How Geometric Torsion at Sub-Planck Lengths Prevents the Existence of Special Classes of Residual D-Branes

1. Polynomials drawing pictures

Algebraic geometry studies geometric shapes that arise as solution sets of polynomial equations. The simplest example is the most familiar:

$x^2 + y^2 - 1 = 0.$

The set of pairs $(x, y) \in \RR^2$ satisfying this equation is a circle of radius one. We have written down a polynomial; it has carved out a curve. That move — from a single algebraic equation to the geometric object it defines — is the seed of the entire subject.

Generalization goes in two directions. More variables and more equations: the equation $x^2 + y^2 + z^2 - 1 = 0$ in three variables defines a sphere, and a pair of equations in three variables generically cuts out a one-dimensional curve, since each equation imposes one constraint and a generic pair imposes two. The second direction is to change the coefficient field. The rest of this article works over the complex numbers, because $\CC$ is algebraically closed — every nonconstant polynomial has as many roots as its degree, and the cohomological machinery we eventually need (Serre duality, derived categories of coherent sheaves, Bridgeland stability) is built on top of that fact. The pictures will continue to be drawn over $\RR$, but the theorems live over $\CC$.

A solution set of polynomial equations carved out inside $\CC^n$ — or, slightly more carefully, inside affine $n$-space $\AA^n_\CC$ — is called an affine algebraic variety. The circle, the sphere, any cubic curve $y^2 = x^3 + ax + b$: all are affine varieties, each cut out as the zero locus of a single polynomial in its ambient space. Crucially, an affine variety is not a topological manifold patched from charts. It is a single algebraic object, defined by a single ideal in a polynomial ring, with a single coordinate ring $\CC[x_1, \dots, x_n] / I$. No gluing is required to define a sphere algebraically; the equation does all the work.

The leap to the modern subject is forced by varieties that genuinely cannot be presented as a single affine piece. The canonical example is projective space $\PP^n_\CC$: the space of one-dimensional linear subspaces of $\CC^{n+1}$, parametrizing all "directions" through the origin. Projective space is proper — the algebraic-geometry analogue of compactness — and a basic theorem says that every global regular function on a connected proper variety is a constant. A closed subvariety of an affine space $\AA^N$ inherits the coordinate functions $x_1, \dots, x_N$ as global regular functions, and these are not all constant unless the subvariety is a single point. So $\PP^n$ for $n \geq 1$ cannot be a closed subvariety of any $\AA^N$: it has too few regular functions to fit. To work with it, you must build it from pieces — $\PP^n$ is glued from $n+1$ copies of affine $n$-space along their overlaps, with explicit transition maps relating the homogeneous coordinates. Elliptic curves are projective varieties for the same reason; their group law and cryptographic structure depend on the point at infinity that affine charts on their own cannot see. Once you have one example that demands gluing, you give up trying to embed every variety in some $\AA^N$ and accept that varieties are objects built by patching affine pieces along common overlaps. The result is the notion of a scheme — a geometric object locally describable by polynomial coordinates, with global topology determined by the gluing data. A scheme is an instruction for gluing affine schemes; the affine ones are the building blocks, projective space is the reason you ever need to build.

Once you have schemes, questions multiply: how do you classify them, when are two equivalent, what objects can live on them. Most of this article concerns objects that live on a particular kind of scheme — a smooth projective surface — and how the structure of such objects can become rich enough to ask categorical questions whose answers tell you something deep about the underlying geometry.

2. Enriques surfaces and a nineteenth-century counterexample

The Italian school — Castelnuovo, Enriques, Severi, and others working in Rome and Bologna in the late nineteenth and early twentieth centuries — set the program of classifying algebraic surfaces. A surface here means a complex algebraic variety of complex dimension two, or equivalently a four-real-dimensional space. The classification problem asked two questions: given two surfaces, can you tell whether one can be transformed into the other by birational maps (rational changes of coordinates invertible on a dense open subset)? And which surfaces are rational, meaning birational to projective space $\PP^2$?

Castelnuovo gave a clean criterion: a smooth projective surface $X$ is rational if and only if two cohomological invariants vanish:

A first encounter with cohomology

Throughout this article, expressions like $H^i(X, \cF)$ appear as vector-space invariants of a variety $X$ and a sheaf $\cF$ on it. The reader who has not seen cohomology before can hold onto a single picture, accurate enough to navigate the rest of the article.

Cohomology is the operation you perform when you have a sequence of vector spaces and linear maps $V^0 \xrightarrow{\;d^0\;} V^1 \xrightarrow{\;d^1\;} V^2 \xrightarrow{\;d^2\;} \cdots$ satisfying $d^{i+1} \circ d^i = 0$ — every composition of two consecutive maps is the zero map. The $i$-th cohomology is then $H^i \;=\; \ker(d^i) \,/\, \operatorname{im}(d^{i-1}),$ the quotient of "things killed by the next map" by "things produced by the previous map." This single move — kernel of a square-zero operator modulo its image — is the universal pattern of cohomology, and it is pure linear algebra at heart.

For a variety $X$ and a sheaf $\cF$, the groups $H^i(X, \cF)$ apply this construction to a complex built from local sections of $\cF$ over an open cover. $H^0(X, \cF)$ is the vector space of global sections — pieces of $\cF$ that exist consistently everywhere on $X$. Higher $H^i$ measure obstructions: $H^1$ counts local-to-global glitches (data defined locally that cannot be glued into one global thing); $H^2$ measures obstructions of obstructions.

The reader who has seen vector calculus already knows a special case. A divergence-free vector field on a punctured plane that is not the gradient of any function — circulating around the puncture — represents a nonzero class in $H^1$ of the punctured plane: closed (killed by the next operator) but not exact (not produced by the previous one). Every flavor of cohomology in this article — sheaf cohomology here, the Ext groups in section 3, derived-category cohomology in section 4, and the BRST $Q$-cohomology of physics in section 5 — is a variation on this single theme.

$p_g(X) \;=\; \dim H^0(X, \omega_X) \;=\; 0, \qquad q(X) \;=\; \dim H^1(X, \cO_X) \;=\; 0.$

Here $\omega_X$ is the canonical line bundle (the bundle of top-degree differential forms) and $\cO_X$ is the structure sheaf, both finite-dimensional complex vector spaces once you take their cohomology. Concretely, $H^0(X, \omega_X)$ is the space of holomorphic top-forms on $X$; $H^1(X, \cO_X)$ measures local-to-global obstructions for holomorphic functions and is harder to describe elementarily, but the Aside above gives the universal picture. The number $p_g$ is the geometric genus; $q$ is the irregularity. Castelnuovo's criterion says that vanishing of these two numbers is necessary and sufficient for $X$ to be rational, provided you also know the surface is regular — which in this situation turns out to be implied.

Enriques produced a counterexample. He constructed a smooth projective surface — now called an Enriques surface — for which $p_g = q = 0$ holds but the surface is not rational. His original construction was a sextic surface in $\PP^3$ passing with multiplicity two through the six edges of the coordinate tetrahedron; the smooth model of that singular surface gave the new example. Vanishing of $p_g$ and $q$ does not in itself force rationality — the Enriques surface filled the loophole that Castelnuovo's criterion left open, and it was the first hint that the classification of surfaces was finer than the classification of curves.

The modern definition is cleaner. Over $\CC$ — and more generally over any algebraically closed field of characteristic different from $2$ — a smooth projective surface $X$ is an Enriques surface if it is minimal (no $(-1)$-curves to blow down), satisfies $p_g(X) = q(X) = 0$, and has canonical bundle $\omega_X$ that is two-torsion:

$\omega_X \;\not\cong\; \cO_X, \qquad \omega_X^{\otimes 2} \;\cong\; \cO_X.$

The non-triviality of $\omega_X$ is what saves Enriques surfaces from being rational; the two-torsion is what makes them distinctive among all surfaces with vanishing $p_g$ and $q$. Everything in this article runs over $\CC$, so this is the working definition from here on.

The characteristic-free definition, and why characteristic 2 is special

The condition $\omega_X^{\otimes 2} \cong \cO_X$ with $\omega_X \not\cong \cO_X$ implicitly assumes that the order-two element of $\mathrm{Pic}(X)$ corresponding to $\omega_X$ is carried by the constant group scheme $\ZZ/2\ZZ$. Over a field of characteristic $\neq 2$ this is automatic — there is nothing else for it to be. In characteristic $2$, the same order-two class can be carried by an infinitesimal group scheme — $\mu_2$ or $\alpha_2$ — and when it is, the canonical bundle becomes trivial, $\omega_X \cong \cO_X$, with $p_g = q = 1$ rather than $0$. These are the singular and supersingular Enriques surfaces; they are bona fide members of the family by every modern test, but our definition above silently excludes them.

