04.10.29 · algebraic-geometry / moduli

Limit linear series (Eisenbud-Harris)

shipped3 tiersLean: none

Anchor (Master): Eisenbud-Harris 1986 *Limit linear series: basic theory* (Invent. Math. 85, 337-371); Eisenbud-Harris 1987 *The Kodaira dimension of the moduli space of curves of genus $\geq 23$* (Invent. Math. 90, 359-387); Griffiths-Harris 1980 *On the variety of special linear systems on a general algebraic curve* (Duke Math. J. 47, 233-272); Kempf 1971 *Schubert methods with an application to algebraic curves* (Math. Centrum Amsterdam); Kleiman-Laksov 1972 *On the existence of special divisors* (Amer. J. Math. 94, 431-436); Harris-Morrison *Moduli of Curves* (Springer GTM 187, 1998) Ch. 5; Osserman 2006 *A limit linear series moduli scheme* (Ann. Inst. Fourier 56, 1165-1205); Esteves-Medeiros 2002 *Limit canonical systems on curves with two components* (Invent. Math. 149, 267-338); Amini-Baker 2015 *Linear series on metrized complexes of algebraic curves* (Math. Ann. 362, 55-106)

Intuition Beginner

A smooth projective curve $C$ of genus $g$ carries linear systems — choices of a line bundle $L$ of degree $d$ together with a vector subspace $V$ of its global sections of dimension $r + 1$ . The classical name for the pair $(L, V)$ is a $g_{d}^{r}$ : a " $g$ upper $r$ lower $d$ ", meaning a $d$ -degree, $r$ -dimensional linear system. Brill-Noether theory answers the central question of curve geometry: for which $(g, r, d)$ does the general curve carry a $g_{d}^{r}$ ? The answer is a single integer, the Brill-Noether number $ρ (g, r, d) = g - (r + 1) (g - d + r)$ . The general curve carries a $g_{d}^{r}$ exactly when $ρ \geq 0$ .

The Eisenbud-Harris construction of limit linear series solves a different problem: what happens to a $g_{d}^{r}$ when the smooth curve $C$ degenerates to a nodal curve $C_{0}$ ? You cannot just restrict the line bundle to the limit, because a line bundle on $C_{0}$ does not always come from a line bundle on a smooth nearby fibre. Eisenbud and Harris in 1986 found the right recipe: on a nodal curve whose components form a tree (a compact-type curve), record a separate $g_{d}^{r}$ on each component, then enforce a compatibility condition at every node that says the vanishing orders of sections on the two sides are large enough to fit together.

The recipe works. A limit linear series on the nodal curve smooths out to a $g_{d}^{r}$ on the general fibre, and the parameter space of limit linear series has the expected Brill-Noether dimension. This lets you prove statements about $g_{d}^{r}$ 's on the general curve by specialising to a carefully chosen nodal curve and counting limit linear series combinatorially.

Visual Beginner

A schematic of a nodal curve $C_{0} = C_{1} \cup_{p} C_{2}$ — two smooth components joined at a single point $p$ — with a linear system drawn on each side. Each component carries its own line bundle $L_{i}$ and section space $V_{i}$ ; at the node $p$ , the vanishing orders of sections in $V_{1}$ and $V_{2}$ have to satisfy a numerical compatibility that lets the two pieces fit together. The compatibility inequality is the central object: it is what makes the pair $(L_{1}, V_{1}), (L_{2}, V_{2})$ behave like the limit of a single linear system on a smoothing.

The picture captures the essential geometry: a linear system on the smooth fibre becomes a tuple of linear systems on the components of the limit, with a compatibility book-keeping the loss of information at the node. The whole moduli theory of curves uses degenerations of this kind, and limit linear series are the foundational technical tool for tracking linear systems through such degenerations.

Worked example Beginner

Compute the limit linear series on a genus-2 curve degenerating to two rational components, and check that smoothing recovers the hyperelliptic pencil.

Step 1. Start with a smooth projective curve $C$ of genus $g = 2$ . Such a curve carries a unique $g_{2}^{1}$ — a degree- $2$ , dimension- $1$ linear system, also called the hyperelliptic pencil. The Brill-Noether number is $ρ (2, 1, 2) = 2 - 2 \cdot (2 - 2 + 1) = 2 - 2 = 0$ , so the $g_{2}^{1}$ exists and is unique up to the projective automorphisms of the image $P^{1}$ .

Step 2. Degenerate $C$ to a nodal curve $C_{0} = C_{1} \cup_{p} C_{2}$ , where $C_{1}$ and $C_{2}$ are smooth rational curves (each a copy of $P^{1}$ ) glued at a single point $p$ . The dual graph is a single edge between two vertices, which is a tree, so $C_{0}$ is of compact type. The arithmetic genus is $g_{a} (C_{0}) = g_{a} (C_{1}) + g_{a} (C_{2}) + (number of nodes) - 1 = 0 + 0 + 1 - 1 + 1 = 2$ , matching the genus of the smooth fibre, as the deformation invariance of arithmetic genus requires.

Step 3. On each $C_{i} ≅ P^{1}$ , the only line bundle of degree $2$ is $O (2)$ , and $H^{0} (P^{1}, O (2))$ has dimension $3$ . So a $g_{2}^{1}$ on $C_{i}$ is the data of a line bundle $L_{i} = O (2)$ together with a $2$ -dimensional subspace $V_{i} \subset H^{0} (C_{i}, O (2))$ . The vanishing sequence of $V_{i}$ at the node $p$ is the strictly increasing sequence $a_{0} < a_{1}$ of orders to which sections of $V_{i}$ vanish at $p$ .

Step 4. Apply the Eisenbud-Harris compatibility at $p$ . For $r = 1$ and $d = 2$ , the inequality reads $a_{i} (V_{1}, p) + a_{1 - i} (V_{2}, p) \geq 2$ for $i = 0, 1$ . Look for refined limit linear systems — those with equality. The unique solution to $a_{0} (V_{1}, p) + a_{1} (V_{2}, p) = 2$ and $a_{1} (V_{1}, p) + a_{0} (V_{2}, p) = 2$ with $0 \leq a_{0} < a_{1} \leq 2$ on each side is $a_{∙} (V_{1}, p) = (0, 2)$ and $a_{∙} (V_{2}, p) = (0, 2)$ . So $V_{1}$ must contain a section not vanishing at $p$ and a section vanishing to order $2$ at $p$ ; the same for $V_{2}$ .

Step 5. There is exactly one such $V_{i}$ on each rational component: take $V_{i}$ to be the span of the constant section $1$ and the section $z^{2}$ in the coordinate $z$ centred at $p$ . This produces a unique limit $g_{2}^{1}$ on $C_{0}$ , matching the uniqueness of the hyperelliptic pencil on the smooth fibre.

What this tells us: the unique $g_{2}^{1}$ on a smooth genus-2 curve smooths up from a unique refined limit $g_{2}^{1}$ on the two-component compact-type curve, and the combinatorics of vanishing sequences at the node is what selects it. This is the simplest substantive example of the Eisenbud-Harris machinery, and the same logic — write down compatibility inequalities, look for refined solutions, count — drives the proof of the full Brill-Noether theorem.

Check your understanding Beginner

Exercise (easy, true-false).

True or false: a nodal curve $C_{0}$ whose dual graph contains a cycle is of compact type, and the Eisenbud-Harris definition of a limit linear series applies directly.

Hint

Compact type means the dual graph is a tree — that is, it has no cycles.

Answer

False. A nodal curve is of compact type exactly when its dual graph is a tree — no cycles. The Eisenbud-Harris 1986 definition of a limit linear series is given on compact-type curves; for curves whose dual graphs have cycles, the definition has to be extended via Osserman 2006 linked linear series or Esteves-Medeiros 2002 limit canonical systems. The compact-type restriction is what makes the Picard group of $C_{0}$ a direct product of the component Picard groups — without cycles, line bundles split component-wise.

Formal definition Intermediate+

Let $k$ be an algebraically closed field, let $C$ be a smooth projective curve over $k$ of genus $g$ , let $L$ be a line bundle on $C$ of degree $d$ , and let $V \subset H^{0} (C, L)$ be a vector subspace of dimension $r + 1$ .

Definition (linear series). A linear series of degree $d$ and dimension $r$ on $C$ , or $g_{d}^{r}$ , is the pair $(L, V)$ . The linear series is complete when $V = H^{0} (C, L)$ , and special when $h^{1} (C, L) > 0$ , equivalently when $h^{0} (C, L) > d - g + 1$ by Riemann-Roch (see 04.04.01).

Definition (vanishing sequence at a point). Given a $g_{d}^{r} = (L, V)$ on $C$ and a point $p \in C$ , the vanishing sequence of $V$ at $p$ is the strictly increasing sequence of integers $$ a_\bullet(V, p) = (a_0(V, p) < a_1(V, p) < \cdots < a_r(V, p)) $$ where $a_{i} (V, p)$ is the $(i + 1)$ -st smallest order of vanishing at $p$ achieved by a section in $V$ . Equivalently, $V \cap H^{0} (C, L (- a_{i} \cdot p)) \neq = V \cap H^{0} (C, L (- (a_{i} + 1) \cdot p))$ for each $i$ , and $a_{0} < \dots < a_{r}$ are precisely the integers at which the dimension drops.

