Michal–Bastiani calculus on locally convex spaces

5 Completeness, Mackey completeness and the existence of integrals

In Chapter 4 the weak integral of a curve was defined, but its existence was always a hypothesis. In a Banach space this hypothesis is superfluous: continuous curves on a compact interval have Riemann integrals. In a general locally convex space nothing of the sort holds, and one has to pay for integrals with some completeness of \(F\). The aim of this chapter is to determine how much completeness is needed, and to draw the consequences for integrals depending on parameters.

Two answers are classical. Weak integrals of continuous curves exist when \(F\) is sequentially complete, and weak integrals of \(C^{1}\) curves exist under the weaker assumption of Mackey completeness. The second one is the natural hypothesis for smooth calculus: it is what makes parameter-dependent integrals of smooth maps smooth. Mackey completeness is also the completeness notion of Kriegl and Michor [ KM97 , § 2 ] ; we stress that here it is only a hypothesis on the model space, inside Michal–Bastiani calculus. Our main reference is [ GN26 , § 1.1 ] .

The chapter is organized as follows. In Section 5.1 we introduce both completeness notions and compare them. In Section 5.2 we study Riemann sums, which in both cases provide the integral as a limit. Section 5.3 contains the two existence theorems, Section 5.4 the second half of the fundamental theorem of calculus, Section 5.5 the regularity of integrals depending on parameters, and Section 5.6 the equivalence of two definitions of Mackey completeness.

5.1 Completeness notions

We state completeness through the topology of \(F\), without fixing a uniform structure.

Definition 5.1 Sequential completeness
✓
#

A sequence \((u_n)\) in \(F\) is Cauchy if for every neighborhood \(W\) of \(0\) there is \(N\) with \(u_n - u_m \in W\) for all \(m, n \geq N\). The space \(F\) is sequentially complete if every Cauchy sequence converges.

Definition 5.2 Mackey completeness
✓
#

A sequence \((u_n)\) in \(F\) is Mackey–Cauchy if there are a bounded set \(B \subseteq F\) and real numbers \(r_k \to 0\) with \(u_l - u_k \in r_k B\) for all \(k \leq l\). The space \(F\) is Mackey complete if every Mackey–Cauchy sequence converges.

Here bounded means bounded in the sense of von Neumann: absorbed by every neighborhood of \(0\). In other words, a Mackey–Cauchy sequence is Cauchy at a controlled rate, measured by a single bounded set; in a normed space one may take for \(B\) the unit ball, and the condition reads \(\| u_l - u_k\| \leq r_k\).

Remark 5.3
#

Definition 5.2 is one of the equivalent descriptions of Mackey completeness in [ GN26 , 1.1.18(e) ] and [ KM97 , 2.2, 2.14 ] . Glöckner and Neeb take as definition the existence of weak integrals of Lipschitz curves; we prove in Section 5.6 that both definitions agree, so the choice is a matter of convenience.

Complete uniform additive groups are sequentially complete, and sequentially complete spaces are Mackey complete.

In particular, every complete locally convex space (for instance, a Fréchet or a Banach space) is Mackey complete, and all the results of this chapter apply to it.

Proof ▶

The first claim amounts to comparing our notion of Cauchy sequence with the uniform one. For the second, it suffices to observe that a Mackey–Cauchy sequence is Cauchy: given a neighborhood \(W\) of \(0\), the bounded set \(B\) is absorbed by \(W\), so \(r_k B \subseteq W\) for large \(k\).

We will need the following elementary fact, which is where local convexity enters.

Lemma 5.5 Convex hulls of bounded sets
✓

In a locally convex space, the convex hull of a bounded set is bounded.

Proof ▶

It is immediate that if \(B\) is absorbed by a convex balanced neighborhood \(W\) of \(0\), so is \(\operatorname {conv}B\); and these neighborhoods form a basis of neighborhoods of \(0\).

5.2 Riemann sums

In both existence theorems the integral will be obtained as the limit of a sequence of Riemann sums. We only need sums over uniform partitions.