The robust formulation, due to Bombieri and Mumford, replaces "$\omega_X$ two-torsion" with the numerical condition that $K_X$ is numerically trivial, together with $b_2(X) = 10$ and $\chi(\cO_X) = 1$. In characteristic $\neq 2$ this recovers exactly the definition above; in characteristic $2$ it also catches the two extra families. None of the arguments in this article touch characteristic $2$, so the cleaner complex-analytic formulation is what we use.

The two-torsion has a concrete geometric consequence. Whenever a line bundle on a smooth variety squares to the trivial bundle, you can build an unramified double cover of the variety. For an Enriques surface this cover is itself smooth, and it turns out to be a K3 surface — a simply connected surface with trivial canonical bundle and nonzero $p_g$. An Enriques surface is therefore a K3 surface modulo a fixed-point-free involution; the K3 cover carries strictly more cohomology than the quotient.

$X_{\text{K3}} \;\twoheadrightarrow\; X_{\text{Enriques}} \;=\; X_{\text{K3}}/\langle \tau \rangle,$

where $\tau$ is the deck involution. Enriques surfaces inherit pieces of K3 structure, but only in quotient form, and a great deal of the subject consists of tracking what survives.

An Enriques surface can, but need not, contain rational curves. A smooth rational curve on a surface has self-intersection $-2$, because it is a smooth $\PP^1$ embedded with normal bundle $\cO(-2)$; such curves are called $(-2)$-curves. A generic Enriques surface contains none at all — these are called unnodal or generic, and they are the cleanest representatives of the family.

3. Vector bundles and coherent sheaves

To do anything with a variety beyond classifying it up to isomorphism, you put stuff on it. The most basic stuff is a vector bundle: a continuous (in our case, holomorphic and algebraic) family of vector spaces parametrized by the points of the variety. The tangent bundle of a smooth variety $X$ assigns to each point the tangent space there; the canonical bundle assigns to each point the one-dimensional space of top-degree differential forms. A line bundle is a vector bundle whose fibers are one-dimensional.

The algebraic way to package a vector bundle is as a locally free sheaf: a sheaf $\cE$ on $X$ such that on each small open subset $U$, the sections $\cE(U)$ form a free module over the ring of regular functions $\cO_X(U)$. "Locally free" expresses the bundle's local triviality; the global structure is captured by how the local trivializations glue.

This is enough for many purposes, but the category of vector bundles has a fundamental defect: it is not abelian. A morphism of vector bundles is a fiberwise linear map, and the kernel and cokernel can fail to be vector bundles — because the rank of the map can jump at certain points. Take a map $\cO_X \to \cO_X$ given by multiplication by a section vanishing along a curve; the cokernel is supported on that curve, where it is one-dimensional, and zero everywhere else. A "vector space that vanishes outside a subvariety" is not a vector bundle.

To fix this we generalize. The right object is a coherent sheaf — and on the schemes we care about, the working description is that a coherent sheaf is, locally, the cokernel of a map between two free $\cO_X$-modules of finite rank. This local-cokernel condition is what is technically called finitely presented; it agrees with coherence on a Noetherian scheme, and every variety considered in this article is Noetherian. Coherent sheaves form an abelian category: kernels and cokernels stay coherent, and exact sequences make perfect sense. Vector bundles sit inside the coherent sheaves as the locally free sheaves of finite rank, but they are no longer the only objects.

A skyscraper sheaf $\cO_p$ supported at a point $p \in X$ assigns the field $\CC$ to any open set containing $p$ and zero to any open set not containing $p$. It is coherent — locally it is the cokernel of the inclusion of the maximal ideal — but not locally free, because its rank jumps from one to zero as you move off the point. An ideal sheaf $\cI_Z$ of a subvariety $Z \subset X$ consists of regular functions vanishing along $Z$; it is the kernel of the surjection $\cO_X \twoheadrightarrow \cO_Z$, and it is coherent and torsion-free, but not locally free unless $Z$ is a divisor (a codimension-one subvariety—like a curve on a surface—that is locally defined by a single function).

These are exactly the sheaves that physicists call $0$-branes (skyscrapers) and $6$-branes wrapping cycles (ideal sheaves), as we will see in section 5. The fact that $\mathrm{Coh}(X)$ contains them all on equal footing is what makes it the right setting for both algebraic geometry and string-theoretic D-brane computations.

4. Complexes and the derived category

Coherent sheaves are abelian, but still not enough. For deeper invariants — cohomology, derived functors, the homological invariants that appear in moduli problems — you need to consider not just individual sheaves but complexes of sheaves.

A complex of coherent sheaves is a sequence

$\cdots \;\to\; \cE^{-1} \;\xrightarrow{\;d^{-1}\;}\; \cE^0 \;\xrightarrow{\;d^0\;}\; \cE^1 \;\xrightarrow{\;d^1\;}\; \cE^2 \;\to\; \cdots$

where the morphisms compose to zero, $d^{i+1} \circ d^i = 0$. This is the same square-zero condition we just met in the cohomology Aside of section 2, now applied to bundle-like objects rather than abstract vector spaces; the differential $d^i$ is a morphism of sheaves, but the formula behind every flavor of cohomology in this article is the same. The cohomology of the complex at degree $i$ is $H^i(\cE^\bullet) = \ker d^i / \im d^{i-1}$, again a coherent sheaf. A complex is bounded if only finitely many of the $\cE^i$ are nonzero. The whole machinery of homological algebra rests on the recognition that the complex carries more information than its cohomology, but the appropriate equivalence relation makes complexes indistinguishable when they have the same cohomology.

That equivalence relation is quasi-isomorphism: a morphism of complexes inducing an isomorphism on every cohomology group. The bounded derived category of coherent sheaves, $D^b(X)$, is the category of bounded complexes with quasi-isomorphisms formally inverted. An object of $D^b(X)$ is a bounded complex; a morphism is a "roof" $\cE^\bullet \xleftarrow{\sim} \cF^\bullet \to \cG^\bullet$ where the left arrow is a quasi-isomorphism. Coherent sheaves embed into $D^b(X)$ as complexes concentrated in a single degree.

The derived category has two structural features distinguishing it from an abelian category. First, the shift functor $[1]$ translates a complex one position to the left:

$\cE^\bullet[1]^i \;=\; \cE^{i+1}, \qquad d_{\cE[1]} \;=\; -d_\cE.$

It is invertible, with inverse $[-1]$, and applying $[1]$ repeatedly gives an action of $\ZZ$ on $D^b(X)$. Second, short exact sequences of complexes get replaced by distinguished triangles:

$\cE^\bullet \;\to\; \cF^\bullet \;\to\; \cG^\bullet \;\to\; \cE^\bullet[1].$

A distinguished triangle is the derived-category analogue of a short exact sequence. Long exact sequences in cohomology come out of distinguished triangles, and most structure theorems of algebraic geometry translate naturally into this language. The derived category equipped with its shift and its distinguished triangles is the prototypical example of a triangulated category, the abstract setting for the rest of this article.

The Serre functor is the categorical incarnation of Serre duality — and it is the operator whose interaction with stability conditions will produce our final contradiction. For an Enriques surface, $\dim X = 2$ and $\omega_X$ is two-torsion, so

$\mathsf{S}_X^{\,2}(\cE^\bullet) \;=\; \cE^\bullet \otimes \omega_X^{\otimes 2}[4] \;=\; \cE^\bullet[4].$

The Serre functor squares to the homological shift by four. On a K3 surface the Serre functor is just $[2]$ (because $\omega_X = \cO_X$), so it commutes with everything; on an Enriques surface it carries genuine torsion information — which is why the Enriques case is rigid in a way the K3 case is not.

Two bookkeeping consequences for $K_0(\mathrm{Coh}(X))$ — the abstract Grothendieck group whose elements are formal $\ZZ$-linear combinations of coherent sheaves modulo short exact sequences — will be used repeatedly. First, a distinguished triangle $A \to B \to C$ gives the additivity relation $[B] = [A] + [C]$ in $K_0$. Second, the class of a bounded complex equals the alternating sum of the classes of its cohomology sheaves:

$[\mathcal{F}^\bullet] \;=\; \sum_{k \in \ZZ} (-1)^k \bigl[\mathcal{H}^k(\mathcal{F}^\bullet)\bigr] \;\in\; K_0(\mathrm{Coh}(X)).$

This recipe is what converts the cohomology-sheaf description of a complex into a numerical class — it will land the final contradiction in section 8.

Why does any of this matter? The mathematician's answer is the homological algebra of varieties. The physicist's answer is more striking — and it is what convinced algebraic geometers they were studying the right object.

