Michal–Bastiani calculus on locally convex spaces

8 Comparison with Fréchet calculus

On normed spaces two calculi are available: the classical Fréchet calculus, in which the derivative is required to depend continuously on the point in the operator norm, and the calculus of Michal and Bastiani, which only asks for joint continuity of \((x, v) \mapsto \mathrm{d}f(x)(v)\). The second notion is weaker; Fréchet \(C^{k}\) maps are Michal–Bastiani \(C^{k}\) (Proposition 2.19). In this chapter we prove the converse, up to the loss of one derivative. As a consequence both notions of smoothness agree, so that for smooth maps between normed spaces one may use whichever calculus is more convenient.

In the second part of the chapter we apply this comparison to maps with bounded below derivative, and show that such maps reflect smoothness. Throughout, \(E\) and \(F\) are real normed spaces; completeness is assumed only where it is stated.

8.1 From Michal–Bastiani to Fréchet

Theorem 8.1 Michal–Bastiani \(C^{k+1}\) implies Fréchet \(C^{k}\)
✓

If \(f\) is \(C^{k+1}\) on \(U\) in the sense of Michal–Bastiani, then \(f\) is Fréchet \(C^{k}\) on \(U\).

The extra derivative is what converts joint continuity into continuity in the operator norm: in the proof below, joint continuity of the next derivative, together with the mean value inclusion, shows that the derivatives are locally Lipschitz in the operator norm.

Proof ▶

We prove, by induction on \(k\) and for all \(m\) simultaneously, the following statement: if \(g \colon U \to \mathcal{L}^m(E; F)\) is such that \(\hat g(x, v_1, \dots , v_m) = g(x)(v_1, \dots , v_m)\) is \(C^{k+1}\) in the sense of Michal–Bastiani, then \(g\) is Fréchet \(C^{k}\). Here \(\mathcal{L}^m(E; F)\) is the space of continuous \(m\)-linear maps with the operator norm. For \(m = 0\) this is the theorem.

The candidate for the derivative of \(g\) at \(y\) is the \((m+1)\)-linear map \(B_y(w, v) = \mathrm{d}\hat g\bigl((y, v), (w, 0)\bigr)\), the derivative of \(\hat g\) in a base direction; it is multilinear by Theorem 2.12. Since \(\mathrm{d}\hat g\) is jointly continuous and vanishes in the direction \(0\), the maps \(B_{y'}\) are uniformly bounded in operator norm for \(y'\) near \(y\). With the mean value inclusion (Lemma 2.8), this shows that \(g\) is locally Lipschitz into \(\mathcal{L}^m(E; F)\); in particular it is continuous, which is the case \(k = 0\).

For the inductive step, suppose that \(\hat g\) is \(C^{k+2}\). Then \((y, w, v) \mapsto B_y(w, v)\) is \(C^{k+1}\), so by the case \((k, m + 1)\) the map \(y \mapsto B_y\) is Fréchet \(C^{k}\), and in particular continuous in operator norm. The mean value inclusion then gives

\[ \| g(y + w) - g(y) - B_y(w, \cdot )\| = o(\| w\| ), \]

so \(g\) is Fréchet differentiable with derivative \(y \mapsto B_y\), and \(g\) is Fréchet \(C^{k+1}\).

Remark 8.2
#

This is [ GN26 , 2.3.17(b) ] . The proof in [ GN26 ] goes through the compact-open topology on \(\mathcal{L}(E, F)\) and the exponential law ( [ GN26 , 1.7.30 ] ). The proof above stays in normed spaces: the derivatives are tracked as continuous multilinear maps with the operator norm, which is closer to the existing Fréchet calculus of Mathlib. The price is the auxiliary statement for all \(m\), which has to be proved simultaneously.

Since the loss of one derivative disappears when \(k = \infty \), we obtain the following.

Corollary 8.3 Smooth maps agree
✓

A map between open subsets of normed spaces is smooth in the sense of Michal–Bastiani if and only if it is smooth in the sense of Fréchet.

Proof ▶

Theorem 8.1 and Proposition 2.19.

8.2 Operator-valued maps on locally convex spaces

Let \(E\) be a locally convex space, \(F_1\), \(F_2\) normed spaces, \(U \subseteq E\) open and \(g \colon U \to \mathcal{L}(F_1, F_2)\). If \(\hat g(y, \xi ) = g(y)\xi \) is \(C^{k+1}\) on \(U \times F_1\) in the sense of Michal–Bastiani, then \(g\) is \(C^{k}\) for the operator norm on \(\mathcal{L}(F_1, F_2)\).