Definition (Brill-Noether locus). The Brill-Noether locus of complete $g_{d}^{r}$ 's on $C$ is $$ W^r_d(C) = {[L] \in \mathrm{Pic}^d(C) : h^0(C, L) \geq r + 1} \subset \mathrm{Pic}^d(C). $$ It is a closed subscheme of the Picard variety, realised as the $(r + 1)$ -st degeneracy locus of a Picard-bundle evaluation map.

Definition (Brill-Noether number). The Brill-Noether number is the integer $$ \rho(g, r, d) = g - (r + 1)(g - d + r). $$ It equals the expected dimension of $W_{d}^{r} (C)$ inside $Pic^{d} (C)$ .

Definition (compact-type nodal curve). A connected reduced projective nodal curve $C_{0} = ⋃_{i} C_{i}$ over $k$ is of compact type when its dual graph — the graph with one vertex per irreducible component and one edge per node — is a tree. Equivalently, the generalised Jacobian $Pic^{0} (C_{0})$ is compact (proper as a $k$ -scheme), and decomposes as the product $\prod_{i} Pic^{0} (C_{i})$ of the Jacobians of the components.

Definition (limit linear series, Eisenbud-Harris 1986). Let $C_{0} = C_{1} \cup_{p} C_{2}$ be a compact-type nodal curve with a single node $p$ , with smooth components $C_{1}, C_{2}$ of genera $g_{1}, g_{2}$ summing to $g_{a} (C_{0}) = g_{1} + g_{2}$ . A limit $g_{d}^{r}$ on $C_{0}$ is the data of a pair of $g_{d}^{r}$ 's $(L_{1}, V_{1}), (L_{2}, V_{2})$ , one on each $C_{i}$ , satisfying the Eisenbud-Harris compatibility inequality at $p$ : $$ a_i(V_1, p) + a_{r - i}(V_2, p) \geq d \quad \text{for every } i = 0, 1, \ldots, r. $$ The limit linear series is refined when all $r + 1$ inequalities are equalities. The general definition for a multi-component compact-type curve imposes the same compatibility at every node.

Counterexamples to common slips

The compatibility inequality is not symmetric in $V_{1}$ and $V_{2}$ at a fixed index $i$ , but symmetric across the index $i \mapsto r - i$ . Setting $a_{0} (V_{1}, p) + a_{0} (V_{2}, p) \geq d$ (constant index) is the wrong condition — the right one pairs the smallest vanishing order on one side with the largest on the other.
A limit linear series is not a linear series on $C_{0}$ in the naive sense. There is no single line bundle $L$ on $C_{0}$ underlying it; the data is the pair $(L_{1}, L_{2})$ with the compatibility, and these pairs do not glue to a line bundle on $C_{0}$ in general.
The Brill-Noether number $ρ$ can be negative. Negative $ρ$ does not make $W_{d}^{r}$ empty in absolute terms — only for the general curve. Special curves (hyperelliptic, trigonal, curves on K3 surfaces, curves with extra Weierstrass points) can carry $g_{d}^{r}$ 's with $ρ < 0$ .
Refined limit linear series are dense in the moduli scheme $\overline{G}_{d}^{r} (C_{0})$ , but not every limit linear series is refined. A non-refined limit linear series fails the smoothing theorem in its full strength and requires the extension of Osserman 2006 to be smoothed.

Key theorem with proof Intermediate+

Theorem (Eisenbud-Harris smoothing theorem; Eisenbud-Harris 1986 Theorem 3.4). Let $π : C \to B$ be a one-parameter family of curves over a smooth one-dimensional base $B$ , with smooth generic fibre and special fibre $C_{0} = C_{1} \cup_{p} C_{2}$ of compact type with one node. Let $(L_{∙}, V_{∙})$ be a refined limit $g_{d}^{r}$ on $C_{0}$ . Then $(L_{∙}, V_{∙})$ extends to a $g_{d}^{r}$ on the generic fibre: there exists a line bundle $L$ on $C$ and a subbundle $\mathcal{V} \subset \pi_ \mathcal{L} $o f r ank$ r + 1 $w h oser es t r i c t i o n t o t h e g e n er i c f ib r e i s a$ g^r_d $o f d e g r ee$ d $an dd im e n s i o n$ r $, an d w h oser es t r i c t i o n t o$ C_0 $r eco n s t r u c t s$ (L_\bullet, V_\bullet)$.*

Proof. The argument has four steps. First, build a line bundle on a partial normalisation of $C$ . Second, twist by components of $C_{0}$ to adjust the bidegree. Third, identify the rank- $(r + 1)$ subbundle of sections via the vanishing-sequence data. Fourth, descend along the normalisation to land on $C$ itself.

Step 1: line bundle on the normalisation. Let $ν : C \to C$ be the normalisation along the section through the node $p$ — that is, the family obtained by separating the two components of $C_{0}$ at $p$ . The special fibre $C_{0}$ is the disjoint union $C_{1} ⊔ C_{2}$ , and the generic fibre is unchanged. On $C_{0}$ a line bundle is the same as a pair $(L_{1}, L_{2})$ . Choose a line bundle $L$ on $C$ whose restriction to $C_{0}$ is the disjoint pair $(L_{1}, L_{2})$ and whose generic fibre has degree $d$ . Such a $L$ exists because $C \to B$ is smooth and the relative Picard scheme has a section in each component-degree.

Step 2: bidegree twist. A line bundle on $C$ has a bidegree on $C_{0}$ , given by the pair $(de g L_{1}, de g L_{2})$ summing to $d$ . Twisting $L$ by a multiple of $C_{2}$ (viewed as a Cartier divisor on $C$ whose pullback to the generic fibre vanishes) shifts the bidegree by $(+ 1, - 1)$ on the special fibre; twisting by $C_{1}$ shifts by $(- 1, + 1)$ . So the bidegree of $L ∣_{C_{0}}$ may be set to any pair summing to $d$ , and the generic fibre is unchanged. Choose the twist so the bidegree matches that of the input limit linear series.

Step 3: rank- $(r + 1)$ subbundle from vanishing data. The pushforward $π_{*} L$ is a coherent sheaf on $B$ , and at the special fibre its stalk equals $H^{0} (C_{0}, L ∣_{C_{0}}) = H^{0} (C_{1}, L_{1}) \oplus H^{0} (C_{2}, L_{2})$ . The pair of vanishing sequences $a_{∙} (V_{1}, p), a_{∙} (V_{2}, p)$ identifies $r + 1$ "compatible" directions: for each $i = 0, \dots, r$ , the section in $V_{1}$ vanishing to order $a_{i} (V_{1}, p)$ at $p$ pairs with the section in $V_{2}$ vanishing to order $a_{r - i} (V_{2}, p) = d - a_{i} (V_{1}, p)$ at $p$ (using the refined-equality assumption). The compatibility is exactly what makes these section-pairs lift to sections of $L$ over a formal neighbourhood of $C_{0}$ in $C$ — the vanishing orders match the orders of the smoothing parameter $t$ along $B$ needed to glue. This produces a rank- $(r + 1)$ subsheaf $V \subset π_{*} L$ on $B$ whose fibre at the special point reconstructs the data of $V_{1}$ and $V_{2}$ encoded by the vanishing sequences, and whose generic fibre is a rank- $(r + 1)$ subspace of $H^{0}$ of the generic fibre of $L$ .

Step 4: descent along normalisation. The normalisation $ν$ is an isomorphism away from $C_{0}$ , so $L$ on $C$ has a canonical descent to a line bundle $L$ on $C$ over the generic fibre. Over $C_{0}$ , the descent requires identifying the fibres at the two preimages of the node $p$ ; the refined compatibility $a_{i} (V_{1}, p) + a_{r - i} (V_{2}, p) = d$ provides exactly the matching identification needed. The result is a line bundle $L$ on $C$ with $de g L ∣_{C_{t}} = d$ for every fibre $C_{t}$ , and a rank- $(r + 1)$ subbundle $V \subset π_{*} L$ with $V ∣_{C_{0}}$ reconstructing the limit linear series and $V ∣_{C_{t}}$ for $t$ generic giving a $g_{d}^{r}$ on the smooth fibre. $□$

Bridge. The Eisenbud-Harris smoothing theorem builds toward the modern study of Brill-Noether theory on stable curves and the geometry of $\overline{M}_{g}$ , and the central insight is that the compatibility inequality $a_{i} (V_{1}, p) + a_{r - i} (V_{2}, p) \geq d$ encodes exactly the condition for a tuple of $g_{d}^{r}$ 's on components to be the limit of a single $g_{d}^{r}$ on a smooth fibre — putting these together identifies the Brill-Noether locus on the general curve with the dense refined-limit locus on a chain-of-elliptic-curves degeneration. The bridge is: limit linear series turn analytic limit statements about linear systems on smooth curves into combinatorial enumeration statements about vanishing sequences on chains of components, and the foundational reason is that the refined compatibility equalities pin down a unique smoothing of each combinatorial datum.

