that two linear operators $L$ and $M$ satisfy the Lax equation $L_{t} = [M, L]$ . This forces the eigenvalues of $L$ to be time-independent — they are the conserved quantities. The nonlinear flow on $u$ becomes a linear, isospectral flow on the spectrum of $L$ , and the nonlinear problem dissolves into the linear machinery of scattering theory.

Formal Statement

Lax Pair

A Lax pair for an evolution equation $u_{t} = K (u)$ is a pair of linear operators $(L, M)$ , depending on the field $u (x, t)$ , such that the equation is equivalent to the Lax equation:

\frac{\partial L}{\partial t} = [M, L] = M L - L M .

⟳Differentiate the eigenvalue equation

L ψ = λ ψ

t

and demand

λ_{t} = 0

; the compatibility of

ψ_{t} = M ψ

with isospectrality yields exactly this commutator equation.

Here $L$ is the spatial (scattering) operator whose spectrum encodes conserved data, and $M$ generates the time evolution of the eigenfunctions $ψ$ .

Korteweg–de Vries Equation and its Lax Operators

The canonical example is the KdV equation:

u_{t} - 6 u u_{x} + u_{xxx} = 0,

⟳Balance nonlinear steepening

- 6 u u_{x}

against dispersion

u_{xxx}

; rescaling absorbs constants. The "

- 6

" is a normalization that makes the scattering coefficient of

u

equal to one in

L

with Lax pair given by the Schrödinger operator and a third-order operator:

L = - \partial_{x}^{2} + u (x, t), M = - 4 \partial_{x}^{3} + 6 u \partial_{x} + 3 u_{x} .

⟳Choose

M

as the most general antisymmetric third-order differential operator; computing

L_{t} - [M, L]

produces a multiplication operator equal to

u_{t} - 6 u u_{x} + u_{xxx}

, which must vanish.

Isospectrality

If $u (x, t)$ solves the KdV equation and $L = - \partial_{x}^{2} + u$ , then the discrete eigenvalues $λ_{n} = - κ_{n}^{2}$ of $L$ (the bound-state energies) are constants of the motion:

\frac{d λ _{n}}{d t} = 0, n = 1, \dots, N .

⟳From

L ψ = λ ψ

, apply

\partial_{t}

, use the Lax equation

L_{t} = [M, L]

, and project onto the normalized eigenfunction

ψ

; the commutator term integrates to zero, leaving

λ_{t} ⟨ ψ, ψ ⟩ = 0

Inverse Scattering Transform — Solution Scheme

For KdV with rapidly decaying initial data $u (x, 0) \to 0$ as $∣ x ∣ \to \infty$ , the scattering data

S (t) = {R (k, t), {κ_{n}}_{n = 1}^{N}, {c_{n} (t)}_{n = 1}^{N}}

(reflection coefficient

R

, bound-state momenta

κ_{n}

, normalization constants

c_{n}

) evolve linearly:

R (k, t) = R (k, 0) e^{8 i k^{3} t}, c_{n} (t) = c_{n} (0) e^{4 κ_{n}^{3} t}, κ_{n} = const .

⟳Insert the asymptotic form of

ψ

∣ x ∣ \to \infty

(where

u \to 0

, so

M \to - 4 \partial_{x}^{3}

) into

ψ_{t} = M ψ

; matching the plane-wave coefficients

e^{\pm ik x}

gives ODEs

\dot{R} = 8 i k^{3} R

etc.

Why It Holds / Derivation

The Three-Stage IST Pipeline

We solve $u_{t} - 6 u u_{x} + u_{xxx} = 0$ with given $u (x, 0)$ .

Step 1 — Direct scattering ( $t = 0$ ). Solve the time-independent Schrödinger equation

- ψ_{xx} + u (x, 0) ψ = k^{2} ψ .

⟳This is

L ψ = λ ψ

with

λ = k^{2}

; the continuous spectrum

λ > 0

gives scattering states,

λ < 0

gives bound states.

Define the Jost solutions $ψ \sim e^{- ik x}$ as $x \to - \infty$ , with asymptotics $a (k) e^{- ik x} + b (k) e^{ik x}$ as $x \to + \infty$ . The scattering data are the transmission/reflection coefficients $T = 1/ a$ , $R = b / a$ , and the bound states are the zeros of $a (k)$ on the positive imaginary axis $k = i κ_{n}$ .

