Singularities of functions: a global point of view

One chooses a projectivization f_X:X\rightarrow\mathbb{A}^{\!1} of f. A theorem of Hironaka [Hironaka77] (see.3em also [L-NB-S20, Th. 6.4.5]) asserts that the underlying complex analytic space X^\mathrm{an} admits a Whitney stratification satisfying Thom’s A_{f_X} condition. In fact, in the quasi-projective context, the strata can be chosen quasi-projective and connected, so that the f_X-image of a stratum is either a point or a Zariski dense open set of \mathbb{A}^{\!1}. The construction can be made in order to ensure that U\smallsetminus f^{-1}(B) is a stratum, for some finite set B of \mathbb{A}^{\!1}. Thom’s first isotopy lemma implies that the restriction of f_X to each stratum S induces a C^\infty fibration S^\mathrm{an}\rightarrow f_X(S)^\mathrm{an}. Applying this to the stratum U\smallsetminus f^{-1}(B) yields the assertion for some (possibly non minimal) finite set B. Minimality can then obviously be achieved.

Let \mathcal{F} be an object of \operatorname{Perv}(\Delta,0;\mathbb{Z}). Assume that {}^{\scriptscriptstyle\mathrm{p}}\psi_t(\mathcal{F}) and {}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_t(\mathcal{F}) are \mathbb{Z}-free, and let us prove that \mathcal{F} is an object of \operatorname{sPerv}(\Delta,0;\mathbb{Z}). Freeness of {}^{\scriptscriptstyle\mathrm{p}}\psi_t(\mathcal{F}) as a \mathbb{Z}-module is equivalent to the property that the restriction of \mathcal{F} to the punctured disc \Delta^* is a sheaf!locally constant –locally constant sheaf of free \mathbb{Z}-modules. According to [M-S22, Prop. 10.2.49], we are reduced to checking that \mathcal{H}^0(i_0^!\mathcal{F}) is free, if i_0:\{0\}\mathrel{\lhook\joinrel\rightarrow}\Delta denotes the inclusion. The distinguished triangle [M-S22, (10.90)] reads

and since {}^{\scriptscriptstyle\mathrm{p}}\psi_t(\mathcal{F}),{}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_t(\mathcal{F}) are concentrated in degree zero, it leads to the long exact sequence

If {}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_t(\mathcal{F}) is free, then so is \mathcal{H}^0(i_0^!\mathcal{F}).

Conversely, let us assume that \mathcal{F} is an object of \operatorname{sPerv}(\Delta,0;\mathbb{Z}). Then the local system (up to a shift) \mathcal{F}|_{\Delta^*} is free, and this implies freeness of {}^{\scriptscriptstyle\mathrm{p}}\psi_t(\mathcal{F}) and therefore that of \operatorname{image}(\operatorname{Var}_t). The assumption implies that \mathcal{H}^0(i_0^!\mathcal{F}) is free, and therefore so is {}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_t(\mathcal{F}).

Since f_X is proper, we have identifications

The result follows by taking the k-th perverse cohomology of both terms and using that {}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_{t-c} is t-exact.

In view of the exact sequence of relative cohomology, it is enough to show the restriction morphism \mathrm{H}^r(U;A)\rightarrow\mathrm{H}^r(f^{-1}(\Delta_\rho);A) is an isomorphism. By pushing forward by f, we are led to showing a similar result for the hypercohomology of \boldsymbol{R} f_*A_U on \mathbb{A}^{\!1}. Let j:\mathbb{A}^{\!1}\smallsetminus\Delta_\rho\mathrel{\lhook\joinrel\rightarrow}\mathbb{A}^{\!1} denote the open inclusion. We are led to showing the vanishing of any hypercohomology space of the complex \boldsymbol{R} j_!j^{-1}(\boldsymbol{R} f_*A_U). Since j^{-1}(\boldsymbol{R} f_*A_U) has sheaf!locally constant –locally constant cohomology on \mathbb{A}^{\!1}\smallsetminus\Delta_\rho by our assumption on \rho, a simple induction reduces the question to the vanishing, for any r\in\mathbb{N}, of the cohomology \mathrm{H}^r(\mathbb{A}^{\!1}\smallsetminus\mathaccent"7017\Delta_\rho;j_!\mathcal{L}) for a locally constant sheaf \mathcal{L} of A-modules on \mathbb{A}^{\!1}\smallsetminus\Delta_\rho. By considering the radial projection \mathbb{A}^{\!1}\smallsetminus\mathaccent"7017\Delta_\rho\rightarrow\partial\Delta_\rho, this follows from the vanishing of the relative cohomology \mathrm{H}^r([\rho,+\infty),\{\rho\};A) for each r\in\mathbb{N}.

