4 · algebraic, analytic, and geometric techniques

matrix algebra, jordan forms and the fredholm alternative

Any linear operator on a finite-dimensional complex vector space admits a Jordan canonical form: a block-diagonal matrix whose diagonal blocks are elementary Jordan blocks associated with the eigenvalues. The sizes of the blocks are uniquely determined by the dimensions of the kernels of the successive powers $(A-\lambda I)^k$. Over the reals the corresponding real canonical form replaces complex conjugate pairs by $2\times 2$ rotation-scaling blocks.

When the linear equation $Ax=b$ is considered, the Fredholm alternative asserts that a solution exists if and only if $b$ is orthogonal to the kernel of the adjoint operator $A^*$. In the language of differential equations this criterion decides the solvability of the variational equation along a periodic orbit or of the linearized bifurcation equations at a multiple eigenvalue. The same alternative guarantees that the bordered matrices arising in numerical continuation remain invertible when phase or unfolding conditions are correctly imposed.

definition (jordan block). for $\lambda\in\mathbb{C}$ and $m\ge 1$, the Jordan block $J_m(\lambda)$ is the $m\times m$ matrix with $\lambda$ on the diagonal, $1$ on the superdiagonal, and $0$ elsewhere.

definition (generalized eigenspace). for $A\in\mathrm{M}_n(\mathbb{C})$ and eigenvalue $\lambda$, $$ E_\lambda =\bigcup_{k\ge 1}\ker(A-\lambda I)^k =\ker(A-\lambda I)^{m_\lambda} $$ for $m_\lambda$ large enough (in fact $m_\lambda\le n$). write $d_k=\dim\ker(A-\lambda I)^k$.

theorem (jordan form over $\mathbb{C}$). there exists an invertible $P$ such that $P^{-1}AP$ is block diagonal, the diagonal blocks being Jordan blocks $J_{m_j}(\lambda_j)$ for the eigenvalues of $A$. for each eigenvalue $\lambda$, the number of blocks of size at least $k$ equals $$ d_k-d_{k-1}, $$ and the number of blocks of exact size $k$ equals $$ (d_k-d_{k-1})-(d_{k+1}-d_k), $$ with $d_0=0$. in particular the block partition for each $\lambda$ is uniquely determined by the sequence $(d_k)$.

proof (uniqueness of block sizes; existence sketched). on the generalized eigenspace $E_\lambda$ the operator $N=A-\lambda I$ is nilpotent. for a nilpotent operator the number of Jordan blocks of size $\ge k$ for eigenvalue $0$ equals $\dim\ker N^k-\dim\ker N^{k-1}$. indeed, each block of size $m$ contributes $1$ to $d_k-d_{k-1}$ precisely when $m\ge k$ (one new linearly independent vector appears in $\ker N^k$ beyond $\ker N^{k-1}$ per such block). therefore $d_k-d_{k-1}$ counts blocks of size at least $k$, and telescoping gives the number of exact size $k$. applying this to each $N=A-\lambda I|_{E_\lambda}$ yields uniqueness of the partition for every eigenvalue.

existence follows from the primary decomposition theorem: $V=\bigoplus_\lambda E_\lambda$, on which $A-\lambda I$ is nilpotent, together with the classical existence of a Jordan basis for a nilpotent operator (choose a basis of $\ker N$ compatible with a flag of images, lift chains of length equal to the Weyr or Segre characteristics; standard linear algebra). conjugating by the change-of-basis matrix $P$ produces $P^{-1}AP$ in Jordan form. $\square$

remark (real canonical form). if $A$ is real, complex conjugate eigenvalues appear in pairs. replacing each pair of conjugate Jordan chains by real and imaginary parts yields real blocks that are $2\times 2$ rotation-scaling matrices inflated to larger real Jordan-type blocks. the same counts $d_k$ still determine the structure after pairing.

definition (adjoint). with respect to the standard Hermitian product $\langle u,v\rangle=v^*u$ on $\mathbb{C}^n$, the adjoint of $A$ is $A^*=\overline{A}^T$, characterized by $\langle Au,v\rangle=\langle u,A^*v\rangle$.

