Past exam of the mathematics course of the University of Cambridge / 2004 / iii / Paper 8 (Introduction to Integrable Systems)

A Poisson structure is a real bilinear operation on smooth functions on $M$,which is antisymmetric, obeys the Jacobi identity, and is a derivation in each argument. Explicitly,Antisymmetry supplies the corresponding Leibniz rule in the first argument. Equivalently it is a smooth Poisson bivector $\Pi$ with $\{f,g\}=\Pi (df,dg)$ and $[\Pi ,\Pi ]=0$ for the Schouten-Nijenhuis bracket. In local coordinates,The last equation is the coordinate Jacobi identity. Nondegeneracy is not part of the definition: a Poisson structure may have Casimir functions and symplectic leaves of different dimensions.

A Hamiltonian system consists of such a Poisson manifold, a Hamiltonian $H\in C^{\infty }(M)$, and the flow of its Hamiltonian vector field. Choose the conventionFor an observable with no explicit time dependence, $df/dt=\{f,H\}$. In particular $H$ is conserved because $\{H,H\}=0$. For the canonical Poisson bracket, this convention gives ${\dot{q}}_i=\partial H/\partial p_i$ and ${\dot{p}}_i=-\partial H/\partial q_i$.

For the Arnold-Liouville theorem, work on a symplectic manifold of dimension $2d$. Let $F_1=H,F_2,\dots ,F_d$ be smooth functions with pairwise vanishing Poisson brackets, and supposeon a connected component $N$ of a common level. If $N$ is compact, then $N$ is a $d$-dimensional torus. Its commuting Hamiltonian vector fields generate translations, and the Hamiltonian motion on it is linear in angular coordinates. Compactness ensures completeness of these restricted vector fields.

Moreover a neighbourhood of $N$ admits action-angle variables $(I_1,\dots ,I_d,{\theta }_1,\dots ,{\theta }_d)$, with each angle defined modulo $2\pi$, for whichOne compatible symplectic form convention is $\omega ={\sum }_{i}d{\theta }_i\wedge d{I}_i$. The equations becomeThe resulting motion is periodic when the frequencies are commensurable and otherwise quasiperiodic on a subtorus. This is Liouville integrability with $d$ independent first integrals in involution. Compactness and regularity cannot be discarded from the torus conclusion. On a general Poisson manifold, apply this theorem on a regular symplectic leaf of dimension $2d$; the ambient dimension alone does not specify the required number of integrals.

A field $u(x,t)$ has infinitely many degrees of freedom: its initial state is a function rather than a finite list of coordinates. For a Hamiltonian evolution, the analogue of the finite-dimensional construction seeks an infinite sequence of independent conserved functionals $H_j[u]$ whose Poisson brackets vanish. Their Hamiltonian vector fields then define commuting time evolutions. For example, a Poisson operator $J$ produces the field equation $u_t=J\delta H/\delta u$, where $\delta H/\delta u$ is the variational derivative.

A Lax equation packages many conservation laws into one operator identity. If $L_t=[A,L]$, cyclicity of an appropriate trace givesFor a formal pseudodifferential operator the relevant trace is the Adler trace; suitable fractional powers of a monic differential operator produce further conserved functionals. A Lenard-Magri recursion can prove that such functionals are Poisson-commuting functions, rather than merely conserved quantities. A Lax equation by itself does not establish all the independence and completeness properties needed for integrability.

The Korteweg-De Vries equation illustrates the mechanism. In one time normalization take $L={\partial }^2+u$ and $A={\partial }^3+\frac{3}{2}u\partial +\frac{3}{4}{u}_x$. Direct composition givesIts spectral problem, together with the inverse scattering transform for decaying data or suitable spectral coordinates for periodic data, turns nonlinear evolution into simple evolution of spectral data. Solitary waves arise from discrete spectral data, while continuous data describe dispersive radiation.