This pattern appears again in 04.04.16 (Lazarsfeld's K3-vector-bundle proof of Petri), where a degeneration to a K3 surface plays the role of the compact-type chain, and in 04.10.22 (stable curves and Deligne-Mumford stability), where the compact-type boundary of $\overline{M}_{g}$ is the natural setting. The Eisenbud-Harris technique generalises the Griffiths-Harris 1980 degeneration to an irreducible nodal curve by allowing iterated nodal degenerations to chains of curves of any length, and is dual to the deformation-theoretic viewpoint of 04.10.20 (deformation theory of smooth curves): smoothings of $(C_{0}, L_{0}, V_{0})$ are controlled by the same first-order obstruction calculus that controls smoothings of $C_{0}$ alone. The construction identifies the smoothing of refined limit linear series with the existence of $g_{d}^{r}$ on the generic fibre, which is exactly what makes the Eisenbud-Harris proof of Brill-Noether work.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Let $C_{0} = C_{1} \cup_{p} C_{2}$ be a compact-type nodal curve of arithmetic genus $g$ with $C_{i}$ smooth of genus $g_{i}$ and $g_{1} + g_{2} = g$ . State and verify the Eisenbud-Harris compatibility inequality for a refined limit $g_{d}^{1}$ on $C_{0}$ , and solve it for the vanishing sequences at $p$ .

Hint

For $r = 1$ the vanishing sequences are pairs $(a_{0}, a_{1})$ with $a_{0} < a_{1}$ . Write down the two compatibility inequalities $a_{0} (V_{1}) + a_{1} (V_{2}) \geq d$ and $a_{1} (V_{1}) + a_{0} (V_{2}) \geq d$ , and impose equality.

Answer

For $r = 1$ , the vanishing sequence of $V_{i}$ at $p$ is $(a_{0} (V_{i}, p), a_{1} (V_{i}, p))$ with $a_{0} < a_{1}$ . The two refined-equality compatibilities are $a_{0} (V_{1}, p) + a_{1} (V_{2}, p) = d$ and $a_{1} (V_{1}, p) + a_{0} (V_{2}, p) = d$ . Subtracting gives $a_{1} (V_{1}, p) - a_{0} (V_{1}, p) = a_{1} (V_{2}, p) - a_{0} (V_{2}, p)$ , so both sides have the same gap $δ = a_{1} - a_{0}$ . Let $a = a_{0} (V_{1}, p)$ and $b = a_{0} (V_{2}, p)$ ; then $a + (b + δ) = d$ and $(a + δ) + b = d$ both hold automatically once $a + b + δ = d$ . So the refined data is one parameter: choose any pair $a, b \geq 0$ and $δ \geq 1$ with $a + b + δ = d$ , giving sequences $(a, a + δ)$ on side 1 and $(b, b + δ)$ on side 2. For a rational component $C_{i} ≅ P^{1}$ and $L_{i} = O (d)$ , the only sections vanishing to high order at $p$ are determined by powers of the local coordinate, so this is a discrete (combinatorial) count. Rubric: full credit for the one-parameter family description and the gap identification $δ_{1} = δ_{2}$ .

Exercise 4 (medium, symbolic).

State the Brill-Noether theorem (Kempf 1971, Kleiman-Laksov 1972, Griffiths-Harris 1980) and verify its prediction for $(g, r, d) = (5, 1, 4)$ , $(7, 2, 5)$ , $(9, 1, 5)$ .

Hint

The theorem says: for every smooth curve $C$ of genus $g$ , $W_{d}^{r} (C) \neq = \emptyset$ when $ρ \geq 0$ (Kempf, Kleiman-Laksov); for a general curve, $W_{d}^{r} (C) = \emptyset$ when $ρ < 0$ (Griffiths-Harris). Compute $ρ$ in each case.

Answer

$(5, 1, 4)$ : $ρ = 5 - 2 \cdot (5 - 4 + 1) = 5 - 4 = 1 \geq 0$ , so every genus- $5$ curve carries a $g_{4}^{1}$ , and $W_{4}^{1} (C)$ has dimension $\geq 1$ . Indeed every smooth genus- $5$ curve has a $g_{4}^{1}$ (it can be realised as a complete intersection of three quadrics in $P^{4}$ , with the $g_{4}^{1}$ given by the projection from a line on one of the quadrics).

$(7, 2, 5)$ : $ρ = 7 - 3 \cdot (7 - 5 + 2) = 7 - 12 = - 5 < 0$ , so a general genus- $7$ curve carries no $g_{5}^{2}$ . Equivalently, no general genus- $7$ curve admits a degree- $5$ map to $P^{2}$ . Special curves can; for instance, a smooth plane curve of degree $5$ has genus $(2 4) = 6$ , and curves of higher genus rarely admit low-degree plane embeddings.

$(9, 1, 5)$ : $ρ = 9 - 2 \cdot (9 - 5 + 1) = 9 - 10 = - 1 < 0$ , so a general genus- $9$ curve carries no $g_{5}^{1}$ . The gonality of a general genus- $9$ curve is therefore at least $6$ . By the Brill-Noether-Severi formula $gon (C) = ⌈(g + 2) /2 ⌉ = ⌈ 11/2 ⌉ = 6$ for a general genus- $9$ curve, consistent with the $g_{5}^{1}$ non-existence.

Rubric: full credit for the three $ρ$ values and the corresponding existence / non-existence statements.

Exercise 5 (medium, short-answer).

Show that on a compact-type curve $C_{0} = C_{1} \cup_{p} C_{2}$ , the Picard group decomposes as $Pic (C_{0}) = Pic (C_{1}) \times Pic (C_{2})$ , and explain why this decomposition fails for non-compact-type curves.

Hint

Use the normalisation sequence $0 \to O_{C_{0}}^{\times} \to O_{C_{0}}^{\times} \to k_{node}^{\times} \to 0$ and take cohomology. The connecting homomorphism vanishes when the dual graph is a tree.

Answer

Let $ν : C_{0} = C_{1} ⊔ C_{2} \to C_{0}$ be the normalisation. The exact sequence of sheaves on $C_{0}$ , $$ 0 \to \mathcal{O}{C_0}^\times \to \nu* \mathcal{O}{\widetilde{C_0}}^\times \to \mathcal{O}p^\times \to 0, $$ identifies the difference between line bundles on $C_{0}$ and pairs of line bundles on $C_{1}, C_{2}$ as a quotient by gluing data at the node $p$ . Taking sheaf cohomology gives the long exact sequence $$ H^0(\widetilde{C_0}, \mathcal{O}{\widetilde{C_0}}^\times) \to k^\times \to H^1(C_0, \mathcal{O}{C_0}^\times) \to H^1(\widetilde{C_0}, \mathcal{O}{\widetilde{C_0}}^\times) \to 0. $$ For compact type, $H^0(\widetilde{C_0}, \mathcal{O}{\widetilde{C_0}}^\times) = k^\times \oplus k^\times $s u r j ec t so n t o$ k^\times $v ia t h e d i f f er e n ce ma p (b ec a u se t h e d u a l g r a p hb e in g a t r ee m e an s t h e g l u in g co n s t r ain t s a r ee x a c tl y w ha tt h e d i f f er e n ce ma p c a pt u r es) . T h eco nn ec t in g ma pt o$ \mathrm{Pic}(C_0) $v ani s h es, an d t h ese q u e n ces h o w s$ \mathrm{Pic}(C_0) \cong \mathrm{Pic}(C_1) \times \mathrm{Pic}(C_2)$.

For a non-compact-type curve $C_{0}$ (dual graph with cycles), the difference map fails to be surjective onto $k^{\times}$ at every node simultaneously — there are cyclical constraints. The kernel of the connecting map picks up a piece of $k^{\times}$ for every independent cycle, contributing an extra $G_{m}$ -factor to $Pic^{0} (C_{0})$ . The generalised Jacobian gains a multiplicative part of rank equal to the first Betti number of the dual graph, and the product decomposition $\prod_{i} Pic (C_{i})$ fails.

Rubric: full credit for the cohomology sequence, the tree-versus-cycle distinction, and the multiplicative-factor description of the failure.

Exercise 6 (medium, short-answer).

State the Petri conjecture and explain how Eisenbud-Harris 1986 reduce its proof to a combinatorial check on a chain of elliptic curves.

Hint

Petri's conjecture: the Petri map $μ_{0} : H^{0} (C, L) \otimes H^{0} (C, K_{C} \otimes L^{- 1}) \to H^{0} (C, K_{C})$ is injective for every line bundle $L$ on a general curve $C$ . Limit linear series let you specialise $C$ to a nodal curve and check injectivity by counting.

Answer

The Petri conjecture (Petri 1923, proved by Gieseker 1982): for every smooth projective curve $C$ in a dense open subset of $M_{g}$ and every line bundle $L$ on $C$ , the Petri map $$ \mu_0(L) : H^0(C, L) \otimes H^0(C, K_C \otimes L^{-1}) \to H^0(C, K_C) $$ is injective. Equivalently, the Brill-Noether locus $W_{d}^{r} (C)$ is smooth of pure dimension $ρ$ for the general curve.

Eisenbud-Harris 1986 give an alternative proof by specialising $C$ to a chain $C_{0} = E_{1} \cup E_{2} \cup \dots \cup E_{g}$ of $g$ elliptic curves glued at general points, then computing the kernel of the limit Petri map combinatorially. The compact-type condition holds because the dual graph is a path (a tree). On each elliptic component, the Riemann-Roch and Petri behaviour is explicit: $H^{0}$ dimensions are dictated by degree alone (every line bundle of positive degree on an elliptic curve has $h^{0}$ equal to the degree, for the degrees that arise). The limit Petri map decomposes as a product of node-by-node compatibility maps, each one a tensor-product calculation in a $2$ -dimensional space, and injectivity reduces to checking that the gluing parameters at the nodes are generic. The specialisation argument transports injectivity on the chain to injectivity on a Zariski-dense open in $M_{g}$ .