It is done by noticing that, for any constructible complex \mathcal{F} on \mathbb{A}^{\!1}, the inclusion \widetilde\Delta_\rho\subset{\widetilde{\mathbb{A}}}^{\!1}_\mathrm{mod}\, induces an isomorphism

We will consider the perverse (r-d)-cohomology of the pushforward complex \boldsymbol{R} f_*A_U and \boldsymbol{R} f_!A_U. At this point, using perverse cohomology sheaves instead of ordinary cohomology sheaves will simplify many arguments.

From Corollary 2 we deduce the expressions

which can be identified with those of Corollary 1, proving thereby Proposition 3.

Let us start with a perverse sheaf \mathcal{F} on \mathbb{A}^{\!1} with singularities on C. We will prove that the cohomology space \mathrm{H}^0_\mathrm{c}(\widetilde\Delta_\rho;\mathcal{F}) has a natural increasing filtration indexed by \{a_1,\dots,a_\ell\} such that

For the sake of simplicity, we assume that each a_i is the real part of a unique c_i\in C (otherwise, one replaces the map \operatorname{Re}:C\rightarrow\mathbb{R} by the map \operatorname{Re}+\varepsilon\operatorname{Im} with \varepsilon>0 small). Let us decompose \widetilde\Delta_\rho as \widetilde\Delta_{\rho,<a_i}\sqcup\widetilde\Delta_{\rho,\geqslant a_i}, denoting thereby the intersection of \widetilde\Delta_\rho with the open half-plane \operatorname{Re}(t)<a_i, respectively closed half-plane \operatorname{Re}(t)\geqslant a_i. These subsets have the same topology as \widetilde\Delta_\rho, and we have a long exact sequence of cohomology with compact support

We first claim that \mathrm{H}^k_\mathrm{c}(\widetilde\Delta_\rho;\mathcal{F})=0 for k\neq0. Indeed, from this exact sequence and by induction on \#C, it is enough to prove this when \#C=1. In such a case, if \widetilde\Delta_c is a small disc centered at c\in C with a closed half-circle \delta_c deleted on its boundary, we have \mathrm{H}^k_\mathrm{c}(\widetilde\Delta_c;\mathcal{F})\simeq\mathrm{H}^k_\mathrm{c}(\widetilde\Delta_\rho;\mathcal{F}). The cohomology long exact sequence associated with the open/closed decomposition \Delta_c=\widetilde\Delta_c\sqcup\delta_c is then identified with that associated with the shifted distinguished triangle of [M-S22, (10.87)]

and one concludes with the property that {}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_c\mathcal{F} is concentrated in degree zero since it is perverse on a point c (because the functor {}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_c is t-exact, see.3em Rem. 10.4.23 in loc.3pt cit.-.5pt).