theorem (fredholm alternative, finite dimensions). let $A\in\mathrm{M}_n(\mathbb{C})$ and $b\in\mathbb{C}^n$. the equation $Ax=b$ admits a solution if and only if $$ \langle b,w\rangle=0 \quad\text{for every }w\in\ker A^*. $$ when a solution exists, the solution set is an affine space modeled on $\ker A$.

proof. if $Ax=b$ and $A^*w=0$, then $$ \langle b,w\rangle=\langle Ax,w\rangle=\langle x,A^*w\rangle=0. $$ conversely, assume $\langle b,w\rangle=0$ for all $w\in\ker A^*$. the fundamental theorem of linear algebra gives $$ \mathrm{im}\,A=(\ker A^*)^\perp. $$ thus $b\in(\ker A^*)^\perp$ if and only if $b\in\mathrm{im}\,A$, i.e. $Ax=b$ is solvable. the solution set is then $x_0+\ker A$ for any particular solution $x_0$. $\square$

remark (dynamical use). along a $T$-periodic orbit of a flow, the variational equation always admits the nontrivial free solution $\dot x_*$. solvability of inhomogeneous variational equations (Melnikov, Floquet corrections, bordering for continuation) is decided exactly by orthogonality of the forcing term to the adjoint kernel (Fredholm on the circle).

manifold geometry as computational foundation

A smooth manifold $M$ is locally Euclidean, yet its global topology and the presence of singularities force all computational procedures to be formulated in a coordinate-free manner. Charts, transition maps, and the tangent bundle supply the data structures in which Newton-type correctors and predictor-corrector continuation can be implemented without leaving the manifold. The Riemannian metric (or any convenient fiber metric on the tangent bundle) converts the abstract Newton equation into a practical linear solve, while parallel transport along the continuation path yields consistent initial guesses for the iterative corrector. Thus the differential geometry of $M$ is not merely an existence framework; it is the natural computational domain in which numerical algorithms operate.

definition (chart and transition). a chart is a diffeomorphism $\phi:U\subset M\to\phi(U)\subset\mathbb{R}^n$. if $(U,\phi)$ and $(V,\psi)$ overlap, the transition map $\psi\circ\phi^{-1}$ is a $C^\infty$ diffeomorphism of open sets in $\mathbb{R}^n$. all computations (Jacobians, Newton steps) are performed in chart coordinates and pushed forward by the transitions.

definition (newton map in a chart). for a smooth equation $F(x)=0$ with $F:M\to E$ taking values in a vector bundle (or, after trivialization, in $\mathbb{R}^m$), the Newton correction $\delta$ in a chart solves the linear equation $$ DF(x)\cdot\delta=-F(x) $$ in the corresponding coordinates, after which $x$ is updated by the exponential of a Riemannian metric or by ordinary addition in the chart.

remark. the abstract Newton equation is chart-independent when $F$ is a section of a vector bundle: $DF(x)$ is a linear map between fibers. choosing a metric identifies duals so that least-squares and Gauss-Newton variants become available on under- or over-determined systems.

numerical location of fixed points and periodic orbits

Isolated zeros of a smooth map $f:X\to X$ (or $F:X\to\mathbb{R}^n$) are located by Newton's method $$ x_{k+1}=x_k-Df(x_k)^{-1}f(x_k) $$ or by its quasi-Newton variants (Broyden's update) that avoid repeated factorizations of the Jacobian. Quadratic convergence is assured once the initial guess lies inside a ball of uniform invertibility of $Df$.