5. B-branes, the topological string, and Q-cohomology

There is a question that mathematicians of the 1990s could not have answered cleanly without help from physics: why is the derived category of coherent sheaves the right invariant of an algebraic variety, rather than just the abelian category of coherent sheaves? Coherent sheaves are abelian, geometrically natural, and sufficient for many purposes. What forces us to enlarge them to a triangulated category?

The answer came from a class of two-dimensional quantum field theories called topological string theories. The story begins with the type II superstring on a Calabi-Yau target, but you can read what follows as a structural argument that any reasonable notion of "what lives on $X$" forces you to the derived category — even without committing to string theory.

$Q$-cohomology: cohomology in physics

Before diving into topological strings, a physical analogy worth pinning to the cohomology Aside in section 2.

In quantum field theories with a fermionic gauge symmetry, one frequently has a square-zero operator $Q$ acting on the Hilbert space of states: $Q^2 = 0$. Physical states are taken to be elements of $\ker(Q)$ — states "$Q$-closed" in the sense of being annihilated by $Q$ — modulo elements of $\operatorname{im}(Q)$, which are deemed gauge-equivalent to zero. The space of physical states is therefore $\mathcal{H}_{\text{phys}} \;=\; \ker(Q) \,/\, \operatorname{im}(Q).$ This is exactly the cohomology construction we have been using. The operator $Q$ plays the role of the differential $d$; the condition $Q^2 = 0$ is the same as $d^{i+1} \circ d^i = 0$; "physical states modulo gauge" is "kernel modulo image."

The classic prototype is electromagnetism. The gauge transformation $A_\mu \mapsto A_\mu + \partial_\mu \chi$ is the image of an exterior derivative, the physical field strengths $F = dA$ are the closed forms, and the gauge group of $A$-fields modulo gradients is, on a topologically nontrivial spacetime, exactly the de Rham cohomology of that spacetime. The BRST formalism systematizes this: any time you have a gauge symmetry, you can construct a $Q$ such that $Q^2 = 0$ and physical states are $Q$-cohomology classes.

In the topological string below, $Q$ is one of the worldsheet supercharges, promoted to a scalar by the topological twist. Open-string states between two D-branes will turn out to be $Q$-cohomology groups, and we will see that mathematically these are the Ext groups $\Ext^i(E, F)$ — yet another flavor of the same kernel-mod-image construction, now applied to coherent sheaves. The thread "cohomology equals kernel mod image of a square-zero operator" runs through every formal object in this article.

A Calabi-Yau threefold $X$ supports a sigma model with $\mathcal{N} = (2,2)$ supersymmetry on the worldsheet. Witten's topological twist of this theory comes in two flavors, the A-twist and the B-twist, distinguished by which combination of worldsheet supercharges you promote to a worldsheet scalar. The B-twist is consistent precisely when $c_1(X) = 0$, the Calabi-Yau condition. After twisting, one supercharge becomes a worldsheet scalar nilpotent operator $Q$, the BRST operator, satisfying

$Q^2 \;=\; 0.$

Physical observables are $Q$-cohomology classes — equivalence classes of states $\psi$ satisfying $Q\psi = 0$ ("$Q$-closed") modulo states of the form $\psi = Q\chi$ ("$Q$-exact"). This is the same kernel-mod-image computation as in section 4; the BRST operator is the $d$, and physical states are the cohomology in the linear-algebra sense above. Correlation functions in the B-model depend only on the complex structure of $X$ and not on its Kähler structure. The closed-string state space is the Dolbeault cohomology $\bigoplus_{p,q} H^q(X, \wedge^p T_X)$, packaged more invariantly as the Hochschild cohomology of $D^b(X)$.

The question sharpens when you allow worldsheets with boundary — which physicists must, because that is what describes open strings ending on extended objects called D-branes.

A worldsheet with boundary needs boundary conditions to make the variational problem well-posed. Witten's analysis of B-model boundary conditions shows that they amount to a choice of complex submanifold of $X$ together with a holomorphic vector bundle on it. The conclusion: a B-brane is a holomorphic vector bundle $E$ on $X$, and the open-string spectrum between two B-branes $E$ and $F$ is

$\mathcal{H}_{\text{open}}(E,F) \;=\; \bigoplus_q \Ext^q(E, F).$

The right-hand side is a finite-dimensional complex vector space — the Ext groups of $E$ and $F$. For $q = 0$, $\Ext^0(E, F) = \Hom(E, F)$ is just the linear bundle maps from $E$ to $F$; for higher $q$, $\Ext^q$ measures "$q$-step extensions" of $F$ by $E$ and is itself an instance of the kernel-mod-image construction we have been tracking. The cohomology groups $H^i(X, \cF)$ from section 2 are the special case $\Ext^i(\cO_X, \cF)$ — Ext is a generalization of cohomology, with the structure sheaf $\cO_X$ in the first slot replaced by an arbitrary sheaf. The open-string ghost number matches the homological degree, and the worldsheet $Q$-cohomology of operators between branes lands exactly on this $\bigoplus_q \Ext^q$.

That would be satisfying if it were complete, but three physical phenomena push the formalism further. Singular branes: a D-brane wrapped on a point (a "$0$-brane") is a skyscraper sheaf, not a bundle, and a brane consisting of an ideal sheaf of a subvariety is a coherent sheaf that is not locally free. Anti-branes: a brane and its anti-brane have opposite orientations, so you need formal additive inverses. Tachyon condensation: a brane $E$ and antibrane $F$ with an open-string tachyon $T : F \to E$ can decay to a bound state, and Sen's tachyon condensation analysis identifies that bound state with the mapping cone $\mathrm{Cone}(T)$ — a complex of branes, not a single brane. Multi-step bound states give complexes of arbitrary length.

Coherent sheaves handle the first phenomenon. Triangulated structure — shifts, cones, quasi-isomorphism as gauge equivalence — handles the second and third. The natural closure of "holomorphic bundle" under these physical operations is exactly $D^b(\mathrm{Coh}(X))$.

Douglas's 2001 proposal is that the category of B-type D-branes on a Calabi-Yau is the bounded derived category of coherent sheaves. This sits inside Kontsevich's 1994 Homological Mirror Symmetry conjecture, which predicts an equivalence

$D^b(\mathrm{Coh}(X)) \;\simeq\; D^\pi \mathrm{Fuk}(X^\vee)$

between the B-side category on $X$ and the Fukaya category on the mirror $X^\vee$, trading complex geometry for symplectic geometry. For mathematicians, the conjecture was the first strong hint that the derived category was the structurally correct object on the algebraic side.

The key dictionary entry for our purposes:

$\text{open-string states between branes } E, F \;\longleftrightarrow\; \Ext^q(E, F) \;\longleftrightarrow\; Q\text{-cohomology}.$

Once you accept that the right object is $D^b(X)$, you can ask physical questions about which branes are stable — which actually exist as BPS states at a given point of moduli space. Douglas formalized this as $\Pi$-stability, with branes carrying a phase determined by the period of the holomorphic three-form. Bridgeland, in 2007, gave a clean mathematical version: a stability condition on any triangulated category. The Enriques surfaces we care about are not Calabi-Yau, but the formalism extends, and what started as a physics motivation for $D^b(X)$ becomes an algebraic-geometric tool for studying Kuznetsov components.

6. Exceptional collections and the Kuznetsov component

Let $X$ be a generic (unnodal) complex Enriques surface with derived category $D^b(X)$. The goal here is to carve $D^b(X)$ into pieces, isolating the component on which the final argument will run.

The cleanest pieces of any derived category are those generated by exceptional objects.

A short calculation shows that on an Enriques surface, every line bundle is exceptional. Given a line bundle $L$, the self-Ext groups are

$\Ext^i(L, L) \;=\; H^i(X, L \otimes L^{-1}) \;=\; H^i(X, \cO_X).$

The first equality uses that $\Ext^i(L, L)$ on a smooth variety equals the cohomology of the bundle $L \otimes L^{-1}$, which is the trivial bundle $\cO_X$. The second equality just rewrites the trivial bundle as $\cO_X$. The right-hand side is now exactly the cohomology Castelnuovo's criterion was about: $\CC$ for $i = 0$ (global constants on a connected variety), zero for $i = 1$ (since $q = 0$), and zero for $i = 2$ (since $p_g = 0$). Every line bundle is exceptional.

What is special to the unnodal case is that one can find a particularly clean exceptional collection.

The construction is lattice-theoretic. The Picard lattice of a generic Enriques surface is the rank-$10$ Enriques lattice $E_{10} = U \oplus E_8(-1)$, and one finds ten isotropic divisor classes $f_1, \dots, f_{10}$ with intersection numbers $f_i \cdot f_j = 1 - \delta_{ij}$; the line bundles built from these classes give the orthogonal collection. The vanishings $H^*(X, L_i \otimes L_j^{-1}) = 0$ for $i \neq j$ rest on Riemann–Roch and vanishing arguments that fail in the presence of $(-2)$-curves — this is exactly where the unnodal hypothesis enters.