It follows that (Proof) is a short exact sequence and thus \mathrm{H}^0_\mathrm{c}(\widetilde\Delta_{\rho,<a_{i+1}};\mathcal{F})\rightarrow\mathrm{H}^0_\mathrm{c}(\widetilde\Delta_\rho;\mathcal{F}) is an inclusion, defining a filtration by the formula \mathrm{H}^0_\mathrm{c}(\widetilde\Delta_\rho;\mathcal{F})_{\leqslant a_i}:=\mathrm{H}^0_\mathrm{c}(\widetilde\Delta_{\rho,<a_{i+1}};\mathcal{F}). The similar exact sequence with \widetilde\Delta_\rho replaced with \widetilde\Delta_{\rho,<a_{i+1}} shows that \mathrm{gr}_{a_i}\mathrm{H}^0_\mathrm{c}(\widetilde\Delta_\rho;\mathcal{F}) is identified with \mathrm{H}^0_\mathrm{c}(\widetilde\Delta_{\rho, [a_i,a_{i+1})};\mathcal{F}). The latter space is isomorphic to \mathrm{H}^0_\mathrm{c}(\widetilde\Delta_{c_i};\mathcal{F}), that we have already identified with {}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_{c_i}\mathcal{F}. This concludes the proof of (Proof).

Let us now come back to the statement of the proposition. The formulas of Corollary 1 lead us to consider the perverse cohomology sheaves \mathcal{F}=\raisebox{6pt}{p}\!R^{r-d}f_*({}^{\scriptscriptstyle\mathrm{p}}\!A_U) and \mathcal{F}=\raisebox{6pt}{p}\!R^{r-d}f_!({}^{\scriptscriptstyle\mathrm{p}}\!A_U), and to apply the previous result to them. However, we are faced to the problem of making the functor of vanishing cycles!functorvanishing cycles commute with pushforward by f when f is not proper. This question is solved by Proposition 1.

One can give two proofs: one by computing on \widetilde X and the other one by computing on \widetilde\mathbb{P}^1 and using the commutation of Verdier duality with proper pushforward. We will sketch the first one. The point is to compute the dualizing complex on the manifold with corners \widetilde X: one finds that it is the extension by zero to \widetilde X of the shifted constant sheaf \mathbb{Q}_U[2d]. Then the computation of the Verdier dual of the shifted sheaf \raisebox{6pt}{p}\mathbb{Q}_{\widetilde U_\mathrm{mod}\,(f)} (recall Notation 5.1.1(?)) is seen to be isomorphic to \raisebox{6pt}{p}\mathbb{Q}_{\widetilde U_\mathrm{rd}(-f)}. The result is obtained by applying Verdier duality to the cohomology of these sheaves.

The second proof of Proposition 5 would yield that, for a perverse sheaf \mathcal{F} on \mathbb{A}^{\!1} with Verdier dual \mathcal{F}^{\scriptscriptstyle\vee}, the natural pairing

is nondegenerate, where \alpha denotes any of the open inclusions \mathbb{A}^{\!1}\mathrel{\lhook\joinrel\rightarrow}{\widetilde{\mathbb{A}}}^{\!1}_\mathrm{mod}\,(t) and \mathbb{A}^{\!1}\mathrel{\lhook\joinrel\rightarrow}{\widetilde{\mathbb{A}}}^{\!1}_\mathrm{mod}\,(-t). The theory of mixed Hodge modules expresses the weight filtration W_{\scriptscriptstyle\bullet}(\mathcal{F}^{\scriptscriptstyle\vee}) in terms of W_{\scriptscriptstyle\bullet}\mathcal{F}, and this leads to the formulas of the theorem.

We start with Item ? of the definition of a Stokes filtration, and we compute the sheaf \mathrm{gr}_c\widetilde\Phi_\infty(\mathcal{F}). We set \widetilde\Delta'_c=\widetilde\Delta_\rho^{\leqslant c}\smallsetminus \widetilde\Delta_\rho^{<c}. (See.3em Figure Proof.)

From the exact sequence of sheaves

we deduce that \mathrm{gr}_c\widetilde\Phi_\infty(\mathcal{F})\simeq p_!(\widetilde\mathcal{F}_{\widetilde\Delta'_c}). An easy computation on the stalks together with the vanishing lemma 4 shows that p_!(A_{\widetilde\Delta_c\smallsetminus\widetilde\Delta'_c}\otimes\widetilde\mathcal{F})=0, leading to the isomorphism \mathrm{gr}_c\widetilde\Phi_\infty(\mathcal{F})\simeq\widetilde{}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_c\mathcal{F}.

