betaFeedback

WorksMathematics

Koszul duality for boundary categories

Abstract

We study how Koszul duality relates two kinds of boundary constructions associated with a smooth category and a proper category. For a smooth category with a proper Koszul-dualizing subcategory, we construct a fully faithful functor from its categorical neighborhood of infinity to the singularity category associated to the proper category. We prove that this functor is an equivalence precisely when the smooth category is reflexive, meaning that it can be reconstructed from its finite-dimensional module theory. Under a Gorenstein condition, satisfied in the Calabi–Yau setting, the functor is induced by restriction of modules. We give algebraic criteria that ensure equivalence and obtain an identification of Hochschild cohomology.

Introduction

Background and motivation

Singularity categories measure the failure of coherent complexes to be perfect and thereby make singularities accessible to homological algebra [18]. Efimov’s categorical formal punctured neighborhood of infinity provides a way to study behavior at infinity directly from a category, without choosing a compactification [Efi17]. A relation between these constructions allows boundary phenomena in smooth geometry to be studied through the singularities of a proper algebraic model. Koszul duality suggests such a relation because, under suitable hypotheses, it exchanges perfect modules with modules having finite-dimensional cohomology. To turn this principle into a comparison of boundary categories, one must identify the subcategories being divided out and determine when the comparison recovers every object of the target. These are the two questions addressed in this paper.

The motivation is also particularly concrete in symplectic geometry. Ganatra–Gao–Venkatesh identify the Rabinowitz Fukaya category of a nondegenerate Liouville manifold with the categorical formal punctured neighborhood of infinity of its wrapped Fukaya category [GGV22, Theorem 1.1]. When compact Lagrangians supply an appropriate Koszul dual, a comparison with a singularity category offers an algebraic description of this Rabinowitz category. The essential-surjectivity question asks whether every object of the singularity category comes from the categorical boundary, so that this algebraic description is an equivalence.

Several earlier results establish the foundations of this comparison. Efimov relates categorical infinity to singularity categories arising from localizations of smooth proper dg categories [Efi17, Theorem 3.2]. For a proper scheme over a perfect field, he identifies the categorical formal punctured neighborhood of its dg category of bounded coherent complexes with the opposite of its singularity category [Efi17, Theorem 9.2]. His Proposition 9.3 gives a quotient description of infinity under categorical Koszul-duality hypotheses. Ganatra–Gao–Venkatesh formulate these hypotheses in terms of a proper Koszul-dualizing subcategory and identify the quotient by right-proper objects with the categorical formal punctured neighborhood [GGV22, Definition 2.12 and {xref}Proposition 2.13]. We use this quotient description to compare the boundary with the singularity category of the Koszul dual.

At the level of modules, the exchange of finiteness conditions is a familiar feature of Koszul duality. For Adams connected A∞A_{\infty} algebras with proper Koszul dual, Lu–Palmieri–Wu–Zhang prove an equivalence between perfect modules on one side and modules with finite-dimensional total cohomology on the other [LPWZ08, Theorem B and {xref}Corollary 7.2(b)]. The question of reconstructing a category from its pseudo-perfect modules is expressed more generally by *reflexivity*, introduced by Kuznetsov–Shinder {citep}[KS25, Definition 3.11]. Goodbody characterizes reflexivity in the Morita homotopy category and studies its consequences for Hochschild cohomology {citep}[Goo26, Theorems 4.4 and 5.1]. Booth–Goodbody–Opper develop its relationship with Koszul duality, generation, and completion, and establish criteria for connective and coconnective dg algebras {citep}[BGO25, Theorem 2.3.8 and § 3]`. These results provide the reconstruction property and the algebraic criteria used below.

Calabi–Yau structures give a further reason to compare the two boundaries. Brav–Dyckerhoff transfer a smooth Calabi–Yau structure to the subcategory of proper objects [BD19, Theorem 3.1], while Gao’s categorical residue produces a duality of dimension one less on the categorical formal punctured neighborhood of infinity [Gao25, Theorem 1.7]. Rivera–Takeda–Wang study the associated products on Hochschild complexes with boundary coefficients using pre-Calabi–Yau structures [19]. In his Oberwolfach report, Takeda conjectures compatibility of the products and duality structures appearing in Rabinowitz Floer theory, categorical infinity, and singular Hochschild cohomology under the relevant identifications [23 p. 3165], which leads to the general form of the following conjecture:

Conjecture 1.1. Let B\mathcal{B} be a proper Koszul dualizing subcategory in a smooth A∞A_{\infty}-category A\mathcal{A}. Then there is an isomorphism of rings:

HHsg∗(B)≅HH∗(A^∞)HH^{*}_{\mathrm{sg}}(\mathcal{B}) \cong HH_{*}(\widehat{\mathcal{A}}_{\infty})

The main purpose of this paper is to address the categorical identification needed for this perspective and give consequences for Hochschild invariants.

Main results

Fix a field k\mathbb{k}. All starting categories are small and k\mathbb{k}-linear, and we use dg module categories, passing to dg models for strictly unital A∞A_{\infty} categories when forming quotients. All tensor products and module Hom functors are computed in these models. For a category C\mathcal{C}, write Mod⁡(C)\operatorname{Mod}(\mathcal{C}) for its dg category of right modules, Perf⁡(C)\operatorname{Perf}(\mathcal{C}) for the full subcategory of perfect modules, and PsPerf⁡(C)\operatorname{PsPerf}(\mathcal{C}) for the full subcategory of pseudo-perfect modules, whose values are perfect k\mathbb{k}-complexes. Smoothness gives PsPerf⁡(C)⊂Perf⁡(C)\operatorname{PsPerf}(\mathcal{C}) \subset\operatorname{Perf}(\mathcal{C}), while properness gives the reverse inclusion. For proper B\mathcal{B}, the dg singularity category is

Dsg(B)=PsPerf⁡(B)/Perf⁡(B).\mathcal{D}_{\mathrm{sg}}(\mathcal{B})=\operatorname{PsPerf}(\mathcal{B})/\operatorname{Perf}(\mathcal{B}).

Let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be a Koszul-dualizing full inclusion in the sense of Definition 4.1, with A\mathcal{A} smooth and B\mathcal{B} proper. In particular, the graph bimodule Γj(a,b)=Hom⁡A(a,jb)\Gamma_j(a,b)=\operatorname{Hom}_{\mathcal{A}}(a,jb) is perfect over k\mathbb{k} at every pair of objects. Writing (−)∨=Hom⁡k(−,k)(-)^{\vee}=\operatorname{Hom}_{\mathbb{k}}(-,\mathbb{k}) for linear duality, define

Φ=−⊗AΓj∨:Perf⁡(A)⟶PsPerf⁡(B).(1)\Phi=-\otimes_{\mathcal{A}}\Gamma_j^{\vee}:\operatorname{Perf}(\mathcal{A})\longrightarrow\operatorname{PsPerf}(\mathcal{B}). \tag*{(1)}

For a perfect module MM, its value at bb is Φ(M)(b)≃Hom⁡A(M,YAr(jb))∨\Phi(M)(b)\simeq\operatorname{Hom}_{\mathcal{A}}(M,Y_{\mathcal{A}}^{r}(jb))^{\vee}, where YAr(jb)=Hom⁡A(−,jb)Y_{\mathcal{A}}^{r}(jb)=\operatorname{Hom}_{\mathcal{A}}(-,jb) is the right Yoneda module. Equivalently,

Φ≃(−)∨∘((jop)∗)op∘(Hom⁡A(−,A))op,(2)\Phi\simeq(-)^{\vee}\circ((j^{\mathrm{op}})^{*})^{\mathrm{op}}\circ(\operatorname{Hom}_{\mathcal{A}}(-,\mathcal{A}))^{\mathrm{op}}, \tag*{(2)}

where the rightmost functor is perfect-module duality and (jop)∗(j^{\mathrm{op}})^{*} is restriction of left modules.

We will need the notion of reflexivity. The category A\mathcal{A} is reflexive if the canonical evaluation functor

ev⁡A:Perf⁡(A)op⟶PsPerf⁡(PsPerf⁡(A)op),\operatorname{ev}_{\mathcal{A}}:\operatorname{Perf}(\mathcal{A})^{\mathrm{op}}\longrightarrow\operatorname{PsPerf}(\operatorname{PsPerf}(\mathcal{A})^{\mathrm{op}}),
M⟼[N⟼Hom⁡A(M,N)]M\longmapsto[N\longmapsto\operatorname{Hom}_{\mathcal{A}}(M,N)]

is a quasi-equivalence; see Definition 4.4 for further details. Write Perf⁡(A)^∞\widehat{\operatorname{Perf}(\mathcal{A})}_{\infty} for the categorical formal punctured neighborhood of infinity of Perf⁡(A)\operatorname{Perf}(\mathcal{A}), using the essential-image construction of Definition 5.1. Under the present hypotheses, it is quasi-equivalent to Perf⁡(A)/PsPerf⁡(A)\operatorname{Perf}(\mathcal{A})/\operatorname{PsPerf}(\mathcal{A}) by Proposition 5.2.

Theorem A. Let A\mathcal{A} be smooth and let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be a Koszul-dualizing full inclusion, with B\mathcal{B} proper. The functor Φ\Phi in (1) is quasi-fully faithful. Its composite with linear duality induces a quasi-fully faithful functor

Perf⁡(A)^∞⟶Dsg(Bop)op.(3)\widehat{\operatorname{Perf}(\mathcal{A})}_{\infty}\longrightarrow\mathcal{D}_{\mathrm{sg}}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}}. \tag*{(3)}

The following conditions are equivalent:

(i) The functor (3) is a quasi-equivalence.

(ii) The functor Φ\Phi is a quasi-equivalence.

(iii) The category A\mathcal{A} is reflexive.

If B\mathcal{B} also satisfies the Gorenstein condition (12), its Nakayama functor νB=−⊗BB∨\nu_{\mathcal{B}}=-\otimes_{\mathcal{B}}\mathcal{B}^{\vee}, where B∨\mathcal{B}^{\vee} is the dual diagonal bimodule, is an autoequivalence of Mod⁡(B)∗preserving∗\operatorname{Mod}(\mathcal{B}) *preserving*\operatorname{PsPerf}(\mathcal{B})$, and

Φ≃νB∘j∗\Phi\simeq\nu_{\mathcal{B}}\circ j^{*}

on Perf⁡(A)\operatorname{Perf}(\mathcal{A}).

Restriction j∗j^{*} is then quasi-fully faithful on Perf⁡(A)\operatorname{Perf}(\mathcal{A}) and induces a quasi-fully faithful functor

Perf⁡(A)^∞⟶Dsg(B).(4)\widehat{\operatorname{Perf}(\mathcal{A})}_{\infty}\longrightarrow\mathcal{D}_{\mathrm{sg}}(\mathcal{B}). \tag*{(4)}

This functor is a quasi-equivalence precisely when (i)–(iii) hold. These conditions are also equivalent to coinduction, the right adjoint of j∗j^{*}, restricting to a quasi-equivalence

j∗:PsPerf⁡(B)→∼Perf⁡(A),j_{*}:\operatorname{PsPerf}(\mathcal{B})\xrightarrow{\sim}\operatorname{Perf}(\mathcal{A}),

with quasi-inverse j∗j^{*}.

