Michal–Bastiani calculus on locally convex spaces

9 Analytic maps

We follow the treatment of Glöckner and Neeb [ GN26 , Ch. 2 ] , which builds on Bochnak and Siciak [ BS71a , BS71b ] and on Glöckner [ Glo02 , §2 ] . The point of this approach is that no completeness of the target space is needed for the basic theory. Bochnak and Siciak assume throughout that the target space \(F\) is sequentially complete; Glöckner removes this hypothesis by working in the completion \(\widetilde F\) and coming back to \(F\) through the derivatives, which lie in \(F\) [ Glo02 , Prop. 2.4, Lemma 2.5 ] .

9.1 Continuous homogeneous polynomials

Definition 9.1 Continuous homogeneous polynomials
✓
#

Let \(E\), \(F\) be topological vector spaces over \(\mathbb {K} \in \{ \mathbb {R}, \mathbb {C}\} \) and \(k \in \mathbb {N}_0\). A map \(p \colon E \to F\) is a continuous homogeneous polynomial of degree \(k\) if \(p(x) = A(x, \dots , x)\) for some continuous \(k\)-linear map \(A \colon E^k \to F\).

Replacing \(A\) by its symmetrization we may take \(A\) symmetric, and then \(A\) is determined by \(p\):

Theorem 9.2 Polarization; Mazur–Orlicz [ BS71a , Theorem A ]
✓

If \(A\) is symmetric and \(k\)-linear, then for \(x_1, \dots , x_k \in E\)

\[ \sum _{\varepsilon \in \{ \pm 1\} ^k} \varepsilon _1 \cdots \varepsilon _k \, A\Bigl(\sum _j \varepsilon _j x_j, \dots , \sum _j \varepsilon _j x_j\Bigr) = 2^k k! \, A(x_1, \dots , x_k). \]

In particular, if \(x \mapsto A(x, \dots , x)\) is continuous, then \(A\) is continuous.

Proof ▶

Expand \(A(\sum _j \varepsilon _j x_j, \dots )\) by multilinearity over all maps \(r \colon \{ 1, \dots , k\} \to \{ 1, \dots , k\} \). The sign sum \(\sum _\varepsilon \prod _j \varepsilon _j^{1 + |r^{-1}(j)|}\) is \(2^k\) if \(r\) is a bijection and \(0\) otherwise, and \(A(x_{r(1)}, \dots , x_{r(k)}) = A(x_1, \dots , x_k)\) for bijective \(r\) by symmetry. The continuity statement follows because the left hand side is a finite combination of values of \(x \mapsto A(x, \dots , x)\). No completeness, local convexity or Hausdorff property is used.

9.2 One complex variable

Let \(F\) be a complex locally convex space, not necessarily complete, and \(U \subseteq \mathbb {C}\) open. A map \(f \colon U \to F\) is complex analytic if around every point it is a power series \(f(z_0 + z) = \sum _n z^n a_n\) with \(a_n \in F\), converging pointwise on a disc [ GN26 , Rem. 2.1.3 ] .

A complex analytic \(f \colon U \to F\) is continuous, its complex derivative exists and is complex analytic. The Cauchy integrals \(\int _0^{2\pi } e^{-int} f(z_0 + \rho e^{it})\, dt = 2\pi \rho ^n a_n\) exist as weak integrals in \(F\), the Cauchy estimates \(q(a_n) \le \rho ^{-n} \sup _{|\zeta | = \rho } q(f(z_0 + \zeta ))\) hold for every continuous seminorm \(q\), and the Taylor series of \(f\) at \(z_0\) converges to \(f\) in \(F\) on every disc whose closure lies in \(U\) ( [ GN26 , Cor. 2.1.8, Lemma 2.1.10 ] ).

Proof ▶

No completion is used. Every sum and integral that appears has its value already in \(F\) (a value of \(f\), a coefficient, a derivative), and convergence is checked seminorm by seminorm against a geometric majorant. Scalar estimates are transferred to \(F\) by the Hahn–Banach theorem: if \(|\ell (v)| \le B\) for every continuous linear \(\ell \) with \(|\ell | \le q\), then \(q(v) \le B\).

Theorem 9.4 Complex analytic \(=\) real smooth with complex linear differential
✓

\(f \colon U \to F\) is complex analytic if and only if it is \(C^\infty \) in the real Michal–Bastiani sense and \(\mathrm{d}f(z)(w) = w\, \mathrm{d}f(z)(1)\) for all \(z \in U\), \(w \in \mathbb {C}\) [ GN26 , Cor. 2.1.9 ] .

Remark 9.5
#

Glöckner and Neeb prove the one variable characterizations with \(C^1\) in place of \(C^\infty \) under Mackey completeness [ GN26 , Prop. 2.1.6 ] , and Corollary 2.1.9 without completeness through the completion \(\widetilde F\). Here the smooth version is proved without completeness and without the completion: smoothness provides all the Taylor coefficients \(\frac{1}{n!}\mathrm{d}^{(n)} f(z)(1, \dots , 1)\) in \(F\). For a single power series on a disc, the derivative at a point other than the center is only known to exist when the differentiated series converges in \(F\); whether this can fail in an incomplete space we have not settled.

9.3 Complex analytic maps

Definition 9.6 Complex analytic maps [ GN26 , Def. 2.1.2 ] , [ BS71b , Def. 5.6 ]
✓
#

Let \(E\), \(F\) be complex locally convex spaces and \(U \subseteq E\) open. A map \(f \colon U \to F\) is complex analytic if it is continuous and for every \(x \in U\) there are continuous homogeneous polynomials \(p_k\) of degree \(k\) such that \(f(x + y) = \sum _{k \ge 0} p_k(y)\), as a pointwise limit, for all \(y\) in a \(0\)-neighborhood \(Y\) with \(x + Y \subseteq U\).

Definition 9.7 Gâteaux analytic maps [ GN26 , Def. 2.1.11 ]
✓
#

\(f\) is Gâteaux analytic if \(z \mapsto f(x + z y)\) is complex analytic (locally a power series with coefficients in \(F\)) on \(\{ z \in \mathbb {C} : x + zy \in U\} \), for all \(x \in U\), \(y \in E\).

For \(f \colon U \to F\) the following are equivalent:

  1. \(f\) is \(C^\infty \) (in the real Michal–Bastiani sense) and \(\mathrm{d}f(x) \colon E \to F\) is complex linear for each \(x \in U\);

  2. \(f\) is continuous and Gâteaux analytic;

  3. \(f\) is complex analytic;

  4. \(f\) is \(C^\infty _{\mathbb {C}}\), with complex difference quotients \(t^{-1}(f(x + ty) - f(x))\), \(t \in \mathbb {C}\).