For Item ?, one notices that, for c'\neq c, the inequality c'<_\theta c is equivalent to the inclusion \Delta_{\rho,\leqslant c',\theta}\subset \Delta_{\rho,<c,\theta}. Since proper pushforward commutes with base change, the exact sequence

together with Lemma 4, leads to the exact sequence (with an obvious notation)

hence the filtration property.

Verification of Item ? follows by induction on c with respect to the order at \theta.

Let us start with Item ?. Let I\neq S^1 be any strict open interval in S^1 and assume that a splitting of \mathscr{L}|_I exists, i.e., a morphism \lambda:\mathop{\textstyle\bigoplus}\displaylimits_c\mathrm{gr}_c\mathscr{L}|_I\stackrel{\sim}{\longrightarrow}\mathscr{L}|_I compatible with the Stokes filtrations and whose associated graded morphism is the identity. Let us set C=\{c_1,\dots,c_n\} and let \lambda_i:\mathrm{gr}_{c_i}\mathscr{L}|_I\rightarrow\mathscr{L}_{\leqslant c_i}|_I be the restriction of \lambda to \mathrm{gr}_{c_i}\mathscr{L}|_I. Another lifting can be obtained as follows. For j\in\{1,\dots,n\} we denote j<_Ii if c_j<_\theta c_i for any \theta\in I. For any i,j with j<_Ii, let \psi_{ji}:\mathrm{gr}_{c_i}\mathscr{L}|_I\rightarrow\mathrm{gr}_{c_j}\mathscr{L}|_I be any morphism of (constant) local systems of A-modules. Then \widetilde\lambda with restriction for each i given by \widetilde\lambda_i=\lambda_i+\sum_{j<_Ii}\lambda_j\circ\psi_{ji} defines another splitting on I, and any other splitting on I is obtained like this. The condition j<_Ii implies that there is no Stokes argument for the pair (c_i,c_j) in I. It follows that, if I contains one Stokes argument for each pair (c_i,c_j), a splitting on I, if it exists, is unique.

We now show that a splitting exists on any interval I which satisfies: (*) I contains at most one Stokes argument for each pair c\neq c'\in C. The proof is by induction on the number of such Stokes arguments in I. One first notes that if a splitting exists on I, it can be extended to any J\supset I provided that J contains the same Stokes argument as I does. Furthermore, the case of one Stokes argument follows from Item ? of Definition 3. Assume then that a splitting \lambda exists on I satisfying (*) with a boundary point being a Stokes argument \theta such that, if I' is a small open neighbourhood of \theta, I\cup I' still satisfies (*). We will prove that a splitting exists on I\cup I'. Let \lambda' be a splitting of \mathscr{L}_{\scriptscriptstyle\bullet} on I' and set J=I\cap I'.

Given any i and any j<_Ji, we can find \eta_{ji},\eta'_{ji}:\mathrm{gr}_{c_i}\mathscr{L}_{|J}\rightarrow\mathrm{gr}_{c_j}\mathscr{L}_{|J} such that

If j<_Ji and j\not<_Ii, then we must have j<_{I'}i. We can write

The right-hand side is a new splitting on I. In the left-hand side, we use the equality \lambda_j|_J\circ\eta_{ji}=\lambda'_j\circ\eta_{ji}+\sum_{k<_Jj}\lambda'_k|_J\circ\eta'_{kj}\circ\eta_{ji} and we find a new splitting on the left-hand side (but possibly not on the right-hand side)

By iterating this process, we get new splittings both on I' and I that coincide on J, defining thus a splitting on I\cup I'.

For Item ?, we can assume that (\mathscr{L},\mathscr{L}_{\scriptscriptstyle\bullet})|_I and \mathscr{L}',\mathscr{L}'_{\scriptscriptstyle\bullet})|_I are graded, according to Item ?, and write \lambda=(\lambda_{ji}) with \lambda_{ji}:\mathrm{gr}_{c_i}\mathscr{L}|_I\rightarrow\mathrm{gr}_{c_j}\mathscr{L}'|_I being constant and zero at any \theta\in I such that c_j<_\theta c_i. Since such a \theta exists by assumption on I for any pair c_i\neq c_j, it follows that \lambda|_I is diagonal. Then, Item ? follows easily.

