Michal–Bastiani calculus on locally convex spaces

3 Higher differentials and rules of calculus

There are two natural ways to say that a map \(f \colon U \to F\) is \(C^{k}\). The one we adopted in Definition 2.3 is recursive: \(f\) is \(C^{k+1}\) when \(\mathrm{d}f\), regarded as a map on the open set \(U \times E\), is \(C^{k}\). This is the form used by Glöckner, and it is the convenient one for proofs by induction. The classical definition, due to Bastiani [ Bas64 ] in 1964, asks instead that the iterated directional derivatives \(\mathrm{d}^{(j)} f(x; y_1, \dots , y_j)\) exist and be jointly continuous in the point and in all the directions, for \(j \leq k\).

In this chapter we prove that both definitions agree, and we derive from this the basic structure of the higher differentials: if \(f\) is \(C^{k+r}\), then \(\mathrm{d}^{(k)} f\) is again \(C^{r}\), and for each \(x\) the map \(\mathrm{d}^{(k)} f(x; \cdot )\) is continuous, symmetric and \(k\)-linear. We first define the iterated differentials and Bastiani’s notion, and prove a formula (the slice lemma) relating the iterated differentials of \(f\) to those of \(\mathrm{d}f\); this formula is the bridge between the two definitions. We then deal with regularity and symmetry, and the last section collects the rules of calculus used in practice: the product rule, Taylor’s formula and the second-order chain rule. Our reference is [ GN26 , § 1.3 ] .

3.1 Iterated differentials

Definition 3.1 Iterated differentials
✓
#

For \(f \colon E \to F\), \(x \in E\) and \(y_1, \dots , y_k \in E\) we define

\[ \mathrm{d}^{(0)} f(x) = f(x), \qquad \mathrm{d}^{(k+1)} f(x; y_1, \dots , y_{k+1}) = \mathrm{d}\bigl(z \mapsto \mathrm{d}^{(k)} f(z; y_1, \dots , y_k)\bigr)(x)(y_{k+1}), \]

that is, \(\mathrm{d}^{(k)} f(x; y_1, \dots , y_k) = (D_{y_k} \cdots D_{y_1} f)(x)\).

So \(\mathrm{d}^{(k)} f(x; y_1, \dots , y_k)\) is obtained by differentiating \(f\) successively in the directions \(y_1, \dots , y_k\), keeping the directions fixed and moving only the point. Note that \(\mathrm{d}^{(1)} f(x; y) = \mathrm{d}f(x)(y)\).

Definition 3.2 Bastiani’s \(C^{k}\) maps
✓

A map \(f\) has continuous iterated derivatives of order \(k\) on \(U\) if, for every \(j \leq k\), the iterated differentials \(\mathrm{d}^{(j)} f(x; y_1, \dots , y_j)\) exist for \(x \in U\) and are continuous on \(U \times E^j\).

Remark 3.3
#

In a Banach space the \(k\)-th Fréchet derivative is a map from \(U\) into the space of continuous \(k\)-linear maps \(E^k \to F\), and its continuity is measured in the operator norm. Here no topology is put on spaces of multilinear maps: what is required is the joint continuity of \(\mathrm{d}^{(j)} f\) on \(U \times E^j\). This is the same choice already made for \(k = 1\) in Definition 2.3, and it is what allows the theory to work without norms.

The bridge between the two definitions is a formula that expresses the iterated differentials of \(f\) through those of \(\mathrm{d}f\), in directions of the form \((y, 0)\).

For \(g \colon E \times E' \to F\) and \(a \in E'\),

\[ \mathrm{d}^{(k)}\bigl(g(\cdot , a)\bigr)(x; v_1, \dots , v_k) = \mathrm{d}^{(k)} g\bigl((x, a); (v_1, 0), \dots , (v_k, 0)\bigr). \]

In particular

\[ \mathrm{d}^{(k+1)} f(x; y_0, \dots , y_k) = \mathrm{d}^{(k)}(\mathrm{d}f)\bigl((x, y_0); (y_1, 0), \dots , (y_k, 0)\bigr). \]

In words: differentiating the partial map \(g(\cdot , a)\) is the same as differentiating \(g\) in directions that do not move the second variable. The second formula is the case \(g = \mathrm{d}f\), \(a = y_0\).

Proof ▶

Both sides are limits of the same difference quotients, and the claim follows by induction on \(k\). The second formula is [ GN26 , (1.36) ] .

Theorem 3.5 The two definitions of \(C^{k}\) agree
✓

A map \(f\) is \(C^{k}\) on \(U\) in the sense of Definition 2.3 if and only if it has continuous iterated derivatives of order \(k\) on \(U\).

From now on we use whichever of the two descriptions is more convenient.