Definition 5.6 Riemann sums
✓
#

For \(\gamma \colon \mathbb {R}\to F\) and \(N \geq 1\), the Riemann sum of \(\gamma \) over \([a, b]\) with \(N\) equal subintervals and left endpoints is

\[ R_N(\gamma ) = \frac{b-a}{N} \sum _{i=0}^{N-1} \gamma \Bigl(a + i\, \frac{b-a}{N}\Bigr). \]

The basic estimate compares a Riemann sum with the sums of its refinements. It is stated with an arbitrary convex set \(V\), so that the same lemma serves both for neighborhoods of \(0\) (the sequentially complete case) and for multiples of a bounded set (the Mackey complete case).

Let \(V \subseteq F\) be convex, and suppose that \(\gamma (s) - \gamma (t) \in V\) whenever \(s, t \in [a, b]\) and \(0 \leq s - t \leq \frac{b - a}{N}\). Then \(R_{Nm}(\gamma ) - R_N(\gamma ) \in (b - a)\, V\) for every \(m \geq 1\). Moreover, for a continuous real function \(g\) the Riemann sums \(R_N(g)\) tend to \(\int _a^b g\).

Proof ▶

The difference \(R_{Nm} - R_N\) is \(b - a\) times an average of increments of \(\gamma \) over intervals of length at most \(\frac{b - a}{N}\); since \(V\) is convex, the average lies in \(V\). The second claim is the classical convergence of Riemann sums of continuous functions.

Lemma 5.8 Limits of Riemann sums
✓

If \(\gamma \) is continuous on \([a, b]\) and the dyadic Riemann sums \(R_{2^n}(\gamma )\) converge to \(y\), then \(y\) is the weak integral of \(\gamma \) over \([a, b]\).

Proof ▶

For each continuous linear functional \(\varphi \) we have \(\varphi (R_{2^n}(\gamma )) = R_{2^n}(\varphi \circ \gamma )\), which converges both to \(\varphi (y)\) and, by Lemma 5.7, to \(\int _a^b \varphi \circ \gamma \).

Remark 5.9
#

The dyadic sums \(R_{2^n}(\gamma )\) form a sequence, each term being a refinement of the previous one. This is why sequential notions of completeness suffice: no nets of partitions are needed.

5.3 Existence of integrals

We can now state the two existence theorems. The construction is the same in both; only the Cauchy estimate for the dyadic sums changes.

Theorem 5.10 Continuous curves, sequentially complete case
✓

If \(F\) is sequentially complete, every curve that is continuous on \([a, b]\) has a weak integral over \([a, b]\).

Proof ▶

By uniform continuity on the compact interval \([a, b]\), for each closed convex neighborhood \(V\) of \(0\) the increments of \(\gamma \) over sufficiently short intervals lie in \(V\). By Lemma 5.7, the dyadic Riemann sums form a Cauchy sequence; its limit is the weak integral by Lemma 5.8. This is [ GN26 , 1.1.15 ] .

If \(F\) is Mackey complete, every curve that is \(C^{1}\) on a neighborhood of \([a, b]\) has a weak integral over \([a, b]\).

Proof ▶

Let \(K\) be the closure of the convex hull of \(\{ s\, \gamma '(t) : s \in [0, 1],\ t \in [a, b]\} \). It is bounded by Lemma 5.5, since \(\gamma '([a, b])\) is compact. By the mean value inclusion (Lemma 2.8), \(\gamma (s) - \gamma (t) \in h\, K\) whenever \(0 \leq s - t \leq h\). Then Lemma 5.7 gives

\[ R_{2^m}(\gamma ) - R_{2^n}(\gamma ) \in \frac{(b - a)^2}{2^n}\, K \qquad (n \leq m), \]

so the dyadic Riemann sums form a Mackey–Cauchy sequence, with bounded set \(K\) and \(r_n = (b - a)^2 2^{-n}\). Its limit is the weak integral by Lemma 5.8. Compare [ GN26 , 1.1.18 ] and [ KM97 , 2.14 ] .

Remark 5.12
#

In the continuous case the Cauchy estimate comes from uniform continuity, and it only says that the differences of dyadic sums are eventually small. In the \(C^{1}\) case it comes from the mean value inclusion, and it says more: the differences lie in multiples of one fixed bounded set, with factors tending to \(0\) at a known rate. This additional control is what allows us to replace sequential completeness by Mackey completeness.

5.4 The fundamental theorem of calculus, second part

The first part of the fundamental theorem (Theorem 4.8) recovers a \(C^{1}\) curve from the integral of its derivative. The second part, which we prove now, differentiates an integral with respect to its upper limit.