By Proposition 8(?), \lambda is graded on each good interval, hence each \mathrm{gr}_c\lambda is also an isomorphism. The assertion follows from Property 3(?).

One can apply Proposition 5 for each \mathrm{e}^{-\texttt{i}\theta}f, by using that \boldsymbol{R} p_! commutes with restriction to \theta.

The equivalence of Items ? and ? follows from Corollary 4. For that of Items ? and ?, one uses that, on any good open interval I, the morphism \mathsf{P}|_I:\mathscr{L}|_I\rightarrow\iota^{-1}\mathscr{L}^{\prime{\scriptscriptstyle\vee}}|_I is graded.

For the first point, we have to show that, for each c,\theta, the pairing \mathsf{P} induces zero on \widetilde\Phi_\infty(\mathcal{F})_{\leqslant c,\theta}\otimes\widetilde\Phi_\infty(\mathcal{F}')_{<c,\theta'} and on \widetilde\Phi_\infty(\mathcal{F})_{<c,\theta}\otimes\widetilde\Phi_\infty(\mathcal{F}')_{\leqslant c,\theta'}. But this is clear since, in the first case for example,

0=– – – 0 we can compute \widetilde\Phi_\infty(\mathcal{F})_{<c,\theta'} with the set \widetilde\Delta_{\rho,\theta'}^{<c+\varepsilon\mathrm{e}^{\texttt{i}\theta}}, according to Lemma 4, 0=– – – 0 the sets \widetilde\Delta_{\rho,\theta}^{\leqslant c} and \widetilde\Delta_{\rho,\theta'}^{<c+\varepsilon\mathrm{e}^{\texttt{i}\theta}} do not intersect.

The second and third points are consequences of the results already obtained.

We apply Proposition 11 to \mathcal{F}=\raisebox{6pt}{p}\!R^{r-d}f_*{}^{\scriptscriptstyle\mathrm{p}}\!A_U and \mathcal{F}'=\raisebox{6pt}{p}\!R^{2d-r}f_!{}^{\scriptscriptstyle\mathrm{p}}\!A_U, and we identify the Stokes-filtered local systems

Setting v=u^{-1}, it is convenient to first consider the complex (\Omega^{\raisebox{1pt}{}}_U[v],\mathrm{d}+v\mathrm{d} f), which also comes equipped with an action of \partial_v defined as \mathrm{e}^{-vf}\circ \partial_v\circ\mathrm{e}^{vf}. Then \mathrm{H}^r_{\mathrm{dR}}(U,vf) is a \mathbb{C}[v]-module equipped with a compatible action of \partial_v, i.e., a module over the Weyl algebra \mathbb{C}[v]\langle\partial_v\rangle of differential operators with polynomial coefficients in the variable v. Let us regard v as the derivation \partial_t with respect to the variable t=-\partial_v. Regarded as a \mathbb{C}[t]\langle\partial_t\rangle-module, \mathrm{H}^r_{\mathrm{dR}}(U,vf) is interpreted as the (d-r)-Gauss-Manin system attached to f. The regularity theorem of Griffiths implies that it has a regular singularities at any of its singularity c\in\mathbb{A}^{\!1} and also at infinity. A standard theorem of differential equations implies that the \mathbb{C}[v]\langle\partial_v\rangle-module \mathrm{H}^r_{\mathrm{dR}}(U,vf), regarded as the Laplace transform of a \mathbb{C}[t]\langle\partial_t\rangle-module with regular singularities, has a regular singularity at v=0, an possibly irregular one at v=\infty, and no other singularity. The statement of the theorem is obtained by tensoring \mathrm{H}^r_{\mathrm{dR}}(U,vf) with \mathbb{C}[u,u^{-1}]=\mathbb{C}[v^{-1},v] over \mathbb{C}[v]. The assertion on the rank is obtained by identifying the restriction to u=1 of the free \mathbb{C}[u,u^{-1}]-module \mathrm{H}^r_{\mathrm{dR}}(U,f/u) with \mathrm{H}^r_{\mathrm{dR}}(U,f).