Eisenbud-Harris 1983 Divisors on general curves uses a one-node version of this. The full chain-of-elliptic-curves argument appears in Eisenbud-Harris 1986 §5 and was independently realised by Welters; Lazarsfeld 1986 gives a third proof via vector bundles on K3 surfaces (see 04.04.16).

Rubric: full credit for the conjecture statement, the chain-of-elliptic-curves specialisation idea, and the reduction to a node-by-node tensor-product calculation.

Exercise 7 (hard, short-answer).

Prove that on a compact-type curve $C_{0} = C_{1} \cup_{p} C_{2}$ with components of genera $g_{1}, g_{2}$ , the dimension of the moduli scheme of refined limit $g_{d}^{r}$ 's equals $ρ (g, r, d)$ when $ρ \geq 0$ , where $g = g_{1} + g_{2}$ .

Hint

On each component, the moduli of $g_{d}^{r}$ 's with prescribed vanishing sequence $a_{∙}$ at $p$ has dimension $g_{i} - \sum_{j} (a_{j} - j)$ (the adjusted Brill-Noether number). Sum the dimensions across components and impose the refined-equality constraint at $p$ .

Answer

On the smooth component $C_{i}$ , the variety $G_{d}^{r} (C_{i}, (a_{∙}))$ parametrising $g_{d}^{r}$ 's on $C_{i}$ with vanishing sequence at $p$ at least $a_{∙} = (a_{0}, \dots, a_{r})$ has expected dimension equal to the adjusted Brill-Noether number $$ \rho(g_i, r, d, a_\bullet) = g_i - \sum_{j = 0}^{r} (a_j - j) - (r + 1)(g_i - d + r) + \sum_{j = 0}^{r} (a_j - j) = \rho(g_i, r, d) - \text{vanishing adjustment} $$ — more precisely, dimension $ρ (g_{i}, r, d) - \sum_{j} (a_{j} (V_{i}, p) - j)$ accounting for the codimension of imposing each vanishing condition.

Summing across $C_{1}$ and $C_{2}$ and using the refined equality $a_{i} (V_{1}, p) + a_{r - i} (V_{2}, p) = d$ , $$ \sum_j (a_j(V_1, p) - j) + \sum_j (a_j(V_2, p) - j) = \sum_j (a_j(V_1, p) + a_{r-j}(V_2, p)) - 2 \sum_j j = (r + 1) d - r(r + 1). $$ Therefore the dimension of the refined limit moduli is $$ \rho(g_1, r, d) + \rho(g_2, r, d) - [(r+1)d - r(r+1)] = (g_1 + g_2) - 2(r+1)(g - d + r) - \text{correction}, $$ and a careful book-keeping (Eisenbud-Harris 1986 Proposition 3.6) shows the result simplifies to $ρ (g, r, d) = g - (r + 1) (g - d + r)$ . The key cancellation is that the over-counted node contribution $(r + 1) (d - r)$ cancels exactly against the adjustment from the vanishing sequences.

Rubric: full credit for the adjusted Brill-Noether dimension on each component, the refined-equality identity $\sum_{j} (a_{j} (V_{1}, p) + a_{r - j} (V_{2}, p)) = (r + 1) d$ , and the final simplification to $ρ (g, r, d)$ .

Exercise 8 (hard, short-answer).

Outline the Eisenbud-Harris 1986 proof of the Brill-Noether existence theorem (Kempf, Kleiman-Laksov) by specialising to a chain of elliptic curves.

Hint

Specialise $C$ to a chain $E_{1} \cup E_{2} \cup \dots \cup E_{g}$ of $g$ elliptic curves. On each $E_{i}$ , a refined limit $g_{d}^{r}$ is constrained by its vanishing sequences at the two adjacent nodes. Count the number of compatible vanishing sequences combinatorially and check that the count is positive whenever $ρ \geq 0$ .

Answer

The Eisenbud-Harris 1986 proof of Brill-Noether existence proceeds by specialisation and combinatorial counting:

Specialisation. Consider a flat family $π : C \to B$ over a smooth one-dimensional base $B$ with smooth generic fibre $C$ of genus $g$ and special fibre $C_{0} = E_{1} \cup E_{2} \cup \dots \cup E_{g}$ a chain of $g$ general elliptic curves glued at general points. The dual graph of $C_{0}$ is a path, hence a tree, so $C_{0}$ is of compact type.
Local existence on each $E_{i}$ . A $g_{d}^{r}$ on an elliptic curve $E$ with prescribed vanishing sequences $a_{∙}, b_{∙}$ at the two attaching nodes $p, q$ exists when the line bundle $L \in Pic^{d} (E)$ is forced by Riemann-Roch and the constraints to lie in a non-empty fibre of $E^{r + 1} \to Pic^{d} (E)$ . On a general $E$ with generic gluing points, this happens whenever an explicit combinatorial inequality holds.
Compatibility along the chain. Refined compatibility at each interior node forces a sequence of "vanishing-sequence transitions" along the chain. Encode each transition combinatorially: at the node $p_{i} = E_{i} \cap E_{i + 1}$ , the vanishing sequence on the right of $E_{i}$ and on the left of $E_{i + 1}$ satisfy refined equality. This produces a lattice-path combinatorial count: vanishing sequences are encoded as Young diagrams inscribed in a $(r + 1) \times (d - r)$ rectangle, and refined compatibilities are encoded as moves between adjacent diagrams in the Bruhat order.
Counting paths. The number of lattice paths in the Young-diagram framework, by a Lindström-Gessel-Viennot-style argument, equals
$i = 0 \prod r \frac{i !}{( g - d + r + i )!} \cdot g!,$
which is the Castelnuovo number — the classical 19th-century count of $g_{d}^{r}$ 's on a general curve when $ρ = 0$ .
Conclusion. The count is positive exactly when each factor is well-defined, that is, when $g - d + r + i \geq 0$ for every $i = 0, \dots, r$ , which is the condition $g - d + r \geq 0$ , equivalent to $g - d + 2 r \geq r$ , and the analysis pins down $ρ \geq 0$ . By Zariski-density of the chain inside its smoothing component, the existence of refined limit $g_{d}^{r}$ 's on $C_{0}$ propagates to existence of $g_{d}^{r}$ 's on the general fibre.

The proof simultaneously gives existence (the chain count is positive when $ρ \geq 0$ ) and the dimension count (the count of $ρ = 0$ instances matches the Castelnuovo formula) and the non-existence for $ρ < 0$ (the chain count is zero, and by upper semi-continuity the smooth-fibre count is zero on the general fibre).

Rubric: full credit for (a) the chain-of-elliptic-curves specialisation, (b) the lattice-path / Young-diagram combinatorial encoding of vanishing sequences, (c) the positivity-when- $ρ \geq 0$ count, and (d) the semi-continuity propagation argument.

Advanced results Master

Theorem (Brill-Noether theorem; Kempf 1971, Kleiman-Laksov 1972, Griffiths-Harris 1980). Let $C$ be a smooth projective curve of genus $g$ over an algebraically closed field. (Existence) If $ρ (g, r, d) \geq 0$ , then $W_{d}^{r} (C) \neq = \emptyset$ and every component has dimension $\geq ρ$ . (Non-existence) If $C$ is general in $M_{g}$ and $ρ < 0$ , then $W_{d}^{r} (C) = \emptyset$ .

The existence proof of Kempf 1971 realises $W_{d}^{r} (C)$ as the $(r + 1)$ -st degeneracy locus of a Picard-bundle evaluation map and applies Fulton's intersection theory to bound the codimension. Kleiman-Laksov 1972 give an independent proof using the Porteous formula, computing the cohomology class $$ [W^r_d(C)] = \prod_{i = 0}^{r} \frac{i!}{(g - d + r + i)!} \cdot \theta^{g - \rho} $$ in $H^{*} (Pic^{d} (C); Z)$ , where $θ$ is the cohomology class of the theta divisor. Positivity of this class forces non-emptiness when $ρ \geq 0$ . The non-existence direction of Griffiths-Harris 1980 specialises $C$ to an irreducible nodal curve and uses a vanishing-order argument on the normalisation; Eisenbud-Harris 1986 strengthens this to a chain-of-elliptic-curves specialisation that proves existence and non-existence simultaneously.

Theorem (Eisenbud-Harris moduli scheme of limit linear series; Eisenbud-Harris 1986 Theorem 3.3). For a compact-type curve $C_{0}$ of arithmetic genus $g$ , the moduli functor of refined limit $g_{d}^{r}$ 's is representable by a quasi-projective scheme $\overline{G}_{d}^{r} (C_{0})$ of dimension at least $ρ (g, r, d)$ at every point, with equality on the dense locus of refined limit linear series. The scheme $\overline{G}_{d}^{r} (C_{0})$ fits into a family $\overline{G}_{d}^{r} \to \overline{M}_{g}$ extending the usual Brill-Noether moduli over the compact-type locus, and the smoothing theorem identifies a dense open of $\overline{G}_{d}^{r} (C_{0})$ with limit fibres of $G_{d}^{r} (C_{t})$ for $C_{t} \to C_{0}$ a smoothing.