Let \(I \subseteq \mathbb {R}\) be an open interval, \(a \in I\) and \(\gamma \) continuous on \(I\). If the weak integrals \(\eta (s) = \int _a^s \gamma \) exist for \(s \in I\), then \(\eta \) is differentiable on \(I\) with \(\eta ' = \gamma \). If \(F\) is sequentially complete, these integrals always exist.

Proof ▶

By additivity in the interval and the affine change of variables \(r \mapsto t + sr\),

\[ \frac{1}{s}\bigl(\eta (t + s) - \eta (t)\bigr) - \gamma (t) = \int _0^1 \bigl(\gamma (t + sr) - \gamma (t)\bigr)\, dr . \]

Let \(W\) be a closed convex neighborhood of \(0\). By continuity of \(\gamma \) at \(t\), for small \(s\) the integrand lies in \(W\), and Lemma 4.5 puts the integral in \(2W\). The last claim follows from Theorem 5.10. This is [ GN26 , 1.1.12 ] .

Remark 5.14
#

Note that the first assertion needs no completeness at all: what is assumed is the existence of the integrals \(\eta (s)\), and the limit of the difference quotients is then controlled by Hahn–Banach through Lemma 4.5. Completeness of \(F\) enters only to guarantee that existence.

5.5 Integrals depending on parameters

Let \(U\) be an open subset of a locally convex space \(E\), let \(T \subseteq \mathbb {R}\) be open with \([a, b] \subseteq T\), and let \(f\) be a map on \(U \times T\). We study the regularity of \(x \mapsto \int _a^b f(x, t)\, dt\). All the results follow from a single induction, in which the existence of the integrals is isolated as a hypothesis.

Theorem 5.15 The induction behind parameter-dependent integrals
✓

Let \(T \subseteq \mathbb {R}\) be open with \([a, b] \subseteq T\), and \(m \in \mathbb {N}\). Suppose that, for every \(C^{m}\) map \(f\) on a product \(U \times T\), the weak integrals \(\int _a^b f(x, t)\, dt\) exist. Then, for every \(C^{k + m}\) map \(f\) on \(U \times T\), the map \(x \mapsto \int _a^b f(x, t)\, dt\) is \(C^{k}\) on \(U\).

Here the hypothesis is understood for all open sets \(U\) of all locally convex spaces \(E\); this uniformity is what makes the induction work.

Proof ▶

We argue by induction on \(k\), for all locally convex spaces \(E\) at once. The case \(k = 0\) is Proposition 4.10. For the inductive step, differentiation under the integral sign (Theorem 4.11) gives

\[ \mathrm{d}\Bigl(\int _a^b f(\cdot , t)\, dt\Bigr)(x)(h) = \int _a^b \mathrm{d}f\bigl((x, t), (h, 0)\bigr)\, dt , \]

and the integrand is a \(C^{k - 1 + m}\) map on the product \((U \times E) \times T\), to which the inductive hypothesis applies.

Remark 5.16
#

This single induction yields all the results on parameter-dependent integrals at once; the sources prove the sequentially complete case and the Mackey complete case separately.

Remark 5.17 Lean
#

The quantification over all spaces \(E\) is the reason for the universe restriction mentioned in Chapter 2.

Combined with the existence theorems of Section 5.3, the induction gives the following statements, with no existence hypotheses left.

Let \(f\) be a map on \(U \times T\) with \([a, b] \subseteq T\), and \(I(x) = \int _a^b f(x, t)\, dt\).

  1. If \(F\) is sequentially complete and \(f\) is \(C^{k}\), then \(I\) is \(C^{k}\), and \(\mathrm{d}I(x)(h) = \int _a^b \mathrm{d}f\bigl((x,t), (h,0)\bigr)\, dt\) for \(k \geq 1\).

  2. If \(F\) is Mackey complete and \(f\) is \(C^{k+1}\), then \(I\) is \(C^{k}\); if \(f\) is \(C^{2}\), the formula for \(\mathrm{d}I\) of (1) holds.

  3. If \(F\) is Mackey complete and \(f\) is smooth, then \(I\) is smooth.

Proof ▶

Apply Theorem 5.15 with \(m = 0\) and Theorem 5.10 in the first case, and with \(m = 1\) and Theorem 5.11 in the other two. This is [ GN26 , 1.3.23 ] .