The next piece of categorical scaffolding is the semiorthogonal decomposition.

With ten orthogonal exceptional line bundles, we get

$D^b(X) \;=\; \langle \mathrm{Ku}(X), L_1, L_2, \dots, L_{10} \rangle,$

where $\mathrm{Ku}(X)$ is everything left over.

Concretely, $\mathrm{Ku}(X)$ consists of complexes whose hypercohomology against every $L_i$ vanishes in every degree — the part of $D^b(X)$ that does not see any of the ten chosen line bundles.

The numerical Grothendieck group $K_{\mathrm{num}}$

For a triangulated category $\mathcal{C}$, the numerical Grothendieck group $K_{\mathrm{num}}(\mathcal{C})$ is the quotient of the abstract Grothendieck group $K_0(\mathcal{C})$ by the kernel of the Euler pairing $\chi(E, F) \;=\; \sum_{i \in \ZZ} (-1)^i \dim_\CC \mathrm{Ext}^i(E, F).$ For any admissible subcategory of the derived category of a smooth proper variety, this quotient is a finite-rank free abelian group — a fact going back to Bayer–Lahoz–Macri–Stellari and made explicit as Lemma 12.7 of Bayer–Lahoz–Macri–Nuer–Perry–Stellari (arXiv:1902.08184). In particular $K_{\mathrm{num}}(\mathrm{Ku}(X)) \cong \ZZ^2$ for an unnodal Enriques surface, sitting inside the rank-$12$ lattice $K_{\mathrm{num}}(D^b(X)) \cong H^*_{\mathrm{alg}}(X, \ZZ)$ as the orthogonal complement (under $-\chi$) of the rank-$10$ sublattice spanned by $[L_1], \dots, [L_{10}]$. The torsion-freeness will turn out to be load-bearing in the contradiction of section 8. A further fact in the same vein: the Euler form $-\chi$ restricted to $K_{\mathrm{num}}(\mathrm{Ku}(X))$ is positive-definite of signature $(2,0)$, computed in the Kuznetsov–Perry framework (see e.g.\ arXiv:1902.08184, Section 12, and the Enriques-specific analysis of Li–Pertusi–Zhao arXiv:2104.13610). This is what eventually forces $T^2 = I$ to give $T \in \{\pm I\}$: a $(1,1)$ lattice would admit hyperbolic involutions, but a positive-definite rank-$2$ lattice does not.

The rank of $K_{\mathrm{num}}(\mathrm{Ku}(X))$ is therefore $2$, sitting as the orthogonal complement of the rank-$10$ sublattice spanned by $[L_1], \ldots, [L_{10}]$ inside the rank-$12$ ambient lattice.

Because each $\langle L_i \rangle$ is admissible, the inclusion $\iota : \mathrm{Ku}(X) \hookrightarrow D^b(X)$ has both a left and a right adjoint. The right adjoint, the projection functor

$\zeta_K^! \;=\; \iota^! \;:\; D^b(X) \;\to\; \mathrm{Ku}(X),$

kills the line-bundle part. For $F \in D^b(X)$, the unit of adjunction fits into a triangle

$\bigoplus_{i=1}^{10} \mathbf{R}\Hom(L_i, F) \otimes L_i \;\longrightarrow\; F \;\longrightarrow\; \zeta_K^!(F),$

so $\zeta_K^!$ is the natural way to manufacture objects of $\mathrm{Ku}(X)$ out of objects of the ambient $D^b(X)$.

We will frequently abbreviate $\langle L_1, L_2, \dots, L_{10}\rangle$ as $\langle L_\bullet\rangle$ when the indexing is clear from context.

The standard reference is Bondal–Kapranov's 1989 paper on Representable functors, Serre functors, and mutations; modern proofs and extensions can be found in Kuznetsov's arXiv:0711.1734 §2 and arXiv:1509.07657 Lemma 2.8. From here we use $\zeta_K^!$ and $\mathsf{R}_{\langle L_\bullet\rangle}$ interchangeably whenever objects already lie in $\mathrm{Ku}(X)$.

The Serre functor of the ambient category restricts and projects to give the intrinsic Serre functor of $\mathrm{Ku}(X)$:

$\mathsf{S}_{\mathrm{Ku}}(E) \;=\; \zeta_K^!\bigl(\, \mathsf{S}_X \, \iota(E) \,\bigr) \;=\; \zeta_K^!\bigl(E \otimes \omega_X[2]\bigr) \;=\; \mathsf{R}_{\langle L_\bullet\rangle}\bigl(E \otimes \omega_X\bigr)[2].$

Squaring this functor in the ambient $D^b(X)$ kills the torsion in $\omega_X$, giving $\mathsf{S}_X^{\,2} \cong [4]$. The intrinsic Serre square on $\mathrm{Ku}(X)$ is more delicate — it equals $[4]$ only after composing with a non-trivial mutation autoequivalence:

$\mathsf{S}_{\mathrm{Ku}}^{\,2} \;\cong\; \Phi \circ [4], \qquad \Phi \;:=\; \mathsf{R}_{\langle L_\bullet\rangle} \circ \mathsf{R}_{\langle L_\bullet \otimes \omega_X\rangle}.$

The autoequivalence $\Phi$ acts as the identity on $K_{\mathrm{num}}(\mathrm{Ku}(X))$ (because $\omega_X$ is two-torsion in Picard, so $-\otimes\omega_X$ acts trivially on numerical classes, and right mutation through an orthogonal collection is the identity on its own orthogonal complement). But $\Phi$ is not isomorphic to the identity functor: the line bundles $L_i$ and $L_i \otimes \omega_X$ are non-isomorphic objects of $D^b(X)$, so the two mutations being composed run through different full subcategories. The gap between numerical triviality of $\Phi$ and its categorical non-triviality is exactly what Section 8 will exploit.

The Kuznetsov component is not strictly Calabi–Yau (a CY-2 category satisfies $\mathsf{S} = [2]$), but its Serre functor squares to a shift up to the residual mutation autoequivalence $\Phi$. Categories with this structure are called Enriques-type categories in the Kuznetsov–Perry framework. The non-triviality of $\mathsf{S}_{\mathrm{Ku}}$ — the fact that it is not literally $[2]$, even after squaring — is the algebraic remnant of the two-torsion of $\omega_X$, and it is what generates the obstruction in the next two sections.

One more object is needed. There is a recipe that produces 3-spherical objects in $\mathrm{Ku}(X)$ from line bundles outside the orthogonal collection. The notion of an $n$-spherical object is due to Seidel and Thomas, and it requires two conditions — a Calabi-Yau-$n$ condition on the Serre functor, and a constraint on the endomorphism algebra that identifies it (as a graded vector space) with the singular cohomology of the $n$-sphere.

See arXiv:1912.04332 for the proof. One ingredient deserves to be stated separately, since it is the precise place where unnodality enters the spherical-object construction:

This strengthens Zube's orthogonality $\Ext^\bullet(L_i, L_j) = \delta_{ij}\cdot\CC$ by also vanishing the $\omega_X$-twisted Ext groups; nodality (the existence of a $(-2)$-curve) would obstruct the underlying Riemann–Roch + vanishing argument and the off-diagonal twisted Exts would generically be nonzero. Section 8 leans on this lemma to kill the cross terms in a line-bundle Hom computation.

The algebraic Mukai lattice and $\mathrm{ch}(\omega_X) = 1$

The numerical Grothendieck group $K_{\mathrm{num}}(D^b(X))$ for an Enriques surface is naturally identified with the algebraic Mukai lattice $H^*_{\mathrm{alg}}(X, \ZZ) \;=\; H^0(X, \ZZ) \,\oplus\, \mathrm{NS}(X)/\mathrm{tors} \,\oplus\, \tfrac{1}{2}\ZZ \cdot \rho_X,$ where $\rho_X$ is the fundamental class. The Mukai vector $v(F) \;=\; \bigl(\mathrm{rk}(F),\, c_1(F),\, \tfrac{\mathrm{rk}(F)}{2} + \mathrm{ch}_2(F)\bigr)$ realises the isomorphism (see Nuer, arXiv:1406.0908). Because $\omega_X$ is two-torsion in $\mathrm{Pic}(X)$, its first Chern class $c_1(\omega_X)$ is killed in the torsion-free quotient $\mathrm{NS}(X)/\mathrm{tors}$, and consequently the entire Chern character collapses: $\mathrm{ch}(\omega_X) \;=\; 1 \quad \text{in } H^*_{\mathrm{alg}}(X, \ZZ).$ So tensoring by $\omega_X$ acts as the identity on $K_{\mathrm{num}}(D^b(X))$ — even though $-\otimes\omega_X$ is a non-trivial autoequivalence of $D^b(X)$. This numerical-vs-categorical asymmetry is the structural fact that makes Enriques surfaces categorically distinctive.