Proof ▶

Let \(D(y, \xi , w) = \mathrm{d}\hat g((y, \xi ), (w, 0))\); it is linear in \(\xi \) and in \(w\), and continuous. Continuity at \((y_0, 0, 0)\) gives a convex neighborhood \(N\) of \(y_0\), a neighborhood \(W\) of \(0\) and \(C\) with \(\| D(y, \xi , w)\| \le C\| \xi \| \) for \(y \in N\), \(w \in W\). If \(y - y_0 \in rW\), the mean value inequality along the segment \([y_0, y] \subseteq N\) gives \(\| g(y) - g(y_0)\| \le rC\): \(g\) is continuous. The operator \(h(y, w) = D(y, \cdot , w)\) is the directional derivative of \(g\) for the operator norm, by the mean value inequality applied to \(s \mapsto g(y + sw)\xi - s\, h(y, w)\xi \) and the continuity of \(h\), which is the case \(k = 0\) applied to \(h\) on \(U \times E\). The induction runs on the domain \(E \times E\) of the derivative.

The statement is meaningless if \(F_1\) is not normable: no vector topology on \(\mathcal{L}(F_1, F_2)\) makes the evaluation continuous. Theorem 8.1 is the case where \(E\) is normed; there the multilinear version is needed for the induction.

8.3 Maps with bounded below derivative

In this section we consider a Banach space \(E\) (complete), a normed space \(F\) (complete or not), and a map \(j \colon U \to F\), smooth on an open set \(U \subseteq E\), with derivative bounded below on \(U\): there is \(m {\gt} 0\) with

\[ m \| v\| \le \| \mathrm{d}j(x)(v)\| \qquad (x \in U,\ v \in E). \]

By Corollary 8.3 it makes no difference in which sense \(j\) is smooth. No complement of the range of \(\mathrm{d}j(x)\) is assumed. The simplest example is a continuous linear map \(j \colon E \to F\) with \(m\| v\| \leq \| j(v)\| \), for instance an isometric embedding; in that case \(\mathrm{d}j(x) = j\) for every \(x\).

Our aim is to show that such a \(j\) reflects smoothness: a continuous map \(\gamma \) into \(U\) is of class \(C^{k}\) as soon as \(j \circ \gamma \) is. The test maps \(\gamma \) are defined on open subsets of arbitrary locally convex spaces. We first prove a linear version, in which the unknown appears linearly.

Lemma 8.5 Linear bootstrap
✓

Let \(W\) be an open subset of a locally convex space \(H\), \(\gamma \colon W \to U\) and \(w \colon W \to F\) of class \(C^{k}\), and \(z \colon W \to E\) continuous with \(\mathrm{d}j(\gamma q)(z q) = w(q)\) for \(q \in W\). Then \(z\) is of class \(C^{k}\).

Proof ▶

By induction on \(k\). For the inductive step, fix \(q \in W\) and a direction \(h\), and write \(A_t = \mathrm{d}j(\gamma (q + th))\). From the equation,

\[ A_0\Bigl(\frac{z(q + th) - z(q)}{t}\Bigr) = \frac{w(q + th) - w(q)}{t} - \frac{A_t - A_0}{t}\, z(q + th). \]

Since \(j\) is Fréchet \(C^2\) (Theorem 8.1), \(t \mapsto A_t\) is differentiable in the operator norm, so the right-hand side converges. As \(A_0\) is bounded below and \(E\) is complete, the difference quotients of \(z\) are Cauchy, hence converge to some \(D\) with \(A_0 D = \mathrm{d}w(q)(h) - \mathrm{d}^2 j(\gamma q)(\mathrm{d}\gamma (q)(h), z q)\).

The derivative \(\mathrm{d}z\) is continuous, because solutions of \(A(p) x(p) = r(p)\) are continuous when \(A\) is continuous in the operator norm and uniformly bounded below. Moreover, it satisfies an equation of the same type on \(W \times H\), with a right-hand side of class \(C^{k}\) by the inductive hypothesis applied to \(z\). So \(\mathrm{d}z\) is \(C^{k}\) by the inductive hypothesis again, and \(z\) is \(C^{k+1}\).

Theorem 8.6 Bounded below derivatives reflect smoothness
✓

Let \(W\) be an open subset of a locally convex space and \(\gamma \colon W \to U\) continuous. If \(j \circ \gamma \) is of class \(C^{k}\), then so is \(\gamma \).