Remark 5.19
#

The loss of one derivative in the Mackey complete case has a simple origin: in that setting we only know that \(C^{1}\) curves have integrals (Theorem 5.11), so the induction has to be started with \(m = 1\). For smooth maps nothing is lost, and this is why Mackey completeness is the natural hypothesis for smooth calculus.

5.6 Mackey completeness and Lipschitz curves

There are two current definitions of Mackey completeness. The one we adopted in Definition 5.2 asks for the convergence of Mackey–Cauchy sequences, as in [ KM97 , 2.2 ] . Glöckner and Neeb, on the other hand, define Mackey completeness through integrals: a space is Mackey complete if every Lipschitz curve has a weak integral ( [ GN26 , 1.1.17(b) ] ). Their definition says directly which curves can be integrated; ours is phrased in terms of sequences alone. In this section we prove that both definitions agree.

The equivalence is stated in [ GN26 , 1.1.18(e) ] and [ KM97 , 2.14 ] . One direction is a variant of the existence theorem of Chapter 5, with difference quotients in place of derivatives. For the converse we give our own argument: a Mackey–Cauchy sequence is encoded as a Lipschitz curve whose integral is the limit we are looking for.

Definition 5.20 Lipschitz curves
✓

A curve \(\gamma \colon [a, b] \to F\) is Lipschitz if the set of difference quotients

\[ \Bigl\{ \tfrac {1}{s - t}\bigl(\gamma (s) - \gamma (t)\bigr) : s, t \in [a, b],\ s \neq t \Bigr\} \]

is bounded in \(F\).

In a normed space this is the usual Lipschitz condition. In general there is no constant to speak of, and its role is played by the bounded set of difference quotients.

Definition 5.21 Mackey completeness after Glöckner and Neeb
✓

We say that \(F\) integrates Lipschitz curves if every Lipschitz curve \([a, b] \to F\), \(a {\lt} b\), has a weak integral over \([a, b]\).

The following lemma plays, for Lipschitz curves, the role that the mean value inclusion plays for \(C^{1}\) curves.

Lemma 5.22 Lipschitz curves are continuous
✓

Let \(\gamma \) be a Lipschitz curve on \([a, b]\) and \(K = \operatorname {conv}(\{ 0\} \cup Q)\), where \(Q\) is the set of difference quotients. Then \(\gamma (s) - \gamma (t) \in h K\) whenever \(0 \leq s - t \leq h\). In particular \(\gamma \) is continuous.

Proof ▶

For \(s \neq t\) we have \(\gamma (s) - \gamma (t) = (s - t)\, q\) with \(q \in Q\), and \((s - t)q \in hK\) since \(K\) is convex and contains \(0\). Continuity follows because the bounded set \(K\) is absorbed by every neighborhood of \(0\).

5.6.1 Mackey complete spaces integrate Lipschitz curves

Proof ▶

The argument is that of Theorem 5.11, with the difference quotients in place of the derivative. The set \(K\) of Lemma 5.22 is bounded by Lemma 5.5. By Lemmas 5.22 and 5.7,

\[ R_{2^m}(\gamma ) - R_{2^n}(\gamma ) \in (b - a)\, \frac{b - a}{2^n}\, K \qquad (n \leq m), \]

so the dyadic Riemann sums form a Mackey–Cauchy sequence. Its limit is the weak integral by Lemma 5.8.

5.6.2 The converse

We start with an elementary observation, which allows us to work with a fast subsequence.

Lemma 5.24 Subsequences
✓

A Mackey–Cauchy sequence with a convergent subsequence converges, to the same limit.

Proof ▶

Let \(u_{n_j} \to x\) and let \(W\) be a neighborhood of \(0\); choose a symmetric neighborhood \(V\) with \(V + V \subseteq W\). For large \(k\) we have \(r_k B \subseteq V\). Choosing \(j\) with \(n_j \geq k\) and \(u_{n_j} - x \in V\), we get

\[ u_k - x = -(u_{n_j} - u_k) + (u_{n_j} - x) \in V + V \subseteq W . \]

The idea of the converse is the following. The increments of a Mackey–Cauchy sequence are placed on disjoint tents, which accumulate at \(t = 1\); the tents are scaled so that the integral of the \(j\)-th one is the \(j\)-th increment. The resulting curve is Lipschitz, and its integral is then the sum of the increments, that is, the limit of the sequence up to a translation.