The proof separates reconstruction from passage to the quotient. Efimov’s generation result identifies Perf⁡(B)\operatorname{Perf}(\mathcal{B}) with PsPerf⁡(A)\operatorname{PsPerf}(\mathcal{A}); under this identification, the factorization (2) expresses Φ\Phi through the evaluation functor defining reflexivity. For the quotients, Φ\Phi sends PsPerf⁡(A)\operatorname{PsPerf}(\mathcal{A}) to the Nakayama image νB(Perf⁡(B))\nu_{\mathcal{B}}(\operatorname{Perf}(\mathcal{B})), which linear duality identifies with Perf⁡(Bop)op\operatorname{Perf}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}}. This determines the target in (3). The essential image of Φ\Phi is thick, so quotienting by the corresponding full subcategories preserves and detects essential surjectivity, by Lemma 2.8. The Gorenstein condition identifies the Nakayama image with Perf⁡(B)\operatorname{Perf}(\mathcal{B}) and gives the restriction functor in (4).

A weak proper nn-Calabi–Yau structure on B\mathcal{B} identifies B∨\mathcal{B}^{\vee} with BΔ[n]\mathcal{B}_{\Delta}[n]. It therefore implies the Gorenstein condition and gives νB≃[n]\nu_{\mathcal{B}}\simeq[n], so that Φ≃[n]∘j∗\Phi\simeq[n]\circ j^{*}. Combining Theorem A with the reflexivity criteria of Booth–Goodbody–Opper yields the following consequence. The degreewise finiteness required by those criteria follows from the Koszul-duality hypotheses: the tensor–Hom map in Definition 4.1(iii) is a biduality map, as in (54), and condition (ii) transfers this finiteness to morphisms between perfect A\mathcal{A}-modules.

Corollary B. Let A\mathcal{A} be smooth and let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be a Koszul-dualizing full inclusion. Suppose that B\mathcal{B} carries a weak proper nn-Calabi–Yau structure and that Perf⁡(A)\operatorname{Perf}(\mathcal{A}) has a split-generator GG whose dg endomorphism algebra A=Hom⁡A(G,G)A=\operatorname{Hom}_{\mathcal{A}}(G,G) satisfies either

H>0(A)=0,(5)H^{>0}(A)=0, \tag*{(5)}

or

H<0(A)=0andH0(A) is semisimple.(6)H^{<0}(A)=0 \qquad\text{and}\qquad H^{0}(A)\text{ is semisimple}. \tag*{(6)}

Then restriction induces a quasi-equivalence

Perf⁡(A)^∞→∼Dsg(B).\widehat{\operatorname{Perf}(\mathcal{A})}_{\infty}\xrightarrow{\sim}\mathcal{D}_{\mathrm{sg}}(\mathcal{B}).

Moreover, j∗j^{*} and j∗j_{*} restrict to mutually inverse quasi-equivalences between Perf⁡(A)\operatorname{Perf}(\mathcal{A}) and PsPerf⁡(B)\operatorname{PsPerf}(\mathcal{B}).

If A\mathcal{A} itself carries a weak smooth nn-Calabi–Yau structure, Theorem 3.5 gives the weak proper structure on B\mathcal{B}. When A\mathcal{A} is reflexive, Theorem 5.5 transports Gao’s categorical residue through the restriction quasi-equivalence and equips Dsg(B)\mathcal{D}_{\mathrm{sg}}(\mathcal{B}) with a weak generalized proper (n−1)(n-1)-Calabi–Yau structure in the sense of Remark 3.3. Corollary 7.2 gives the corresponding isomorphisms of Hochschild cohomology algebras, preserving cup products.

Structure of the paper

Structure of the paper. In Section 2, we recall module categories, the Nakayama functor, and the properties of dg quotients used below. In Section 3, we review weak smooth and weak proper Calabi–Yau structures and their relation to the Gorenstein condition. In Section 4, we prove that the comparison functor Φ\Phi is quasi-fully faithful and characterize its essential surjectivity by reflexivity. In Section 5, we recall the two boundary constructions and prove Theorem A and its Calabi–Yau refinement, Theorem 5.5. In Section 6, we give algebraic criteria for the comparison to be a quasi-equivalence, apply them to Koszul duals of proper algebras, and discuss geometric examples. In Section 7, we deduce comparisons of Hochschild cohomology algebras and relate them to singular Hochschild cohomology under Keller’s hypotheses.

Acknowledgments

Acknowledgments. We thank Alex Takeda for sharing his inspiring perspective, and Zhengfang Wang for helpful discussion, especially for suggesting the reflexivity property for consideration.

Categorical preliminaries

We recall the finiteness conditions, duality functors, and properties of dg quotients. All starting categories are small (except for module categories) and kk-linear, complexes are cohomologically graded, and modules are right modules unless otherwise specified. Tensor products and Hom functors are derived. We use the standard module and Morita theory of dg and strictly unital A∞A_\infty categories; see [4, 5, 24, 20].

Basic definitions

Definition 2.1. Let C\mathcal{C} be a small dg or strictly unital A∞A_\infty category.

  1. Mod⁡(C)\operatorname{Mod}(\mathcal{C}) is the dg category of right C\mathcal{C}-modules, defined as the dg category of dg or A∞A_\infty functors Fun⁡(Cop⁡,Mod⁡(k))\operatorname{Fun}(\mathcal{C}^{\operatorname{op}},\operatorname{Mod}(k)).

  1. Perf⁡(C)\operatorname{Perf}(\mathcal{C}) is the split-closed pretriangulated envelope of the right Yoneda modules YCr(X):=Hom⁡C(−,X)Y_{\mathcal{C}}^{r}(X):=\operatorname{Hom}_{\mathcal{C}}(-,X). Equivalently, it is the full subcategory of compact objects of Mod⁡(C)\operatorname{Mod}(\mathcal{C}), i.e. Perf⁡(C)=Mod⁡(C)ω\operatorname{Perf}(\mathcal{C})=\operatorname{Mod}(\mathcal{C})^\omega.

  1. PsPerf⁡(C)\operatorname{PsPerf}(\mathcal{C}) is the full subcategory of modules MM for which M(X)M(X) is a perfect kk-complex for every X∈CX\in\mathcal{C}. Such modules are called or pseudo-perfect or proper.

  1. An object X∈CX\in\mathcal{C} is right proper if its right Yoneda module YCr(X)Y_{\mathcal{C}}^{r}(X) is proper. The full subcategory of right-proper objects is denoted Cprop⁡\mathcal{C}^{\operatorname{prop}}.

There are following quasi-equivalences of module categories:

Hom⁡C(−,C):Perf⁡(C)op⁡→∼Perf⁡(Cop⁡).(7)\operatorname{Hom}_{\mathcal{C}}(-,\mathcal{C}) : \operatorname{Perf}(\mathcal{C})^{\operatorname{op}} \xrightarrow{\sim} \operatorname{Perf}(\mathcal{C}^{\operatorname{op}}). \tag*{(7)}
(−)∨:PsPerf⁡(C)op⁡→∼PsPerf⁡(Cop⁡).(8)(-)^\vee: \operatorname{PsPerf}(\mathcal{C})^{\operatorname{op}} \xrightarrow{\sim} \operatorname{PsPerf}(\mathcal{C}^{\operatorname{op}}). \tag*{(8)}

Remark 2.2. For a strictly unital A∞A_\infty category, the Yoneda embedding gives an A∞A_\infty quasi-equivalence from C\mathcal{C} to the full dg subcategory of Mod⁡(C)\operatorname{Mod}(\mathcal{C}) on its representable modules [20 Sections 2f–2g]. We use these dg models when forming tensor products of dg categories and Drinfeld quotients.

Set Ce=Cop⁡⊗kC\mathcal{C}^{\mathrm e}=\mathcal{C}^{\operatorname{op}}\otimes_k\mathcal{C}, and use the diagonal convention

CΔ(X,Y)=Hom⁡C(Y,X).\mathcal{C}_{\Delta}(X,Y)=\operatorname{Hom}_{\mathcal{C}}(Y,X).

Thus the diagonal bimodule CΔ\mathcal{C}_{\Delta} is a right Ce\mathcal{C}^{\mathrm e}-module, covariant in XX and contravariant in YY.

Definition 2.3. The category C\mathcal{C} is smooth if CΔ∈Perf⁡(Ce)\mathcal{C}_{\Delta} \in\operatorname{Perf}(\mathcal{C}^{\mathrm{e}}). It is proper if Hom⁡C(X,Y)∈Perf⁡(k)\operatorname{Hom}_{\mathcal{C}}(X,Y) \in\operatorname{Perf}(k) for all X,Y∈CX,Y \in\mathcal{C}, or equivalently if CΔ∈PsPerf⁡(Ce)\mathcal{C}_{\Delta} \in\operatorname{PsPerf}(\mathcal{C}^{\mathrm{e}}). Over the field kk, the latter condition means that each morphism complex has finite-dimensional total cohomology.

The following elementary facts relate these finiteness conditions and will also be used to check preservation of pseudo-perfect modules.

Lemma 2.4. A module MM is pseudo-perfect if and only if Hom⁡C(P,M)∈Perf⁡(k)\operatorname{Hom}_{\mathcal{C}}(P,M) \in\operatorname{Perf}(k) for every P∈Perf⁡(C)P \in\operatorname{Perf}(\mathcal{C}). Moreover,

C proper⟹Perf⁡(C)⊂PsPerf⁡(C),\mathcal{C}\ \mathrm{proper} \quad\Longrightarrow\quad\operatorname{Perf}(\mathcal{C}) \subset\operatorname{PsPerf}(\mathcal{C}),
C smooth⟹PsPerf⁡(C)⊂Perf⁡(C).\mathcal{C}\ \mathrm{smooth} \quad\Longrightarrow\quad\operatorname{PsPerf}(\mathcal{C}) \subset\operatorname{Perf}(\mathcal{C}).

Proof. The first assertion follows from Hom⁡C(YCr(X),M)≃M(X)\operatorname{Hom}_{\mathcal{C}}(Y_{\mathcal{C}}^{r}(X),M) \simeq M(X) and closure of Perf⁡(k)\operatorname{Perf}(k) under finite cones and direct summands. If C\mathcal{C} is proper, its right Yoneda modules are pseudo-perfect, and the same closure argument gives the first inclusion. If C\mathcal{C} is smooth, tensor a finite resolution of CΔ\mathcal{C}_{\Delta} from representable bimodules with a pseudo-perfect module MM. Since

M⊗C(YCl(X)⊗kYCr(Y))≃M(X)⊗kYCr(Y)M \otimes_{\mathcal{C}} \left(Y_{\mathcal{C}}^{l}(X) \otimes_k Y_{\mathcal{C}}^{r}(Y)\right) \simeq M(X) \otimes_k Y_{\mathcal{C}}^{r}(Y)

is perfect, so is M≃M⊗CCΔM \simeq M \otimes_{\mathcal{C}} \mathcal{C}_{\Delta}.

The Nakayama functor

The dual diagonal bimodule is defined to be C∨(X,Y)=Hom⁡C(X,Y)∨\mathcal{C}^{\vee}(X,Y) = \operatorname{Hom}_{\mathcal{C}}(X,Y)^{\vee}. The Nakayama functor is

νC=−⊗CC∨:Mod⁡(C)⟶Mod⁡(C).(9)\nu_{\mathcal{C}} = - \otimes_{\mathcal{C}} \mathcal{C}^{\vee} : \operatorname{Mod}(\mathcal{C}) \longrightarrow\operatorname{Mod}(\mathcal{C}). \tag*{(9)}

For P∈Perf⁡(C)P \in\operatorname{Perf}(\mathcal{C}), tensor–Hom duality gives

νC(P)≃Hom⁡C(P,C)∨.\nu_{\mathcal{C}}(P) \simeq\operatorname{Hom}_{\mathcal{C}}(P,\mathcal{C})^{\vee}.