If \(F\) is Mackey complete, they are also equivalent to: \(f\) is \(C^1_{\mathbb {C}}\); \(f\) is \(C^1\) and \(\mathrm{d}f(x)\) is complex linear for each \(x \in U\).

Proof ▶

The equivalence of (1)–(4) is formalized without any completeness of \(F\) and without the completion. (4) \(\Rightarrow \) (1): complex difference quotients restrict to real ones, and substituting \(t \mapsto ct\) gives complex homogeneity. (3) \(\Rightarrow \) (4): the complex derivative of \(z \mapsto f(x + zy)\) at \(0\) is \(\mathrm{d}f(x, y)\), and \(\mathrm{d}f\) is again complex analytic. For Mackey complete \(F\), a \(C^1\) map with complex linear differentials is weakly holomorphic along lines, hence Gâteaux analytic by the weak-to-strong theorem. For every \(k \geq 1\), \(C^k_{\mathbb {C}}\) maps are the \(C^k\) maps with complex linear differentials [ GN26 , Lemma 1.4.17 ] : writing \(t = su\) with \(s = |t|\) and \(|u| = 1\), the real difference quotients converge to \(\mathrm{d}f(x, uy)\) uniformly on the unit circle, and the second differential is complex linear in the direction by symmetry. (3) \(\Rightarrow \) (2) restricts the expansion to lines. (1) \(\Rightarrow \) (3): the Taylor polynomials \(\frac{1}{n!}\mathrm{d}^{(n)} f(x)(y, \dots , y)\) are continuous homogeneous polynomials over \(\mathbb {C}\) (complex multilinearity of the higher differentials from homogeneity and symmetry), and the one variable Taylor expansion along \(z \mapsto f(x + zy)\) gives the series on a balanced \(Y\) with \(x + 2Y \subseteq U\). (2) \(\Rightarrow \) (1): \(\mathrm{d}f\) is a locally uniform limit of difference quotients, by seminorm estimates of Taylor remainders along lines; it is Gâteaux analytic on \(U \times E\) by a two variable Osgood lemma, and a discrete polarization over roots of unity writes its Taylor coefficients as finite combinations of coefficients of \(f\) along lines, so they lie in \(F\). Induction on \(k\) gives smoothness.

Let \(E\), \(F\) be complex normed spaces and \(U \subseteq E\) open. A complex analytic map \(f \colon U \to F\) is complex Fréchet differentiable; if \(F\) is complete, the converse holds. In particular, if \(F\) is complete, a locally uniform limit of complex differentiable maps on \(U\) is complex differentiable. If \(F\) is complete, complex differentiable maps on \(U\) are \(C^{1}\) over \(\mathbb {C}\).

Proof ▶

A complex analytic map is real Fréchet smooth (Theorem 9.8 and Proposition 2.3.17 of [ GN26 ] ), with complex linear derivative. Conversely, along each complex line a complex differentiable map is holomorphic in one variable, hence complex analytic along lines; being continuous, it is complex analytic by Theorem 9.8. For limits, the restrictions to complex lines are locally uniform limits of holomorphic functions of one variable, which are holomorphic, and the limit is continuous.

Remark 9.10
#

A word of caution on the hypotheses. No completeness is needed for the equivalence of (1)–(3): in the proof of (2) \(\Rightarrow \) (1), the Cauchy integral \(\mathrm{d}f(x, y) = \frac{1}{2\pi i} \oint _{|\zeta | = 1} f(x + \zeta y)\, \zeta ^{-2}\, d\zeta \) is a weak integral which exists in \(F\), because along the line it is a coefficient of a power series with coefficients in \(F\). Completeness only enters when one starts from a \(C^1\) map, whose higher derivatives are a priori only in the completion. Since (1) is stated with the real calculus, we do not need a separate complex Michal–Bastiani calculus: a complex \(C^k\) map is the same as a real \(C^k\) map with complex linear differentials.

9.4 Real analytic maps

Definition 9.11 Complexification [ GN26 , Def. 2.2.1 ]
✓
#

For a real topological vector space \(E\), \(E_{\mathbb {C}} := E \times E\) with the product topology and \((x + iy)(u, v) := (xu - yv, xv + yu)\). It is locally convex if \(E\) is, and \(E = E \times \{ 0\} \).

Definition 9.12 Real analytic maps; Milnor [ Mil84 ] , [ Glo02 , Def. 2.3 ] , [ GN26 , Def. 2.2.2 ]
✓

Let \(E\), \(F\) be real locally convex spaces and \(U \subseteq E\) open. A map \(f \colon U \to F\) is real analytic if it extends to a complex analytic map \(V \to F_{\mathbb {C}}\) on an open neighborhood \(V\) of \(U\) in \(E_{\mathbb {C}}\).

Definition 9.13 Analytic maps in the sense of Bochnak and Siciak [ BS71b , Def. 5.6 ]
✓
#

\(f \colon U \to F\) is BS analytic if it is continuous and for every \(x \in U\) there are continuous homogeneous polynomials \(p_k\) of degree \(k\) such that \(f(x + y) = \sum _{k \ge 0} p_k(y)\), as a pointwise limit, for all \(y\) in a \(0\)-neighborhood. Over \(\mathbb {C}\) this is Definition 9.6.

Compositions of complex analytic maps are complex analytic: by Theorem 9.8 this is the chain rule of the real Michal–Bastiani calculus, and no completeness is involved. In the same way complex analyticity is a local property.

If every point of \(U\) has a neighborhood \(W\) such that \(f|_{U \cap W}\) is real analytic, then \(f\) is real analytic.

Proof ▶

Each \(x \in U\) has a convex open \(0\)-neighborhood \(C_x\) with \(x + C_x \subseteq U\) and a complex analytic extension \(g_x\) of \(f\) on the convex set \(B_x := (x + C_x) + i C_x\). Two such extensions agree on \(B_x \cap B_y\). Glöckner and Neeb deduce this from the identity theorem in \(E_{\mathbb {C}}\) [ GN26 , Thm. 2.1.16(c) ] . We use one variable instead: for \(a + ib \in B_x \cap B_y\), the set \(D\) of \(z \in \mathbb {C}\) with \(a + zb \in B_x \cap B_y\) is convex and contains \(0\), \(i\) and a real interval around \(0\), on which \(g_x(a + zb) = f(a + zb) = g_y(a + zb)\). For each continuous complex linear functional \(\lambda \), the holomorphic functions \(\lambda \circ g_x(a + zb)\) and \(\lambda \circ g_y(a + zb)\) on \(D\) then agree, by the one variable identity theorem, in particular at \(z = i\); and functionals separate points. The extensions glue to a complex analytic map on \(\bigcup _x B_x\).