The results of [E-S-Y13] allow one to replace the logarithmic complexes on X with the Kontsevich complexes on \overline X, as in Theorem 5.4.3. In loc.3pt cit.-.5pt, it is also proved that the dimension of the cohomologies when restricted to any u_o\in\mathbb{C} does not depend on u_o. This leads to the \mathbb{C}[u]-freeness property. The duality property has been considered in [C-Y16].

From Proposition 12 one deduces that for each \alpha\in\mathcal{A}, the filtration F_{\mathrm{irr},\alpha}^{\raisebox{1pt}{}}\mathrm{H}_{\mathrm{dR}}^r(U,f/u) is a filtration by subbundles. For any u_o\in\mathbb{C}^*, let i_{u_o}:\{u_o\}\mathrel{\lhook\joinrel\rightarrow}\nobreak\mathbb{G}_\mathrm{m} denote the inclusion. Since F_{\mathrm{irr}}^{>\mathsf{p}}\mathrm{H}_{\mathrm{dR}}^r(U,f/u) is a subbundle of F_{\mathrm{irr}}^{>\mathsf{p}-1}\mathrm{H}_{\mathrm{dR}}^r(U,f/u), the composed morphism

is injective. Therefore, i_{u_o}^*F_{\mathrm{irr}}^{>\mathsf{p}}\mathrm{H}_{\mathrm{dR}}^r(U,f/u)\rightarrow i_{u_o}^*F_{\mathrm{irr}}^\mathsf{p}\mathrm{H}_{\mathrm{dR}}^r(U,f/u) is injective and the dimension of i_{u_o}^*\mathrm{gr}_{F_{\mathrm{irr}}}^\mathsf{p}\mathrm{H}_{\mathrm{dR}}^r(U,f/u) is the difference between the ranks of the bundles F_{\mathrm{irr}}^\mathsf{p}\mathrm{H}_{\mathrm{dR}}^r(U,f/u) and F_{\mathrm{irr}}^{>\mathsf{p}}\mathrm{H}_{\mathrm{dR}}^r(U,f/u), so is independent of u_o. It follows that \mathrm{gr}_{F_{\mathrm{irr}}}^\mathsf{p}\mathrm{H}_{\mathrm{dR}}^r(U,f/u) is \mathbb{C}[u,u^{-1}]-free.

The notion of micro-support is useful for explaining the latter property (see.3em [K-S90, Dimca04, M-S22]): Katz’ condition is equivalent to the property that the micro-support of the cone \mathcal{C} of the natural morphism \boldsymbol{R} f_!{}^{\scriptscriptstyle\mathrm{p}}\!A_U\rightarrow\boldsymbol{R} f_*{}^{\scriptscriptstyle\mathrm{p}}\!A_U is reduced to the zero section of the cotangent bundle of \mathbb{A}^{\!1}, and this, in turn, is equivalent to the property that its perverse cohomology sheaves are constant. The long exact sequence of perverse cohomology attached to the corresponding distinguished triangle, together with Artin’s vanishing result (see.3em [K-S90, Prop. 10.3.17] or [Dimca04, Prop. 5.2.13]), implies that the assertions of Proposition 15 are satisfied, and the converse holds in the same way.

It remains to identify the constant local systems. This is obtained by analyzing the perverse Leray spectral sequence

and the compact support analogue with \raisebox{6pt}{p}\!R^qf_!{}^{\scriptscriptstyle\mathrm{p}}\!A_U. Due to the vanishing and constancy statements in Items ? and ?, one obtains that it degenerates at E_2. Then for r<d, the fiber of the shifted constant sheaf \raisebox{6pt}{p}\!R^{r-d}f_*{}^{\scriptscriptstyle\mathrm{p}}\!A_U, which is isomorphic to \mathrm{H}^{-1}(\mathbb{A}^{\!1},\raisebox{6pt}{p}\!R^{r-d}f_*{}^{\scriptscriptstyle\mathrm{p}}\!A_U), is thus isomorphic to \mathrm{H}^{r-1}(U,A_U), and we have a similar argument for Item ?.