Proof ▶

Both implications go by induction on \(k\), through Lemma 3.4. If \(f\) is \(C^{k+1}\), the iterated differentials of \(f\) are slices of those of \(\mathrm{d}f\), which is \(C^{k}\); hence they exist and are continuous. For the converse, the derivatives of \(\mathrm{d}f\) in a general direction \((h, w)\) split, by linearity in the second variable and Theorem 2.16, into a derivative in the direction \((h, 0)\), which is an iterated differential of \(f\), and one in the direction \((0, w)\), which is again of the form \(\mathrm{d}f\). The details can be found in [ GN26 , 1.3.10 and 1.3.12 ] .

3.2 Regularity and symmetry

Theorem 3.6 Regularity of the differentials
✓

If \(f\) is \(C^{k + r}\) on \(U\), then \(\mathrm{d}^{(k)} f\) is \(C^{r}\) on \(U \times E^k\).

Thus each differentiation costs exactly one order of regularity, as in the Banach case; in particular, the iterated differentials of a smooth map are smooth.

Proof ▶

We argue by induction on \(k\), using that \(\mathrm{d}^{(k+1)} f\) is, up to reordering the variables, the derivative of \(\mathrm{d}^{(k)} f\) in directions of the form \((y, 0)\). This is [ GN26 , 1.3.13 ] .

If \(f\) is \(C^{k}\) on \(U\) and \(x \in U\), then \(\mathrm{d}^{(k)} f(x; \cdot )\) is a continuous symmetric \(k\)-linear map \(E^k \to F\).

This is what allows us to regard \(\mathrm{d}^{(k)} f(x)\) as the \(k\)-th differential of \(f\) at \(x\), in the same sense as in finite dimension: the order in which we differentiate is irrelevant, and the result depends linearly on each direction.

Proof ▶

We begin with symmetry. It suffices to consider transpositions of adjacent variables, and these reduce to Theorem 2.18 applied to an iterated differential of lower order. Linearity in the last variable is Theorem 2.12 applied to \(\mathrm{d}^{(k-1)} f\); by symmetry, the map is linear in every variable. Continuity follows from the joint continuity of \(\mathrm{d}^{(k)} f\). This is [ GN26 , 1.3.17 ] .

Remark 3.8
#

A word of caution: the continuity in Theorem 3.7 is continuity of the \(k\)-linear map \(\mathrm{d}^{(k)} f(x; \cdot )\) on \(E^k\), for a fixed \(x\). Nothing is claimed about the dependence of this multilinear map on \(x\) in a topology on a space of multilinear maps; the regularity in \(x\) is expressed by the joint statement of Theorem 3.6.

3.3 Rules of calculus

With the chain rule (Theorem 2.14) and the structure of the higher differentials (Chapter 3) at hand, we can now derive the remaining rules of calculus that are used in practice. All of them are classical in the Banach setting, and their statements are the familiar ones; what has to be checked is that the proofs survive without norms. In each case the argument either reduces to the chain rule or, through a continuous linear functional, to the calculus of real functions of one variable.

We begin with continuous bilinear maps and the product rule. We then compute the derivatives of a map along a line, which leads to Taylor’s formula in a weak form, valid without any completeness assumption on \(F\). The section closes with the chain rule for second differentials. The references are the results of [ GN26 ] quoted in each proof.

3.3.1 Bilinear maps and the product rule

Let \(\beta \colon A \times B \to F\) be continuous and bilinear. Then \(\beta \) is \(C^{k}\) for every \(k\), with \(\mathrm{d}\beta \bigl((a, b), (h, w)\bigr) = \beta (a, w) + \beta (h, b)\).

Proof ▶

The difference quotient of \(\beta \) at \((a,b)\) in the direction \((h,w)\) is \(\beta (a, w) + \beta (h, b) + t\, \beta (h, w)\), and letting \(t \to 0\) gives the formula. The differential is therefore a sum of continuous bilinear maps composed with continuous linear maps, and induction on \(k\) completes the proof. This is [ GN26 , 1.2.3 and 1.3.5 ] .

Note that continuity of \(\beta \) means joint continuity on \(A \times B\). This cannot be weakened to separate continuity: for \(b = 0\) and \(h = 0\) the differential above reduces to \(\beta (a, w)\), which must be jointly continuous in \((a, w)\).

If \(f \colon U \to A\) and \(g \colon U \to B\) are \(C^{k}\) and \(\beta \) is continuous and bilinear, then \(x \mapsto \beta (f(x), g(x))\) is \(C^{k}\) and

\[ \mathrm{d}\bigl(\beta (f, g)\bigr)(x)(h) = \beta \bigl(f(x), \mathrm{d}g(x)(h)\bigr) + \beta \bigl(\mathrm{d}f(x)(h), g(x)\bigr). \]
Proof ▶

It is immediate from the chain rule applied to \(\beta \circ (f, g)\), together with Theorem 3.9.

Example 3.11
#

Let \(\beta \colon E \times E \to F\) be continuous and bilinear, and let \(q(x) = \beta (x, x)\). Taking \(f = g = \mathrm{id}_E\) in Corollary 3.10, we obtain that \(q\) is smooth, with \(\mathrm{d}q(x)(h) = \beta (x, h) + \beta (h, x)\). When \(\beta \) is symmetric this is the familiar \(2\beta (x, h)\).