Thus infinite-dimensional Hamiltonian integrability is more than an infinite list of formulas: one wants sufficiently complete commuting invariants and a reconstruction mechanism that solves the evolution. Boundary conditions and the function space are essential. A formal hierarchy need not converge, and an infinite family of conserved quantities does not automatically give a compact infinite-dimensional torus or an unrestricted analogue of the Arnold-Liouville theorem.

First let $n\ge 0$. The ordinary Leibniz rule, applied to a test function $h$, givesEquivalently, the composition of differential operators isFor completeness, this formula follows by induction: applying $\partial$ differentiates either the coefficient or the remaining derivative of $h$, and the two contributions combine by $\left(\genfrac{}{}{0ex}{}{n}{r}\right)+\left(\genfrac{}{}{0ex}{}{n}{r-1}\right)=\left(\genfrac{}{}{0ex}{}{n+1}{r}\right)$. The initial case is multiplication by $f$.

For all integers $n$, define $\left(\genfrac{}{}{0ex}{}{n}{0}\right)=1$ and $\left(\genfrac{}{}{0ex}{}{n}{r}\right)=n(n-1)\cdots (n-r+1)/r!$. The formal pseudodifferential composition rule isFor $n=-s<0$, $\left(\genfrac{}{}{0ex}{}{-s}{r}\right)=(-1{)}^r(\genfrac{}{}{0ex}{}{s+r-1}{r})$, so this is usually an infinite series. In particularApplying $\partial$ on the left gives $m_f$: all later coefficient derivatives cancel in pairs. This uniquely determines the normal-ordered inverse expansion. Repeating the inverse construction, or using the same binomial recurrence downward in $n$, gives every negative power. More explicitly, applying $\partial$ to the proposed $n$-series yields coefficient $\left(\genfrac{}{}{0ex}{}{n}{r}\right)+\left(\genfrac{}{}{0ex}{}{n}{r-1}\right)=\left(\genfrac{}{}{0ex}{}{n+1}{r}\right)$ at $f^{(r)}{\partial }^{n+1-r}$, so it is consistent with the already determined $(n+1)$-series. Negative powers here are formal; differentiation on all of $C^{\infty }(\mathbb{R})$ has a nontrivial kernel and has no genuine two-sided inverse there.

Use the formal pseudodifferential composition rule on each coefficient $w_i(x)$ and then the exponential eigenfunction convention. For every integer $n$,ConsequentlyCollecting all terms of a fixed power of the spectral parameter gives the useful answerThe sum defining each $v_s$ is finite. For example,andThis is the action on a formal pseudodifferential wave function. It is a formal expansion near $k=\infty$, with $k^{\beta }$ carried as a formal prefactor. It does not assert convergence for finite $k$, and negative powers are not evaluated at $k=0$.

Write the two formal pseudodifferential operators in normal order aswhere $p={\alpha }_1$ and $q={\alpha }_2$. The formal pseudodifferential composition rule givesThus $\boxed{\mathrm{ord}(P_1{P}_2)=p+q}$ whenever $ab$ is not identically zero, in particular when both leading coefficients are nowhere zero. In general the rigorous statement over the specified coefficient algebra is $\mathrm{ord}(P_1{P}_2)\le p+q$.

The leading scalar coefficients commute. Subtracting the reversed composition also cancels the terms involving $a_1,b_1$, leavingTherefore the order filtration of formal pseudodifferential operators giveswith equality precisely when the displayed coefficient is not identically zero. The bound can be strict: constant-coefficient powers of $\partial$ commute, giving the zero operator, whose order can be assigned $-\infty$.

There is also a genuine qualification to product-order addition in $C^{\infty }(\mathbb{R})$. Choose nonzero smooth functions $a,b$ with disjoint compact supports and take $P_1=m_a,P_2=m_b$. Each has order zero, but their product is zero. Nonzero coefficients as elements of this algebra need not have a nonzero product. Thus an unqualified assertion of exact product order requires an additional leading-symbol hypothesis.