In the linear example above, the theorem says that a continuous map into \(E\) is \(C^{k}\) as soon as it is \(C^{k}\) when regarded, through \(j\), as a map into the larger space \(F\).

Proof ▶

By induction on \(k\). Fix \(q \in W\), a direction \(h\), and write \(A = \mathrm{d}j(\gamma q)\), \(d_t = \gamma (q + th) - \gamma (q)\). Since \(j\) is strictly differentiable at \(\gamma q\), for any \(c {\gt} 0\) we have

\[ \| j(\gamma (q + th)) - j(\gamma q) - A d_t\| \le c \| d_t\| \]

for small \(t\). With \(c \le m/2\) the lower bound gives \(\| d_t / t\| \le (2/m) \| (j(\gamma (q+th)) - j(\gamma q))/t\| \), so \(A(d_t/t) \to \mathrm{d}(j\circ \gamma )(q)(h)\), and as in Lemma 8.5 the quotients \(d_t/t\) converge. The derivative \(\mathrm{d}\gamma \) then solves

\[ \mathrm{d}j(\gamma q)\bigl(\mathrm{d}\gamma (q)(h)\bigr) = \mathrm{d}(j \circ \gamma )(q)(h), \]

and Lemma 8.5 applies.

Remark 8.7
#

A word of caution on the scope of this result. The proof uses that \(x \mapsto \mathrm{d}j(x)\) is continuous in the operator norm, to keep the lower bound near a point. The Michal–Bastiani definition only gives the joint continuity of \((x, v) \mapsto \mathrm{d}j(x)(v)\); on normed spaces the operator norm continuity is recovered through Theorem 8.1. For locally convex model spaces, Section 8.4 proves the statement under a locally uniform lower bound, which is automatic in normed spaces; with only a pointwise lower bound it is open.

8.4 Uniformly injective derivatives on locally convex spaces

We now let \(E\), \(F\) be Hausdorff locally convex spaces, \(E\) sequentially complete (Definition 5.1), and \(j \colon U \to F\) of class \(C^{2}\) on an open set \(U \subseteq E\). For a single continuous linear map \(L \colon E \to F\) the analogue of being bounded below is being a topological embedding: for every continuous seminorm \(p\) on \(E\) there are a continuous seminorm \(q\) on \(F\) and \(C \ge 0\) with \(p(w) \le C\, q(Lw)\). The proof of Theorem 8.6 needs this near each point, uniformly, and also for convex combinations of derivatives. We found no source for the results of this section; they were written for this library by adapting the Banach proof, and we do not claim they are new.

Definition 8.8 Uniformly injective derivatives
✓

The derivatives of \(j\) are uniformly injective on \(V \subseteq U\) if for every continuous seminorm \(p\) on \(E\) there are a continuous seminorm \(q\) on \(F\) and \(C \ge 0\) with

\[ p(w) \le C\, q(z) \qquad \text{for all } w \in E \text{ and } z \in \overline{\operatorname {conv}}\, \{ \mathrm{d}j(y)(w) : y \in V\} . \]

We say that this holds near \(x_0\) if it holds on some neighborhood of \(x_0\).

Equivalently, \(p(w) \le C\, q\bigl(\sum _i \lambda _i \mathrm{d}j(y_i)(w)\bigr)\) for all finite convex combinations with \(y_i \in V\). Taking a single point, each \(\mathrm{d}j(y)\), \(y \in V\), is a topological embedding with constants independent of \(y\). The hypothesis does not follow from each \(\mathrm{d}j(x)\) being an embedding: in the Fréchet space of rapidly decreasing sequences there is a smooth family of linear automorphisms \(L(t)\) with \(L(0) = \mathrm{id}\) whose inverses are not equicontinuous near \(t = 0\) (Example 16.4).

Lemma 8.9 The normed case
✓

If \(E\), \(F\) are normed and \(\mathrm{d}j(x_0)\) is bounded below, then the derivatives of \(j\) are uniformly injective near \(x_0\).

Proof ▶

Let \(m\| v\| \le \| \mathrm{d}j(x_0) v\| \). By Theorem 8.1, \(x \mapsto \mathrm{d}j(x)\) is continuous in the operator norm, so \(\| \mathrm{d}j(y) - \mathrm{d}j(x_0)\| {\lt} m/2\) on a neighborhood \(V\) of \(x_0\). The closed ball of radius \(\frac m2\| w\| \) around \(\mathrm{d}j(x_0) w\) is closed and convex and contains every \(\mathrm{d}j(y) w\), \(y \in V\), hence every \(z\) in their closed convex hull, and then \(\| z\| \ge \frac m2 \| w\| \). A continuous seminorm satisfies \(p \le c\| \cdot \| \).