Compositions of real analytic maps are real analytic. Real analytic maps are smooth, and \(\mathrm{d}^{(k)} f(x, y_1, \dots , y_k) = \mathrm{d}^{(k)} g(x, y_1, \dots , y_k)\) for any complex analytic extension \(g\), \(x \in U\) and \(y_i \in E\). A real analytic map is locally the sum of its Taylor series \(\sum _k \frac{1}{k!} \mathrm{d}^{(k)} f(x)(y, \dots , y)\); in particular it is BS analytic.

Proof ▶

Composition: compose the extensions, after shrinking the domain of the inner one. Smoothness: \(f = \mathrm{Re} \circ g \circ \iota \) on \(U\), with \(\iota \colon E \to E_{\mathbb {C}}\) and \(\mathrm{Re} \colon F_{\mathbb {C}} \to F\) continuous and real linear (Glöckner and Neeb argue instead that \(F\) is closed in \(F_{\mathbb {C}}\)). The derivatives of \(g \circ \iota \) at \(x\) are those of \(g\) in real directions, by an identity between difference quotients. The Taylor expansion is that of \(g\) (Theorem 9.8), restricted to real points.

Pairs, sums, scalar multiples and finite sums of complex (resp. real) analytic maps are complex (resp. real) analytic, and so is \(x \mapsto \beta (f(x), g(x))\) for a jointly continuous complex (resp. real) bilinear map \(\beta \).

Proof ▶

In the complex case, by Theorem 9.8 it suffices to see that the maps are smooth with complex linear differentials, which follows from the rules of Section 3.3. In the real case, \((E \times F)_{\mathbb {C}} \cong E_{\mathbb {C}} \times F_{\mathbb {C}}\) as complex locally convex spaces, so the pair of two complex analytic extensions extends the pair; for \(\beta \) use its complexification \(\beta _{\mathbb {C}}(a + ia', b + ib') = \beta (a, b) - \beta (a', b') + i(\beta (a, b') + \beta (a', b))\), which is jointly continuous and complex bilinear.

Theorem 9.17 The differential of a real analytic map
✓

If \(f \colon U \to F\) is real analytic on an open set \(U \subseteq E\), then \(\mathrm{d}f \colon U \times E \to F\) is real analytic. (Glöckner and Neeb use this, for \(r = \omega \), in [ GN26 , Lemma 3.3.8(d) ] ; the complex case is [ Glo02 , Lemma 2.6 ] .)

Proof ▶

Let \(g\) be a complex analytic extension of \(f\) on \(V \subseteq E_{\mathbb {C}}\). Then \(\mathrm{d}g\) is complex analytic on \(V \times E_{\mathbb {C}}\) (Theorem 9.8), and composed with \((E \times E)_{\mathbb {C}} \cong E_{\mathbb {C}} \times E_{\mathbb {C}}\) it extends \(\mathrm{d}f\), since \(\mathrm{d}g(x, y) = \mathrm{d}f(x, y)\) for \(x \in U\), \(y \in E\) (Theorem 9.15).

Proposition 9.18 Complex analytic maps are real analytic [ GN26 , Prop. 2.2.11 ]
✓

Let \(E\), \(F\) be complex locally convex spaces. Every complex analytic map \(f \colon U \to F\) is real analytic, as a map between the underlying real spaces.

Proof ▶

Let \(\pi _\pm (u + iv) := u \pm J v\) on \(E_{\mathbb {C}} = E \oplus iE\), where \(J\) is multiplication by \(i\) in \(E\), and \(\alpha (y) := \frac12 (y - i J y)\), \(\beta (y) := \frac12 (y + i J y)\) from \(F\) to \(F_{\mathbb {C}}\). Then \(\pi _+\) and \(\alpha \) are complex linear, \(\pi _-\) and \(\beta \) conjugate linear, and \(g := \alpha \circ f \circ \pi _+ + \beta \circ f \circ \pi _-\) is smooth on \(\pi _+^{-1}(U) \cap \pi _-^{-1}(U)\) with complex linear differentials, hence complex analytic by Theorem 9.8. Since \(\pi _\pm (x) = x\) and \(\alpha (y) + \beta (y) = y\) for real points, \(g\) extends \(f\). This is the extension of Glöckner and Neeb, \(\psi ^{-1} \circ (f \times \bar f) \circ \varphi \), written without the conjugate spaces \(\bar E\), \(\bar F\).

Let \(E\) be a real normed space. Then \(\| u + iv\| _{\mathbb {C}} := \sup _{|w| = 1} \| \mathrm{Re}(w(u + iv))\| \) is a complex norm on \(E_{\mathbb {C}}\) with \(\max (\| u\| , \| v\| ) \le \| u + iv\| _{\mathbb {C}} \le \| u\| + \| v\| \); it extends the norm of \(E\), defines the product topology, and is complete if \(E\) is.

Proof ▶

The norm is induced by the injective map \(p \mapsto (w \mapsto \mathrm{Re}(wp))\) into \(C(S^1, E)\), and \(\mathrm{Re}(w c p) = |c| \, \mathrm{Re}(w (c/|c|) p)\) gives \(\| cp\| _{\mathbb {C}} \le |c| \| p\| _{\mathbb {C}}\). The bounds follow from \(w = 1\), \(w = -i\) and \(|\mathrm{Re}\, w|, |\mathrm{Im}\, w| \le 1\). This is Taylor’s complexification of the norm, the smallest reasonable complexification norm (Muñoz, Sarantopoulos and Tonge, Studia Math. 134 (1999), Proposition 3); Glöckner and Neeb use the largest complex norm extending \(\| \cdot \| \) instead [ GN26 , Exercise 2.3.7 ] .

Continuous linear maps and constants are real analytic.

Let \(A\) be a continuous inverse algebra: a unital locally convex algebra whose unit group \(A^\times \) is open, with inversion \(\iota \) continuous on it (complete normed algebras are examples). Then \(\iota \) is smooth with \(\mathrm{d}\iota (x, h) = -x^{-1} h x^{-1}\); it is complex analytic if \(A\) is complex, and real analytic if \(A\) is real. No completeness is assumed.