For \(j \in \mathbb {N}\) let \(l_j = 1 - 2^{-j}\) and \(w_j = 2^{-(j+1)}\), so that the intervals \([l_j, l_j + w_j] = [l_j, l_{j+1}]\) tile \([0, 1)\). The \(j\)-th tent is the function

\[ \tau _j(t) = \max \Bigl(0,\ \frac{w_j}{2} - \Bigl|t - l_j - \frac{w_j}{2}\Bigr|\Bigr), \]

supported on \([l_j, l_{j+1}]\), with slopes \(\pm 1\) and height \(w_j / 2\). For a sequence \((e_j)\) in \(F\) we consider the curve \(\gamma (t) = \sum _j \tau _j(t)\, e_j\), which is a finite sum for each \(t\).

The slopes \(\pm 1\) are what makes \(\gamma \) Lipschitz, provided the vectors \(e_j\) stay in a bounded set; the fact that the tents become small near \(t = 1\) is what allows the \(e_j\) to be large multiples of the increments.

Each \(\tau _j\) is continuous, takes values in \([0, 1]\), satisfies \(|\tau _j(s) - \tau _j(t)| \leq |s - t|\), and \(\int _0^1 \tau _j = w_j^2 / 4\). For each \(t\), \(\gamma (t) = \sum _{j {\lt} N} \tau _j(t)\, e_j\) for all large \(N\).

Proof ▶

Elementary. The last claim holds because each \(t\) meets the support of at most finitely many tents.

If \(F\) integrates Lipschitz curves, then \(F\) is Mackey complete.

Proof ▶

Let \((u_n)\) be Mackey–Cauchy, with \(u_l - u_k \in r_k B\) for \(k \leq l\). Let \(K = \operatorname {conv}(\{ 0\} \cup B \cup -B)\), which is convex, symmetric and bounded by Lemma 5.5. Put \(c_j = 2^{-(j+1)}\), so that \(\sum _j c_j = 1\).

A fast subsequence. Since \(r_k \to 0\), we can choose \(n_0 {\lt} n_1 {\lt} \cdots \) with \(|r_{n_j}| \leq c_j w_j^2 / 4\). Let \(d_j = u_{n_{j+1}} - u_{n_j}\) and \(e_j = (4 / w_j^2)\, d_j\). Then \(e_j \in c_j K\).

The curve is Lipschitz. Let \(\gamma = \sum _j \tau _j e_j\). A difference quotient of \(\gamma \) is a finite sum \(\sum _j \frac{\tau _j(s) - \tau _j(t)}{s - t}\, e_j\) with coefficients of absolute value at most \(1\), hence it lies in \(\sum _j c_j K \subseteq K\). So \(\gamma \) is a Lipschitz curve on \([0, 1]\), and by hypothesis it has a weak integral \(y\).

The integral is the limit. Since \(\int _0^1 \tau _j e_j = d_j\), the partial sum \(\gamma _N = \sum _{j {\lt} N} \tau _j e_j\) has weak integral \(u_{n_N} - u_{n_0}\). On the other hand, \(\gamma (t) - \gamma _N(t) = \sum _{j \geq N} \tau _j(t) e_j \in 2^{-N} K\) for every \(t\). By Lemma 4.5, applied with \(W\) the closure of \(2^{-N} K\), we get \(y - (u_{n_N} - u_{n_0}) \in 2^{1-N}\, \overline{K}\), and this tends to \(0\) since \(K\) is bounded. Hence \(u_{n_N} \to u_{n_0} + y\), and by Lemma 5.24 the whole sequence converges.

Remark 5.28
#

The curve \(\gamma \) has corners at the ends and at the top of every tent, so it is not \(C^{1}\) as soon as some \(e_j \neq 0\). The argument therefore uses the full class of Lipschitz curves, and not only \(C^{1}\) curves.

Corollary 5.29 The two definitions agree
✓

A locally convex space is Mackey complete if and only if it integrates Lipschitz curves.

Proof ▶

Theorems 5.23 and 5.27.

In particular, the results of Chapter 5 stated under Mackey completeness hold verbatim with the definition of Glöckner and Neeb, and the references to [ GN26 ] can be compared directly.