Its right adjoint is

νCR=Hom⁡C(C∨,−):Mod⁡(C)⟶Mod⁡(C).(10)\nu_{\mathcal{C}}^{R} = \operatorname{Hom}_{\mathcal{C}}(\mathcal{C}^{\vee},-) : \operatorname{Mod}(\mathcal{C}) \longrightarrow\operatorname{Mod}(\mathcal{C}). \tag*{(10)}

Suppose now that C\mathcal{C} is proper, and put

IC=thick⁡Mod⁡(C){YCl(X)∨:X∈C}⊂PsPerf⁡(C).(11)\mathcal{I}_{\mathcal{C}} = \operatorname{thick}_{\operatorname{Mod}(\mathcal{C})}\{Y_{\mathcal{C}}^{l}(X)^{\vee}: X \in\mathcal{C}\} \subset\operatorname{PsPerf}(\mathcal{C}). \tag*{(11)}

This is the image of the perfect modules under the Nakayama functor, which we will show in Lemma 2.7. We also recall the following definition.

Definition 2.5. A proper category C\mathcal{C} satisfies the Gorenstein condition if

IC=Perf⁡(C)(12)\mathcal{I}_{\mathcal{C}} = \operatorname{Perf}(\mathcal{C}) \tag*{(12)}

as full subcategories of Mod⁡(C)\operatorname{Mod}(\mathcal{C}). For proper dg algebras, this is the condition of [1 Definition 5.2].

Remark 2.6. For an ordinary finite-dimensional algebra BB, one has IB=thick⁡(B∨)=K⁡b(inj⁡-B)\mathcal{I}_{B} = \operatorname{thick}(B^{\vee}) = \operatorname{K}^{\mathrm{b}}(\operatorname{inj}\text{-}B), where inj⁡-B\operatorname{inj}\text{-}B denotes the finitely generated injective right modules. Gorenstein condition (12) is equivalent to finite injective dimension of both regular modules BB{}_{B}B and BBB_{B}; see [Che11, Corollary 3.9] and [1 Remark 5.3].

Lemma 2.7. If C\mathcal{C} is proper, the Nakayama functor restricts to a quasi-equivalence

νC:Perf⁡(C)→∼IC.\nu_{\mathcal{C}} : \operatorname{Perf}(\mathcal{C}) \xrightarrow{\sim} \mathcal{I}_{\mathcal{C}}.

For P∈Perf⁡(C)P \in\operatorname{Perf}(\mathcal{C}) and M∈Mod⁡(C)M \in\operatorname{Mod}(\mathcal{C}), there is a natural quasi-isomorphism

Hom⁡C(P,M)∨≃Hom⁡C(M,νCP).(13)\operatorname{Hom}_{\mathcal{C}}(P,M)^{\vee} \simeq\operatorname{Hom}_{\mathcal{C}}(M,\nu_{\mathcal{C}}P). \tag*{(13)}

If C\mathcal{C} satisfies the Gorenstein condition, then νC\nu_{\mathcal{C}} is a Serre functor on Perf⁡(C)\operatorname{Perf}(\mathcal{C}) and an autoequivalence of Mod⁡(C)\operatorname{Mod}(\mathcal{C}) preserving PsPerf⁡(C)\operatorname{PsPerf}(\mathcal{C}). Its inverse is νCR\nu_{\mathcal{C}}^{R}.

Proof. On perfect modules, νC\nu_{\mathcal{C}} is the composite of the dualities in (7), using the inclusion Perf⁡(Cop⁡)⊂PsPerf⁡(Cop⁡)\operatorname{Perf}(\mathcal{C}^{\operatorname{op}}) \subset\operatorname{PsPerf}(\mathcal{C}^{\operatorname{op}}) since C\mathcal{C} is proper. It is therefore quasi-fully faithful. Since it sends YCr(X)Y_{\mathcal{C}}^{r}(X) to YCl(X)∨Y_{\mathcal{C}}^{l}(X)^{\vee} and its source is idempotent complete, its essential image is precisely JC\mathcal{J}_{\mathcal{C}}. For representable PP, (13) is the tensor–Hom adjunction

Hom⁡C(M,YCl(X)∨)≃(M⊗CYCl(X))∨≃M(X)∨;\operatorname{Hom}_{\mathcal{C}}(M,Y_{\mathcal{C}}^{l}(X)^{\vee}) \simeq(M\otimes_{\mathcal{C}}Y_{\mathcal{C}}^{l}(X))^{\vee} \simeq M(X)^{\vee};

the general case follows by finite cones and direct summands.

Under (12), this gives a Serre autoequivalence of Perf⁡(C)\operatorname{Perf}(\mathcal{C}). The bar tensor functor νC\nu_{\mathcal{C}} preserves direct sums and cones and induces a quasi-equivalence on the compact objects, so its colimit-preserving extension gives an autoequivalence of Mod⁡(C)\operatorname{Mod}(\mathcal{C}) [5 Lemma 3.10]. By Lemma 2.4, pseudo-perfectness is characterized by perfectness over kk of the Hom complexes from all perfect modules. This condition is preserved by νC\nu_{\mathcal{C}} and its inverse, proving the last assertion. □

Drinfeld quotient

Let N⊂C\mathcal{N} \subset\mathcal{C} be a full dg subcategory. The dg quotient is a dg category C/N\mathcal{C}/\mathcal{N} equipped with a functor

q:C⟶C/Nq:\mathcal{C}\longrightarrow\mathcal{C}/\mathcal{N}

that annihilates N\mathcal{N} and is universal with this property in the homotopy category of dg categories [Dri04, Theorem 1.6.2], [5 Theorem 4.8]. More precisely, precomposition with qq identifies morphisms C/N→D\mathcal{C}/\mathcal{N}\to\mathcal{D} in Ho⁡(dgCat⁡k)\operatorname{Ho}(\operatorname{dgCat}_{k}) with morphisms C→D\mathcal{C}\to\mathcal{D} that send every object of N\mathcal{N} to a contractible object. Over the field kk, a concrete model has the same objects as C\mathcal{C} and freely adjoins, for each N∈NN\in\mathcal{N}, an endomorphism εN\varepsilon_{N} of degree −1-1 with dεN=id⁡Nd\varepsilon_{N}=\operatorname{id}_{N} [Dri04, Section 3.1]. Its morphism complexes admit the description

Hom⁡C/N(X,Y)≃cone⁡(YCr(Y)∣N⊗NYCl(X)∣N⟶Hom⁡C(X,Y)).(14)\operatorname{Hom}_{\mathcal{C}/\mathcal{N}}(X,Y)\simeq\operatorname{cone}\left(Y_{\mathcal{C}}^{r}(Y)|_{\mathcal{N}}\otimes_{\mathcal{N}}Y_{\mathcal{C}}^{l}(X)|_{\mathcal{N}}\longrightarrow\operatorname{Hom}_{\mathcal{C}}(X,Y)\right). \tag*{(14)}

where the arrow is induced by composition [Dri04, Section 1.3.1].

If C\mathcal{C} and N\mathcal{N} are pretriangulated, then C/N\mathcal{C}/\mathcal{N} is pretriangulated and there is an equivalence of triangulated categories

H0(C/N)≃H0(C)/H0(N)H^{0}(\mathcal{C}/\mathcal{N})\simeq H^{0}(\mathcal{C})/H^{0}(\mathcal{N})

with the Verdier quotient [Dri04, Theorem 3.4]. In particular,

q(X)≃0⟺X∈thick⁡H0(C)H0(N)q(X)\simeq0 \quad\Longleftrightarrow\quad X\in\operatorname{thick}_{H^{0}(\mathcal{C})}H^{0}(\mathcal{N})

by [Dri04, Remark A.3(ii)] or [8 Proposition 4.6.2(3)]. Consequently, replacing N\mathcal{N} by its thick closure inside C\mathcal{C} gives a quasi-equivalent dg quotient. Quasi-equivalences of dg pairs also induce quasi-equivalences of their quotients [Dri04, Remark 3.6(ii)]. For A∞A_{\infty} categories, one may use dg models or the quotient constructions of [17, 16].

The following lemma shows that quotienting by a common subcategory preserves full faithfulness and, under a thickness hypothesis, detects essential surjectivity; see [Dri04, Lemma A.5], [8 Proposition 4.6.2, Lemma 4.7.1] for the arguments for triangulated categories.

Lemma 2.8. Let N⊂C⊂D\mathcal{N}\subset\mathcal{C}\subset\mathcal{D} be full pretriangulated dg subcategories. The induced functor

jˉ:C/N⟶D/N\bar{j}:\mathcal{C}/\mathcal{N}\longrightarrow\mathcal{D}/\mathcal{N}

is quasi-fully faithful. If H0(C)H^{0}(\mathcal{C}) is thick in H0(D)H^{0}(\mathcal{D}), then jˉ\bar{j} is essentially surjective if and only if the inclusion j:C↪Dj:\mathcal{C}\hookrightarrow\mathcal{D} is essentially surjective.

Proof. For X,Y∈CX,Y \in\mathcal{C}, the cone descriptions (14) in C/N\mathcal{C}/\mathcal{N} and D/N\mathcal{D}/\mathcal{N} agree because both inclusions are full. Thus jˉ\bar{j} is quasi-fully faithful.

Suppose that H0(C)H^0(\mathcal{C}) is thick in H0(D)H^0(\mathcal{D}). The quotient D→D/C\mathcal{D} \to\mathcal{D}/\mathcal{C} factors through D/N\mathcal{D}/\mathcal{N}. By the kernel description above, an object represented by X∈DX \in\mathcal{D} vanishes in H0(D/C)H^0(\mathcal{D}/\mathcal{C}) precisely when XX is isomorphic in H0(D)H^0(\mathcal{D}) to an object of C\mathcal{C}. Hence

ess im⁡H0(jˉ)=ker⁡(H0(D/N)⟶H0(D/C)).(15)\operatorname{ess\,im} H^0(\bar{j}) = \ker\left(H^0(\mathcal{D}/\mathcal{N}) \longrightarrow H^0(\mathcal{D}/\mathcal{C})\right). \tag*{(15)}

If jˉ\bar{j} is essentially surjective, every object of D\mathcal{D} therefore belongs to H0(C)H^0(\mathcal{C}) up to isomorphism. Conversely, essential surjectivity of jj passes to the quotients, which have the same objects as their respective source categories. □

Remark 2.9. Under the thickness hypothesis of Lemma 2.8, the kernel as in (15) shows that the essential image of H0(jˉ)H^0(\bar{j}) is thick. A quasi-fully faithful functor between pretriangulated dg categories is a Morita equivalence precisely when its image generates the target under shifts, cones, and direct summands [5 Lemma 3.10]. Thus jˉ\bar{j} is a Morita equivalence if and only if it is a quasi-equivalence.

Calabi–Yau structures

We recall weak smooth and weak proper Calabi–Yau structures, which will be used in Theorem 5.5 and Corollary B, respectively. Lemma 3.4 shows that a weak proper Calabi–Yau structure implies the Gorenstein condition (12) required in Theorem A for the functor (4).

Write C∗(C)=C∗(C,CΔ)C_\ast(\mathcal{C}) = C_\ast(\mathcal{C},\mathcal{C}_{\Delta}) for the Hochschild chain complex with diagonal coefficients, computed by the cyclic bar construction, and set

HHm(C)=H−m(C∗(C)),m∈Z.HH_m(\mathcal{C}) = H^{-m}(C_\ast(\mathcal{C})), \qquad m \in\mathbb{Z}.