7. Bridgeland stability conditions

The reason this notion exists in the first place is physics. Section 5 explained why D-branes on a Calabi–Yau live in $D^b(X)$; the next question, the one that drove the subject in the late 1990s, was: which of those objects actually exist as physical states? Not all of them. A D-brane wrapping a cycle has a mass; for the brane to be a stable particle in the four-dimensional effective theory, that mass has to be locked in place by the supersymmetry algebra, not just by dynamics. The objects for which this happens are the BPS states, and the algebra of stability conditions is the language that catches them.

In type II string theory compactified on a Calabi–Yau threefold, the resulting four-dimensional theory has $\mathcal{N}=2$ supersymmetry — eight supercharges, with a complex central charge $Z(\gamma) \in \CC$ extending the algebra and depending on the charge $\gamma$ of the state. On the IIB side $Z(\gamma) = \int_\gamma \Omega$ is the period of the holomorphic three-form over a 3-cycle; on the IIA side $Z$ is built from the complexified Kähler class together with the brane charge. Either way, a representation-theoretic calculation gives the BPS bound

$M \;\geq\; |Z(\gamma)|,$

with equality on short multiplets. The states that saturate this bound — annihilated by half of the supercharges — are the BPS states. They cannot decay into lighter states of the same total charge because the bound forbids it; their stability is built into the algebra rather than the dynamics.

Bound states obey a triangle inequality. If a BPS state $E$ of charge $\gamma = \gamma_1 + \gamma_2$ is composed of constituents $A_1, A_2$ of charges $\gamma_i$, the central charge is additive but the mass is sub-additive:

$|Z(\gamma_1 + \gamma_2)| \;\leq\; |Z(\gamma_1)| + |Z(\gamma_2)|,$

with equality precisely when the two phases align. The deficit between the two sides is the binding energy. As the moduli of $X$ vary — Kähler class on the IIA side, complex structure on IIB — the central charges $Z(\gamma_i)$ rotate in $\CC$, and across real-codimension-one walls of marginal stability their phases align. On one side of the wall the bound state is BPS; on the other it has decayed into its constituents. The natural angular variable to track is the BPS phase $\phi(\gamma) = \tfrac{1}{\pi}\arg Z(\gamma)$, and the discontinuous reorganization of the spectrum across walls is the phenomenon called wall-crossing.

Michael Douglas, in a sequence of papers culminating in his ICM 2002 lecture, translated this picture into the language of triangulated categories. A distinguished triangle $A \to E \to B$ in $D^b(X)$ describes $E$ as a tachyon condensation bound state of $A$ and $B$. The state $E$ is $\Pi$-stable (the $\Pi$ stands for period) when, for every such triangle with $A, B$ nonzero, the BPS phases satisfy

$\phi(A) \;<\; \phi(E) \;<\; \phi(B).$

This single inequality re-encodes the mass-deficit picture entirely: a sub-brane of smaller phase contributes mass that locks into the sum constructively, leaving binding energy on the table; if the phases ever cross, the bond breaks. $\Pi$-stability varies continuously across the Kähler moduli space, the spectrum jumps at walls, and the worldvolume gauge theory on $E$ undergoes a corresponding change of quiver and superpotential.

Bridgeland's 2007 axiomatization is the rigorous mathematical version. The dictionary is exact: the heart of a bounded t-structure is the categorical incarnation of "which objects count as particles versus antiparticles" at a given point of moduli space; the central charge is a linear function on the Grothendieck group, realized for Calabi–Yau examples by the same period integrals that appear in the physics; the phase is the BPS phase; the Harder–Narasimhan filtration is the unique decomposition of any object into elementary semistable factors of strictly decreasing phase, mathematically formalizing the existence of a well-defined BPS spectrum; and the support property is what makes the moduli space of stability conditions — denoted $\mathrm{Stab}(D^b(X))$ — a complex manifold rather than a wild set, so that one can deform $\sigma$ continuously in the way the physical Kähler moduli demand. Conjecturally, a connected component of the space $\mathrm{Stab}(D^b(X))$ is the universal cover of the stringy Kähler moduli space of $X$ — the moduli that string theory says is the true parameter space for the B-model. Donaldson–Thomas invariants count $\sigma$-semistable objects and recover BPS state counts; the Kontsevich–Soibelman wall-crossing formula matches the physical spectrum jumps to the categorical operations of tilting a heart at a torsion pair.

With the physical picture as backdrop, we now state the axioms for a triangulated category $\mathcal{D}$ — applied throughout to $\mathcal{D} = \mathrm{Ku}(X)$, but everything generalizes.

For nonzero $E \in \mathcal{A}$, the phase is

$\phi(E) \;=\; \frac{1}{\pi} \arg Z(E) \;\in\; (0, 1].$

The phase is the angular position of $Z(E)$ in the upper half-plane, normalized so that the negative real axis sits at $\phi = 1$. An object $E$ is $\sigma$-semistable if $\phi(F) \leq \phi(E)$ for every proper subobject $0 \neq F \subsetneq E$ in $\mathcal{A}$.

The Harder–Narasimhan filtration extends the phase to objects of the full triangulated category $\mathrm{Ku}(X)$, not just the heart. For any nonzero $E \in \mathrm{Ku}(X)$, there is a uniquely determined filtration whose factors are semistable with strictly decreasing phases; the largest phase appearing is $\phi^+(E)$ and the smallest is $\phi^-(E)$ — and it is $\phi^+$ that will eventually produce a contradiction.

There is an equivalent reformulation in terms of slicings. A slicing $\mathcal{P}$ assigns to each real number $\phi$ the abelian subcategory $\mathcal{P}(\phi)$ of semistable objects of phase $\phi$, satisfying $\mathcal{P}(\phi + 1) = \mathcal{P}(\phi)[1]$ and $\Hom(\mathcal{P}(\phi_1), \mathcal{P}(\phi_2)) = 0$ for $\phi_1 > \phi_2$. The slicing and heart formulations are equivalent, with the heart recovered as $\mathcal{A} = \mathcal{P}\bigl((0, 1]\bigr)$.

The support property — a quadratic-form condition due to Kontsevich and Soibelman, equivalent to the bound $\|v(E)\| \leq C |Z(E)|$ for all semistable $E$ — upgrades local injectivity to local biholomorphism. We treat it as part of the definition.

The space $\mathrm{Stab}(\mathrm{Ku}(X))$ carries a right action of $\widetilde{\mathrm{GL}}^+(2,\RR)$, the universal cover of the orientation-preserving general linear group on $\RR^2$. An element of $\widetilde{\mathrm{GL}}^+(2,\RR)$ is a pair $(T, f)$ where $T \in \mathrm{GL}^+(2,\RR)$ is a real $2 \times 2$ matrix with positive determinant, and $f : \RR \to \RR$ is an increasing function with $f(\phi + 1) = f(\phi) + 1$, compatible with the action of $T$ on the unit circle. The action on a stability condition is

$\sigma \cdot (T, f) \;=\; (\mathcal{P}', Z'), \qquad Z' \;=\; T^{-1} \circ Z, \qquad \mathcal{P}'(\phi) \;=\; \mathcal{P}(f(\phi)).$

So $T$ shears the central charge as a real-linear map and $f$ relabels phases. The shift functor $[1]$ takes the slicing to $\mathcal{P}'(\phi) = \mathcal{P}(\phi + 1)$ and multiplies $Z$ by $-1$ (since $[1]_* = -1$ on $K$), corresponding to the universal-cover element $(T, f) = (-I, \phi \mapsto \phi + 1)$.

There is also a left action of $\mathrm{Aut}(\mathrm{Ku}(X))$. For an autoequivalence $\Phi$,

$\Phi \cdot \sigma \;=\; \bigl(\Phi(\mathcal{A}), Z \circ \Phi_*^{-1}\bigr).$

The Serre functor is one such autoequivalence. The two actions commute, and the natural compatibility one can ask between $\mathsf{S}_{\mathrm{Ku}}$ and $\sigma$ is that the left action of $\mathsf{S}_{\mathrm{Ku}}$ lies in the same $\widetilde{\mathrm{GL}}^+(2,\RR)$-orbit as $\sigma$.

Serre-invariant stability conditions exist on many Kuznetsov components — cubic threefolds, cubic fourfolds, Gushel–Mukai threefolds and fourfolds — and where they exist they are essentially unique up to the $\widetilde{\mathrm{GL}}^+(2,\RR)$-action; this near-uniqueness is what makes them so powerful for moduli theory. The question is whether any exist on $\mathrm{Ku}(X)$ for $X$ a generic Enriques surface.