Step 2 — Time evolution. By the Isospectrality Theorem and the linear ODEs of equation [5], propagate the scattering data forward to time

t

. The bound-state momenta

κ_{n}

are frozen; only phases and normalizations move. Step 3 — Inverse scattering (reconstruction). Form the kernel

B (ξ; t) = n = 1 \sum N c_{n}^{2} (t) e^{- κ_{n} ξ} + \frac{1}{2 π} \int_{- \infty}^{\infty} R (k, t) e^{ik ξ} d k,

⟳

B

is the Fourier-plus-bound-state transform of the scattering data; it serves as the inhomogeneity of the Marchenko equation below.

and solve the Gelfand–Levitan–Marchenko (GLM) integral equation for $K (x, y; t)$ , $y > x$ :

K (x, y; t) + B (x + y; t) + \int_{x}^{\infty} K (x, z; t) B (z + y; t) d z = 0.

⟳

K

is the kernel of the transformation operator

ψ = (1 + K) e^{- ik x}

mapping free solutions to Jost solutions; substituting into the Schrödinger equation and using completeness yields this linear equation.

Finally, recover the solution by differentiating along the diagonal:

u (x, t) = - 2 \frac{d}{d x} K (x, x; t) .

⟳Expand the Schrödinger equation for the transformation operator to leading order; the potential is recovered as twice the

x

-derivative of the diagonal of

K

Because every step after the nonlinear Step 1 is linear, the nonlinear PDE has been completely solved. $■$

Peter Lax | 1926–

Hungarian-American mathematician at the Courant Institute. In his 1968 paper "Integrals of nonlinear equations of evolution and solitary waves," Lax abstracted the GGKM discovery into the operator formalism $L_{t} = [M, L]$ that now bears his name — the Lax pair. This single algebraic reformulation revealed integrability as a structural property: any equation expressible as an isospectral deformation is solvable by inverse scattering. Lax received the Abel Prize (2005) in part for this work, which unified soliton theory, spectral theory, and Hamiltonian mechanics.

Building Intuition

Think of the inverse scattering transform as a nonlinear Fourier transform. The ordinary Fourier transform diagonalizes linear constant-coefficient PDEs: each mode $e^{ik x}$ evolves independently. The IST does the same for special nonlinear PDEs, but the "modes" are now of two species: a continuous spectrum (the reflection coefficient $R (k)$ , behaving like dispersive radiation) and a discrete spectrum (the bound states $κ_{n}$ , which are the solitons). In this nonlinear Fourier space, the dynamics trivializes — modes just rotate by phases. The deep magic is that the discrete eigenvalues are conserved: each soliton carries a fixed eigenvalue $- κ_{n}^{2}$ like a quantum number, so solitons survive collisions and merely exchange phase shifts. The nonlinear PDE is thus secretly a collection of non-interacting oscillators in the right coordinates.

The IST commuting diagram: nonlinear evolution is conjugated to linear flow in scattering space

Worked Examples

Single-Soliton from a Reflectionless Potential

Given: Initial data with one bound state,

R (k, 0) = 0

(reflectionless),

κ_{1} = κ

, normalization

c_{1}^{2} (0) = 2 κ e^{2 κ x_{0}}

. Step 1. Time-evolve:

c_{1}^{2} (t) = 2 κ e^{2 κ x_{0}} e^{8 κ^{3} t}

. Step 2. Kernel (no continuous part):

B (ξ; t) = c_{1}^{2} (t) e^{- κ ξ}

. Step 3. With a single exponential kernel, the GLM equation [8] has a degenerate (separable) solution

K (x, y) = - c_{1}^{2} e^{- κ y} / (1 + \frac{c _{1}^{2}}{2 κ} e^{- 2 κ x})

. Setting

y = x

K (x, x; t) = - \frac{c _{1}^{2} ( t ) e ^{- 2 κ x}}{1 + \frac{c _{1}^{2} ( t )}{2 κ} e ^{- 2 κ x}} .

Step 4. Apply equation [9],

u = - 2 \partial_{x} K (x, x; t)

. After simplification:

u (x, t) = - 2 κ^{2} sech^{2} [κ (x - 4 κ^{2} t - x_{0})] .

