Michal–Bastiani calculus on locally convex spaces

16 Counterexamples

Many theorems of the Banach theory fail on Fréchet spaces, and most of the hypotheses in this document are there because of an example. This chapter collects the examples that are formalized. Each one sits next to the positive result it limits, which is also stated in Lean, so that the boundary between the two settings can be read off from the formal statements. Two further examples are stated where they belong: the counterexamples for series of homogeneous polynomials (Remark 9.27) and Bognár’s closed positive definite subspace of a Krein space (Example 15.5).

Throughout, \(s\) denotes the Fréchet space of rapidly decreasing real sequences, with the seminorms \(p_k(x) = \sup _n (n + 1)^k |x_n|\), and \(\mathbb {R}^{\mathbb {N}}\) the Fréchet space of all real sequences with the product topology.

16.1 Algebras

Example 16.1 Units need not form an open set [ GN26 , Example 5.3.9(b) ]
✓
#

In a complete normed ring the units form an open set. In the complete locally convex algebra \(C(\mathbb {R}, \mathbb {R})\), with the compact-open topology, they do not: the plateau functions, equal to \(1\) on \([-n, n]\) and to \(0\) outside \((-n-1, n+1)\), are not units, and they converge to \(1\).

16.2 Inverse functions

Example 16.2 The inverse function theorem fails on \(\mathbb {R}^{\mathbb {N}}\), after [ Ham82 , Counterexample 5.5.1 ]
✓
#

On a Banach space, a map strictly differentiable at \(a\) with derivative the identity sends neighbourhoods of \(a\) onto neighbourhoods of \(f(a)\). On \(\mathbb {R}^{\mathbb {N}}\) the map \(P(a)_k = a_k - \sum _{j \le k} j\, a_{k-j} a_j\) is smooth, with \(P(0) = 0\) and \(\mathrm{d}P(0) = \mathrm{id}\), but its image is not a neighbourhood of \(0\): the sequence \(g_n = \frac{1}{n} e_0 + e_n\) tends to \(0\), and \(g_n \notin P(\mathbb {R}^{\mathbb {N}})\) for \(n \ge 2\).

Remark 16.3
#

Hamilton’s operator is \(Pf = f - x f f'\) on \(C^\infty [-1, 1]\). The map above is its action on the coefficients of formal power series at \(0\), which is where his argument takes place; we state it on \(\mathbb {R}^{\mathbb {N}}\) to avoid power series.

Example 16.4 Lower bounds are not locally uniform
✓
#

On normed spaces, if \((y, v) \mapsto L(y)v\) is \(C^{1}\) and \(L(y_0)\) is a topological embedding, then the \(L(y)\) are bounded below, with a common constant, for \(y\) near \(y_0\). On \(s\) there is a smooth family \(L(t)\) of linear automorphisms, \(L(t)x = (a_n(t) x_n)_n\), with \(L(0) = \mathrm{id}\), whose lower bound is not locally uniform near \(t = 0\): the inverses \(L(t)^{-1}\) are not equicontinuous on any neighbourhood of \(0\). This is why the hypothesis of Definition 8.8 is needed.

Example 16.5 Invertible derivatives do not reflect smoothness
✓
#

There are a smooth bijection \(j\) of \(\mathbb {R}\times s\), whose derivative at every point is a continuous linear automorphism, and a continuous curve \(\gamma \) such that \(j \circ \gamma \) is smooth but \(\gamma \) is not \(C^{1}\) on any neighbourhood of \(0\). Moreover \(j\) satisfies the hypothesis of Theorem 8.12 at each single point but not near every point, \((p, v) \mapsto \mathrm{d}j(p)^{-1} v\) is not continuous, and \(j^{-1}\) is not \(C^{1}\) near \((0, c)\). Compare Theorem 8.6 and Theorem 8.12.

16.3 Differential equations

Example 16.6 Uniqueness fails for linear equations [ GN26 , Example 2.4.3 ]
✓
#

On a normed space, the only solution of \(y' = Ay\), \(y(0) = 0\), for a continuous linear \(A\), is \(y = 0\) [ GN26 , Proposition 2.4.11 ] . On \(\mathbb {R}^{\mathbb {N}}\), with \(A\) the left shift, the curve \(t \mapsto (h(t), h'(t), h''(t), \dots )\) is a nonzero solution with \(y(0) = 0\), where \(h\) is the smooth function \(h(t) = e^{-1/t}\) for \(t {\gt} 0\), \(h(t) = 0\) for \(t \le 0\), all of whose derivatives vanish at \(0\).

Example 16.7 Existence fails for linear equations
✓
#

On a Banach space, \(t \mapsto e^{tA} y_0\) solves \(y' = Ay\), \(y(0) = y_0\), for all \(t\). On \(s\), with \(D x = (n^2 x_n)_n\), the equation \(y' = Dy\) has no solution on any interval \([0, \varepsilon )\) with \(y(0) = (e^{-(n+1)})_n\). This is the heat equation with reversed time on \(C^\infty (S^1)\), written on the Fourier side; the example goes back to Milnor [ Mil84 , § 3, exercise after Warning 3.6, p. 1025 ] , and is presented in [ Nee06 , Example II.3.12 ] (numbering of the arXiv version).

Example 16.8 The Frobenius theorem fails on \(s\) [ KM97 , 32.12, Example 1 ]
✓
#

There are a continuous linear operator \(D\) of \(s\) and a point \(a\) with \(Da \neq 0\) such that no curve through \(a\) is tangent to the rank one distribution \(x \mapsto \mathbb {R}\, Dx\); compare Theorem 11.2.

16.4 Analytic maps

Example 16.9 Complex analytic maps need not be locally bounded [ GN26 , Exercise 2.1.10 ]
✓
#

A complex analytic map into a normed space is bounded near each point. The map \(f(x) = \sum _k 2^{k+1} x_k^{2k+2}\) is complex analytic on \(\ell ^1(\mathbb {N}, \mathbb {C})\) but unbounded on the unit ball, and \(g(x) = (f((k + 1)x))_k\) is a complex analytic map from \(\ell ^1\) to the Fréchet space \(\mathbb {C}^{\mathbb {N}}\) which is bounded on no neighbourhood of \(0\).

16.5 Symplectic forms

Example 16.10 Darboux’s theorem fails for weak symplectic forms [ Mar72 ]
✓
#

By the Darboux–Weinstein theorem a strong symplectic form on a Banach space is constant in suitable charts [ Wei69 ] (formalized in lc-manifolds). On \(H = \ell ^2(\mathbb {N}, \mathbb {R})\), with \(S = \mathrm{diag}(1/(n+1))\), Marsden’s form \(\Omega = -\mathrm{d}\theta \) on \(H \times H\), where \(\theta _{(x, e)}(e_1, e_2) = g_x(e, e_1)\) and \(g_x(e, f) = \langle (S + \| x\| ^2 I) e, f\rangle \), is smooth, closed and weakly nondegenerate, but it admits no Darboux chart at the origin; it cannot even be made constant by a vector bundle chart of the fibres, as observed by the referee of [ Mar72 ] .

Proof ▶

\(\Omega \) is strongly nondegenerate at \((x, 0)\) for \(x \neq 0\), since \(S + \| x\| ^2 I\) is onto, and not at \((0, 0)\), since \(S\) is not onto. A chart carrying \(\Omega \) to a constant form at every point would make both nondegeneracy conditions equivalent.