3.3.2 Taylor’s formula

To pass from the differentials of \(f\) to the values of \(f\) we restrict \(f\) to a line and compose with a continuous linear functional. The resulting real function of one variable is as regular as \(f\), and its derivatives are the iterated differentials of \(f\) along the line.

Let \(f\) be \(C^{k}\) on \(U\), \(j {\lt} k\) and \(\psi \) a continuous linear functional on \(F\). Then the real function \(s \mapsto \psi \bigl(\mathrm{d}^{(j)} f(x + sh; h, \dots , h)\bigr)\) has derivative \(\psi \bigl(\mathrm{d}^{(j+1)} f(x + sh; h, \dots , h)\bigr)\) wherever \(x + sh \in U\).

Proof ▶

By the chain rule along the line \(s \mapsto (x + sh, h, \dots , h)\), whose derivative is \((h, 0, \dots , 0)\).

Theorem 3.13 Taylor’s formula, weak form
✓

Let \(f\) be \(C^{k+1}\) on a set containing the segment \([x, x+h]\). For every continuous linear functional \(\psi \) on \(F\),

\[ \psi \bigl(f(x + h)\bigr) = \sum _{j=0}^{k} \frac{1}{j!}\, \psi \bigl(\mathrm{d}^{(j)} f(x; h, \dots , h)\bigr) + \int _0^1 \frac{(1-t)^k}{k!}\, \psi \bigl(\mathrm{d}^{(k+1)} f(x + th; h, \dots , h)\bigr)\, dt . \]
Proof ▶

We apply Taylor’s formula with integral remainder to the real function \(s \mapsto \psi (f(x + sh))\), whose derivatives are given by Lemma 3.12. Compare [ GN26 , 1.6.1 and 1.6.14 ] , where the formula is stated for the integrals in \(F\); the weak form needs no completeness of \(F\).

Remark 3.14
#

The remainder in Theorem 3.13 is a real integral, one for each functional \(\psi \). Writing it as a vector of \(F\) requires the weak integral of \(t \mapsto \frac{(1-t)^k}{k!}\, \mathrm{d}^{(k+1)} f(x + th; h, \dots , h)\) to exist, which in general depends on a completeness property of \(F\) (Chapters 4 and 5). Since continuous linear functionals separate points, the weak form determines \(f(x + h)\) all the same.

3.3.3 The second-order chain rule

Theorem 3.15 Second-order chain rule
✓
#

Let \(\tau \colon U \to V\) and \(g \colon V \to F\) be \(C^{2}\), with \(U \subseteq E\) and \(V \subseteq E'\) open. For \(y \in U\) and \(a, b \in E\),

\[ \mathrm{d}^{(2)} (g \circ \tau )(y; a, b) = \mathrm{d}^{(2)} g\bigl(\tau y; \mathrm{d}\tau (y)(a), \mathrm{d}\tau (y)(b)\bigr) + \mathrm{d}g(\tau y)\bigl(\mathrm{d}^{(2)} \tau (y; a, b)\bigr). \]

The first term is what one would guess by applying the chain rule twice naively; the second involves the second differential of \(\tau \), and vanishes when \(\tau \) is the restriction of a continuous linear map, since then \(\mathrm{d}\tau (y)(b)\) does not depend on \(y\).

Proof ▶

On \(U \times E\), the first-order chain rule gives \(\mathrm{d}(g \circ \tau ) = \mathrm{d}g \circ \Phi \) with \(\Phi (y', b') = (\tau y', \mathrm{d}\tau (y')(b'))\). We differentiate this once more with the first-order chain rule. The derivative of \(\Phi \) at \((y, b)\) in the direction \((a, 0)\) is \((\mathrm{d}\tau (y)(a), \mathrm{d}^{(2)}\tau (y; a, b))\). Splitting it as \((\mathrm{d}\tau (y)(a), 0) + (0, \mathrm{d}^{(2)}\tau (y; a, b))\), the first term gives \(\mathrm{d}^{(2)} g\), and the second gives \(\mathrm{d}g\), because \(\mathrm{d}g(\tau y)\) is linear.

3.3.4 Operations on \(C^{k}\) maps

Directional derivatives and \(C^{k}\) maps behave as expected under the elementary operations: maps into products are differentiated componentwise, and finite sums, multiples, differences, products \(f \cdot g\) of a real function with a vector-valued one, and products of maps \((x, y) \mapsto (g(x), h(y))\) of \(C^{k}\) maps are \(C^{k}\). A \(C^{k}\) map with values in a closed subspace is \(C^{k}\) into the subspace [ GN26 , compare Lemma 1.3.19 ] . The differential depends only on the germ of the map, and on normed spaces the first and second Michal–Bastiani differentials of \(C^{1}\) and \(C^{2}\) maps are the Fréchet derivatives.

Proof ▶

Each statement reduces to the chain rule and continuous linear maps, or to the product rule for the continuous bilinear map \((t, v) \mapsto tv\); the closed subspace case uses that difference quotients stay in the subspace.