Periodic orbits of a flow $\varphi_t$ are zeros of the map $$ x\mapsto\varphi_T(x)-x $$ on the extended space $X\times\mathbb{R}_+$. Because every point on the orbit is a zero, the Jacobian is singular. An integral phase condition $$ \int_0^T\langle\dot x_*(t),x(t)-x_*(t)\rangle\,dt=0 $$ (or its discrete analog) restores uniqueness and makes the bordered Jacobian invertible. Continuation of the resulting system with respect to a parameter traces branches of periodic orbits, detects period-doubling points, and follows the boundaries of Arnold tongues.

theorem (local quadratic convergence of newton). let $F:\mathbb{R}^n\to\mathbb{R}^n$ be $C^2$ near a root $x_*$ with $DF(x_*)$ invertible. then there exist a neighborhood $U$ of $x_*$ and a constant $C$ such that the Newton iteration $$ x_{k+1}=x_k-DF(x_k)^{-1}F(x_k) $$ is well-defined on $U$ and satisfies $$ \|x_{k+1}-x_*\|\le C\|x_k-x_*\|^2 $$ for all $x_0\in U$.

proof. by continuity of $DF$ and invertibility of $DF(x_*)$, there is a ball about $x_*$ on which $\|DF(x)^{-1}\|$ is bounded by some $M$ (Neumann series about $DF(x_*)$). Taylor's theorem with integral remainder gives $$ F(x_k) =F(x_*)+DF(x_*)(x_k-x_*)+R_k =DF(x_*)(x_k-x_*)+R_k, $$ where $\|R_k\|\le K\|x_k-x_*\|^2$ for a constant $K$ depending on $\|D^2F\|$ near $x_*$. the Newton step rearranges as $$ x_{k+1}-x_* =x_k-x_*-DF(x_k)^{-1}F(x_k) =DF(x_k)^{-1}\Bigl( DF(x_k)(x_k-x_*)-F(x_k) \Bigr). $$ writing $DF(x_k)=DF(x_*)+E_k$ with $\|E_k\|\le L\|x_k-x_*\|$ and substituting the Taylor expansion of $F$ produces a residual of order $\|x_k-x_*\|^2$ only. absorbing the uniform bound on $\|DF(x_k)^{-1}\|$ yields $\|x_{k+1}-x_*\|\le C\|x_k-x_*\|^2$. shrinking $U$ so that $C\|x_0-x_*\|<1$ keeps the sequence inside $U$ by induction. $\square$

definition (periodic orbit as zero of the shooting map). a $T$-periodic orbit through $x$ satisfies $\Phi_T(x)=x$ where $\Phi_T=\varphi_T$ is the time-$T$ map of the flow. the map $$ G(x,T)=\varphi_T(x)-x $$ has zeros precisely at periodic points with period $T$.

theorem (kernel of the monodromy). if $x(t)$ is a nonconstant $T$-periodic solution of an autonomous ODE $\dot x=v(x)$, then $D_x\varphi_T\bigl(x(0)\bigr)\cdot\dot x(0)=\dot x(0)$. hence $DG$ is singular at every periodic orbit of an autonomous vector field.

proof. the curve $t\mapsto x(t+\varepsilon)$ is again a $T$-periodic solution with initial condition $x(\varepsilon)$. differentiating the identity $$ \varphi_T\bigl(x(\varepsilon)\bigr)=x(\varepsilon) $$ at $\varepsilon=0$ gives $D\varphi_T(x(0))\dot x(0)=\dot x(0)$. $\square$

remark (phase condition). an integral phase constraint of the form $\langle\dot x_*,x-x_*\rangle_{L^2([0,T])}=0$ cuts transversely through the orbit circle and makes the bordered linearization invertible under generic nondegeneracy (simple Floquet multiplier $1$). this is the finite-dimensional Fredholm alternative applied to the bordered Jacobian of $(G,\mathrm{phase})$.

algebraic decomposition of linear time-invariant systems

A linear differential-algebraic equation $$ E\dot x=Ax+Bu $$ is transformed by the Weierstrass-Kronecker algorithm into a direct sum of a purely differential subsystem (ordinary differential equation of finite order) and a purely algebraic nilpotent subsystem. The algorithm produces invertible matrices $P$ and $Q$ such that $$ PEQ=\operatorname{diag}(I,N),\qquad PAQ=\operatorname{diag}(J,I), $$ where $N$ is nilpotent. Each cyclic block of the Jordan or Kronecker form corresponds to an independent chain of integrators or differentiators. This decomposition isolates the dynamic modes from the constraints, supplies the correct regularity index, and yields a minimal realization of the input-output behavior. In network theory the same block structure organizes the port variables of interconnected mechanical or electrical systems into independent cyclic subsystems.