Eisenbud-Harris construct $\overline{G}_{d}^{r} (C_{0})$ as a closed subscheme of $\prod_{i} G_{d}^{r} (C_{i})$ cut out by the compatibility inequalities at every node, with the refined locus an open subscheme. Osserman 2006 gives an alternative construction via linked Grassmannians that extends to non-compact-type curves, producing a relative moduli scheme over $\overline{M}_{g}$ whose fibres on the compact-type locus match the Eisenbud-Harris construction.

Theorem (Eisenbud-Harris-Harris-Mumford on Kodaira dimension of $\overline{M}_{g}$ ; Eisenbud-Harris 1987 + Harris-Mumford 1982). For $g \geq 24$ , the moduli space $\overline{M}_{g}$ is of general type — its Kodaira dimension equals $dim \overline{M}_{g} = 3 g - 3$ .

The proof uses Brill-Noether divisor classes on $\overline{M}_{g}$ constructed from limit-linear-series data, computes their slopes (the ratio of $K_{\overline{M}_{g}}$ -coefficient to boundary-coefficient in the Néron-Severi group), and shows that the slope of the Brill-Noether divisor stays below the slope of the canonical divisor for $g \geq 24$ . Harris-Mumford 1982 had previously proved this for $g$ odd; Eisenbud-Harris 1987 extends to the even case and to $g = 24$ . The cases $g = 22, 23$ were resolved by Farkas in 2009.

Theorem (refined Brill-Noether for special curves; Pflueger 2017, Jensen-Ranganathan 2021). For curves of fixed gonality $k$ — that is, curves admitting a degree- $k$ map to $P^{1}$ — the Brill-Noether locus $W_{d}^{r} (C)$ has dimension equal to the refined Brill-Noether number $$ \rho_k(g, r, d) = \max_{\ell = 0, \ldots, r'} \rho(g, r - \ell, d) - \ell k $$ where $r^{'} = min (r, g - d + r)$ , for the general $k$ -gonal curve.

Pflueger 2017 conjectured the formula via tropical methods; Jensen-Ranganathan 2021 proved it using a refined chain-of-loops degeneration and limit-linear-series combinatorics. The refined Brill-Noether number quantifies how the gonality stratification of $M_{g}$ corrects the Brill-Noether dimension count, and reduces to the classical $ρ$ for curves of maximal gonality.

Theorem (Amini-Baker specialisation; Amini-Baker 2015). Let $C$ be a smooth curve over a complete discretely valued field $K$ with residue field $k$ , with semistable reduction $C$ whose special fibre is a metrised complex of algebraic curves $Σ$ . For every $g_{d}^{r}$ on $C$ , the specialisation gives a divisor of rank $\geq r$ on $Σ$ in the Baker-Norine sense, and the resulting specialisation map from $g_{d}^{r}$ 's on $C$ to combinatorial divisors on $Σ$ recovers and extends Eisenbud-Harris in the compact-type case.

Amini-Baker 2015 unify Eisenbud-Harris compact-type limit linear series with Baker's specialisation theorem on tropical curves by interpreting both as instances of a single divisor theory on metrised complexes — a hybrid object that records both the algebraic data on each component (line bundles, sections) and the combinatorial data of the dual graph (edge lengths, gluing). The Berkovich-analytic viewpoint identifies the metrised complex with the skeleton of the Berkovich analytification of $C$ , and the specialisation map factors through the Berkovich retraction.

Theorem (Esteves-Medeiros for non-compact-type curves; Esteves-Medeiros 2002). For a curve $C_{0}$ with two components meeting at multiple points (not of compact type), there is an extended notion of limit canonical system whose moduli is a positive-dimensional closed subscheme of $\prod_{i} G_{d}^{r} (C_{i})$ extending the Eisenbud-Harris construction. The smoothing theorem in this setting requires a stronger compatibility — the vanishing sequences must satisfy a glueing condition at every node, not just the per-node inequality.

The Esteves-Medeiros 2002 construction was the first systematic extension of limit linear series outside compact type, and inspired the linked-Grassmannian framework of Osserman 2006 that gives a uniform moduli scheme. The construction is foundational to modern Brill-Noether theory on $\overline{M}_{g}$ and to refined and tropical Brill-Noether.

Synthesis. The Eisenbud-Harris construction of limit linear series is the foundational technical tool for tracking linear systems through degenerations of smooth curves to compact-type nodal curves, and the central insight is that the analytic question "what is the limit of a $g_{d}^{r}$ as $C \to C_{0}$ ?" reduces to the combinatorial book-keeping of vanishing sequences satisfying the compatibility $a_{i} (V_{1}, p) + a_{r - i} (V_{2}, p) \geq d$ at each node. Three apparently distinct constructions — the line-bundle bidegree on a partial normalisation, the rank- $(r + 1)$ subbundle of sections selected by vanishing-order matching, and the descent along the normalisation morphism — fit into one smoothing theorem. Putting these together, limit linear series turn the Brill-Noether existence question into a lattice-path count on a chain of elliptic curves, turn the Petri conjecture into a node-by-node tensor-product injectivity check, and turn the question of the Kodaira dimension of $\overline{M}_{g}$ into a slope-of-divisor calculation using Brill-Noether divisor classes on the moduli. The bridge is: limit linear series identify the analytic limit of a $g_{d}^{r}$ with a combinatorial datum of vanishing sequences, and this is exactly what makes the Eisenbud-Harris-Harris-Mumford machine compute slopes of Brill-Noether divisors and prove $\overline{M}_{g}$ is general type for $g \geq 24$ . The foundational reason is that refined compatibility equalities pin down a unique smoothing for each combinatorial datum, so the count of refined limit linear series on a chain equals the count of $g_{d}^{r}$ 's on the smoothed-out general curve.

The construction also generalises in three directions. To non-compact-type curves, the Esteves-Medeiros 2002 linked-Picard-bundle formalism and the Osserman 2006 linked-Grassmannian moduli scheme produce a uniform construction valid on every stable curve, and through this the foundational tool of the modern Brill-Noether theory on $\overline{M}_{g}$ . To metrised complexes and tropical curves, the Amini-Baker 2015 specialisation theorem identifies limit linear series with divisors on metrised complexes in the Baker-Norine sense, and through this with the tropical Riemann-Roch theorem of Baker-Norine 2007 on chip-firing-equivalence classes. To refined Brill-Noether on special curves, the Pflueger-Jensen-Ranganathan formula $ρ_{k}$ corrects the classical $ρ$ by the gonality stratification, and limit linear series on chains of loops give the proof. The Eisenbud-Harris technique thus extends from the original 1986 compact-type definition to the modern foundational tool for studying linear series in degenerative geometry, and identifies the algebraic geometry of $\overline{M}_{g}$ with the combinatorics of metrised complex divisor theory.

The synthesis is structural: every modern result in the geometry of $\overline{M}_{g}$ — Eisenbud-Harris on Kodaira dimension, Farkas on slopes of effective divisors, the Pflueger-Jensen-Ranganathan refined Brill-Noether, Amini-Baker tropical specialisation — uses limit linear series as the bridge between analytic limit statements and combinatorial counting statements.

Full proof set Master

Theorem (Eisenbud-Harris smoothing theorem), proof. Given in the Intermediate-tier section: a refined limit $g_{d}^{r}$ on a compact-type curve $C_{0} = C_{1} \cup_{p} C_{2}$ lifts to a $g_{d}^{r}$ on the generic fibre of a smoothing $C \to B$ via the four-step construction (line bundle on the normalisation, bidegree twist, vanishing-sequence subbundle, descent along normalisation). The refined-equality assumption $a_{i} (V_{1}, p) + a_{r - i} (V_{2}, p) = d$ provides exactly the matching identification at the node needed for descent. $□$

Proposition (compatibility inequality is necessary). Let $π : C \to B$ be a smoothing of a compact-type curve $C_{0} = C_{1} \cup_{p} C_{2}$ . If $(L, V)$ is a relative $g_{d}^{r}$ on $C$ with $V ∣_{C_{0}}$ giving the pair $(L_{1}, V_{1}), (L_{2}, V_{2})$ on the components, then the vanishing sequences satisfy $a_{i} (V_{1}, p) + a_{r - i} (V_{2}, p) \geq d$ for every $i = 0, \dots, r$ .

Proof. Let $t$ be a local coordinate on $B$ at the special point, so the smoothing parameter governs the gluing at $p$ . A section $σ$ of $V$ over a neighbourhood of the special point restricts on each $C_{i}$ to a section vanishing to some order $α_{i}$ at $p$ . Compatibility with the smoothing parameter requires $α_{1} + α_{2} = d$ when $σ$ does not vanish on a whole component (this is the local picture of the smoothing of a node: the line bundle $O_{C} (C_{2})$ has degree $d$ along the central fibre with $1$ on $C_{2}$ and $- 1$ on $C_{1}$ , and the section equation $x y = t$ near the node forces vanishing orders to sum to $d$ ).

For each $i = 0, \dots, r$ , the section of $V$ achieving the $(i + 1)$ -st smallest vanishing order at $p$ on $C_{1}$ — that is, $α_{1} = a_{i} (V_{1}, p)$ — must have $α_{2} \leq a_{r - i} (V_{2}, p)$ , since the $r + 1$ pairs $(α_{1}, α_{2})$ across a basis of $V$ have decreasing $α_{1}$ and increasing $α_{2}$ along the basis. Combining $α_{1} + α_{2} = d$ with the bounds gives $a_{i} (V_{1}, p) + a_{r - i} (V_{2}, p) \geq d$ . $□$

Proposition (Kempf-Kleiman-Laksov dimension lower bound). For every smooth projective curve $C$ of genus $g$ and every $(r, d)$ with $ρ = g - (r + 1) (g - d + r) \geq 0$ , every irreducible component of $W_{d}^{r} (C)$ has dimension $\geq ρ$ .