Proof ▶

From \(\iota (x + th) - \iota (x) = -t\, \iota (x + th) h \iota (x)\) we get the derivative, and \(\mathrm{d}\iota (x, h) = -\iota (x) h \iota (x)\) is \(C^k\) when \(\iota \) is (product rule), so \(\iota \) is smooth. In the complex case \(\mathrm{d}\iota (x, \cdot )\) is complex linear, hence \(\iota \) is complex analytic by Theorem 9.8. For real \(A\), the complexification \(A_{\mathbb {C}}\) with \((x_1 + iy_1)(x_2 + iy_2) = (x_1x_2 - y_1y_2) + i(x_1y_2 + y_1x_2)\) is a complex continuous inverse algebra: near \(1\), \((a + ib)^{-1} = (1 + c^2)^{-1}(1 - ic)a^{-1}\) with \(c = a^{-1}b\) depends continuously on \((a, b)\), and units forming a neighbourhood of \(1\), with inversion continuous at \(1\), suffice by translation. Inversion in \(A_{\mathbb {C}}\) extends \(\iota \).

Let \(U\) be open and connected, and \(f\), \(g\) complex (resp. real) analytic on \(U\). Then \(f = g\) if (a) they agree on a nonempty open subset, or (b) \(\mathrm{d}^{(n)} f(x, y, \dots , y) = \mathrm{d}^{(n)} g(x, y, \dots , y)\) for all \(n\), \(y\) and some \(x \in U\), or, in the complex case, (c) they agree on \((x + H) \cap U\) for a real subspace \(H\) with \(H + iH = E\). In particular, two complex analytic maps on a connected open subset of \(E_{\mathbb {C}}\) which agree at its real points agree. No completeness is needed.

Proof ▶

The set of points near which \(f = g\) is open; it contains \(x\) under (b), by the Taylor expansions (Theorems 9.8 and 9.15); it is relatively closed in \(U\), since the differentials are continuous and vanish on it. Glöckner and Neeb treat the real case through a complex extension; the real Taylor expansion suffices. (a) implies (b). Under (c), for \(y \in H\) the functions \(z \mapsto \ell (f(x + zy))\) and \(\ell (g(x + zy))\) agree on a real interval, hence near \(0\) by the one variable identity theorem, so the differentials agree on the diagonal of \(H^n\); real polarization and complex multilinearity give (b).

Proposition 9.22 Complex extensions with values in a complex space [ GN26 , Prop. 2.2.12 ]
✓

A real analytic map \(f \colon U \to F\) into a complex locally convex space extends to a complex analytic map \(V \to F\) on an open set \(V \subseteq E_{\mathbb {C}}\) containing \(U\).

Proof ▶

Compose an extension \(V \to F_{\mathbb {C}}\) with the complex linear map \(F_{\mathbb {C}} \to F\), \(v + iw \mapsto v + iw\).

Let \(F_1\), \(F_2\) be normed spaces and \(g \colon E \to \mathcal{L}(F_1, F_2)\), with \(\mathcal{L}(F_1, F_2)\) carrying the operator norm.

  1. If \(E\) is a complex locally convex space, \(F_1\), \(F_2\) are complex, \(U \subseteq E\) is open and \((y, \xi ) \mapsto g(y)\xi \) is complex analytic on \(U \times F_1\), then \(g \colon U \to \mathcal{L}_{\mathbb {C}}(F_1, F_2)\) is complex analytic.

  2. If \(E\) is a real locally convex space, \(F_1\), \(F_2\) are real, \(U \subseteq E\) is any subset and \((y, \xi ) \mapsto g(y)\xi \) is real analytic on \(U \times F_1\), then \(g \colon U \to \mathcal{L}(F_1, F_2)\) is real analytic. Iterating, if \((y, \xi , \eta ) \mapsto \Omega (y)\xi \eta \) is real analytic on \(U \times F_1 \times F_2\), then \(y \mapsto \Omega (y) \in \mathcal{L}(F_1, \mathcal{L}(F_2, F_3))\) is real analytic (e.g. the local representatives of a differential \(2\)-form).

No completeness is assumed.

Proof ▶

Core step. Let \(Z\) be complex locally convex, \(F_1\) real normed, \(F_2\) complex normed and \(K \colon Z \to \mathcal{L}_{\mathbb {R}}(F_1, F_2)\) such that \(\hat K(z, \xi ) = K(z)\xi \) is smooth on \(U \times F_1\) with \(\partial _{(ch, 0)}\hat K = c\, \partial _{(h, 0)}\hat K\) for \(c \in \mathbb {C}\). By Theorem 8.4, \(K\) is smooth for the operator norm and \(\mathrm{d}K(z, h)\xi = \partial _{(h, 0)} \hat K(z, \xi )\), so \(\mathrm{d}K(z, \cdot )\) is complex linear and \(K\) is complex analytic by Theorem 9.8.

(1) Apply the core step to \(g(y)\) viewed as a real operator, and compose with the continuous complex linear projection \(P T = \frac12(T - iTi)\) of \(\mathcal{L}_{\mathbb {R}}(F_1, F_2)\) onto \(\mathcal{L}_{\mathbb {C}}(F_1, F_2)\), which fixes complex linear operators.

(2) Let \(H\) be a complex analytic extension of \(\hat g\) on \(V \subseteq (E \times F_1)_{\mathbb {C}}\). Since \(V\) need not contain \(W \times F_1\) for an open \(W \ni y\), we use the differential of \(H\) at \(\xi = 0\): \(K(z)\xi := \mathrm{d}H((z, 0), (0, \xi ))\) for \(z \in W := \{ z \mid (z, 0) \in V\} \) and \(\xi \in F_1\). Then \(K(z) \in \mathcal{L}_{\mathbb {R}}(F_1, (F_2)_{\mathbb {C}})\), with the norm of Lemma 9.19, and the core step applies since \(\mathrm{d}H\) is complex analytic (Theorem 9.17). For \(y \in U\) the line \(t \mapsto (y, t\xi )\) consists of real points, where \(H\) agrees with \(\hat g\), so \(K(y)\xi = g(y)\xi + i0\). The map \(\mathcal{L}_{\mathbb {R}}(F_1, (F_2)_{\mathbb {C}}) \to \mathcal{L}(F_1, F_2)_{\mathbb {C}}\), \(T \mapsto \mathrm{Re} \circ T + i\, \mathrm{Im} \circ T\), is continuous and complex linear, and carries \(K\) to a complex analytic extension of \(g\). The bilinear statement follows by applying (2) to \((y, \xi ) \mapsto \Omega (y)\xi \) and then to \(\Omega \). We do not claim this argument is new.

Remark 9.24 Two notions of real analyticity
#

Real analytic maps are BS analytic (Theorem 9.15). The converse holds when \(E\) is a Fréchet space [ GN26 , Rem. 2.2.7 ] , and more generally when \(E_{\mathbb {C}}\) is a Baire space [ BS71b , Thm. 7.1 ] , [ Glo02 , 2.9 ] ; see Theorem 9.36. For BS analyticity, closedness under composition is only known when the spaces are Baire [ BS71b , Thm. 7.3 ] ; this is why the theory of analytic manifolds and Lie groups uses Milnor’s definition.