Lemma 8.10 Difference quotients
✓

Assume the derivatives of \(j\) are uniformly injective near \(x \in U\). If \(d_i \to 0\), \(\tau _i \ne 0\) and \(\tau _i^{-1}(j(x + d_i) - j(x)) \to R\) along a countably generated filter, then \(\tau _i^{-1} d_i \to D\) with \(\mathrm{d}j(x) D = R\).

Proof ▶

Write \(g_i = \tau _i^{-1} d_i\) and \(\Delta _i = \tau _i^{-1}(j(x + d_i) - j(x))\). By the mean value inclusion (Lemma 2.8), \(\Delta _i\) lies in the closed convex hull of \(\{ \mathrm{d}j(y)(g_i) : y \in [x, x + d_i]\} \), and the segment lies in \(V\) for large \(i\); so \(p(g_i) \le C q(\Delta _i)\) is eventually bounded, for each continuous seminorm \(p\).

Joint continuity of \(\mathrm{d}^2 j\) at \((x, 0, 0)\) and homogeneity give, for each continuous seminorm \(q\) on \(F\), a continuous seminorm \(r\) on \(E\) with \(q(\mathrm{d}^2 j(z)(u, v)) \le r(u) r(v)\) for \(z\) near \(x\). The mean value inclusion, twice, gives \(q(\Delta _i - \mathrm{d}j(x) g_i) \le r(d_i)\, r(g_i) \to 0\). Hence \(\mathrm{d}j(x) g_i \to R\). As \(\mathrm{d}j(x)\) is an embedding, \((g_{i_n})\) is Cauchy along a sequence \(i_n\) tending to the filter, and converges to some \(D\) because \(E\) is sequentially complete; then \(\mathrm{d}j(x) D = R\), and \(p(g_i - D) \le C q(\mathrm{d}j(x) g_i - R) \to 0\).

Lemma 8.11 Linear bootstrap, locally convex case
✓

Assume the derivatives of \(j\) are uniformly injective near every point of \(U\) and \(j\) is \(C^{k+1}\). Let \(W\) be an open subset of a locally convex space, \(\gamma \colon W \to U\) and \(w \colon W \to F\) of class \(C^{k}\), and \(z \colon W \to E\) continuous with \(\mathrm{d}j(\gamma q)(z q) = w(q)\). Then \(z\) is of class \(C^{k}\).

Proof ▶

As in Lemma 8.5. The second order term \(t^{-1}\bigl(\mathrm{d}j(\gamma (q + th))(z(q+th)) - \mathrm{d}j(\gamma q)(z(q + th))\bigr)\) converges to \(\mathrm{d}^2 j(\gamma q)(\mathrm{d}\gamma (q)(h), z q)\) by Glöckner’s lemma (Lemma 2.11) for \(\mathrm{d}j\), and the limit of the quotients of \(z\) is produced as in Lemma 8.10. Continuity of \(\mathrm{d}z\): if \(\mathrm{d}j(\gamma a)(z a) = r(a)\), then \(p(z a' - z a) \le C q\bigl(r(a') - \mathrm{d}j(\gamma a')(z a)\bigr) \to 0\) by joint continuity of \(\mathrm{d}j\).

Theorem 8.12 Uniformly injective derivatives reflect smoothness
✓

Assume the derivatives of \(j\) are uniformly injective near every point of \(U\), and that \(j\) is \(C^{2}\) and \(C^{k}\). Let \(W\) be an open subset of a locally convex space and \(\gamma \colon W \to U\) continuous. If \(j \circ \gamma \) is of class \(C^{k}\), then so is \(\gamma \).

Proof ▶

By induction on \(k\). Lemma 8.10, with \(d_t = \gamma (q + th) - \gamma (q)\) and \(\tau _t = t\), gives the derivative \(\mathrm{d}\gamma \), which solves \(\mathrm{d}j(\gamma q)(\mathrm{d}\gamma (q)(h)) = \mathrm{d}(j \circ \gamma )(q)(h)\); then Lemma 8.11 applies.

Corollary 8.13 Pointwise lower bounds in Banach spaces
✓

Theorem 8.6 holds when \(\mathrm{d}j(x)\) is bounded below at every \(x \in U\), with a constant depending on \(x\).

Proof ▶

Lemma 8.9 and Theorem 8.12; a Banach space is sequentially complete.

Whether Theorem 8.12 holds when each \(\mathrm{d}j(x)\) is only assumed to be a topological embedding is open.