definition (matrix pencil). a pencil is a pair $(E,A)$ of $n\times n$ matrices, written $\lambda E-A$. the pencil is regular if $\det(\lambda E-A)$ is not identically zero as a polynomial in $\lambda$.

definition (weierstrass form). a regular pencil is in Weierstrass form if $$ E=\operatorname{diag}(I_r,N),\qquad A=\operatorname{diag}(J,I_{n-r}), $$ with $N$ nilpotent and $J$ in Jordan form (or any convenient matrix of size $r$).

theorem (weierstrass decomposition of a regular pencil). if $(E,A)$ is regular, there exist invertible $P,Q$ such that $(PEQ,PAQ)$ is in Weierstrass form. the size of the nilpotent block and the Segre characteristics of $N$ are uniquely determined; the finite eigenvalues are the roots of $\det(\lambda E-A)=0$, and the nilpotency index of $N$ is the Kronecker index of the DAE $E\dot x=Ax$.

proof (structure theory; construction in outline). regularity means that the rational matrix $(\lambda E-A)^{-1}$ exists as a meromorphic function of $\lambda$. the finite poles correspond to ordinary eigenvalues of the slow subsystem; the polar part at infinity is encoded by a nilpotent. more elementary: over an algebraically closed field one can find bases that split the generalized null chains of the pencil. work with matrices over the principal ideal domain $\mathbb{C}[\lambda]$ and apply the Smith normal form algorithm to $\lambda E-A$; the invariant factors determine elementary divisors for the finite spectrum, while the infinite elementary divisors determine the sizes of the nilpotent blocks of $N$. transforming by the unimodular factors yields invertible constant (in $\lambda$) matrices $P,Q$ realizing the Weierstrass normal form when the pencil is square and regular. uniqueness of the invariant factors of a matrix over a PID gives uniqueness of the block data. $\square$

remark (systems with input). for $E\dot x=Ax+Bu$ the same $P,Q$ act on the state; the input matrix becomes $PB$, partitioned conformally into slow and fast channels. algebraic constraints appear exactly in the nilpotent block, where formal differentiations of the input of order less than the index may be required for classical solutions.

exterior differential systems and interaction geometry

An exterior differential system (EDS) on a manifold $M$ is a differentially closed ideal $\mathcal{I}\subset\Omega^*(M)$. An integral manifold is a submanifold on which every form of $\mathcal{I}$ pulls back to zero. Cartan's existence theorem (in the real-analytic category) guarantees local integral manifolds once the polar spaces satisfy a dimensional regularity condition; the Cartan-Kahler theory supplies the precise count of free data.

In constrained mechanical networks the EDS is generated by the kinematic constraints and the balance laws. The maximal integral manifolds are precisely the admissible configurations and velocities of the network. Kron-Hoffman interaction geometry realizes the interconnection of subsystems as a fiber product of the individual EDS; the resulting ideal encodes both the topological incidence relations and the constitutive laws. The same formalism recovers the Dirac structure of a non-holonomic system as the annihilator of a suitably chosen Pfaffian system, thereby linking the algebraic decomposition of the linear theory to the nonlinear geometry of constraints.

definition (exterior differential system). a graded ideal $\mathcal{I}\subset\Omega^*(M)$ is an EDS if it is closed under exterior differentiation: $d\mathcal{I}\subset\mathcal{I}$. an immersed submanifold $i:N\hookrightarrow M$ is an integral manifold if $i^*\theta=0$ for every $\theta\in\mathcal{I}$.

definition (algebraic ideal vs differential ideal). the algebraic ideal generated by a set of forms need not be differentially closed; one closes it by adjoining exteriors and algebraic consequences. Pfaffian systems of chapter 3 are the rank-one generator case of this construction.