Put every formal pseudodifferential operator in normal order, with all coefficient functions on the left. Its formal pseudodifferential residue is the coefficient of ${\partial }^{-1}$. The Adler trace isHere and below, use periodic coefficients and integrate over a period, or impose convergence and vanishing boundary terms on the real line. Equivalently, the algebraic integration functional must annihilate total derivatives. This convention is needed for the requested trace property; arbitrary elements of $C^{\infty }(\mathbb{R})$ do not supply it automatically.

It suffices to compute with monomials $A=a{\partial }^p$, $B=b{\partial }^q$. Write $r=p+q+1$. If $r<0$, no term in either product has power ${\partial }^{-1}$, so both residues are zero. If $r\ge 0$, the formal pseudodifferential composition rule givesSince $q=r-1-p$, reversal of the factors in the numerator provesAfter $r$ integrations by parts, $\int a{b}^{(r)}dx=(-1{)}^r\int b{a}^{(r)}dx$. Hence the two product traces agree. More explicitly, for $r\ge 1$ their residue difference is the total derivativeFor $r=0$ the residue difference is $ab-ba=0$. In products of operators bounded above in order, only finitely many monomial pairs can contribute to a residue: the required $r$ is nonnegative, so the two downward coefficient indices have a bounded sum. Summing the monomial result therefore proves

The boundary convention is substantive. Without it, take $A={\partial }^2$, $B=(\tanh x){\partial }^{-2}$. Both individual residues vanish, but $\mathrm{res}[A,B]=2(\tanh x{)}^{\prime }=2{\mathrm{sech}}^{2}x$ and its integral over $\mathbb{R}$ is $4$, not zero. The primitive has unequal limits at the two ends. Also a generic smooth residue need not have a convergent integral at all. Thus the proof supplies the intended cyclic trace under its necessary analytic or formal integration convention.

Write $P={\sum }_{j\le \alpha }{p}_j(x){\partial }^j$, setting unused coefficients to zero. Split the Adler trace into its fixed leading part and variable parts:Cyclicity from part (i) givesThus this Hamiltonian trace functional on monic differential operators has the required formTo make each coefficient explicit, expand ${\partial }^l{p}_j$ by the Leibniz rule. A term with $k$ derivatives of $p_j$ contributes to the residue exactly when $l-k+j=-1$. Consequently,These are smooth functions depending on $P$, not on the variable coefficients of $L$. In particular $a_0=p_{-1}$ and $a_1=p_{-2}+p_{-1}^{\prime }$. The coefficient $a_0$ must be included: it appears in the displayed sum even though the printed coefficient list starts at $a_1$.

The monic operators form an affine space, not a vector space. Accordingly $l_P$ is affine in the coordinates $u_l$, with constant $c$, and its differential is linear in variations. Its variational derivatives are $\delta l_P/\delta u_l=a_l$. It is the restriction of a linear trace pairing on the full operator space, which explains the terminology without discarding the constant term.

Use the canonical representatives of functionals of $u$: $P_f=f{\partial }^{-1}$, $P_g=g{\partial }^{-1}$. The formal pseudodifferential composition rule givesand the reversed product has $f,g$ exchanged. ThereforeMultiplying by $L={\partial }^2+u$ on the left, only the leading displayed commutator term can produce a residue. Differentiation of its coefficient produces powers at most ${\partial }^{-2}$, and multiplication by $u$ produces powers at most ${\partial }^{-3}$. Thusand, with the trace convention of part (i),There is no $u$ dependence in this bracket.