8. The contradiction

The argument runs by contradiction. Suppose $\sigma = (\mathcal{A}, Z)$ is a Serre-invariant Bridgeland stability condition on $\mathrm{Ku}(X)$, so there exists a cover element $(T, f) \in \widetilde{\mathrm{GL}}^+(2,\RR)$ realising

$\mathsf{S}_{\mathrm{Ku}} \cdot \sigma \;=\; \sigma \cdot (T, f).$

The plan is to extract enough information from this single relation, combined with the explicit categorical structure of $\mathrm{Ku}(X)$, to produce a numerical impossibility — a class that must equal both $+v$ and $-v$ in a torsion-free lattice.

8.1. Recasting $\mathsf{S}_{\mathrm{Ku}}^{\,2}$ as $\Phi \circ [4]$

The intrinsic Serre functor of $\mathrm{Ku}(X)$ is given, by the standard formula for the Serre functor of an admissible subcategory ($\mathsf{S}_{\mathrm{Ku}} = \zeta_K^! \circ \mathsf{S}_X \circ \iota$), as

$\mathsf{S}_{\mathrm{Ku}}(F) \;=\; \zeta_K^!\bigl(F \otimes \omega_X[2]\bigr) \;=\; \mathsf{R}_{\langle L_\bullet\rangle}\bigl(F \otimes \omega_X\bigr)[2],$

where $\mathsf{R}_{\langle L_\bullet\rangle}$ denotes right mutation through the orthogonal collection $\langle L_1, \ldots, L_{10}\rangle$. Before iterating, we record a small conjugation identity for right mutations through a twisted orthogonal collection. Since $-\otimes \omega_X$ is an autoequivalence of $D^b(X)$ that carries the orthogonal collection $\{L_i\}$ to $\{L_i \otimes \omega_X\}$, conjugating the unit triangle for right mutation by this autoequivalence gives

$\mathsf{R}_{\langle L_\bullet \otimes \omega_X\rangle} \;=\; (-\otimes \omega_X) \;\circ\; \mathsf{R}_{\langle L_\bullet\rangle} \;\circ\; (-\otimes \omega_X^{-1})$

as functors $D^b(X) \to D^b(X)$. Using $\omega_X^{-1} \cong \omega_X$ (forced by $\omega_X^{\otimes 2} \cong \mathcal{O}_X$), evaluating on an object $G$ rearranges to

$\mathsf{R}_{\langle L_\bullet \otimes \omega_X\rangle}(G) \;\cong\; \mathsf{R}_{\langle L_\bullet\rangle}(G \otimes \omega_X) \otimes \omega_X.$

Now iterate the Serre formula. Applying $\mathsf{S}_{\mathrm{Ku}}$ twice and pulling out the shifts,

$\mathsf{S}_{\mathrm{Ku}}^{\,2}(F) \;=\; \mathsf{R}_{\langle L_\bullet\rangle}\bigl(\mathsf{R}_{\langle L_\bullet\rangle}(F \otimes \omega_X) \otimes \omega_X\bigr)[4].$

The conjugation identity, applied with $G = F$, identifies the inner expression $\mathsf{R}_{\langle L_\bullet\rangle}(F \otimes \omega_X) \otimes \omega_X$ with $\mathsf{R}_{\langle L_\bullet \otimes \omega_X\rangle}(F)$. Substituting,

$\mathsf{S}_{\mathrm{Ku}}^{\,2} \;\cong\; \Phi \circ [4], \qquad \Phi \;:=\; \mathsf{R}_{\langle L_\bullet\rangle} \circ \mathsf{R}_{\langle L_\bullet \otimes \omega_X\rangle}.$

The autoequivalence $\Phi$ is a composition of two right mutations through different orthogonal collections — the original one and its $\omega_X$-twist. Although the individual classes $[L_i]$ and $[L_i \otimes \omega_X]$ coincide in the numerical Grothendieck group $K_{\mathrm{num}}(D^b(X))$ (the canonical bundle is two-torsion in $\mathrm{Pic}(X)$), the full subcategories $\langle L_\bullet\rangle$ and $\langle L_\bullet \otimes \omega_X\rangle$ are genuinely different inside $D^b(X)$ because the line bundles $L_i$ and $L_i \otimes \omega_X$ are not isomorphic as objects. So $\Phi$ is a non-trivial autoequivalence of $\mathrm{Ku}(X)$ — and on a generic test object it will turn out to be not isomorphic to the shift $[2]$.

8.2. $\Phi$ acts trivially on $K_{\mathrm{num}}(\mathrm{Ku}(X))$

Both ingredients of $\Phi$ — the outer mutation $\mathsf{R}_{\langle L_\bullet\rangle}$ and the inner mutation $\mathsf{R}_{\langle L_\bullet \otimes \omega_X\rangle}$ — act as the identity on the numerical Grothendieck group $K_{\mathrm{num}}(\mathrm{Ku}(X))$. The argument has two parts.

First, the tensor $- \otimes \omega_X$ acts on $K_{\mathrm{num}}(D^b(X))$ via multiplication by the Chern character $\mathrm{ch}(\omega_X)$. Since $\omega_X$ is two-torsion in $\mathrm{Pic}(X)$, its first Chern class $c_1(\omega_X)$ vanishes in $\mathrm{NS}(X)/\mathrm{tors}$, and so $\mathrm{ch}(\omega_X) = 1$ in the algebraic cohomology lattice $H^*_{\mathrm{alg}}(X, \ZZ) \cong K_{\mathrm{num}}(D^b(X))$ (this is the Mukai-lattice computation at the end of Section 6). Hence tensoring by $\omega_X$ is the identity on numerical classes; in particular $[L_i] = [L_i \otimes \omega_X]$ in $K_{\mathrm{num}}(D^b(X))$ for each $i$.

Second — and this is what handles the inner mutation — the equality $[L_i] = [L_i \otimes \omega_X]$ implies that the two sublattices $\mathrm{span}_\ZZ\{[L_i]\}$ and $\mathrm{span}_\ZZ\{[L_i \otimes \omega_X]\}$ are literally equal inside $K_{\mathrm{num}}(D^b(X))$. Their Euler-pairing orthogonal complements therefore coincide as well, so the two right mutations $\mathsf{R}_{\langle L_\bullet\rangle}$ and $\mathsf{R}_{\langle L_\bullet \otimes \omega_X\rangle}$ induce the same numerical projection — orthogonal projection onto $\langle L_\bullet\rangle^\perp = K_{\mathrm{num}}(\mathrm{Ku}(X))$. Restricted to classes already in $K_{\mathrm{num}}(\mathrm{Ku}(X))$, both projections are the identity.

Taking the composition, $\Phi$ acts as the identity on $K_{\mathrm{num}}(\mathrm{Ku}(X))$. Consequently the squared Serre functor $\mathsf{S}_{\mathrm{Ku}}^{\,2} = \Phi \circ [4]$ also acts trivially on $K_{\mathrm{num}}$, since the shift $[4]$ acts as $(-1)^4 = +1$ on numerical classes.

8.3. The cover-element matrix is $-I$

Squaring the Serre-invariance relation:

$\mathsf{S}_{\mathrm{Ku}}^{\,2} \cdot \sigma \;=\; \sigma \cdot (T^2, f \circ f).$

The action on the central charge $Z: K_{\mathrm{num}}(\mathrm{Ku}(X)) \to \CC$ is $Z(\mathsf{S}_{\mathrm{Ku}}^{\,2}(F)) = T^{-2} \cdot Z(F)$. But $\mathsf{S}_{\mathrm{Ku}}^{\,2}$ acts as the identity on $K_{\mathrm{num}}$, so $T^{-2} \cdot Z = Z$ as linear maps. Since $K_{\mathrm{num}}(\mathrm{Ku}(X))$ is a rank-$2$ lattice (positive-definite under $-\chi$, so signature $(2,0)$ — as established in the $K_{\mathrm{num}}$ Aside in Section 6) and $Z$ identifies $K_{\mathrm{num}} \otimes \RR \cong \CC \cong \RR^2$:

$T^2 \;=\; I \quad \text{in } \mathrm{GL}^+(2,\RR), \qquad \text{so} \qquad T \in \{I, -I\}.$

(Reflections, which also satisfy $T^2 = I$, have $\det T = -1$ and are excluded by the orientation-preserving constraint $T \in \mathrm{GL}^+(2,\RR)$.)