9.5 Series of homogeneous polynomials

Where does the sum of a series \(\sum _n p_n\) of continuous homogeneous polynomials, converging pointwise on an open set, define an analytic map? Bochnak and Siciak [ BS71b , § 5 ] assume \(F\) sequentially complete; the results below use no completeness of \(F\).

Lemma 9.25 Moving the centre of a bound
✓

Let \(p\) be a continuous homogeneous polynomial of degree \(n\) between complex vector spaces, \(q\) a seminorm on \(F\) and \(W \subseteq E\) balanced. If \(q(p(c + w)) \le M\) for \(w \in W\), then \(q(p(w)) \le M\) for \(w \in W\).

Proof ▶

With \(\omega = e^{2\pi i/(n+1)}\), expanding the symmetric \(n\)-linear map gives \(\sum _{j \le n} p(\omega ^j c + w) = (n+1)\, p(w)\), and \(p(\omega ^j c + w) = \omega ^{jn} p(c + \omega ^{-j} w)\) with \(\omega ^{-j} w \in W\). Over \(\mathbb {R}\) or \(\mathbb {C}\), [ BS71a , Lemma 2 (ii) ] gives the bound on \((2e)^{-1} W\).

Proposition 9.26 Normal convergence at zero [ BS71b , Prop. 5.2 ]
✓

Let \(E\) be a complex Baire space and \(p_n\) continuous homogeneous polynomials of degree \(n\) such that \(\sum _n p_n(a)\) converges for \(a\) in an absorbing set. For every continuous seminorm \(q\) on \(F\) and every \(\theta {\gt} 0\) there are a \(0\)-neighbourhood \(W\) and \(C\) with \(q(p_n(w)) \le C \theta ^n\) for all \(n\) and \(w \in W\). If the series converges on a \(0\)-neighbourhood, its sum is continuous at \(0\).

Proof ▶

The closed sets \(\{ x : q(p_n(x)) \le N k^n \text{ for all } n\} \) cover \(E\), so one of them has an interior point \(x_0\); Lemma 9.25 moves the bound from \(x_0 + B\) to the balanced core \(B\), and homogeneity rescales it. Continuity at \(0\) follows from the geometric tail estimate and the continuity of the partial sums.

Remark 9.27 Counterexamples
✓
#

\(E\) not Baire [ BS71b , Counterexample 5.1 ] : on the finitely supported sequences with the norm of \(\ell ^2\), \(\sum _n (n x_n)^n\) converges everywhere, but its sum is not continuous at \(0\). \(F\) not complete, one variable: in \(F = \operatorname {span}\{ e_n, (z^n)_n : |z| {\lt} 1\} \subseteq \ell ^2\), the power series \(\sum _n z^n e_n\) converges on the unit disc, but its sum has no complex derivative at \(1/2\), since \(\sum _n n 2^{1-n} e_n \notin F\). A disc of convergence need not be a disc of analyticity. The partial sums of \(\sum _n n 2^{1-n} e_n\) are Mackey–Cauchy in \(F\), so \(F\) is not Mackey complete (Definition 5.2).

Let \(g_n \colon \mathbb {C} \to F\) be continuous with \(\ell \circ g_n\) entire for every continuous functional \(\ell \), and \(q\) a continuous seminorm. If \(q(g_n) \le e^{nB}\) on a closed disc \(\bar D(z_0, R)\) and \((q(g_n(z)))_n\) is bounded at every point of the disc, then for every \(\varepsilon {\gt} 0\) we have \(q(g_n) \le e^{n\varepsilon }\) on \(\bar D(z_0, R/2)\) for \(n\) large.

Proof ▶

The functions \(v_n = n^{-1} \log ^+ q(g_n)\) satisfy the sub-mean value inequality on circles: by Hahn–Banach \(q(g_n(z)) = |\ell (g_n(z))|\) for a functional \(\ell \le q\), and Jensen’s formula gives \(\log |\ell g_n(z)| \le \) the circle average of \(\log |\ell \circ g_n|\). Integrating in polar coordinates, \(\pi s^2 v_n(z) \le \int _{B(z,s)} v_n \le \int _{B(z_0,R)} v_n\) for \(|z - z_0| \le s = R/2\). Since \(0 \le v_n \le B\) and \(v_n \to 0\) pointwise, dominated convergence makes the right-hand side small.

Bochnak and Siciak use instead Siciak’s polynomial lemma [ BS71b , § 4 ] , which rests on potential theory (Siciak’s extremal function, Fekete points); only lines are needed here.

Proposition 9.29 Bounds along lines [ BS71b , Prop. 4.1 ]
✓

If \(\sum _n p_n\) converges pointwise on an open set \(H\), then for every compact set \(K \subseteq \mathbb {C}\) with \(x + K y \subseteq H\) there is \(\theta {\lt} 1\) such that for every continuous seminorm \(q\) there is \(C\) with \(q(p_n(x + z y)) \le C \theta ^n\) for \(z \in K\). The ratio \(\theta \) does not depend on \(q\).

Proof ▶

Near \(z_0\), choose \(\lambda {\gt} 1\) with \(\lambda (x + z y) \in H\) on a disc. Then \(g_n(z) = p_n(\lambda (x + zy))\) is pointwise bounded; Baire’s theorem in \(\mathbb {C}\) gives a small disc where it is uniformly bounded, and the Cauchy estimates for polynomials (by discrete Fourier inversion over roots of unity) give the Bernstein–Walsh bound \(q(g_n) \le e^{nB}\) on the whole disc. Hartogs’ lemma with \(\varepsilon {\lt} \log \lambda \) gives \(q(p_n(x + zy)) = \lambda ^{-n} q(g_n(z)) \le (e^{\varepsilon }/\lambda )^n\) for large \(n\). Only the constant and the threshold depend on \(q\): \(\lambda \) depends on \(H\) along the line and \(\varepsilon = \tfrac 12 \log \lambda \). Cover \(K\) by finitely many discs.

Theorem 9.30 Pointwise limits of polynomials [ BS71a , Thm. 2 ]
✓

Over a Baire space, a pointwise limit of continuous homogeneous polynomials of degree \(k\) is a continuous homogeneous polynomial of degree \(k\). No completeness of \(F\) is needed.

Proof ▶

The multilinear map is the limit of the polarizations. By Baire and Lemma 9.25 the limit \(P\) is bounded on a balanced \(0\)-neighbourhood; the Cauchy estimates for \(t \mapsto P(x_0 + t h)\) then give continuity at every \(x_0\).