A proof for the last assertions can be found in [D-S02a] for M-tameness and in [Bibi97b] for cohomological tameness.

The last assertion follows from Corollary 3. The vanishing property and the isomorphism property are consequences of properties of the perverse sheaves described in Proposition 15, according to the expressions (Proof) and the vanishing property \mathrm{H}^k({\widetilde{\mathbb{A}}}^{\!1}_\mathrm{mod}\,;\mathcal{F})=0 for all k if \mathcal{F} is constant (Lemma 4).

The freeness property is local on S^1, so we can identify \widetilde\Phi_\infty^d(U,f;A) with \mathop{\textstyle\bigoplus}\displaylimits_{c\in C}\widetilde{}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_{f-c}{}^{\scriptscriptstyle\mathrm{p}}\!A_U. Since the Milnor fiber of f at a critical point x\in f^{-1}(c) is a bouquet of \mu(f,x) spheres, each \widetilde{}^{\scriptscriptstyle\mathrm{p}}\mkern-1mu\phi_{f-c}{}^{\scriptscriptstyle\mathrm{p}}\!A_U is locally A-free of rank \sum_{x\in f^{-1}(c)}\mu(f,x).

The first assertion follows from the behavior of spectrum with respect to duality (see.3eme.g..3em[Bibi97b, Prop. 3.3]) or also from the behavior of the irregular Hodge filtration with respect to duality provided by Theorem 5.3.6 and the identification (?). The second assertion follows similarly from the remark after Theorem 5.3.6 (another argument has been given in [Bibi97b]).

We only indicate the main ideas of the proof, showing the interplay between algebraic geometry and singularity theory.

0=– – – 0 From Kouchnirenko’s formula, \mu(f_{\boldsymbol{a}}) is independent of \boldsymbol{a}. 0=– – – 0 A first step is to show that, separately, the left-hand side and the right-hand side of (?) are independent of \boldsymbol{a} as long as no entry of \boldsymbol{a} vanishes. For the left-hand side, according to (?), one can argue with the spectrum, and one shows that the spectrum at infinity of f_{\boldsymbol{a}} is independent of \boldsymbol{a}, due to the semi-continuity property of [N-S97] and the constancy of \mu(f_{\boldsymbol{a}}). For the right-hand side, one shows that the family \mathrm{H}^d_{{\mathrm{dR}}}(U,f_{\boldsymbol{a}}/u) is an algebraic vector bundle on \mathbb{G}_\mathrm{m}\times(\mathbb{G}_\mathrm{m})^{V(\Delta^{\scriptscriptstyle\vee})} with a flat connection having regular singularities along \{0\}\times(\mathbb{G}_\mathrm{m})^{V(\Delta^{\scriptscriptstyle\vee})}. 0=– – – 0 The second step is to prove the result when a_v=1 for all v\in V(\Delta^*). We set f=f_\mathbf{1}. One knows from [Batyrev94] that X is a smooth Fano projective variety (i.e., the canonical bundle K_X is anti-ample), and (see.3em[Fulton93]), one can interpret f as the first Chern class c_1(TX) of X. In particular, by anti-ampleness of K_X, the cup product with c_1(TX) satisfies the Hard Lefschetz property on the \mathbb{Q}-Chow ring of X. Due to a formula of Borisov-Chen-Smith [B-C-S05], this Chow ring is identified with the graded ring with respect to the Newton filtration:

\mathrm{gr}_{\scriptscriptstyle\bullet}^\mathcal{N}(\mathbb{Q}[x,x^{-1}]/(\partial f).

Then one identifies multiplication by c_1(TX)=f on this ring tensored by \mathbb{C} with the action of the monodromy on \mathrm{H}^d_{\mathrm{dR}}(U,f), by an argument similar to that used by Varchenko [Varchenko81]. The Hard Lefschetz property implies then the desired equality (?).