theorem (integral elements for a pfaffian system). if $\mathcal{I}$ is generated as a differential ideal by a regular Pfaffian system $\mathcal{I}^1$ of constant rank, then $p$-dimensional integral elements at $x$ are precisely the $p$-planes contained in $\mathcal{D}_x=(\mathcal{I}^1_x)^\perp$ that are annihilated by the exterior derivatives of the generators restricted to that plane. complete integrability ($\mathcal{I}'=\mathcal{I}$ in the sense of chapter 3) means that every plane in $\mathcal{D}_x$ is an integral element, and Frobenius supplies integral manifolds of dimension $\mathrm{rank}\,\mathcal{D}$.

proof. for $\theta\in\mathcal{I}^1$ and a $p$-plane $E\subset T_x M$ spanned by $X_1,\dots,X_p$, one has $i^*\theta=0$ for immersions with $T_x N=E$ only if $\theta|_E=0$, i.e. $E\subset\mathcal{D}_x$. the condition $i^*d\theta=0$ becomes $d\theta(X_j,X_\ell)=0$ for all $j,\ell$, which is the restriction of $d\theta$ to $E$. if $d\theta\equiv 0\pmod{\mathcal{I}^1}$ then this holds automatically for every $E\subset\mathcal{D}_x$. freeness of the annihilator and the Frobenius theorem of chapter 3 then produce integral manifolds of full dimension $\mathrm{rank}\,\mathcal{D}$. $\square$

remark (cartan-kahler). for general real-analytic EDS that are not completely integrable Pfaffian systems, free data for integral manifolds are counted by Cartan characters; the theorem guarantees existence of integral manifolds through admissible integral elements of ordinary status. the full proof is analytic and uses the Cauchy-Kowalevski theorem in adapted flags of polar spaces.

remark (networks). a fiber product of EDS for subsystems realizes Kirchhoff-type incidence by pullback constraints on shared port variables; the combined ideal is again differentially closed. this is the geometric content of interconnection before one chooses coordinates or a state-space realization.

summary

Taken together, these methods provide the concrete computational and algebraic infrastructure on which the geometric theories of dynamical systems, symplectic reduction, and control are built: Jordan structure and the Fredholm alternative govern linearized analysis, manifold-adapted Newton methods enable numerical continuation, Kronecker-Weierstrass decomposition isolates dynamic modes, and exterior differential systems encode nonlinear constraints in a form ready for both theoretical reduction and practical simulation.

exercises

exercise 1 (jordan data). for $$ A=\begin{pmatrix}2&1&0\\0&2&0\\0&0&2\end{pmatrix} $$ compute $d_k=\dim\ker(A-2I)^k$ for $k=1,2,3$ and read off the Jordan block sizes for $\lambda=2$.

exercise 2 (fredholm solvability). let $$ A=\begin{pmatrix}1&0\\0&0\end{pmatrix},\qquad b=\begin{pmatrix}b_1\\b_2\end{pmatrix}. $$ determine $\ker A^*$ and state precisely when $Ax=b$ is solvable. describe the solution set when it is.

exercise 3 (newton on a quadratic). apply Newton's method to $F(x)=x^2-2$ starting from $x_0=2$. compute $x_1,x_2$ exactly (as rationals) and verify that the errors $|x_k-\sqrt{2}|$ decrease roughly quadratically.

exercise 4 (singular monodromy). for the planar harmonic oscillator $\dot q=p$, $\dot p=-q$, verify explicitly that the time-$2\pi$ map is the identity, so every point is periodic, and that $D\varphi_{2\pi}=I$. explain why a phase condition is still needed if one searches for a single orbit representative numerically.

exercise 5 (rank-nullity and cokernel dimension). for a linear map $A:\mathbb{R}^n\to\mathbb{R}^m$, prove $\dim\ker A-\dim\operatorname{coker} A=n-m$ (equivalently $\operatorname{index}A=n-m$). check the formula on the Jordan matrix of exercise 1 restricted to its generalized eigenspace for $\lambda=2$.