Let \(E\) be a complex locally convex Baire space and \(F\) a Mackey complete complex locally convex space (for instance a sequentially complete one). If a series \(\sum _n p_n\) of continuous homogeneous polynomials converges on an open set \(H\), then its sum is complex analytic on \(H\).

Proof ▶

By Proposition 9.29 the series converges locally uniformly with geometric rate on every complex line, so the sum is Gâteaux analytic, with Taylor coefficients \(T_k(h) = \sum _n (\text{coefficient of } t^k \text{ in } p_n(x + th))\). These series have geometric bounds \(C_q \theta ^n\) with \(\theta \) independent of \(q\), so their partial sums are Mackey–Cauchy and the \(T_k(h)\) lie in \(F\). By Theorem 9.30 each \(T_k\) is a continuous homogeneous polynomial, and \(f(x + h) = \sum _k T_k(h)\) near \(h = 0\), so \(f\) is continuous at \(x\) by Proposition 9.26. Theorem 9.8 concludes.

Both hypotheses are needed (Remark 9.27). Bochnak and Siciak assume \(F\) sequentially complete; Mackey completeness is strictly weaker.

9.6 Series of real homogeneous polynomials

Let \(E\), \(F\) be real locally convex spaces. A continuous real \(n\)-linear map \(A\) extends to a continuous complex \(n\)-linear map \(A_{\mathbb {C}}(z_1, \dots , z_n) = \sum _{S \subseteq \{ 1, \dots , n\} } i^{|S|} A(w_S)\), where \((w_S)_j = \operatorname {Im} z_j\) for \(j \in S\) and \(\operatorname {Re} z_j\) otherwise; the diagonal of \(A_{\mathbb {C}}\) is the complexification \(p_{\mathbb {C}}\) of the homogeneous polynomial \(p\) of \(A\).

\(A_{\mathbb {C}}\) is complex multilinear and continuous, and extends \(A\). Hence \(p_{\mathbb {C}}\) is a continuous complex homogeneous polynomial of degree \(n\) on \(E_{\mathbb {C}}\) with \(p_{\mathbb {C}}(x) = p(x)\) for \(x \in E\).

Proof ▶

Real multilinearity and continuity are clear. In the \(j\)-th slot, \(A_{\mathbb {C}}(\dots , u, \dots ) = R(\operatorname {Re} u) + i R(\operatorname {Im} u)\) with \(R\) real linear (pair \(S \not\ni j\) with \(S \cup \{ j\} \)), which commutes with multiplication by \(i\).

For a homogeneous polynomial \(p\) of degree \(n\) and any \(c\), \(v\), \(\sum _{j=0}^n (-1)^{n-j} \binom {n}{j} p(c + jv) = n!\, p(v)\). Consequently, if \((q(p_n(sx + ty)))_n\) is bounded at each point of the square \(|s|, |t| {\lt} 2\), then \(q(p_n(sx + ty)) \le M K^n\) for \(|s|, |t| \le 1\), with \(M\), \(K\) depending on \(q\).

Proof ▶

The identity is the \(n\)-th forward difference of \(t \mapsto p(c + tv)\). Baire’s theorem in \(\mathbb {R}^2\) gives a disc \(B(c, \rho )\) in the \((s, t)\)-plane where \(q \circ p_n \le m\) for all \(n\); the identity bounds \(q(p_n(v)) \le 2^n m / n!\) for \(|v| {\lt} \rho /(n+1)\), and homogeneity and \((n+1)^n / n! \le e^{n+1}\) give the claim.

The key observation replaces Siciak’s polynomial lemma. For \(x, y \in E\), the complex line \(\lambda \mapsto (x + iy) + \tfrac {\lambda }{2}(x - iy)\) meets the complex multiples of real vectors on the circle \(|\lambda | = 2\):

\[ (x + iy) + e^{i\theta }(x - iy) = 2 e^{i\theta /2} \big(\cos (\theta /2) x + \sin (\theta /2) y\big). \]

On this circle \(p_{\mathbb {C}}\) is therefore determined by the values of \(p\) at real points: \(\| p_{\mathbb {C}}(\cdots )\| = 2^n \| p(\cos (\theta /2) x + \sin (\theta /2) y)\| \).

Let \(p\) be a continuous real homogeneous polynomial of degree \(n\), \(q\) a continuous seminorm on \(F\) and \(\kappa {\gt} 0\). If \(v\) is the real or the imaginary part of \(p_{\mathbb {C}}(x + iy)\), then

\[ \log ^+ \frac{q(v)}{\kappa ^n} \le \frac{1}{2\pi } \int _0^{2\pi } \log ^+ \Big( \big(\tfrac {2}{\kappa }\big)^n q\big(p(\cos (\theta /2) x + \sin (\theta /2) y)\big) \Big) \, d\theta . \]
Proof ▶

Take a real functional \(\varphi \) with \(\varphi (v) = q(v)\) and \(|\varphi | \le q\) (Hahn–Banach) and its complexification \(\varphi _{\mathbb {C}}\), so that \(q(v) \le |\varphi _{\mathbb {C}}(p_{\mathbb {C}}(x + iy))|\). Apply the sub-mean value inequality for \(\log ^+\) of the polynomial \(\lambda \mapsto \kappa ^{-n} \varphi _{\mathbb {C}}(p_{\mathbb {C}}((x + iy) + \tfrac {\lambda }{2}(x - iy)))\) on the circle \(|\lambda | = 2\), and the identity above.

Let \(\sum _n p_n\) be a series of continuous real homogeneous polynomials such that \((q(p_n(y)))_n\) is bounded for every continuous seminorm \(q\) and every \(y\) in a \(0\)-neighborhood \(Y\). Then there is a \(0\)-neighborhood \(Z\) in \(E_{\mathbb {C}}\) such that \(q(\operatorname {Re} p_{n,\mathbb {C}}(z)), q(\operatorname {Im} p_{n,\mathbb {C}}(z)) \le C (9/10)^n\) for \(z \in Z\), with \(C\) depending on \(z\) and \(q\) only. If \(F\) is Mackey complete, the complexified series converges on \(Z\). No Baire assumption on \(E\) is needed.