We regard C∗(C)C_\ast(\mathcal{C}) as a cochain complex; the chain maps defining Calabi–Yau structures have degree zero. Thus a class in HHn(C)HH_n(\mathcal{C}) is represented by a chain map k[n]→C∗(C)\Bbbk[n] \to C_\ast(\mathcal{C}). For the cyclic bar construction, see [Gao25, Section 2.4].

The definition of a weak smooth Calabi–Yau structure uses the inverse dualizing bimodule C!\mathcal{C}^{!}, the Ce\mathcal{C}^{e}-linear dual of CΔ\mathcal{C}_{\Delta}, regarded as a right Ce\mathcal{C}^{e}-module by exchanging the factors of Ce=Cop⊗kC\mathcal{C}^{e} = \mathcal{C}^{\mathrm{op}} \otimes_{\Bbbk} \mathcal{C}. Explicitly,

C!(X,Y)=Hom⁡Ce(CΔ,YCℓ(Y)⊗kYCr(X)).\mathcal{C}^{!}(X,Y) = \operatorname{Hom}_{\mathcal{C}^{e}}\left(\mathcal{C}_{\Delta},Y_{\mathcal{C}}^{\ell}(Y) \otimes_{\Bbbk} Y_{\mathcal{C}}^{r}(X)\right).

The tensor product in the second argument is the bimodule whose value at (U,V)(U,V) is Hom⁡C(Y,U)⊗kHom⁡C(V,X)\operatorname{Hom}_{\mathcal{C}}(Y,U) \otimes_{\Bbbk} \operatorname{Hom}_{\mathcal{C}}(V,X). If C\mathcal{C} is smooth, perfect duality for CΔ\mathcal{C}_{\Delta} gives a natural quasi-isomorphism

C∗(C)≃Hom⁡Ce(C!,CΔ);(16)C_\ast(\mathcal{C}) \simeq\operatorname{Hom}_{\mathcal{C}^{e}}(\mathcal{C}^{!},\mathcal{C}_{\Delta}); \tag*{(16)}

see [BD19, Section 3.2] and [Gan23, Remark 6.7].

Definition 3.1. Suppose that C\mathcal{C} is smooth. A weak smooth nn-Calabi–Yau structure is a class [ωC]∈HHn(C)[\omega_{\mathcal{C}}] \in HH_n(\mathcal{C}), represented by

ωC:k[n]⟶C∗(C),(17)\omega_{\mathcal{C}}:\Bbbk[n] \longrightarrow C_\ast(\mathcal{C}), \tag*{(17)}

for which the bimodule morphism corresponding under (16) is a quasi-isomorphism

C![n]→∼CΔ.(18)\mathcal{C}^{!}[n] \xrightarrow{\sim} \mathcal{C}_{\Delta}. \tag*{(18)}

See [Gan23, Definition 6.5 and Remark 6.7].

The tensor–Hom adjunction gives a natural quasi-isomorphism

Hom⁡k(C∗(C),k)≃Hom⁡Ce(CΔ,C∨),(19)\operatorname{Hom}_{k}(C_{*}(\mathcal{C}),k)\simeq\operatorname{Hom}_{\mathcal{C}^{e}}(\mathcal{C}_{\Delta},\mathcal{C}^{\vee}), \tag*{(19)}

where the right-hand side is the morphism complex in Mod⁡(Ce)\operatorname{Mod}(\mathcal{C}^{e}); see [BD19, Section 3.1] and [Gan23, Remark 6.2].

Definition 3.2. Suppose that C\mathcal{C} is proper. A weak proper nn-Calabi–Yau structure is a chain map, called a Hochschild trace,

tr⁡:C∗(C)⟶k[−n],(20)\operatorname{tr}:C_{*}(\mathcal{C})\longrightarrow k[-n], \tag*{(20)}

taken up to chain homotopy, for which the bimodule morphism corresponding under (19) is a quasi-isomorphism

CΔ⟶∼C∨[−n].(21)\mathcal{C}_{\Delta}\overset{\sim}{\longrightarrow}\mathcal{C}^{\vee}[-n]. \tag*{(21)}

For a trace as in (20), this quasi-isomorphism condition is equivalent to requiring that composition followed by the trace on length-zero Hochschild chains induce perfect pairings

HiHom⁡C(X,Y)⊗kHn−iHom⁡C(Y,X)⟶k(22)H^{i}\operatorname{Hom}_{\mathcal{C}}(X,Y)\otimes_{k}H^{n-i}\operatorname{Hom}_{\mathcal{C}}(Y,X)\longrightarrow k \tag*{(22)}

for every X,Y∈CX,Y\in\mathcal{C} and i∈Zi\in\mathbb{Z}; see [Gan23, Definition 6.1 and Remark 6.2].

Remark 3.3. In Definition 3.2, one can drop the a priori condition of properness on C\mathcal{C}. The perfect pairing, induced by the Hochschild trace (20) which determines the bimodule quasi-isomorphism (21) via the derived tensor–Hom adjunction

Hom⁡k(C∗(C),k[−n])≃Hom⁡Ce(CΔ,C∨[−n]),\operatorname{Hom}_{k}(C_{*}(\mathcal{C}),k[-n])\simeq\operatorname{Hom}_{\mathcal{C}^{e}}(\mathcal{C}_{\Delta},\mathcal{C}^{\vee}[-n]),

implies the finiteness of Hom spaces on homology level at each degree

Hi(Hom⁡C(X,Y))→∼Hn−i(Hom⁡C(Y,X))∨H^{i}(\operatorname{Hom}_{\mathcal{C}}(X,Y))\xrightarrow{\sim}H^{n-i}(\operatorname{Hom}_{\mathcal{C}}(Y,X))^{\vee}

(whereas the total dimension of homology can nevertheless be infinite), and this means C\mathcal{C} is a locally proper dg or A∞A_{\infty} category. In this spirit, we refer to this milder notion as weak (absolute) locally proper Calabi–Yau structure on C\mathcal{C}; this is also named a componentwise proper Calabi–Yau structure by Keller–Wang. There is an even weaker notion in [Gao25, Definition 2.27], obtained by requiring that the pairings (22) be nondegenerate rather than perfect. The corresponding bimodule morphism CΔ→C∨[−n]\mathcal{C}_{\Delta}\to\mathcal{C}^{\vee}[-n] is then injective on cohomology, with no surjectivity requirement, and the morphism cohomology groups may be infinite-dimensional in each degree. We call such a structure a weak generalized proper nn-Calabi–Yau structure.

The next lemma verifies the Gorenstein hypothesis of Theorem A for the category B\mathcal{B} in Corollary B.

Lemma 3.4. If C\mathcal{C} carries a weak proper nn-Calabi–Yau structure, then νC≃[n]\nu_{\mathcal{C}}\simeq[n] on Mod⁡(C)\operatorname{Mod}(\mathcal{C}). In particular, C\mathcal{C} satisfies the Gorenstein condition (12), and the Serre functor on Perf⁡(C)\operatorname{Perf}(\mathcal{C}) is the shift [n][n].

Proof. The quasi-isomorphism (21) gives C∨≃CΔ[n]\mathcal{C}^{\vee}\simeq\mathcal{C}_{\Delta}[n]. Substituting this identification into the definition (9) of the Nakayama functor yields

νC=−⊗CC∨≃−⊗CCΔ[n]≃[n].\nu_{\mathcal{C}}=-\otimes_{\mathcal{C}}\mathcal{C}^{\vee}\simeq-\otimes_{\mathcal{C}}\mathcal{C}_{\Delta}[n]\simeq[n].

By Lemma 2.7, the essential image of νC\nu_{\mathcal{C}} on Perf⁡(C)\operatorname{Perf}(\mathcal{C}) is IC\mathcal{I}_{\mathcal{C}}. Since [n][n] is an autoequivalence of Perf⁡(C)\operatorname{Perf}(\mathcal{C}), we obtain IC=Perf⁡(C)\mathcal{I}_{\mathcal{C}}=\operatorname{Perf}(\mathcal{C}). The Serre duality for [n][n] is (13) restricted to perfect modules. ∎

We also recall the transfer from weak smooth to weak proper Calabi–Yau structures underlying [BD19, Theorem 3.1], which will be used in Theorem 5.5.

Theorem 3.5 ([BD19, Theorem 3.1]). Let C\mathcal{C} be a smooth dg or strictly unital A∞A_\infty category. A weak smooth nn-Calabi–Yau structure on C\mathcal{C}, in the sense of Definition 3.1, induces a weak proper nn-Calabi–Yau structure on Cprop⁡\mathcal{C}^{\operatorname{prop}}, and hence on every full subcategory of Cprop⁡\mathcal{C}^{\operatorname{prop}}.

For a Koszul-dualizing inclusion j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A}, condition (i) in Definition 4.1 says that every jbjb is right proper in A\mathcal{A}. A weak smooth nn-Calabi–Yau structure on A\mathcal{A} in the sense of Definition 3.1 therefore induces a weak proper nn-Calabi–Yau structure on B\mathcal{B} by Theorem 3.5. Lemma 3.4 then verifies the Gorenstein hypothesis in Proposition 4.3 and Theorem A.

Koszul duality and reflexivity

In this section, we prove that the functor Φ\Phi in (1), for a Koszul-dualizing inclusion j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A}, is quasi-fully faithful, and furthermore that it is essentially surjective precisely when A\mathcal{A} is reflexive. Under the Gorenstein condition (12), we also identify Φ\Phi with νB∘j∗\nu_{\mathcal{B}}\circ j^* and obtain the restriction and coinduction functors appearing in Theorem A.

Koszul-dualizing inclusions

Let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be a full inclusion. The standard induction, restriction, and coinduction functors satisfy

j!⊣j∗⊣j∗;(23)j_! \dashv j^* \dashv j_*; \tag*{(23)}

see [5 Section 3.8]. With the right-module convention of § 2, these take the following explicit forms:

(j∗M)(b)=M(jb),(j!P)(a)=P⊗BΓj(a,−),(j∗P)(a)=Hom⁡B(j∗YA(a),P),(24)\begin{aligned} (j^*M)(b) &= M(jb),\\ (j_!P)(a) &= P\otimes_{\mathcal{B}}\Gamma_j(a,-),\\ (j_*P)(a) &= \operatorname{Hom}_{\mathcal{B}}(j^*\mathsf{Y}_{\mathcal{A}}(a),P), \tag*{(24)} \end{aligned}

where M∈Mod⁡(A)M\in\operatorname{Mod}(\mathcal{A}), P∈Mod⁡(B)P\in\operatorname{Mod}(\mathcal{B}), and Γj(a,b)=Hom⁡A(a,jb)\Gamma_j(a,b)=\operatorname{Hom}_{\mathcal{A}}(a,jb) as in the introduction. Thus induction uses the left B\mathcal{B}-module Hom⁡A(a,j(−))\operatorname{Hom}_{\mathcal{A}}(a,j(-)), whereas coinduction uses the right B\mathcal{B}-module Hom⁡A(j(−),a)\operatorname{Hom}_{\mathcal{A}}(j(-),a). Fullness of jj implies that the unit Id⁡→j∗j!\operatorname{Id}\to j^*j_! and the counit j∗j∗→Id⁡j^*j_*\to\operatorname{Id} are quasi-isomorphisms. In particular, j!j_! and j∗j_* are quasi-fully faithful.