Moreover $\mathrm{res}(P_{f}L)=fu$, so $l_{P_f}=\int fudx$ and $\delta l_{P_f}/\delta u=f$. If $K(x,y)=\{u(x),u(y)\}$ is the distributional field bracket, it is characterized bySince $\int {\partial }_x\delta (x-y)g(y)dy=g^{\prime }(x)$, the answer isThe Dirac delta is periodic in the periodic setting, or the ordinary Dirac delta on the real line with admissible test functions. This is the first Hamiltonian structure of the KdV equation, with Poisson operator $-2{\partial }_x$. Its skewness follows by integration by parts; its coefficients are constant, so the functional Jacobi identity has no coefficient-variation terms.

A qualification is needed if the displayed trace formula is interpreted as a bracket on the reduced space for arbitrary operator representatives. Multiplication by a function $h$ gives $l_{m_h}({\partial }^2+u)=0$, yet direct expansion givesWith $h=g=\cos x$ on a period $[0,2\pi ]$, the latter is $\pi$. Hence the formula does not descend through every possible representative of a functional on this reduced affine space. The requested $f{\partial }^{-1},g{\partial }^{-1}$ representatives, used consistently for variational gradients, give the well-defined constant field bracket just computed; a general restriction from the ambient operator algebra needs an appropriate reduction.

The coefficients of the formal series must be smooth functions $H_n\in C^{\infty }(M)$, rather than points of $M$. Interpret complex parameter values in the complexified function algebra. The Casimir function condition for the Poisson pencil is a coefficientwise formal identity:Equality of the formal coefficients provesThese are the Lenard-Magri recursion relations, valid for every smooth function $f$.

For $i\ge 1$ and $j\ge 0$, use the recursion in the first argument, then antisymmetry and the recursion in the second:Each step lowers the first index while raising the second. Iterating $i$ times givesbecause $H_0$ is a Casimir function of the first Poisson structure. This includes $i=0$ directly. The recursion one step higher then givesThus the commuting coefficients of a Poisson pencil Casimir satisfyThey are therefore Poisson-commuting functions for every member of the Poisson pencil. The argument proves involution, not functional independence; an infinite sequence can contain repetitions or zero coefficients.

The intended construction is to solve the Lenard-Magri recursion one step at a time. With $H_0$ a Casimir function of the first Poisson structure, seek $H_n$ satisfyingHere $X_H^{(a)}f=\{f,H{\}}_a$, consistently with part 1(i). If every step can be solved by a globally defined smooth function, the formal series ${\sum }_{n\ge 0}{H}_n{\lambda }^n$ is a Casimir function for the Poisson pencil, and part (i) proves that all the $H_n$ are Poisson-commuting functions for every parameter value after complexification.

The recursive step is a concrete potential problem. Put $Y=-X_{H_{n-1}}^{(2)}$. It must first be tangent to the symplectic leaves of ${\Pi }_1$. On a regular leaf use canonical coordinates $(q_i,p_i)$. The required equations areThus the leaf one-formmust be exact. Where it is closed on a contractible coordinate domain, define $H_n$ by integrating ${\alpha }_Y$ from a base point; path independence follows from closedness. A function of the leaf labels can be added, corresponding to a Casimir function of the first Poisson structure. Global solvability also requires vanishing periods and smooth matching between leaves. This describes how to perform each solvable step without pretending that a degenerate Poisson structure can simply be inverted.

Compatibility alone does not ensure these solvability conditions. There is an immediate obstruction to a Lenard-Magri recursion even in two dimensions. On $M={\mathbb{R}}^2$ takeBoth are Poisson structures, and their linear combination is a scalar multiple of the canonical Poisson bracket, so they are compatible. Every function, including $H_0$, is a Casimir function for the first bracket. But the first recursion, evaluated at $f=y$, would requirea contradiction. An arbitrary initial Casimir need not extend even to $H_1$. Consequently the unconditional existence assertion requires an additional hypothesis: the indicated vector field must be Hamiltonian for ${\Pi }_1$ at every step. Under that hypothesis the recursive construction and involution proof above complete the intended result; without it the counterexample rules out the requested infinite extension.