To choose between the two branches, apply the cover-element relation to one of the 3-spherical objects $S_i = \zeta_K^!(L_i \otimes \omega_X)$ from Section 6. Each $S_i$ has well-defined upper and lower Harder–Narasimhan phases $\phi^\pm(S_i)$ (these exist for any nonzero object under any Bridgeland stability condition), and the cover-element relation gives

$f\bigl(\phi^\pm(S_i)\bigr) \;=\; \phi^\pm\bigl(\mathsf{S}_{\mathrm{Ku}}(S_i)\bigr) \;=\; \phi^\pm(S_i[3]) \;=\; \phi^\pm(S_i) + 3,$

using $\mathsf{S}_{\mathrm{Ku}}(S_i) \cong S_i[3]$ from the previous section and the fact that the homological shift $[k]$ adds exactly $k$ to every phase in the Harder–Narasimhan filtration. Combine with the two cases for $T$:

• Case $T = I$. The lifts of the identity to $\widetilde{\mathrm{GL}}^+(2,\RR)$ are translations by even integers — namely $f(\phi) = \phi + 2k$ for $k \in \ZZ$ (these are the powers of $[2]$). The constraint $2k = 3$ has no integer solution. Excluded.
• Case $T = -I$. The lifts of $-I$ are translations by odd integers — $f(\phi) = \phi + (2k+1)$ for $k \in \ZZ$ (odd powers of the shift $[1]$). The constraint $2k + 1 = 3$ gives $k = 1$, hence $f(\phi) = \phi + 3$.

So $T = -I$ and $f(\phi) = \phi + 3$ uniformly on the universal cover. The numerical consequence — the Serre sign relation — follows immediately: $T = -I$ acting on the central charge $Z$ via $Z(\mathsf{S}_{\mathrm{Ku}}(F)) = T^{-1} \cdot Z(F) = -Z(F)$ gives, by injectivity of $Z$ on $K_{\mathrm{num}}(\mathrm{Ku}(X)) \otimes \RR$ — a consequence of Bridgeland's support property, which forces $Z$ to be a real isomorphism $K_{\mathrm{num}}(\mathrm{Ku}(X)) \otimes \RR \xrightarrow{\sim} \CC$ on the rank-$2$ lattice (see Bayer–Macri–Stellari, arXiv:1410.1934) —

$\boxed{\;[\mathsf{S}_{\mathrm{Ku}}(F)] \;=\; -[F] \quad \text{in } K_{\mathrm{num}}(\mathrm{Ku}(X)) \text{ for every } F \in \mathrm{Ku}(X).\;}$

This sign relation is the constraint we will violate.

8.4. The test object — the projected point

To produce a contradiction we exhibit a single object $E \in \mathrm{Ku}(X)$ on which the Serre sign relation fails by direct categorical computation. The natural candidate is the right-adjoint projection of a generic skyscraper sheaf:

The right-projection unit triangle for $\zeta_K^!$ reads

$\bigoplus_{i=1}^{10} L_i \otimes \mathbf{R}\mathrm{Hom}(L_i, \mathcal{O}_p) \;\xrightarrow{\;\mathrm{ev}_p\;}\; \mathcal{O}_p \;\to\; E \;\to\; \bigoplus_i L_i[1].$

The Hom-spaces are computed via Serre duality on $X$ (with Serre functor $\mathsf{S}_X = - \otimes \omega_X[2]$):

$\mathrm{Ext}^k(L_i, \mathcal{O}_p) \;\cong\; \mathrm{Ext}^{2-k}(\mathcal{O}_p, L_i \otimes \omega_X)^\vee.$

Since $\mathcal{O}_p$ is a skyscraper at a smooth point, the depth-$2$ Koszul resolution of $\mathcal{O}_p$ at $p$ (a regular local ring of dimension 2) gives $\mathrm{Ext}^j(\mathcal{O}_p, L_i \otimes \omega_X) = \delta_{j, 2} \cdot \CC$ — the right-hand side is concentrated in degree 2 with dimension 1. Dualising:

$\mathbf{R}\mathrm{Hom}(L_i, \mathcal{O}_p) \;\cong\; \CC \quad \text{(in degree 0)}.$

The unit triangle therefore simplifies to

$\bigoplus_{i=1}^{10} L_i \;\xrightarrow{\;\mathrm{ev}_p\;}\; \mathcal{O}_p \;\to\; E \;\to\; \bigoplus_i L_i[1].$

Each component $L_i \to \mathcal{O}_p$ surjects onto the fibre $L_i|_p \cong \CC$, so the global evaluation $\mathrm{ev}_p$ is surjective and the long exact sequence in cohomology gives

$\mathcal{H}^0(E) \;=\; \mathrm{coker}(\mathrm{ev}_p) \;=\; 0, \qquad \mathcal{H}^{-1}(E) \;=\; K \;:=\; \ker(\mathrm{ev}_p).$

The kernel sheaf $K$ is locally free of rank $10$ on $X \setminus \{p\}$ (where $\mathcal{O}_p$ vanishes, so the kernel of $\bigoplus L_i \to 0$ equals $\bigoplus L_i$ itself) and torsion-free of rank $10$ at $p$ — the kernel of a rank-$10$ free $\mathcal{O}_{X,p}$-module surjecting onto the residue field $\CC \cong \mathcal{O}_{X,p}/\mathfrak{m}_p$. Globally, $K$ is a torsion-free sheaf of rank $10$. In short:

$E \;\cong\; K[1] \quad \text{is a single sheaf placed in cohomological degree } -1.$

Its class in the numerical lattice is

$[E] \;=\; -[K] \;=\; [\mathcal{O}_p] - \sum_{i=1}^{10} [L_i] \;\in\; K_{\mathrm{num}}(\mathrm{Ku}(X)),$

the orthogonal projection of the point class onto $\langle L_\bullet\rangle^\perp$. The rank component of $[E]$ is $-10$ (zero from the skyscraper minus ten from the line bundles), so $[E] \neq 0$ in the rank-$2$ lattice $K_{\mathrm{num}}(\mathrm{Ku}(X))$ — this non-vanishing will drive the contradiction.

8.5. Computing $\mathsf{S}_{\mathrm{Ku}}(E)$ — the two-degree spread

By the formula for the intrinsic Serre,

$\mathsf{S}_{\mathrm{Ku}}(E) \;=\; \mathsf{R}_{\langle L_\bullet\rangle}(E \otimes \omega_X)[2].$

Tensor the unit triangle for $E$ by $\omega_X$, using $\mathcal{O}_p \otimes \omega_X \cong \mathcal{O}_p$ (the canonical bundle trivialises locally at any smooth point):

$\bigoplus_i (L_i \otimes \omega_X) \;\to\; \mathcal{O}_p \;\to\; E \otimes \omega_X.$

To right-mutate $E \otimes \omega_X$ through $\langle L_\bullet\rangle$, compute $\mathbf{R}\mathrm{Hom}(L_j, E \otimes \omega_X)$. By the tensor-Hom adjunction $\mathbf{R}\mathrm{Hom}(A, B \otimes L) \cong \mathbf{R}\mathrm{Hom}(A \otimes L^{-1}, B)$ for an invertible sheaf $L$, applied with $L = \omega_X$ and using $\omega_X^{-1} \cong \omega_X$ (the canonical bundle is its own inverse, since $\omega_X^{\otimes 2} \cong \mathcal{O}_X$):

$\mathbf{R}\mathrm{Hom}(L_j, E \otimes \omega_X) \;\cong\; \mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, E).$

Apply $\mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, -)$ to the unit triangle for $E$:

$\bigoplus_i \mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, L_i) \;\to\; \mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, \mathcal{O}_p) \;\to\; \mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, E).$

The two anchor terms:

• Line-bundle term. For $i \neq j$, the unnodality hypothesis (Li–Nuer–Stellari–Zhao Lemma 4.4) gives $\mathrm{Ext}^k(L_i, L_j \otimes \omega_X) = 0$ for all $k$, dualised via Serre duality on $X$ to $\mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, L_i) = 0$. For $i = j$,

$\mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, L_j) \;=\; \mathbf{R}\Gamma\bigl(L_j \otimes L_j^{-1} \otimes \omega_X^{-1}\bigr) \;=\; \mathbf{R}\Gamma(\omega_X) \;\cong\; \CC[-2],$

where the last identification uses the Enriques cohomology of the canonical bundle, computed via Serre duality on $X$: $h^0(\omega_X) = h^2(\mathcal{O}_X) = 0$ (equivalently, $\omega_X \neq \mathcal{O}_X$ and $\chi(\mathcal{O}_X) = 1$), $h^1(\omega_X) = h^1(\mathcal{O}_X) = q = 0$ by Serre duality on $X$, and $h^2(\omega_X) = h^0(\mathcal{O}_X) = 1$ likewise.

• Skyscraper term. $\mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, \mathcal{O}_p) \cong \CC$, concentrated in degree 0 — the same evaluation calculation as in the unit triangle.