We recall the notion of a Koszul-dualizing subcategory from [GGV22, Definition 2.12], based on the hypotheses of [Efi17, Proposition 9.3].

Definition 4.1. Suppose that A\mathcal{A} is smooth and B\mathcal{B} is proper. A full inclusion j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} is Koszul-dualizing if the following conditions hold:

(i) Each jbjb is right proper in A\mathcal{A}, in the sense of Definition 2.1; equivalently, Γj(a,b)∈Perf⁡(k)\Gamma_j(a,b)\in\operatorname{Perf}(k) for every a∈Aa\in\mathcal{A} and b∈Bb\in\mathcal{B}.

(ii) Restriction of left modules is quasi-fully faithful on perfect modules:

(jop⁡)∗:Perf⁡(Aop⁡)⟶PsPerf⁡(Bop⁡).(25)(j^{\operatorname{op}})^*:\operatorname{Perf}(\mathcal{A}^{\operatorname{op}})\longrightarrow\operatorname{PsPerf}(\mathcal{B}^{\operatorname{op}}). \tag*{(25)}

(iii) For P,Q∈PsPerf⁡(Bop⁡)P,Q\in\operatorname{PsPerf}(\mathcal{B}^{\operatorname{op}}), evaluation of module morphisms induces a quasi-isomorphism

P∨⊗BQ⟶Hom⁡Bop⁡(Q,P)∨.(26)P^\vee\otimes_{\mathcal{B}}Q\longrightarrow\operatorname{Hom}_{\mathcal{B}^{\operatorname{op}}}(Q,P)^\vee. \tag*{(26)}

Condition (i) ensures that the target in (25) is well defined. For the same reason, the formula

Φ(M)(b)≃Hom⁡A(M,YA(jb))∨,M∈Perf⁡(A),(27)\Phi(M)(b)\simeq\operatorname{Hom}_{\mathcal{A}}(M,\mathsf{Y}_{\mathcal{A}}(jb))^\vee,\qquad M\in\operatorname{Perf}(\mathcal{A}), \tag*{(27)}

defines a pseudo-perfect right B\mathcal{B}-module. By the dualities (7) and (8), condition (ii) makes the functor Φ\Phi of (2) quasi-fully faithful.

Lemma 4.2. Let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be Koszul-dualizing. Induction and restriction give mutually inverse quasi-equivalences

j!:Perf⁡(B)→∼PsPerf⁡(A),j∗:PsPerf⁡(A)→∼Perf⁡(B).(28)j_!:\operatorname{Perf}(\mathcal{B}) \xrightarrow{\sim} \operatorname{PsPerf}(\mathcal{A}),\qquad j^*:\operatorname{PsPerf}(\mathcal{A}) \xrightarrow{\sim} \operatorname{Perf}(\mathcal{B}). \tag*{(28)}

Consequently,

PsPerf⁡(A)=thick⁡Perf⁡(A){YAr(jb):b∈B},(29)\operatorname{PsPerf}(\mathcal{A})=\operatorname{thick}_{\operatorname{Perf}(\mathcal{A})}\{Y^{\mathrm{r}}_{\mathcal{A}}(jb):b\in\mathcal{B}\}, \tag*{(29)}

and Φ\Phi restricts to a quasi-equivalence

Φ:PsPerf⁡(A)→∼IB,(30)\Phi:\operatorname{PsPerf}(\mathcal{A})\xrightarrow{\sim}\mathcal{I}_{\mathcal{B}}, \tag*{(30)}

where IB\mathcal{I}_{\mathcal{B}} is defined in (11).

Proof. By Definition 4.1(i), Γj\Gamma_j is pseudo-perfect, and fullness of jj makes j!j_! quasi-fully faithful. Conditions (ii)–(iii) give the remaining hypotheses of [Efi17, Proposition 9.3], whose categorical proof applies over any field. It follows that j!:Perf⁡(B)→PsPerf⁡(A)j_!:\operatorname{Perf}(\mathcal{B})\to\operatorname{PsPerf}(\mathcal{A}) is a quasi-equivalence, with quasi-inverse j∗j^* by (23). Since j!YBr(b)≃YAr(jb)j_!Y^{\mathrm{r}}_{\mathcal{B}}(b)\simeq Y^{\mathrm{r}}_{\mathcal{A}}(jb), the generation statement (29) follows.

By fullness of jj and (27),

Φ(YAr(jb))≃YBl(b)∨≃νBYBr(b).(31)\Phi(Y^{\mathrm{r}}_{\mathcal{A}}(jb))\simeq Y^{\mathrm{l}}_{\mathcal{B}}(b)^\vee\simeq\nu_{\mathcal{B}}Y^{\mathrm{r}}_{\mathcal{B}}(b). \tag*{(31)}

Since Φ\Phi is quasi-fully faithful, (29) identifies its essential image on PsPerf⁡(A)\operatorname{PsPerf}(\mathcal{A}) with the thick subcategory generated by these modules, which is IB\mathcal{I}_{\mathcal{B}} as defined in (11).

Relation to the Nakayama functor

The Gorenstein condition allows us to recover restriction from Φ\Phi by the inverse Nakayama functor. This identifies the additional functors in Theorem A without changing the essential-surjectivity question.

Proposition 4.3. Let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be Koszul-dualizing, and suppose that B\mathcal{B} satisfies (12). There is a natural quasi-isomorphism

Φ≃νB∘j∗on Perf⁡(A).(32)\Phi\simeq\nu_{\mathcal{B}}\circ j^* \quad\text{on }\operatorname{Perf}(\mathcal{A}). \tag*{(32)}

In particular, restriction gives a quasi-fully faithful functor

j∗:Perf⁡(A)⟶PsPerf⁡(B),j^*:\operatorname{Perf}(\mathcal{A})\longrightarrow\operatorname{PsPerf}(\mathcal{B}),

whose restriction to PsPerf⁡(A)\operatorname{PsPerf}(\mathcal{A}) is the quasi-inverse of j!j_! in (28). The functor Φ\Phi is essentially surjective if and only if coinduction restricts to a quasi-equivalence

j∗:PsPerf⁡(B)→∼Perf⁡(A).(33)j_*:\operatorname{PsPerf}(\mathcal{B})\xrightarrow{\sim}\operatorname{Perf}(\mathcal{A}). \tag*{(33)}

When these conditions hold, j∗j^* and j∗j_* are mutually inverse on these subcategories.

Proof. For M∈Perf⁡(A)M\in\operatorname{Perf}(\mathcal{A}) and b∈Bb\in\mathcal{B}, quasi-full faithfulness of Φ\Phi, (31), and the adjunction νB⊣νBR\nu_{\mathcal{B}}\dashv\nu_{\mathcal{B}}^{\mathrm{R}} give

(j∗M)(b)≃Hom⁡A(YAr(jb),M)≃Hom⁡B(νBYBr(b),Φ(M))≃Hom⁡B(YBr(b),νBRΦ(M))≃(νBRΦ(M))(b).\begin{aligned} (j^*M)(b)&\simeq\operatorname{Hom}_{\mathcal{A}}(Y^{\mathrm{r}}_{\mathcal{A}}(jb),M)\\ &\simeq\operatorname{Hom}_{\mathcal{B}}(\nu_{\mathcal{B}}Y^{\mathrm{r}}_{\mathcal{B}}(b),\Phi(M))\\ &\simeq\operatorname{Hom}_{\mathcal{B}}(Y^{\mathrm{r}}_{\mathcal{B}}(b),\nu_{\mathcal{B}}^{\mathrm{R}}\Phi(M))\\ &\simeq(\nu_{\mathcal{B}}^{\mathrm{R}}\Phi(M))(b). \end{aligned}

These identifications are natural in MM and bb, so j∗≃νBR∘Φj^*\simeq\nu_{\mathcal{B}}^{\mathrm{R}}\circ\Phi. By Lemma 2.7, the Gorenstein condition makes νB\nu_{\mathcal{B}} an autoequivalence of Mod⁡(B)\operatorname{Mod}(\mathcal{B}) with inverse νBR\nu_{\mathcal{B}}^{\mathrm{R}}, and both functors preserve PsPerf⁡(B)\operatorname{PsPerf}(\mathcal{B}). This proves (32) and quasi-full faithfulness of j∗j^* on Perf⁡(A)\operatorname{Perf}(\mathcal{A}). Its assertion on pseudo-perfect modules is (28).

Since νB\nu_{\mathcal{B}} restricts to an autoequivalence of PsPerf⁡(B)\operatorname{PsPerf}(\mathcal{B}), (32) shows that Φ\Phi is essentially surjective precisely when j∗j^* is. For M∈Perf⁡(A)M \in\operatorname{Perf}(\mathcal{A}), the unit M→j∗j∗MM \to j_*j^*M is a quasi-isomorphism: at a∈Aa \in\mathcal{A}, this follows from

(j∗j∗M)(a)≃Hom⁡B(j∗YA(a),j∗M)≃Hom⁡A(YA(a),M)≃M(a).(j_*j^*M)(a) \simeq\operatorname{Hom}_{\mathcal{B}}(j^*Y_{\mathcal{A}}(a),j^*M) \simeq\operatorname{Hom}_{\mathcal{A}}(Y_{\mathcal{A}}(a),M) \simeq M(a).

If j∗j^* is essentially surjective, every P∈PsPerf⁡(B)P \in\operatorname{PsPerf}(\mathcal{B}) has the form j∗Mj^*M up to quasi-isomorphism, and hence j∗P≃Mj_*P \simeq M is perfect. The adjunction j∗⊣j∗j^* \dashv j_* therefore restricts to the quasi-equivalences asserted in (33). Conversely, if (33) holds, then j∗j∗P≃Pj^*j_*P \simeq P for every pseudo-perfect PP by fullness of jj, proving essential surjectivity of j∗j^* and of Φ\Phi. □

Reflexivity

To characterize essential surjectivity of Φ\Phi, we relate it to the notion of reflexivity in the sense of [BGO25, Definition 2.1.1 and Remark 2.1.5(2)]. For a small dg or strictly unital A∞A_\infty category C\mathcal{C}, Lemma 2.4 ensures that the evaluation functor

ev⁡C:Perf⁡(C)op⁡⟶PsPerf⁡(PsPerf⁡(C)op⁡)\operatorname{ev}_{\mathcal{C}}:\operatorname{Perf}(\mathcal{C})^{\operatorname{op}}\longrightarrow\operatorname{PsPerf}(\operatorname{PsPerf}(\mathcal{C})^{\operatorname{op}})
M⟼[N⟼Hom⁡C(M,N)](34)M\longmapsto[N\longmapsto\operatorname{Hom}_{\mathcal{C}}(M,N)] \tag*{(34)}

is well defined.

Definition 4.4. The category C\mathcal{C} is reflexive if (34) is a quasi-equivalence, and semireflexive if it is quasi-fully faithful.

Reflexivity is Morita invariant [2 Proposition 3.8] and invariant under passage to the opposite category [BGO25, Remark 2.1.5(1)]. For A∞A_\infty categories, we use the dg models fixed in Section 2; the induced equivalences of module categories identify their evaluation functors.

Proposition 4.5. Let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be Koszul-dualizing. Then A\mathcal{A} is semireflexive, and the functor Φ\Phi in (1) is essentially surjective if and only if A\mathcal{A} is reflexive.