Proof ▶

Choose a continuous seminorm \(N\) and \(r {\gt} 0\) with \(\{ N {\lt} r\} \subseteq Y\), and let \(Z\) be the set of \(x + iy\) with \(N(x), N(y) {\lt} r/5\). Then \(sx + ty \in Y\) for \(|s|, |t| {\lt} 2\), and \(\tfrac 52(\cos (\theta /2) x + \sin (\theta /2) y) \in Y\). Apply Lemma 9.34 with \(\kappa = 4/5\): the functions \(f_n(\theta ) = \frac1n \log ^+ \big((5/2)^n q(p_n(\cos (\theta /2) x + \sin (\theta /2) y))\big)\) tend to \(0\) pointwise, and are bounded uniformly by Lemma 9.33. By dominated convergence their averages tend to \(0\), so \(q(v_n) \le (4/5)^n (9/8)^n = (9/10)^n\) for large \(n\). The ratio does not depend on \(q\), so the partial sums are Mackey–Cauchy in \(F_{\mathbb {C}}\), which is Mackey complete with \(F\).

Bochnak and Siciak prove Proposition 9.35 with Siciak’s polynomial lemma for a real segment [ BS71b , § 4 ] : a family of polynomials which is pointwise bounded on \([-1, 1]\) is bounded by \(M \omega ^n\) on a neighborhood of \([-1, 1]\) depending only on \(\omega {\gt} 1\). Its proof needs potential theory (extremal functions of increasing unions of compact sets). The auxiliary line avoids it: the homogeneity of \(p_n\) in two real variables brings the complex points into play.

Let \(E\) be a real locally convex space such that \(E_{\mathbb {C}} = E \times E\) is a Baire space (e.g. a Fréchet space) and \(F\) a Mackey complete real locally convex space. Then a map \(f \colon U \to F\) is real analytic if and only if it is BS analytic.

Proof ▶

Around \(a \in U\), the BS series converges on a \(0\)-neighborhood, so its terms are bounded there. By Proposition 9.35 the complexified series converges on a \(0\)-neighborhood \(Z\) of \(E_{\mathbb {C}}\), and its sum is complex analytic on the interior of \(Z\) by Theorem 9.31. Translated to \(a\), it extends \(f\) near \(a\); Lemma 9.14 glues the local extensions. The converse is Theorem 9.15.

9.7 Separately holomorphic maps

Hartogs’ theorem asserts that a function of two complex variables which is holomorphic in each variable separately is continuous, hence holomorphic; no continuity is assumed. We prove the continuity for maps with values in a complex locally convex space \(F\), without completeness, following Mujica [ Muj86 , § 36 ] one continuous seminorm \(q\) at a time: continuity in \(F\) is continuity for every \(q\).

Lemma 9.37 Baire’s theorem for separately continuous maps [ Muj86 , Lemma 36.2 ]
✓

If \(g(\zeta , \sigma )\) is continuous in \(\zeta \in B(a, R)\) and in \(\sigma \in \bar B(c, S)\) separately, then \(q \circ g\) is bounded on \(B(b, \rho ) \times \bar B(c, S)\) for some ball \(B(b, \rho ) \subseteq B(a, R)\).

Proof ▶

The closed sets \(\{ \zeta : q(g(\zeta , \sigma )) \le m \text{ for } \sigma \in \bar B(c, S)\} \) cover \(B(a, R)\).

Let \(A \colon B(c, 3r) \to F\) be such that \(\ell \circ A\) is holomorphic for every continuous functional \(\ell \) with \(|\ell | \le q\), and \(q \circ A \le M\). Then \(A\) is \(q\)-Lipschitz on \(\bar B(c, r)\) with constant \(M/r\), and \(\log ^+ q(A)\) has the sub-mean value property. Hartogs’ lemma (Lemma 9.28) holds for nonnegative continuous functions on a disc with the sub-mean value property.

Proof ▶

Hahn–Banach reduces both statements to the functionals \(\ell \) dominated by \(q\): the Cauchy estimate for \((\ell \circ A)'\), and Jensen’s inequality for \(\ell \circ A\) with \(\ell \) attaining \(q\) at the centre. For Hartogs’ lemma, extend the functions continuously to \(\mathbb {C}\) (Tietze) and argue as before.

Theorem 9.39 Hartogs [ Muj86 , Lemma 36.3, Thm. 36.1 ]
✓

Let \(U \subseteq \mathbb {C}^2\) be open and \(g \colon U \to F\) complex analytic in each variable separately. Then \(g\) is continuous. No completeness of \(F\) is needed.

Proof ▶

Fix \(q\) and a bidisc of radius \(4R\) about \((a, c)\) in \(U\). By Lemma 9.37, \(q(g) \le m\) on \(B(b, \rho ) \times \bar B(c, 3R)\) with \(|b - a| {\lt} R\). Let \(A_k(\sigma )\) be the Taylor coefficients of \(g(\cdot , \sigma )\) at \(b\). The Cauchy estimates on a small circle give \(q(A_k) \le m/r_0^k\); on the circle of radius \(2R\), \((2R)^k q(A_k(\sigma )) \) is bounded for each \(\sigma \). For \(\ell \) dominated by \(q\), \(\ell \circ A_k\) is holomorphic by Osgood’s lemma, so Lemma 9.38 applies to \(v_k = k^{-1} \log ^+((2R)^k q(A_k))\): eventually \(q(A_k(\sigma )) \le (2/(3R))^k\) for \(|\sigma - c| \le R/2\). Hence the Taylor series \(\sum _k (\zeta - b)^k A_k(\sigma )\) has terms bounded by \((5/6)^k\) near \((a, c)\), and each term is \(q\)-continuous, so \(g\) is \(q\)-continuous at \((a, c)\).

Proposition 9.40 Weak holomorphy in Mackey complete spaces
✓

Let \(F\) be Mackey complete and \(f \colon U \to F\) continuous on an open set \(U \subseteq \mathbb {C}\), with \(\ell \circ f\) holomorphic for every continuous functional \(\ell \). Then \(f\) is complex analytic.

Proof ▶

With \(\omega = e^{2\pi i/N}\) and \(j {\lt} N\), the average \(S_N = N^{-1} \sum _{i{\lt}N} \omega ^{i(N-j)} f(x + \tfrac {\rho }{2}\omega ^i) \in F\) satisfies \(|\ell (S_N) - (\rho /2)^j c_j^\ell | \le 2 M_\ell 2^{-N}\), where \(c_j^\ell \) is the \(j\)-th Taylor coefficient of \(\ell \circ f\) at \(x\) and \(M_\ell \) bounds \(|\ell \circ f|\) on the circle of radius \(\rho \): only the coefficients of index \(\ge N\) alias onto \(c_j\). By Hahn–Banach, \(q(S_{N+1} - S_N) \le 4 M_q 2^{-N}\) for every continuous seminorm \(q\), with the ratio \(1/2\) independent of \(q\), so \((S_N)\) converges in \(F\). Its limit gives the coefficient \(a_j \in F\) with \(\ell (a_j) = c_j^\ell \), and the Taylor series converges in \(F\) (WeakTaylorOn.hasPowerSeriesOnDisc).