Proof. Realise $W_{d}^{r} (C)$ as the $(r + 1)$ -st degeneracy locus of the Picard-bundle evaluation map. Specifically, choose a Poincaré line bundle $P$ on $C \times Pic^{d} (C)$ , and consider the morphism of vector bundles on $Pic^{d} (C)$ $$ \phi : R^0 q_* (\mathcal{P}) \to R^0 q_* (\mathcal{P} \otimes \mathcal{O}{D}) $$ where $q : C \times Pic^{d} (C) \to Pic^{d} (C)$ is the projection and $D$ is a sufficiently positive effective divisor on $C$ used to trivialise $R^1 q* \mathcal{P} $. A f t er s t an d a r d co h o m o l o g i c a l r e d u c t i o n s (H a r t s h or n e I I I .12, t h e R i e mann - R oc hi d e n t i t y, ba se - c han g e f or co h o m o l o g y), t h er ank o f$ \phi $a t$ [L] \in \mathrm{Pic}^d(C) $e q u a l s$ d - g + 1 - h^1(L) $, an dd r o pp in g r ank b y$ r $cor r es p o n d s t o$ h^0(L) \geq r + 1 $, t ha t i s,$ [L] \in W^r_d(C) $. T h e$ (r+1) $- s t d e g e n er a cy l oc u so f am or p hi s mb e tw ee nb u n d l eso f r ank s$ a, b $ha se x p ec t e d co d im e n s i o n$ (a - r)(b - r) $(F u l t o n, * I n t er sec t i o n T h eor y * C h .14) . P l ug g in g in$ a = h^0(L) $f or$ L $g e n er i c,$ b = d - g + 1 $, an d t h er ank - d r o p co n d i t i o n$ h^0(L) \geq r + 1 $, t h ee x p ec t e d co d im e n s i o n o f$ W^r_d(C) $in$ \mathrm{Pic}^d(C) $i s$ (r + 1)(g - d + r) $, so i t se x p ec t e dd im e n s i o ni s$ g - (r+1)(g - d + r) = \rho(g, r, d) $. F u l t o n^{'} s g e n er a l t h eor e m o n d e g e n er a cy l oc i g i v es t h e d im e n s i o n l o w er b o u n d : e v er y co m p o n e n t ha s d im e n s i o n$ \geq \rho $w h e n$ \rho \geq 0 $.$ \square$

Proposition (Kleiman-Laksov Porteous formula for $[W_{d}^{r}]$ ). The class of $W_{d}^{r} (C)$ in $H^(\mathrm{Pic}^d(C); \mathbb{Z})$ is $$ [W^r_d(C)] = \prod_{i = 0}^{r} \frac{i!}{(g - d + r + i)!} \cdot \theta^{(r+1)(g - d + r)}, $$ where $θ$ is the cohomology class of the theta divisor and the formula is interpreted via the Porteous determinantal formula.*

Proof. The Porteous formula (see 04.04.13) computes the cohomology class of the rank-deficiency locus of a map of vector bundles as a Schur polynomial in the Chern classes of the bundles. Applied to the Picard-bundle map $ϕ$ above, the Chern class data reduces to the theta-divisor pulled back along the Abel-Jacobi map, and the Schur-polynomial expansion gives the displayed formula. The key cohomology input is Poincaré's identity $H^{*} (Pic^{d} (C); Z) = Λ^{*} H^{1} (C; Z)$ with $θ = \sum e_{i} \land e_{i}^{*}$ for a symplectic basis of $H^{1}$ . Positivity of the class — every factor $i! / (g - d + r + i)!$ is non-negative, and $θ$ is ample — proves $W_{d}^{r} (C) \neq = \emptyset$ whenever the codimension does not exceed the dimension, that is, when $ρ \geq 0$ . $□$

Proposition (worked example: genus-2 hyperelliptic pencil), proof. On a smooth genus-2 curve $C$ , the unique $g_{2}^{1}$ is $∣ K_{C} ∣$ , the canonical pencil. Specialising to $C_{0} = C_{1} \cup_{p} C_{2}$ with $C_{i} ≅ P^{1}$ , the unique refined limit $g_{2}^{1}$ on $C_{0}$ smooths to $∣ K_{C} ∣$ .

Proof. Step 1. On a smooth genus-2 curve, the canonical divisor $K_{C}$ has degree $2 g - 2 = 2$ and $h^{0} (K_{C}) = g = 2$ by Riemann-Roch, so $∣ K_{C} ∣$ is a $g_{2}^{1}$ . Uniqueness: any other $g_{2}^{1} = (L, V)$ with $de g L = 2, h^{0} (L) \geq 2$ has $h^{1} (L) = h^{0} (K_{C} \otimes L^{- 1}) \geq 1$ by Riemann-Roch, so $K_{C} \otimes L^{- 1}$ is effective of degree $0$ , hence the structure sheaf $O_{C}$ , hence $L = K_{C}$ . So the $g_{2}^{1}$ is unique on every smooth genus-2 curve.

Step 2. On $C_{0} = C_{1} \cup_{p} C_{2}$ with $C_{i} ≅ P^{1}$ , a line bundle of degree $d = 2$ on $C_{i}$ is uniquely $O_{P^{1}} (2)$ , and a $g_{2}^{1}$ on $C_{i}$ is a $2$ -dimensional subspace $V_{i} \subset H^{0} (P^{1}, O (2))$ . The space $H^{0} (P^{1}, O (2))$ is $3$ -dimensional, so $V_{i}$ is a hyperplane.

Step 3. The vanishing sequence $a_{∙} (V_{i}, p) = (a_{0}, a_{1})$ with $a_{0} < a_{1}$ . Possible sequences on a hyperplane $V_{i} \subset H^{0} (P^{1}, O (2))$ at a point $p$ : $(0, 1)$ (generic), $(0, 2)$ (one section vanishes to order $2$ at $p$ ), $(1, 2)$ ( $V_{i}$ has a base point at $p$ ). Refined compatibility $a_{0} (V_{1}) + a_{1} (V_{2}) = 2$ and $a_{1} (V_{1}) + a_{0} (V_{2}) = 2$ has solutions: $(0, 2)$ - $(0, 2)$ (matched gap), and $(1, 2)$ - $(0, 1)$ (gap- $1$ , gap- $1$ ) and its reverse.

Step 4. Checking smoothability: the $(0, 2)$ - $(0, 2)$ choice corresponds to $V_{i}$ spanned by the constant section $1$ and the section $z^{2}$ vanishing to order $2$ at $p$ (with $z$ a local coordinate centred at $p$ ). The other choices have base points or non-refined data, and on smoothing produce either a $g_{2}^{1}$ with a base point (not a valid limit of a base-point-free pencil) or a non-refined limit that fails the smoothing theorem in its base form. Only the $(0, 2)$ - $(0, 2)$ choice produces a valid smoothing to the canonical pencil $∣ K_{C} ∣$ .

Step 5. Smoothing the $(0, 2)$ - $(0, 2)$ refined limit: by the smoothing theorem proved above, the unique refined limit on $C_{0}$ smooths to a unique $g_{2}^{1}$ on a smoothing, and this $g_{2}^{1}$ has all the right properties (degree $2$ , dimension $1$ , base-point-free) to coincide with $∣ K_{C} ∣$ on the general fibre. The match between "unique refined limit on $C_{0}$ " and "unique $g_{2}^{1}$ on $C$ " verifies the smoothing recipe in this elementary case. $□$

Theorem (Eisenbud-Harris 1986 on Brill-Noether existence), stated without proof here — full proof in Eisenbud-Harris 1986 §5 and Harris-Morrison Ch. 5 ^{[source pending]}. The proof specialises $C$ to a chain $E_{1} \cup \dots \cup E_{g}$ of $g$ elliptic curves, encodes refined limit $g_{d}^{r}$ 's as lattice paths in a Young diagram, and counts paths to obtain the Castelnuovo number. Positivity of the count for $ρ \geq 0$ and vanishing for $ρ < 0$ propagate to the smooth fibre by upper semi-continuity.

Theorem (Eisenbud-Harris 1987 on Kodaira dimension), stated without proof here — full proof in Eisenbud-Harris 1987 Invent. Math. 90, 359-387 and Harris-Morrison Ch. 6F ^{[source pending]}. The proof computes the slope of the Brill-Noether divisor $\overline{B}_{d}^{r}$ on $\overline{M}_{g}$ via limit-linear-series methods, compares to the slope of the canonical class $K_{\overline{M}_{g}}$ , and shows the comparison gives positivity of $K$ as an effective combination of $\overline{B}_{d}^{r}$ and boundary divisors for $g \geq 24$ .

Theorem (Amini-Baker 2015 metrised complex specialisation), stated without proof here — full proof in Amini-Baker 2015 Math. Ann. 362, 55-106 ^{[source pending]}. The proof constructs the metrised complex as the skeleton of the Berkovich analytification, defines the divisor-theoretic specialisation, and verifies that it factors through Eisenbud-Harris on compact-type curves.