Proof. By Lemma 4.2, the functor

i:B⟶PsPerf⁡(A),b⟼YAr(jb),i:\mathcal{B}\longrightarrow\operatorname{PsPerf}(\mathcal{A}),\qquad b\longmapsto Y_{\mathcal{A}}^{r}(jb),

is a Morita equivalence. Restriction along iop⁡i^{\operatorname{op}} therefore induces a quasi-equivalence

(iop⁡)∗:PsPerf⁡(PsPerf⁡(A)op⁡)⟶∼PsPerf⁡(Bop⁡);(35)(i^{\operatorname{op}})^*:\operatorname{PsPerf}(\operatorname{PsPerf}(\mathcal{A})^{\operatorname{op}})\overset{\sim}{\longrightarrow}\operatorname{PsPerf}(\mathcal{B}^{\operatorname{op}}); \tag*{(35)}

see [2 Proposition 3.5]. For M∈Perf⁡(A)M \in\operatorname{Perf}(\mathcal{A}), the composite (iop⁡)∗ev⁡A(i^{\operatorname{op}})^*\operatorname{ev}_{\mathcal{A}} takes the value

((iop⁡)∗ev⁡A(M))(b)=Hom⁡A(M,YAr(jb)).((i^{\operatorname{op}})^*\operatorname{ev}_{\mathcal{A}}(M))(b)=\operatorname{Hom}_{\mathcal{A}}(M,Y_{\mathcal{A}}^{r}(jb)).

It is thus naturally identified with

Perf⁡(A)op⁡→∼Hom⁡A(−,A)Perf⁡(Aop⁡)→(iop⁡)∗PsPerf⁡(Bop⁡).(36)\operatorname{Perf}(\mathcal{A})^{\operatorname{op}}\xrightarrow[\sim]{\operatorname{Hom}_{\mathcal{A}}(-,\mathcal{A})}\operatorname{Perf}(\mathcal{A}^{\operatorname{op}})\xrightarrow{(i^{\operatorname{op}})^*}\operatorname{PsPerf}(\mathcal{B}^{\operatorname{op}}). \tag*{(36)}

The first arrow is the duality (7), and the second is quasi-fully faithful by Definition 4.1(ii). Together with (35), this proves that ev⁡A\operatorname{ev}_{\mathcal{A}} is quasi-fully faithful. Moreover, ev⁡A\operatorname{ev}_{\mathcal{A}} is essentially surjective precisely when (36) is. Taking the opposite of (36) and applying the linear duality (8) gives Φ\Phi, as in (2). These dualities preserve essential surjectivity, proving the assertion. □

Boundary categories and Koszul duality

We describe the two quotient categories in Theorem A, then use the results of Section 4 to construct the comparison functor and characterize when it is a quasi-equivalence. We also prove Theorem 5.5 by transporting the categorical residue through this equivalence. The final subsection recalls a trace on the singularity category obtained from Hochschild localization.

Categorical formal punctured neighborhood of infinity

The k\mathbb{k}-linear Calkin dg category is the Drinfeld quotient

q:Mod⁡(k)⟶Calk⁡(k):=Mod⁡(k)/Perf⁡(k);q:\operatorname{Mod}(\mathbb{k})\longrightarrow\operatorname{Calk}(\mathbb{k}):=\operatorname{Mod}(\mathbb{k})/\operatorname{Perf}(\mathbb{k});

see [Efi17, Proposition 1.1] or [GGV22, Section 2.2]. For a small dg or strictly unital A∞A_{\infty} category A\mathcal{A}, composition with qq gives the Calkin-valued Yoneda functor

Y‾:A⟶Fun⁡(Aop⁡,Calk⁡(k)),\overline{Y}:\mathcal{A}\longrightarrow\operatorname{Fun}(\mathcal{A}^{\operatorname{op}},\operatorname{Calk}(\mathbb{k})),
a⟼q∘YAop⁡(a).(37)a\longmapsto q\circ Y_{\mathcal{A}}^{\operatorname{op}}(a). \tag*{(37)}

Here Fun⁡\operatorname{Fun} denotes the dg category of strictly unital A∞A_{\infty} functors and A∞A_{\infty} natural transformations, which models the homotopy internal Hom; see [Efi17, Section 1 and Proposition 3.6] and [GGV22, Sections 2.1 and 2.3].

Definition 5.1. The categorical formal punctured neighborhood of infinity of A\mathcal{A} is the full essential image

A^∞:=ess im⁡(Y‾)⊂Fun⁡(Aop⁡,Calk⁡(k)).\widehat{\mathcal{A}}_{\infty}:=\operatorname{ess\,im}(\overline{Y})\subset\operatorname{Fun}(\mathcal{A}^{\operatorname{op}},\operatorname{Calk}(\mathbb{k})).

We use the full essential image before pretriangulated or idempotent completion, as in [Efi17, Definition 3.3(3)].

Since qq annihilates precisely the perfect k\mathbb{k}-complexes, an object a∈Aa\in\mathcal{A} satisfies

Y‾(a)≃0⟺a∈Aprop⁡,\overline{Y}(a)\simeq0\quad\Longleftrightarrow\quad a\in\mathcal{A}^{\operatorname{prop}},

where right-proper objects are defined in Definition 2.1; see also [GGV22, Lemma 2.8]. When A\mathcal{A} is smooth, the right-proper objects of Perf⁡(A)\operatorname{Perf}(\mathcal{A}) are exactly PsPerf⁡(A)\operatorname{PsPerf}(\mathcal{A}) by Lemma 2.4; compare [Efi17, Proposition 4.1]. Under the Koszul-duality hypotheses, the functor induced on the corresponding quotient is a quasi-equivalence.

Proposition 5.2 ([Efi17, Proposition 9.3],[GGV22, Proposition 2.13]). Let A\mathcal{A} be smooth and let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be Koszul-dualizing in the sense of Definition 4.1. The Calkin-valued Yoneda functor induces quasi-equivalences

A/B→∼A^∞(38)\mathcal{A}/\mathcal{B}\xrightarrow{\sim}\widehat{\mathcal{A}}_{\infty} \tag*{(38)}

and

Perf⁡(A)/PsPerf⁡(A)→∼Perf⁡(A)^∞.(39)\operatorname{Perf}(\mathcal{A})/\operatorname{PsPerf}(\mathcal{A})\xrightarrow{\sim}\widehat{\operatorname{Perf}(\mathcal{A})}_{\infty}. \tag*{(39)}

The induced functor A^∞→Perf⁡(A)^∞\widehat{\mathcal{A}}_{\infty}\to\widehat{\operatorname{Perf}(\mathcal{A})}_{\infty} is a Morita equivalence.

Singularity category

Let B\mathcal{B} be a proper dg or strictly unital A∞A_{\infty} category, so that Perf⁡(B)⊂PsPerf⁡(B)\operatorname{Perf}(\mathcal{B})\subset\operatorname{PsPerf}(\mathcal{B}) by Lemma 2.4. We use the dg version of the singularity category recalled in [Efi17, Section 2.1]. In that reference, local properness is the properness condition of Definition 2.3.

Definition 5.3. The dg singularity category of B\mathcal{B} is the Drinfeld quotient

Dsg⁡(B):=PsPerf⁡(B)/Perf⁡(B).(40)\mathcal{D}_{\operatorname{sg}}(\mathcal{B}):=\operatorname{PsPerf}(\mathcal{B})/\operatorname{Perf}(\mathcal{B}). \tag*{(40)}

Its homotopy category is the Verdier quotient

H0(Dsg⁡(B))≃H0(PsPerf⁡(B))/H0(Perf⁡(B))H^{0}(\mathcal{D}_{\operatorname{sg}}(\mathcal{B}))\simeq H^{0}(\operatorname{PsPerf}(\mathcal{B}))/H^{0}(\operatorname{Perf}(\mathcal{B}))

by [Dri04, Theorem 3.4]. Since Perf⁡(B)\operatorname{Perf}(\mathcal{B}) is thick in PsPerf⁡(B)\operatorname{PsPerf}(\mathcal{B}), the quotient vanishes precisely when every pseudo-perfect B\mathcal{B}-module is perfect. In particular, it vanishes when B\mathcal{B} is smooth, by Lemma 2.4.

For a Koszul-dualizing inclusion j:B↪Aj : \mathcal{B} \hookrightarrow\mathcal{A}, the comparison functor Φ\Phi sends PsPerf⁡(A)\operatorname{PsPerf}(\mathcal{A}) to JB\mathcal{J}_{\mathcal{B}} by (30). The following consequence of linear duality identifies the quotient by this subcategory.

Lemma 5.4. Linear duality induces a quasi-equivalence

PsPerf⁡(B)/JB→∼Dsg(Bop)op,[N]⟼[N∨],(41)\operatorname{PsPerf}(\mathcal{B})/\mathcal{J}_{\mathcal{B}} \xrightarrow{\sim} \mathcal{D}_{\mathrm{sg}}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}}, \qquad[N] \longmapsto[N^{\vee}], \tag*{(41)}

where JB\mathcal{J}_{\mathcal{B}} is defined in (11).

Proof. The duality (8) identifies PsPerf⁡(B)\operatorname{PsPerf}(\mathcal{B}) with PsPerf⁡(Bop)op\operatorname{PsPerf}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}}. It sends each generator YBl(b)∨Y_{\mathcal{B}}^{\mathrm{l}}(b)^{\vee} of JB\mathcal{J}_{\mathcal{B}} to the representable left module YBl(b)Y_{\mathcal{B}}^{\mathrm{l}}(b), since B\mathcal{B} is proper. It therefore identifies JB\mathcal{J}_{\mathcal{B}} with Perf⁡(Bop)op\operatorname{Perf}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}} and induces (41) on dg quotients. Pointwise double duality gives the inverse. □

If B\mathcal{B} satisfies the Gorenstein condition (12), then JB=Perf⁡(B)\mathcal{J}_{\mathcal{B}} = \operatorname{Perf}(\mathcal{B}), so Lemma 5.4 gives a linear-duality quasi-equivalence Dsg(Bop)op≃Dsg(B)\mathcal{D}_{\mathrm{sg}}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}} \simeq\mathcal{D}_{\mathrm{sg}}(\mathcal{B}). For a weak proper nn-Calabi–Yau structure, Lemma 3.4 further gives νB−1≃[−n]\nu_{\mathcal{B}}^{-1} \simeq[-n]. Composing linear duality with this inverse Nakayama functor gives the normalization

(−)∨[−n]:Dsg(Bop)op→∼Dsg(B).(42)(-)^{\vee}[-n] : \mathcal{D}_{\mathrm{sg}}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}} \xrightarrow{\sim} \mathcal{D}_{\mathrm{sg}}(\mathcal{B}). \tag*{(42)}

Comparison under Koszul duality

We now apply Φ\Phi to the quotient presentations (39) and (41) to prove Theorem A.

Proof of Theorem A. By Definition 4.1(ii) and the factorization (2), the functor Φ\Phi is quasi-fully faithful. Lemma 4.2 identifies its restriction to pseudo-perfect modules with a quasi-equivalence

Φ:PsPerf⁡(A)→∼JB.\Phi: \operatorname{PsPerf}(\mathcal{A}) \xrightarrow{\sim} \mathcal{J}_{\mathcal{B}}.

Thus, after identifying Perf⁡(A)\operatorname{Perf}(\mathcal{A}) with the full essential image of Φ\Phi, we have full pretriangulated subcategories

JB⊂ess im⁡(Φ)⊂PsPerf⁡(B).\mathcal{J}_{\mathcal{B}} \subset\operatorname{ess\,im}(\Phi) \subset\operatorname{PsPerf}(\mathcal{B}).