If \(F\) is Mackey complete, a map \(g \colon U \to F\) on an open set \(U \subseteq \mathbb {C}^2\) which is complex analytic in each variable separately is complex analytic.

Proof ▶

By Theorem 9.39, \(g\) is continuous. For each functional \(\ell \), \(\ell \circ g\) is continuous and separately holomorphic, so by Osgood’s lemma it has a double power series and is holomorphic along every complex line (swap the variables if the slope exceeds \(1\)). By Proposition 9.40, \(g\) is complex analytic along every complex line, i.e. Gâteaux analytic, and Theorem 9.8 concludes.

Example 9.42 Separate does not imply joint analyticity without completeness
✓
#

Let \(F \subseteq \mathbb {C}^{\mathbb {N} \times \mathbb {N}}\) (product topology) be the span of the matrices \(\alpha (n) \sigma ^m\) and \(\zeta ^n \beta (m)\), and \(G(\zeta , \sigma ) = (\zeta ^n \sigma ^m)\). Then \(G \colon \mathbb {C}^2 \to F\) is complex analytic in each variable (and continuous), but not complex analytic. In particular \(F\) is not Mackey complete.

Proof ▶

The Taylor coefficients of \(G\) in \(\zeta \) at any point, \(\binom {n}{k} \zeta _0^{n-k} \sigma ^m\), and those in \(\sigma \), lie in \(F\), and the series converge coordinatewise. The second Taylor coefficient of \(t \mapsto G(t, t)\) is \(e_{20} + e_{11} + e_{02}\). If it were a combination of matrices \(\alpha _i(n) \sigma _i^m\) and \(\zeta _j^n \beta _j(m)\), take \(P = X \prod _{\sigma _i \ne 0} (X - \sigma _i)\) and \(Q = X \prod _{\zeta _j \ne 0} (X - \zeta _j)\). The form \(x \mapsto \sum _{s, r} q_s p_r x(s, r)\) vanishes on \(\alpha (n)\sigma _i^m\) (it is \(P(\sigma _i) \sum _s q_s \alpha (s)\)) and on \(\zeta _j^n \beta (m)\), but equals \(q_1 p_1 \ne 0\) on \(e_{20} + e_{11} + e_{02}\), since \(p_0 = q_0 = 0\). The last claim follows from Theorem 9.41.

So Mackey completeness enters only through the Taylor coefficients along oblique lines, and some completeness is needed there; continuity needs none.

A Gâteaux analytic map which is locally bounded for every continuous seminorm is continuous, hence complex analytic; no completeness is needed. If \(F\) is Mackey complete, a separately complex analytic map on an open subset of \(E_1 \times E_2\) is Gâteaux analytic; if it is also locally bounded, it is complex analytic.

Proof ▶

If \(q(f) \le M\) on \(a + W\), take a balanced \(V\) with \(a + \lambda h \in W\) for \(h \in V\), \(|\lambda | \le 3\); along \(\lambda \mapsto f(a + \lambda h)\) the Cauchy estimates give \(q(f(a + t h) - f(a)) \le M |t|\) for \(|t| \le 1\) (Lemma 9.38), so \(f\) is \(q\)-continuous at \(a\). For the products, \((\zeta , \sigma ) \mapsto f(p_1 + \zeta u, p_2 + \sigma v)\) is separately analytic on an open subset of \(\mathbb {C}^2\), hence complex analytic by Theorem 9.41, and \(f(p + z(u, v))\) is its restriction to the diagonal.

Let \(E\) be a Baire space, \(f\) Gâteaux analytic on \(U\), \(q\) a continuous seminorm on \(F\), and let the \(N\)-ball of radius \(8\rho \) about \(c\) lie in \(U\). If \(q \circ f\) is bounded near \(c - e\) with \(N(e) {\lt} \rho /16\), then it is bounded near \(c\). No completeness of \(F\) is needed.

Proof ▶

Let \(T_y(k, h)\) be the coefficient of \(\lambda ^k\) in \(f(y + \lambda h)\). On each complex plane, \(f\) is continuous (Theorem 9.39), so by Osgood’s lemma \(\lambda \mapsto \ell (T_y(m, h + \lambda v))\) is a polynomial of degree \(\le m\) whose coefficients are the double Taylor coefficients \(c_{ab}\) of \((\zeta , \sigma ) \mapsto \ell f(y + \zeta h + \sigma v)\), and \(\ell (T_{y+h}(k, v)) = \sum _b c_{kb}\). Hence: if \(T_y(m, \cdot )\) is \(q\)-bounded on a ball, it is \(q\)-continuous (its coefficients along planes are bounded); and a bound on \(t_0 + V\) passes to \(V\) by averaging over roots of unity (shift lemma). With \(x = c - e\): the \(T_x(m, \cdot )\) are \(q\)-bounded on a ball by the Cauchy estimates, hence \(q\)-continuous. The maps \(Q_m(t) = (m+1)^{-1} \sum _j \omega ^{j(m+1-k)} T_x(m, e + \omega ^j t)\), \(\omega = e^{2\pi i/(m+1)}\), satisfy \(\ell (Q_m(t)) = c_{k, m-k}\), so their partial sums are \(q\)-continuous, pointwise bounded, and converge weakly to \(T_c(k, t)\). Baire’s theorem and the shift lemma bound each \(T_c(k, \cdot )\) on a ball, so they are \(q\)-continuous. A second application of Baire’s theorem to \((2^k T_c(k, \cdot ))_k\) and the shift lemma give \(q(T_c(k, v)) \le n 2^{-k}\) near \(0\), hence \(q(f(c + v)) \le 2n\).

Let \(E_1\) and \(E_1 \times E_2\) be Baire spaces, \(E_2\) first countable (e.g. Fréchet spaces), and \(F\) Mackey complete. A separately complex analytic map on an open subset of \(E_1 \times E_2\) is complex analytic.

Proof ▶

Fix \((a, b)\) and \(q\). With a countable decreasing basis \((V_n)\) at \(b\), the closed sets \(\{ x : q(f(x, y)) \le n \text{ for } y \in V_n\} \) cover a neighbourhood of \(a\), so by Baire’s theorem \(q \circ f\) is bounded near some \((x', b)\) with \(x'\) as close to \(a\) as needed. Theorem 9.44 gives a bound near \((a, b)\), and Theorem 9.43 concludes.

9.8 Comparison with Mathlib on Banach spaces

On Banach spaces, Mathlib’s AnalyticOnNhd (convergent power series in norm) should be compared with the definitions above, as Theorem 8.1 compares the Michal–Bastiani and Fréchet calculi. This is planned once the locally convex theory is in place.