Connections Master

Riemann-Roch for curves 04.04.01. Limit linear series build on the Riemann-Roch theorem for smooth curves: $h^{0} (L) - h^{0} (K_{C} \otimes L^{- 1}) = d - g + 1$ for every line bundle $L$ of degree $d$ . The definition of a $g_{d}^{r}$ uses Riemann-Roch through the equivalent characterisation of speciality $h^{1} (L) > 0 \Leftrightarrow h^{0} (L) > d - g + 1$ , and the Brill-Noether number $ρ = g - (r + 1) (g - d + r)$ involves the Riemann-Roch indices on both sides. Without Riemann-Roch, $g_{d}^{r}$ has no setting; with it, the theory of linear series and limit linear series is the cohomological refinement.
Petri map and Gieseker-Petri theorem 04.04.08. The Petri conjecture — injectivity of the Petri map $μ_{0}$ on the general curve — is one of the central applications of Eisenbud-Harris 1986: the specialisation to a chain of elliptic curves reduces the injectivity of $μ_{0}$ to a node-by-node tensor-product calculation. Eisenbud-Harris 1983 Divisors on general curves gives the one-node version; the chain version completes the proof. The two units sit in a tight loop: limit linear series are the technical tool for Petri, and Petri is the central application of limit linear series.
Stable curve and Deligne-Mumford stability 04.10.22. Limit linear series are defined on stable curves of compact type, and the smoothing theorem extends them to the general fibre of a smoothing in $\overline{M}_{g}$ . The compact-type locus is a dense open subset of $\overline{M}_{g}$ , and the limit-linear-series moduli scheme extends the Brill-Noether moduli over this locus. Without Deligne-Mumford stability, the moduli problem has no setting; with it, limit linear series give the natural Brill-Noether theory on the boundary.
Determinantal varieties and the Porteous formula 04.04.13. The Kleiman-Laksov 1972 existence proof of Brill-Noether uses the Porteous formula to compute the cohomology class of $W_{d}^{r} (C)$ as the rank-deficiency class of a Picard-bundle map. The Porteous formula gives both non-emptiness (positivity of the class) and the dimension count (codimension of the degeneracy locus). The Eisenbud-Harris chain-of-elliptic-curves argument is the geometric counterpart of the same combinatorial count, with Porteous on the analytic side and lattice paths on the combinatorial side.
Deformation theory of smooth curves 04.10.20. The smoothing theorem of Eisenbud-Harris is a deformation-theoretic statement: a limit linear series on $C_{0}$ is the special fibre of a relative linear series on a smoothing $C \to B$ . The first-order deformation theory of $(C, L, V)$ — controlled by Ext groups generalising $H^{1} (T_{C})$ — provides the infinitesimal framework, and the smoothing theorem is the formal-integrability statement that compatible first-order data extends. The refined-equality assumption is what makes the obstruction class vanish.
Lazarsfeld K3-vector-bundle proof of Petri 04.04.16. Lazarsfeld 1986 gives an alternative proof of the Petri conjecture via vector bundles on K3 surfaces. The Lazarsfeld and Eisenbud-Harris proofs of Petri are dual in a structural sense: Lazarsfeld degenerates to a K3 surface and uses Bogomolov-stability of a bundle, while Eisenbud-Harris degenerates to a nodal curve and uses combinatorial counting of limit linear series. Both proofs target the same theorem and reveal complementary geometric content.
Determinantal varieties and the Porteous formula 04.04.13. The Porteous formula provides the cohomological dimension count for Brill-Noether loci, while limit linear series provide the geometric content of the Brill-Noether degeneration. Together they form the analytic-combinatorial duality at the heart of modern curve geometry.
Tropical curves 04.12.02. The Amini-Baker 2015 specialisation theorem identifies limit linear series with divisors on metrised complexes in the Baker-Norine sense, unifying Eisenbud-Harris with the tropical Riemann-Roch theorem of Baker-Norine 2007. The tropical specialisation gives a Berkovich-analytic viewpoint that extends Eisenbud-Harris beyond compact type and is the technical foundation for the modern tropical Brill-Noether theory of Cools-Draisma-Payne-Robeva, Pflueger, and Jensen-Ranganathan.

Historical & philosophical context Master

The Brill-Noether problem dates to the foundational 1873 paper of Alexander von Brill and Max Noether, Ueber die algebraischen Functionen und ihre Anwendung in der Geometrie (Math. Ann. 7, 269-310) ^{[source pending]}, where the authors posed the existence question for special linear systems on smooth curves and gave the heuristic dimension count $ρ (g, r, d) = g - (r + 1) (g - d + r)$ as the expected dimension of the locus of $g_{d}^{r}$ 's in $Pic^{d} (C)$ . The heuristic was a dimension count for a parameter space cut out by $(r + 1) (g - d + r)$ vanishing conditions on a $g$ -dimensional Picard variety, but the rigorous proof of non-emptiness when $ρ \geq 0$ had to wait nearly a century. The intervening period saw the work of Hermann Weyl and the Italian school — Severi in particular formulated and partially proved Brill-Noether non-existence for $ρ < 0$ on a general curve, but the proof was incomplete and required modern degenerative methods to be fully rigorous.

The 1971-1972 work of George Kempf and of Steven Kleiman and Dan Laksov gave the first rigorous proofs of Brill-Noether existence. Kempf 1971 ^{[source pending]} realised $W_{d}^{r} (C)$ as a degeneracy locus of a Picard-bundle morphism and applied Fulton's intersection theory. Kleiman-Laksov 1972 ^{[source pending]}, independently, used the Porteous formula to compute the cohomology class of $W_{d}^{r} (C)$ and deduce non-emptiness from positivity. Both proofs rely on the modern algebraic-geometric apparatus of intersection theory on the Picard variety. The Griffiths-Harris 1980 paper On the variety of special linear systems on a general algebraic curve (Duke Math. J. 47, 233-272) ^{[source pending]} proved the harder converse direction — non-existence for $ρ < 0$ on a general curve — by a degeneration to an irreducible nodal curve with $g$ nodes and a vanishing-order count anticipating the Eisenbud-Harris machinery.

David Eisenbud and Joseph Harris's Limit linear series: basic theory (Invent. Math. 85, 337-371, 1986) ^{[source pending]} introduced the modern definition: rather than degenerating to a single irreducible nodal curve, degenerate to a tree of smooth components — a curve of compact type — and track linear systems by recording vanishing sequences at every node, subject to the compatibility inequality. The Eisenbud-Harris moduli scheme of limit linear series is a quasi-projective variety extending the usual Brill-Noether moduli over the compact-type boundary of $\overline{M}_{g}$ , and the smoothing theorem identifies the dense refined-limit locus with the limit fibre of $G_{d}^{r} (C_{t})$ for a smoothing. The companion paper Eisenbud-Harris 1987 The Kodaira dimension of the moduli space of curves of genus $\geq 23$ (Invent. Math. 90, 359-387) ^{[source pending]} applied the technique to prove that $\overline{M}_{g}$ is of general type for $g \geq 24$ — a dramatic strengthening of Harris-Mumford 1982 and a foundational result in the geometry of moduli of curves.

The post-1986 development extended Eisenbud-Harris in three directions. To non-compact-type curves: Eduardo Esteves and Nivaldo Medeiros 2002 Limit canonical systems on curves with two components (Invent. Math. 149, 267-338) ^{[source pending]} introduced the linked-Picard-bundle framework; Brian Osserman 2006 A limit linear series moduli scheme (Ann. Inst. Fourier 56, 1165-1205) ^{[source pending]} gave the universal moduli construction via linked Grassmannians. To tropical / Berkovich settings: Omid Amini and Matthew Baker 2015 Linear series on metrized complexes of algebraic curves (Math. Ann. 362, 55-106) ^{[source pending]} unified Eisenbud-Harris with Baker's specialisation theorem (Adv. Math. 215, 2007) for tropical curves, giving a single divisor theory on metrised complexes. To refined Brill-Noether for special curves: Nathan Pflueger 2017 and David Jensen-Dhruv Ranganathan 2021 proved the refined Brill-Noether formula $ρ_{k} (g, r, d)$ for $k$ -gonal curves via tropical Brill-Noether on chains of loops, completing a programme initiated by Cools-Draisma-Payne-Robeva 2012.

The original Petri conjecture appears in the 1923 paper of Karl Petri, Ueber die invariante Darstellung algebraischer Funktionen einer Veränderlichen (Math. Ann. 88, 242-289) ^{[source pending]}, where Petri conjectured that the canonical ideal of a non-hyperelliptic curve of genus $\geq 4$ is generated by quadrics and cubics (the form-theoretic content); a modern reformulation in terms of injectivity of the Petri map $μ_{0}$ appeared in the 1970s, and the conjecture was proved by Gieseker 1982. Eisenbud-Harris 1986 gives an independent limit-linear-series proof and is the standard modern reference. The Lazarsfeld 1986 vector-bundle proof on K3 surfaces (J. Diff. Geom. 23, 299-307) provides a third independent route.