The essential image is thick on homotopy categories: exactness and full faithfulness give closure under shifts and cones, while idempotents lift to H0(Perf⁡(A))H^{0}(\operatorname{Perf}(\mathcal{A})) and split there. Lemma 2.8 therefore gives a quasi-fully faithful functor on the corresponding dg quotients. Using Proposition 5.2 and Lemma 5.4, we obtain

Perf⁡^(A)∞≃Perf⁡(A)/PsPerf⁡(A)→Φ‾PsPerf⁡(B)/JB→(−)∨Dsg(Bop)op,(43)\begin{aligned} \widehat{\operatorname{Perf}}(\mathcal{A})_{\infty} &\simeq\operatorname{Perf}(\mathcal{A})/\operatorname{PsPerf}(\mathcal{A}) \\ &\xrightarrow{\overline{\Phi}} \operatorname{PsPerf}(\mathcal{B})/\mathcal{J}_{\mathcal{B}} \xrightarrow{(-)^{\vee}} \mathcal{D}_{\mathrm{sg}}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}}, \tag*{(43)} \end{aligned}

where Φ‾\overline{\Phi} is induced by Φ\Phi. The composite is the functor (3).

The same application of Lemma 2.8 shows that Φ‾\overline{\Phi} is essentially surjective if and only if Φ\Phi is, proving (i) ⟺\Longleftrightarrow (ii). Proposition 4.5 identifies essential surjectivity of Φ\Phi with reflexivity of A\mathcal{A}, giving (ii) ⟺\Longleftrightarrow (iii).

Suppose now that B\mathcal{B} satisfies the Gorenstein condition (12). Then JB=Perf⁡(B)\mathcal{J}_{\mathcal{B}} = \operatorname{Perf}(\mathcal{B}), and Lemma 2.7 makes νB\nu_{\mathcal{B}} an autoequivalence of Mod⁡(B)\operatorname{Mod}(\mathcal{B}) preserving both Perf⁡(B)\operatorname{Perf}(\mathcal{B}) and PsPerf⁡(B)\operatorname{PsPerf}(\mathcal{B}). Proposition 4.3 gives

Φ≃νB∘j∗,j∗≃νB−1∘Φon Perf⁡(A).\Phi\simeq\nu_{\mathcal{B}} \circ j^{*}, \qquad j^{*} \simeq\nu_{\mathcal{B}}^{-1} \circ\Phi\quad\text{on } \operatorname{Perf}(\mathcal{A}).

Hence j∗j^{*} is quasi-fully faithful, and (28) identifies its restriction to PsPerf⁡(A)\operatorname{PsPerf}(\mathcal{A}) with a quasi-equivalence onto Perf⁡(B)\operatorname{Perf}(\mathcal{B}). On the dg quotients, the functor induced by j∗j^{*} is therefore Φ‾\overline{\Phi} followed by the autoequivalence induced by νB−1\nu_{\mathcal{B}}^{-1}. This proves the assertions about (4).

Finally, Proposition 4.3 identifies condition (ii) with the assertion that j∗j_* restricts to the quasi-equivalence (33), with quasi-inverse j∗j^*.

We next use the restriction functor of Theorem A to transport the categorical residue of [Gao25, Theorem 1.7] to the singularity category.

Theorem 5.5. Suppose that A\mathcal{A} carries a weak smooth nn-Calabi–Yau structure and is reflexive, and let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be a Koszul-dualizing full inclusion. Restriction induces a quasi-equivalence

j^∗:Perf⁡(A^)∞→∼Dsg(B).(44)\widehat{j}^{*}:\operatorname{Perf}(\widehat{\mathcal{A}})_{\infty}\xrightarrow{\sim}\mathcal{D}_{\mathrm{sg}}(\mathcal{B}). \tag*{(44)}

The categorical residue transports through this quasi-equivalence to a weak generalized proper (n−1)(n-1)-Calabi–Yau structure on Dsg(B)\mathcal{D}_{\mathrm{sg}}(\mathcal{B}), in the sense of Remark 3.3.

Proof. Condition (i) in Definition 4.1 places jBj\mathcal{B} in Aprop\mathcal{A}^{\mathrm{prop}}. Theorem 3.5 therefore gives a weak proper nn-Calabi–Yau structure on B\mathcal{B}. By Lemma 3.4, the category B\mathcal{B} satisfies the Gorenstein condition and νB≃[n]\nu_{\mathcal{B}}\simeq[n]. Since A\mathcal{A} is reflexive, Theorem A gives the quasi-equivalence (44). Moreover, Φ≃[n]∘j∗\Phi\simeq[n]\circ j^* by (32), so composing (3) with the duality (42) gives this same restriction functor: double duality recovers Φ\Phi, and the shift [−n][-n] cancels its Nakayama shift. The weak smooth structure extends to Perf⁡(A)\operatorname{Perf}(\mathcal{A}) by Morita invariance. By [Gao25, Theorem 1.7 and equation (3.67)], it determines a residue trace

res⁡:C∗(Perf⁡(A^)∞)⟶k[1−n](45)\operatorname{res}:C_*(\operatorname{Perf}(\widehat{\mathcal{A}})_{\infty})\longrightarrow\mathbb{k}[1-n] \tag*{(45)}

whose pairings on morphism cohomology are nondegenerate. The quasi-isomorphism C∗(j∗)C_*(j^*) identifies homotopy classes of traces by precomposition. There is therefore a unique homotopy class of traces on Dsg(B)\mathcal{D}_{\mathrm{sg}}(\mathcal{B}) whose pullback is the class of (45). Its pairings are nondegenerate because j∗j^* identifies morphism cohomology and composition. This gives the asserted weak generalized proper Calabi–Yau structure. □

For a proper scheme XX over a perfect field kk, let Coh⁡(X)\operatorname{Coh}(X) denote the dg category of complexes with bounded coherent cohomology and let ωX\omega_X be its dualizing complex. Efimov’s geometric comparison is

Coh⁡(X^)∞≃Coh⁡(X)ωX⊗OXPerf⁡(X)≃Dsg(X)op;(46)\operatorname{Coh}(\widehat{X})_{\infty}\simeq\frac{\operatorname{Coh}(X)}{\omega_X\otimes_{\mathcal{O}_X}\operatorname{Perf}(X)}\simeq\mathcal{D}_{\mathrm{sg}}(X)^{\mathrm{op}}; \tag*{(46)}

see [Efi17, Theorem 9.2 and Proposition B.1] and [Gao25, Section 4]. The denominator is the full essential image of tensoring perfect complexes with ωX\omega_X, and the second quasi-equivalence is induced by Grothendieck–Verdier duality. For Gorenstein XX, the denominator is Perf⁡(X)\operatorname{Perf}(X); if moreover XX has dimension nn and ωX≃OX[n]\omega_X\simeq\mathcal{O}_X[n], then [Gao25, Theorem 1.6] gives the corresponding weak generalized proper (n−1)(n-1)-Calabi–Yau structure on Dsg(X)\mathcal{D}_{\mathrm{sg}}(X). In characteristic zero, the chosen trivialization supplies the weak smooth structure on Coh⁡(X)\operatorname{Coh}(X) by [BD19, Section 5.2 and Proposition 5.12].

Traces on the singularity category

For a proper category B\mathcal{B} with a weak proper nn-Calabi–Yau structure, Hochschild localization gives a trace directly on Dsg(B)\mathcal{D}_{\mathrm{sg}}(\mathcal{B}). We recall this construction and the finiteness condition under which its pairings are perfect. By Morita invariance, regard the trace (20) as a trace on Perf⁡(B)\operatorname{Perf}(\mathcal{B}). Localization for Hochschild chains gives an exact triangle

C∗(Perf⁡(B))⟶C∗(PsPerf⁡(B))⟶C∗(Dsg(B))⟶C∗(Perf⁡(B))[1];(47)C_*(\operatorname{Perf}(\mathcal{B}))\longrightarrow C_*(\operatorname{PsPerf}(\mathcal{B}))\longrightarrow C_*(\mathcal{D}_{\mathrm{sg}}(\mathcal{B})) \longrightarrow C_*(\operatorname{Perf}(\mathcal{B}))[1]; \tag*{(47)}

see [5 Theorem 5.2(a)]. Thus the induced trace is the composite

tr⁡sg:=tr⁡[1]∘∂:C∗(Dsg(B))⟶k[1−n].(48)\operatorname{tr}_{\mathrm{sg}} := \operatorname{tr}[1] \circ\partial: C_*(\mathcal{D}_{\mathrm{sg}}(\mathcal{B})) \longrightarrow\mathbb{k}[1-n]. \tag*{(48)}

The following nondegeneracy statement is a specialization of [3 Theorem 4.3.1(b)].

Theorem 5.6. Let B\mathcal{B} carry a weak proper nn-Calabi--Yau structure, and suppose that

dim⁡kHiHom⁡B(M,N)<∞for all M,N∈PsPerf⁡(B) and i∈Z.(49)\dim_{\mathbb{k}} H^i\operatorname{Hom}_{\mathcal{B}}(M,N)<\infty\qquad\text{for all } M,N\in\operatorname{PsPerf}(\mathcal{B})\text{ and }i\in\mathbb{Z}. \tag*{(49)}

Then the trace (48) induces a bimodule quasi-isomorphism

Dsg(B)Δ≃Dsg(B)∨[1−n].\mathcal{D}_{\mathrm{sg}}(\mathcal{B})_{\Delta}\simeq\mathcal{D}_{\mathrm{sg}}(\mathcal{B})^{\vee}[1-n].

It therefore defines a weak locally proper (n−1)(n-1)-Calabi--Yau structure in the sense of Remark 3.3.

Proof. Put C=PsPerf⁡(B)\mathcal{C}=\operatorname{PsPerf}(\mathcal{B}) and P=Perf⁡(B)\mathcal{P}=\operatorname{Perf}(\mathcal{B}). Condition (49) makes the canonical map CΔ→CΔ∨\mathcal{C}_{\Delta}\to\mathcal{C}^{\vee}_{\Delta} a quasi-isomorphism. Moreover, the restricted Yoneda functor

C⟶Mod⁡(P),M⟼[P⟼Hom⁡B(P,M)],\mathcal{C}\longrightarrow\operatorname{Mod}(\mathcal{P}),\qquad M\longmapsto[P\longmapsto\operatorname{Hom}_{\mathcal{B}}(P,M)],

is quasi-fully faithful, since it is the restriction of the quasi-equivalence of dg module categories induced by B→Perf⁡(B)\mathcal{B}\to\operatorname{Perf}(\mathcal{B}). These verify the two hypotheses of [3 Theorem 4.3.1(b)] for P⊂C\mathcal{P}\subset\mathcal{C}. Apply that theorem with the shifted diagonal CΔ[n]\mathcal{C}_{\Delta}[n], corresponding to the shift autoequivalence [−n][-n] in its notation. The weak proper Calabi--Yau structure gives PΔ[n]≃P∨\mathcal{P}_{\Delta}[n]\simeq\mathcal{P}^{\vee}, and the theorem identifies its image under the connecting morphism with a quasi-isomorphism

Dsg(B)Δ[n]≃Dsg(B)∨[1].\mathcal{D}_{\mathrm{sg}}(\mathcal{B})_{\Delta}[n]\simeq\mathcal{D}_{\mathrm{sg}}(\mathcal{B})^{\vee}[1].

Shifting by [−n][-n] gives the assertion. □

The finiteness hypothesis (49) cannot be omitted from this general statement.