The connecting morphism $\CC[-2] \to \CC$ in $D^b(X)$ corresponds to a class in $\mathrm{Hom}_{D^b}(\CC[-2], \CC) = \mathrm{Ext}^2_\CC(\CC, \CC) = 0$, hence vanishes. The triangle therefore splits, giving

$\mathbf{R}\mathrm{Hom}(L_j \otimes \omega_X, E) \;\cong\; \CC \,\oplus\, \CC[-1] \quad \text{for each } j = 1, \ldots, 10.$

The mutation source then takes the explicit form

$\bigoplus_{j=1}^{10} L_j \otimes \mathbf{R}\mathrm{Hom}(L_j, E \otimes \omega_X) \;\cong\; \biggl(\bigoplus_{j=1}^{10} L_j\biggr) \;\oplus\; \biggl(\bigoplus_{j=1}^{10} L_j\biggr)[-1].$

Plugging into the right-mutation triangle and writing $\Psi := \mathsf{R}_{\langle L_\bullet\rangle}(E \otimes \omega_X)$ for brevity (note: $\Psi$ is not the global autoequivalence $\Phi$ of §8.1 applied to $\mathcal{O}_p$ — it is only the right-mutation through $\langle L_\bullet\rangle$ of $E \otimes \omega_X$, missing the inner mutation through $\langle L_\bullet \otimes \omega_X\rangle$):

$\biggl(\bigoplus_j L_j\biggr) \,\oplus\, \biggl(\bigoplus_j L_j\biggr)[-1] \;\xrightarrow{\;\alpha\;}\; (K \otimes \omega_X)[1] \;\to\; \Psi \;\to\; \cdots.$

The long exact sequence in cohomology — using that the source has $\mathcal{H}^0 = \bigoplus L_j$ and $\mathcal{H}^1 = \bigoplus L_j$, while the target $E \otimes \omega_X = (K \otimes \omega_X)[1]$ has $\mathcal{H}^{-1} = K \otimes \omega_X$ and zero in all other degrees — yields a short exact sequence describing $\mathcal{H}^{-1}(\Psi)$:

$0 \;\to\; K \otimes \omega_X \;\to\; \mathcal{H}^{-1}(\Psi) \;\to\; \bigoplus_{j=1}^{10} L_j \;\to\; 0,$

together with

$\mathcal{H}^0(\Psi) \;\cong\; \bigoplus_{j=1}^{10} L_j, \qquad \mathcal{H}^k(\Psi) \;=\; 0 \text{ for } k \notin \{-1, 0\}.$

Writing $M := \mathcal{H}^{-1}(\Psi)$ for the rank-$20$ extension (rank $10$ from $K \otimes \omega_X$ plus rank $10$ from $\bigoplus L_j$), and shifting by $[2]$:

$\mathsf{S}_{\mathrm{Ku}}(E) \;\cong\; \Psi[2],$

with cohomology

$\mathcal{H}^{-3}\bigl(\mathsf{S}_{\mathrm{Ku}}(E)\bigr) \;=\; M, \qquad \mathcal{H}^{-2}\bigl(\mathsf{S}_{\mathrm{Ku}}(E)\bigr) \;=\; \bigoplus_{j=1}^{10} L_j,$

and zero in all other degrees.

The cohomology of $\mathsf{S}_{\mathrm{Ku}}(E)$ is concentrated in two adjacent degrees, with sheaves of substantially different geometric character — $M$ has rank $20$ with a torsion defect at $p$ (inherited from the $K$-summand), while $\bigoplus L_j$ is locally free of rank $10$ everywhere on $X$. This two-degree categorical signature is the explicit witness to the non-uniformity of $\mathsf{S}_{\mathrm{Ku}}$ that the Serre sign relation cannot accommodate.

8.6. The numerical contradiction

Compute the class of $\mathsf{S}_{\mathrm{Ku}}(E)$ in $K_{\mathrm{num}}(\mathrm{Ku}(X))$ via the alternating sum over cohomology sheaves — the standard recipe for converting a complex into a class in $K_0$:

$[\mathsf{S}_{\mathrm{Ku}}(E)] \;=\; (-1)^{-3} [M] \;+\; (-1)^{-2} \biggl[\bigoplus_{j} L_j\biggr] \;=\; -[M] + \sum_{j} [L_j].$

The short exact sequence $0 \to K \otimes \omega_X \to M \to \bigoplus_j L_j \to 0$ makes $[M] = [K \otimes \omega_X] + \sum_j [L_j]$ additive in $K_0$, and $[K \otimes \omega_X] = [K]$ in $K_{\mathrm{num}}$ (since $\mathrm{ch}(\omega_X) = 1$ on the algebraic Mukai lattice, by the same argument as in Section 8.2). Substituting:

$[\mathsf{S}_{\mathrm{Ku}}(E)] \;=\; -\bigl([K] + \sum_{j} [L_j]\bigr) + \sum_{j} [L_j] \;=\; -[K] \;=\; [E],$

where the final equality $-[K] = [E]$ is the class identity from Section 8.4 ($E \cong K[1]$ shifts the class by a sign), and now both sides are bona fide classes in $K_{\mathrm{num}}(\mathrm{Ku}(X))$ since $E \in \mathrm{Ku}(X)$. But the Serre sign relation derived in Section 8.3 demands

$[\mathsf{S}_{\mathrm{Ku}}(E)] \;=\; -[E].$

Equating the two values:

$[E] \;=\; -[E] \quad \text{in } K_{\mathrm{num}}(\mathrm{Ku}(X)), \qquad \text{i.e.,} \qquad 2[E] \;=\; 0.$

The numerical Grothendieck group $K_{\mathrm{num}}(\mathrm{Ku}(X))$ is a finite-rank free abelian group — this is a standard structural fact for the numerical Grothendieck group of any admissible subcategory of the derived category of a smooth projective variety, going back to Bayer–Lahoz–Macri–Stellari and made explicit in Lemma 12.7 of Bayer–Lahoz–Macri–Nuer–Perry–Stellari. Concretely, $K_{\mathrm{num}}(\mathrm{Ku}(X)) \cong \ZZ^2$, generated by classes such as $\lambda_1 = \pi([\mathcal{O}_X])$ and $\lambda_2 = \pi([\mathcal{O}_p]) = [E]$, where $\pi$ denotes orthogonal projection onto $\langle L_\bullet\rangle^\perp$.

Since $K_{\mathrm{num}}(\mathrm{Ku}(X))$ is torsion-free, the relation $2[E] = 0$ forces $[E] = 0$. To see that this is impossible, view $[E] \in K_{\mathrm{num}}(\mathrm{Ku}(X))$ through the inclusion $K_{\mathrm{num}}(\mathrm{Ku}(X)) \hookrightarrow K_{\mathrm{num}}(D^b(X))$ — injective because the semiorthogonal decomposition $D^b(X) = \langle L_1, \ldots, L_{10}, \mathrm{Ku}(X)\rangle$ induces a direct-sum decomposition of $K_{\mathrm{num}}$. The image is $[E] = [\mathcal{O}_p] - \sum_{i=1}^{10} [L_i]$, whose rank component is $0 - 10 = -10 \neq 0$. So $[E] \neq 0$ in $K_{\mathrm{num}}(D^b(X))$, hence in $K_{\mathrm{num}}(\mathrm{Ku}(X))$. Contradiction.

Two remarks. First, unnodality is essential — not merely a technical convenience: if $X$ contains a $(-2)$-curve, the orthogonal collection $L_1, \dots, L_{10}$ fails to be completely orthogonal, the projection $\zeta_K^!$ becomes more delicate, the Lemma-4.4 vanishings used in computing $\mathsf{S}_{\mathrm{Ku}}(E)$ break down, and the 3-spherical objects $S_i$ that pin the cover-element matrix $T$ need not be available. The non-existence statement holds in the unnodal locus; for nodal Enriques surfaces, an analogous non-existence is conjectured but not yet established (see the discussion in the introduction of Li–Pertusi–Zhao, arXiv:2104.13610).

Second, the obstruction is Serre-invariance specifically, not the existence of stability conditions outright. Bridgeland stability conditions on the ambient category $D^b(X)$ for Enriques surfaces are known to exist (Bayer–Macri–Stellari); whether they exist on $\mathrm{Ku}(X)$ for unnodal $X$ is conjecturally false and remains open. What we have ruled out here is the natural compatibility with Serre duality that, for cubic threefolds, cubic fourfolds, and Gushel–Mukai cases, has driven recent moduli-of-sheaves geometry. The Enriques case is genuinely different: the two-torsion of $\omega_X$ — the same algebraic feature that distinguishes Enriques surfaces from K3 surfaces in the Kodaira classification — reaches into the categorical structure of $\mathrm{Ku}(X)$ via the non-trivial mutation autoequivalence $\Phi$, and prevents any choice of stability that would treat Serre duality symmetrically.