Bibliography Master

@article{EisenbudHarris1986,
  author  = {Eisenbud, David and Harris, Joseph},
  title   = {Limit linear series: basic theory},
  journal = {Inventiones Mathematicae},
  volume  = {85},
  number  = {2},
  year    = {1986},
  pages   = {337--371}
}

@article{EisenbudHarris1987,
  author  = {Eisenbud, David and Harris, Joseph},
  title   = {The {K}odaira dimension of the moduli space of curves of genus $\geq 23$},
  journal = {Inventiones Mathematicae},
  volume  = {90},
  number  = {2},
  year    = {1987},
  pages   = {359--387}
}

@article{GriffithsHarris1980,
  author  = {Griffiths, Phillip and Harris, Joseph},
  title   = {On the variety of special linear systems on a general algebraic curve},
  journal = {Duke Mathematical Journal},
  volume  = {47},
  number  = {1},
  year    = {1980},
  pages   = {233--272}
}

@incollection{Kempf1971,
  author    = {Kempf, George},
  title     = {Schubert methods with an application to algebraic curves},
  booktitle = {Publications of the Mathematisch Centrum},
  publisher = {Mathematisch Centrum Amsterdam},
  year      = {1971}
}

@article{KleimanLaksov1972,
  author  = {Kleiman, Steven L. and Laksov, Dan},
  title   = {On the existence of special divisors},
  journal = {American Journal of Mathematics},
  volume  = {94},
  number  = {2},
  year    = {1972},
  pages   = {431--436}
}

@book{HarrisMorrison,
  author    = {Harris, Joseph and Morrison, Ian},
  title     = {Moduli of Curves},
  publisher = {Springer-Verlag},
  series    = {Graduate Texts in Mathematics},
  volume    = {187},
  year      = {1998}
}

@book{ArbarelloCornalbaGriffithsII,
  author    = {Arbarello, Enrico and Cornalba, Maurizio and Griffiths, Phillip A.},
  title     = {Geometry of Algebraic Curves, Volume II},
  publisher = {Springer-Verlag},
  series    = {Grundlehren der mathematischen Wissenschaften},
  volume    = {268},
  year      = {2011}
}

@article{Osserman2006,
  author  = {Osserman, Brian},
  title   = {A limit linear series moduli scheme},
  journal = {Annales de l'Institut Fourier},
  volume  = {56},
  number  = {4},
  year    = {2006},
  pages   = {1165--1205}
}

@article{EstevesMedeiros2002,
  author  = {Esteves, Eduardo and Medeiros, Nivaldo},
  title   = {Limit canonical systems on curves with two components},
  journal = {Inventiones Mathematicae},
  volume  = {149},
  number  = {2},
  year    = {2002},
  pages   = {267--338}
}

@article{AminiBaker2015,
  author  = {Amini, Omid and Baker, Matthew},
  title   = {Linear series on metrized complexes of algebraic curves},
  journal = {Mathematische Annalen},
  volume  = {362},
  number  = {1--2},
  year    = {2015},
  pages   = {55--106}
}

@article{BrillNoether1873,
  author  = {Brill, Alexander von and Noether, Max},
  title   = {{U}eber die algebraischen {F}unctionen und ihre {A}nwendung in der {G}eometrie},
  journal = {Mathematische Annalen},
  volume  = {7},
  year    = {1873},
  pages   = {269--310}
}

@article{Petri1923,
  author  = {Petri, Karl},
  title   = {{U}eber die invariante {D}arstellung algebraischer {F}unktionen einer {V}er{\"a}nderlichen},
  journal = {Mathematische Annalen},
  volume  = {88},
  year    = {1923},
  pages   = {242--289}
}

@article{Gieseker1982,
  author  = {Gieseker, David},
  title   = {Stable curves and special divisors: {P}etri's conjecture},
  journal = {Inventiones Mathematicae},
  volume  = {66},
  year    = {1982},
  pages   = {251--275}
}

@article{Lazarsfeld1986,
  author  = {Lazarsfeld, Robert},
  title   = {{B}rill-{N}oether-{P}etri without degenerations},
  journal = {Journal of Differential Geometry},
  volume  = {23},
  year    = {1986},
  pages   = {299--307}
}

@article{BakerNorine2007,
  author  = {Baker, Matthew and Norine, Serguei},
  title   = {{R}iemann-{R}och and {A}bel-{J}acobi theory on a finite graph},
  journal = {Advances in Mathematics},
  volume  = {215},
  number  = {2},
  year    = {2007},
  pages   = {766--788}
}

@article{HarrisMumford1982,
  author  = {Harris, Joseph and Mumford, David},
  title   = {On the {K}odaira dimension of the moduli space of curves},
  journal = {Inventiones Mathematicae},
  volume  = {67},
  year    = {1982},
  pages   = {23--88}
}

@article{Pflueger2017,
  author  = {Pflueger, Nathan},
  title   = {{B}rill-{N}oether varieties of $k$-gonal curves},
  journal = {Advances in Mathematics},
  volume  = {312},
  year    = {2017},
  pages   = {46--63}
}

@article{JensenRanganathan2021,
  author  = {Jensen, David and Ranganathan, Dhruv},
  title   = {{B}rill-{N}oether theory for curves of a fixed gonality},
  journal = {Forum of Mathematics, Pi},
  volume  = {9},
  year    = {2021},
  pages   = {e1}
}

@article{CoolsDraismaPayneRobeva2012,
  author  = {Cools, Filip and Draisma, Jan and Payne, Sam and Robeva, Elina},
  title   = {A tropical proof of the {B}rill-{N}oether theorem},
  journal = {Advances in Mathematics},
  volume  = {230},
  number  = {2},
  year    = {2012},
  pages   = {759--776}
}

Prerequisites

04.04.01
04.04.08
04.04.13
04.10.20
04.10.22

Tier anchors

beginner: Harris-Morrison *Moduli of Curves* Ch. 5 informal opening — degenerating a smooth curve to a chain and tracking a linear system through the limit by recording vanishing orders at the nodes
intermediate: Harris-Morrison *Moduli of Curves* (Springer GTM 187, 1998) Ch. 5; Arbarello-Cornalba-Griffiths *Geometry of Algebraic Curves* Vol. II Ch. XXI
master: Eisenbud-Harris 1986 *Limit linear series: basic theory* (Invent. Math. 85, 337-371); Eisenbud-Harris 1987 *The Kodaira dimension of the moduli space of curves of genus $\geq 23$* (Invent. Math. 90, 359-387); Griffiths-Harris 1980 *On the variety of special linear systems on a general algebraic curve* (Duke Math. J. 47, 233-272); Kempf 1971 *Schubert methods with an application to algebraic curves* (Math. Centrum Amsterdam); Kleiman-Laksov 1972 *On the existence of special divisors* (Amer. J. Math. 94, 431-436); Harris-Morrison *Moduli of Curves* (Springer GTM 187, 1998) Ch. 5; Osserman 2006 *A limit linear series moduli scheme* (Ann. Inst. Fourier 56, 1165-1205); Esteves-Medeiros 2002 *Limit canonical systems on curves with two components* (Invent. Math. 149, 267-338); Amini-Baker 2015 *Linear series on metrized complexes of algebraic curves* (Math. Ann. 362, 55-106)

References

Eisenbud-Harris 1986 — Limit linear series: basic theory · Invent. Math. 85, 337-371; originator paper for the compact-type definition, the smoothing theorem, and the Brill-Noether proof
Eisenbud-Harris 1987 — The Kodaira dimension of the moduli space of curves of genus at least 23 · Invent. Math. 90, 359-387; first application: $\overline{\mathcal{M}}_g$ is of general type for $g \geq 24$, via limit-linear-series Brill-Noether divisor classes
Griffiths-Harris 1980 — On the variety of special linear systems on a general algebraic curve · Duke Math. J. 47, 233-272; Brill-Noether non-existence for $\rho < 0$ on a general curve, with the first rigorous degeneration argument anticipating Eisenbud-Harris
Kempf 1971 — Schubert methods with an application to algebraic curves · Math. Centrum Amsterdam preprint; Brill-Noether non-emptiness $W^r_d \ne \emptyset$ for $\rho \geq 0$ via degeneracy loci of a Picard-bundle map
Kleiman-Laksov 1972 — On the existence of special divisors · Amer. J. Math. 94, 431-436; independent Brill-Noether existence proof via the Porteous formula
Harris-Morrison — Moduli of Curves · Springer GTM 187, 1998; Chapter 5 (limit linear series, the smoothing theorem, applications to Brill-Noether and Gieseker-Petri)
Arbarello-Cornalba-Griffiths — Geometry of Algebraic Curves Volume II · Springer Grundlehren 268, 2011; Ch. XXI (limit linear series, refined limit linear series, the moduli scheme)
Osserman 2006 — A limit linear series moduli scheme · Ann. Inst. Fourier 56, 1165-1205; alternative construction of the moduli scheme via linked Grassmannians, valid for non-compact-type curves
Esteves-Medeiros 2002 — Limit canonical systems on curves with two components · Invent. Math. 149, 267-338; extension of the limit-linear-series formalism to curves outside compact type
Amini-Baker 2015 — Linear series on metrized complexes of algebraic curves · Math. Ann. 362, 55-106; Berkovich-tropical unification of Eisenbud-Harris with Baker's specialisation theorem
Brill-Noether 1873 — Ueber die algebraischen Functionen und ihre Anwendung in der Geometrie · Math. Ann. 7, 269-310; the original dimension-counting heuristic for special divisors that later became the Brill-Noether theorem
Petri 1923 — Ueber die invariante Darstellung algebraischer Funktionen einer Veränderlichen · Math. Ann. 88, 242-289; the Petri conjecture and the canonical-ideal generation theorem for non-hyperelliptic curves

Estimated time

beginner: 16m
intermediate: 45m
master: 80m