⟳Substitute

c_{1}^{2} (t)

, complete the algebra, and recognize

sech^{2}

via

\frac{d}{d x} tanh = sech^{2}

; the phase shift

x_{0}

absorbs

e^{2 κ x_{0}}

and the velocity

4 κ^{2}

emerges from the

e^{8 κ^{3} t}

factor.

u (x, t) = - 2 κ^{2} sech^{2} [κ (x - 4 κ^{2} t - x_{0})]

Physical reading: amplitude

= 2 κ^{2}

, speed

= 4 κ^{2}

— taller solitons travel faster. (Units: with

x

in length

L

and

t

in time,

u \sim L^{- 2}

κ \sim L^{- 1}

, speed

\sim L^{2} / t

after dimensional rescaling.)

Conserved Quantities of KdV

Given: KdV

u_{t} = 6 u u_{x} - u_{xxx}

(sign-flipped convention) with

u \to 0

\pm \infty

. Find the first three conserved densities. Step 1. Mass: integrate the PDE directly. Since

u_{t} = (3 u^{2} - u_{xx})_{x}

\frac{d}{d t} \int u d x = [3 u^{2} - u_{xx}]_{- \infty}^{\infty} = 0 \Rightarrow I_{1} = \int u d x .

Step 2. Momentum: multiply by

u

and integrate; the flux is again a perfect

x

-derivative, giving

I_{2} = \int u^{2} d x, \frac{d I _{2}}{d t} = 0.

Step 3. Energy (Hamiltonian): the third invariant is

I_{3} = \int (u^{3} + \frac{1}{2} u_{x}^{2}) d x .

⟳Construct the density via the Miura-linked Gardner expansion or by guessing

α u^{3} + β u_{x}^{2}

and fixing

α, β

\partial_{t} (density)

is a total

x

-derivative; this Hamiltonian generates KdV under the Gardner bracket.

I_{1} = \int u d x, I_{2} = \int u^{2} d x, I_{3} = \int (u^{3} + \frac{1}{2} u_{x}^{2}) d x

These are the first three of an infinite hierarchy $I_{1}, I_{2}, I_{3}, \dots$ , equivalently encoded as the coefficients in the large- $k$ asymptotic expansion of $ln a (k)$ (the trace formula).

Common Misconceptions

Integrability is rare and non-generic. Adding an arbitrary nonlinear term to KdV almost always destroys the Lax structure and the infinite conservation laws. Integrable PDEs are measure-zero islands in the space of nonlinear equations — do not expect a "nearby" perturbed equation to be solvable by IST. Likewise, "integrable" in the soliton sense (Liouville/Lax integrability, exact IST solvability) is distinct from numerical solvability or from "integrable" meaning merely "has a solution." The IST is not a Fourier transform of $u$ . Step 1 (direct scattering) is genuinely nonlinear in

u

— the map

u \mapsto S

depends on

u

in a complicated way. Only the evolution and the inverse steps are linear. People often mistakenly think solitons can be superposed; they cannot be added linearly, though their scattering data combine additively in the discrete spectrum. Burgers' equation is NOT integrable in the IST sense — but is still exactly solvable. Consider

u_{t} + u u_{x} = ν u_{xx} .

The Cole–Hopf transformation

u = - 2 ν (ln ϕ)_{x}

linearizes it to the heat equation

ϕ_{t} = ν ϕ_{xx}

. This is exact solvability, yet Burgers has no solitons, no infinite conservation-law hierarchy, and no Lax pair with isospectral discrete spectrum — its dissipative term smears structure rather than balancing it. The lesson: "exactly linearizable"

\neq =

"integrable in the soliton/IST sense." KdV's conservative dispersion is what permits solitons; Burgers' viscosity forbids them.

Connections

Integrable systems sit at the crossroads of analysis, geometry, and algebra. The Lax equation $L_{t} = [M, L]$ is a special case of the Zero-Curvature Representation $\partial_{t} U - \partial_{x} V + [U, V] = 0$ , which generalizes IST to the AKNS scheme covering the Nonlinear Schrödinger Equation and Sine-Gordon Equation. The conserved quantities and Poisson structure connect to Liouville Integrability and bi-Hamiltonian geometry (Magri's theorem). The Gelfand–Levitan–Marchenko equation is a Riemann–Hilbert Problem in disguise, which is the modern analytic engine of the field. On the algebraic side, the hierarchy structure links to Kac–Moody Algebras and tau functions of the Sato Grassmannian, while finite-gap (algebro-geometric) solutions connect to Riemann Theta Functions on hyperelliptic curves.

Real-World Applications

Shallow-water and tsunami waves — KdV models long surface waves in canals (the original Korteweg–de Vries 1895 setting) and certain internal ocean waves; soliton trains are observed in straits like Gibraltar.
Fiber-optic communications — the Nonlinear Schrödinger equation governs pulse propagation; optical solitons (Hasegawa–Tappert 1973) maintain shape over thousands of km, enabling soliton-based and "eigenvalue communication" transmission schemes that use IST scattering data as the information carrier.
Bose–Einstein condensates — the Gross–Pitaevskii (NLS) equation produces bright and dark matter-wave solitons, experimentally created and controlled in ultracold atomic gases.
Plasma physics — ion-acoustic solitons in collisionless plasmas are described by KdV; Zabusky and Kruskal's 1965 numerical discovery of solitons arose from a plasma-lattice (FPU) model.
Josephson junctions and condensed matter — the sine-Gordon equation describes magnetic flux quanta (fluxons) propagating as solitons in long Josephson junctions, and dislocations in crystals.
Nonlinear lattices — the Toda lattice (discrete integrable system) models energy transport in anharmonic chains and underlies the resolution of the Fermi–Pasta–Ulam–Tsingou recurrence paradox.
Random matrix theory & statistics — Painlevé equations (integrable ODE reductions) govern eigenvalue gap distributions (Tracy–Widom law), connecting IST to growth processes and the longest-increasing-subsequence problem.

Historical Development

1834 — John Scott Russell observes the "Wave of Translation," a solitary hump persisting unchanged along the Union Canal, Scotland — the first soliton sighting.
1895 — Diederik Korteweg and Gustav de Vries derive the KdV equation and find its solitary-wave $sech^{2}$ solution, vindicating Russell.
1955 — Fermi, Pasta, Ulam & Tsingou numerically study an anharmonic chain expecting thermalization; instead observe recurrence — energy returns to initial modes (the FPUT paradox).
1965 — Norman Zabusky and Martin Kruskal simulate KdV as the continuum limit of FPUT, discover that solitary waves pass through each other intact, and coin the term "soliton."
1967 — Gardner, Greene, Kruskal, and Miura (GGKM) solve KdV exactly via the inverse scattering method, linking it to the Schrödinger operator's spectrum.
1968 — Peter Lax introduces the Lax pair formalism $L_{t} = [M, L]$ , abstracting integrability into operator language.
1971–72 — Vladimir Zakharov and Aleksei Shabat solve the Nonlinear Schrödinger equation by IST; Ablowitz, Kaup, Newell & Segur (AKNS, 1973–74) unify these via the $2 \times 2$ zero-curvature scheme.
1974 — Faddeev, Zakharov, and Gardner reveal KdV as a completely integrable Hamiltonian system with action-angle variables (the scattering data).
1970s–80s — Sato, Date, Jimbo, Kashiwara, Miwa develop the tau-function / Grassmannian theory; algebro-geometric (finite-gap) solutions via theta functions (Novikov, Dubrovin, Matveev, Krichever) emerge.

Subtleties & Limits

Where the theory strains and what remains open. (1) Beyond rapidly decaying data: IST for KdV requires

u \to 0

fast enough that the scattering problem is well-posed; for periodic or algebro-geometric initial data the appropriate objects are Bloch spectra and Riemann surfaces, and the analysis is far subtler. The Deift–Zhou nonlinear steepest descent method (1990s) rigorously extracts long-time asymptotics from the Riemann–Hilbert formulation, but uniform asymptotics in transition regions remain delicate. (2) Higher dimensions: genuine integrable PDEs in

2 + 1

dimensions exist (KP, Davey–Stewartson) requiring the $\overset{ˉ}{\partial}$ (d-bar) method, but a clean IST in

3 + 1

dimensions is essentially absent — integrability appears to be a low-dimensional miracle. (3) Perturbed and "near-integrable" systems: how solitons respond to small non-integrable perturbations (radiation, soliton decay) is handled by perturbation theory built on IST, but its rigorous justification and the fate of the FPUT recurrence under the thermodynamic limit connect to deep open questions in statistical mechanics and nonlinear wave turbulence. (4) Classification: there is still no complete, constructive classification of all integrable PDEs; the symmetry approach (Mikhailov, Sokolov) provides powerful necessary conditions but no exhaustive list.