Example 5.7. Let B\mathcal{B} have objects bib_i, i≥1i\geq1, with Hom⁡B(bi,bj)=kδij\operatorname{Hom}_{\mathcal{B}}(b_i,b_j)=\mathbb{k}\delta_{ij} concentrated in degree zero. It is proper, and the trace

C∗(B)≃⨁i≥1k⟶k,(λi)i⟼∑iλi,C_*(\mathcal{B})\simeq\bigoplus_{i\geq1}\mathbb{k}\longrightarrow\mathbb{k},\qquad(\lambda_i)_i\longmapsto\sum_i\lambda_i,

defines a weak proper 00-Calabi--Yau structure. A pseudo-perfect module is a sequence of perfect k\mathbb{k}-complexes; it is perfect over B\mathcal{B} precisely when only finitely many components are nonzero up to quasi-isomorphism. Thus P=(k,k,…)P=(\mathbb{k},\mathbb{k},\ldots) is pseudo-perfect, and (14) gives

Hom⁡Dsg(B)(P,P)≃cone⁡(⨁i≥1k↪∏i≥1k).\operatorname{Hom}_{\mathcal{D}_{\mathrm{sg}}(\mathcal{B})}(P,P)\simeq\operatorname{cone}\left(\bigoplus_{i\geq1}\mathbb{k}\hookrightarrow\prod_{i\geq1}\mathbb{k}\right).

Its cohomology is concentrated in degree zero, where it is the nonzero vector space ∏ik/⨁ik\prod_i\mathbb{k}/\bigoplus_i\mathbb{k}. A weak generalized proper (−1)(-1)-Calabi--Yau structure would give a nondegenerate pairing between H0Hom⁡Dsg(B)(P,P)≠0H^0\operatorname{Hom}_{\mathcal{D}_{\mathrm{sg}}(\mathcal{B})}(P,P)\neq0 and H−1Hom⁡Dsg(B)(P,P)=0H^{-1}\operatorname{Hom}_{\mathcal{D}_{\mathrm{sg}}(\mathcal{B})}(P,P)=0. Hence no such structure exists.

Algebraic criteria

We express the hypotheses of Theorem A in terms of endomorphism algebras, prove Corollary B, and apply the comparison to Koszul duals of proper algebras. We use the dg module models fixed in Section 2, and write Hom for their morphism complexes. For an algebra BB, the category PsPerf⁡(B)\operatorname{PsPerf}(B) consists of modules with finite-dimensional total cohomology.

Split generators and endomorphism algebras

Let j:B↪Aj:\mathcal{B}\hookrightarrow\mathcal{A} be a full inclusion, with A\mathcal{A} smooth and B\mathcal{B} proper. Choose split-generators G∈Perf⁡(A)G\in\operatorname{Perf}(\mathcal{A}) and K∈Perf⁡(B)K\in\operatorname{Perf}(\mathcal{B}), and put

A=Hom⁡A(G,G),B=Hom⁡B(K,K),M=Hom⁡B(K,Φ(G)).(50)A=\operatorname{Hom}_{\mathcal{A}}(G,G),\qquad B=\operatorname{Hom}_{\mathcal{B}}(K,K),\qquad M=\operatorname{Hom}_{\mathcal{B}}(K,\Phi(G)). \tag*{(50)}

Smoothness supplies GG by [Efi10, Lemma 2.5]; the existence of KK is an additional assumption. The Morita equivalences induced by GG and KK identify Φ\Phi with

−⊗AM:Perf⁡(A)⟶PsPerf⁡(B),-\otimes_A M:\operatorname{Perf}(A)\longrightarrow\operatorname{PsPerf}(B),

where MM is an AA–BB-bimodule with action αmβ=Φ(α)∘m∘β\alpha m\beta=\Phi(\alpha)\circ m\circ\beta; see [5 Section 3].

Proposition 6.1. Assume that each jbj b is right proper in A\mathcal{A}, as in Definition 4.1(i). Then Φ\Phi takes values in PsPerf⁡(B)\operatorname{PsPerf}(\mathcal{B}), and it is quasi-fully faithful if and only if its action morphism

A⟶Hom⁡B(M,M),α⟼[m⟼Φ(α)∘m](51)A\longrightarrow\operatorname{Hom}_B(M,M),\qquad\alpha\longmapsto\left[m\longmapsto\Phi(\alpha)\circ m\right] \tag*{(51)}

is a quasi-isomorphism. In this case its essential image, under the preceding Morita equivalence, is thick⁡B(M)\operatorname{thick}_B(M).

Suppose that (51) is a quasi-isomorphism. If

PsPerf⁡(B)=thick⁡B(M),(52)\operatorname{PsPerf}(B)=\operatorname{thick}_B(M), \tag*{(52)}

then Φ\Phi is essentially surjective and hence a quasi-equivalence. If, moreover,

dim⁡kHd(A)<∞for every d∈Z,(53)\dim_k H^d(A)<\infty\qquad\text{for every }d\in\mathbb{Z}, \tag*{(53)}

then jj is Koszul-dualizing and A\mathcal{A} is reflexive.

Proof. The right-properness assumption and (27) imply that Φ\Phi takes values in PsPerf⁡(B)\operatorname{PsPerf}(\mathcal{B}). The map (51) is the map on endomorphism complexes of the generator AA. Closure under shifts, finite cones, and direct summands in both variables shows that Φ\Phi is quasi-fully faithful precisely when this map is a quasi-isomorphism. In that case its essential image is thick⁡B(M)\operatorname{thick}_B(M). By (2), quasi-full faithfulness of Φ\Phi is equivalent to condition (ii) of Definition 4.1.

Assume that (51) is a quasi-isomorphism and that (52) holds. The description of the essential image then shows that Φ\Phi is essentially surjective. Thus it induces a quasi-equivalence

Perf⁡(A)⟶∼PsPerf⁡(B).\operatorname{Perf}(A)\overset{\sim}{\longrightarrow}\operatorname{PsPerf}(B).

Assume in addition (53). Closure under shifts, finite cones, and direct summands extends degreewise finiteness from AA to all morphism complexes in Perf⁡(A)\operatorname{Perf}(A). The preceding quasi-equivalence and linear duality therefore imply this finiteness for all morphism complexes in PsPerf⁡(Bop⁡)\operatorname{PsPerf}(B^{\operatorname{op}}). For P,Q∈PsPerf⁡(Bop⁡)P,Q\in\operatorname{PsPerf}(B^{\operatorname{op}}), tensor–Hom adjunction gives

(P∨⊗BQ)∨≃Hom⁡Bop⁡(Q,P).(54)(P^{\vee}\otimes_B Q)^{\vee}\simeq\operatorname{Hom}_{B^{\operatorname{op}}}(Q,P). \tag*{(54)}

Under this identification, the map (26) is the canonical biduality map for P∨⊗BQP^{\vee}\otimes_B Q. Degreewise finiteness makes this map a quasi-isomorphism, proving condition (iii). Condition (i) holds by assumption, and condition (ii) follows from quasi-full faithfulness of Φ\Phi. Hence jj is Koszul-dualizing. Since Φ\Phi is essentially surjective, Proposition 4.5 now implies that A\mathcal{A} is reflexive. □

The following reflexivity criteria supply the two cases in Corollary B.

Proposition 6.2. ([BG025, Corollaries 3.4.3, 3.3.2, 3.3.6]). A dgdg algebra AA, or a strictly unital A∞A_{\infty} algebra through its dgdg model, is reflexive in either of the following cases:

(i) H>0(A)=0H^{>0}(A)=0 and dim⁡kHd(A)<∞\dim_k H^d(A)<\infty for every dd;

(ii) H<0(A)=0H^{<0}(A)=0, H0(A)H^0(A) is semisimple, and either AA is proper, or AA is smooth with dim⁡kHd(A)<∞\dim_k H^d(A)<\infty for every dd.

Now we are ready to prove Corollary B.

Proof of Corollary B. Apply Proposition 6.1, Proposition 6.2, Morita invariance [2 Proposition 3.8] and Theorem A. □

To check the generation condition in (52), let R=krR = k^r, with standard idempotents e1,…,ere_1,\ldots,e_r, and put Ri=eiRR_i = e_iR. A strictly unital minimal A∞A_\infty algebra BB over RR is positive if B<0=0B^{<0} = 0 and B0=RB^0 = R; projection to RR is an augmentation [22 Section 4.2].

Proposition 6.3. One has PsPerf⁡(B)=thick⁡B(R)\operatorname{PsPerf}(B) = \operatorname{thick}_B(R) in either of the following cases:

(i) BB is augmented over RR, H>0(B)=0H^{>0}(B) = 0, H0(B)H^0(B) is finite-dimensional, and its augmentation identifies H0(B)/rad⁡H0(B)H^0(B)/\operatorname{rad} H^0(B) with RR;

(ii) BB is positive.

In either case, a module whose cohomology is Ri[−d]R_i[-d] as a graded RR-module is quasi-isomorphic to the corresponding shifted augmentation module.

Proof. Case (i) follows by truncation and composition series over H0(B)H^0(B) [BGO25, Proposition 2.2.2 and Corollary 2.2.3]. Case (ii) and the assertion about module structures are [22 Lemmas 4.2–4.3]. □

Consequently, once (51) is a quasi-isomorphism, essential surjectivity is equivalent to R∈thick⁡B(M)R \in\operatorname{thick}_B(M) in these cases. For example, suppose G=⨁jGjG = \bigoplus_j G_j and K=⨁iKiK = \bigoplus_i K_i, with eie_i corresponding to KiK_i. Writing Mj=Hom⁡B(K,Φ(Gj))M_j = \operatorname{Hom}_B(K,\Phi(G_j)), formula (27) gives

Mjei≃Hom⁡A(Gj,j!Ki)∨.M_j e_i \simeq\operatorname{Hom}_A(G_j,j!K_i)^\vee.

If H∗(Mj)H^*(M_j) is Rij[−dj]R_{i_j}[-d_j] as a graded RR-module and every index 1,…,r1,\ldots,r occurs among the iji_j, Proposition 6.3 gives

M≃⨁jRij[−dj],thick⁡B(M)=PsPerf⁡(B).(55)M \simeq\bigoplus_j R_{i_j}[-d_j], \qquad\operatorname{thick}_B(M) = \operatorname{PsPerf}(B). \tag*{(55)}

Koszul duals of proper algebras

The following proposition identifies the subcategory that must be divided out when starting from the other direction of the Koszul duality functor.

Proposition 6.4. Let BB be proper and let PP split-generate PsPerf⁡(B)\operatorname{PsPerf}(B). Suppose that A=Hom⁡B(P,P)A = \operatorname{Hom}_B(P,P) is smooth and dim⁡kHd(A)<∞\dim_k H^d(A) < \infty for every dd. Then AA is reflexive, and the quasi-equivalence Ψ=Hom⁡B(P,−)\Psi= \operatorname{Hom}_B(P,-) identifies the following pairs of dg categories:

Ψ:(PsPerf⁡(B),thick⁡B(B∨))→∼(Perf⁡(A),PsPerf⁡(A)).(56)\Psi: \bigl(\operatorname{PsPerf}(B),\operatorname{thick}_B(B^\vee)\bigr) \xrightarrow{\sim} \bigl(\operatorname{Perf}(A),\operatorname{PsPerf}(A)\bigr). \tag*{(56)}

Its inverse is −⊗AP-\otimes_A P, and it induces quasi-equivalences

Perf⁡(A)‾∞≃PsPerf⁡(B)/thick⁡B(B∨)≃Dsg(Bop)op.(57)\overline{\operatorname{Perf}(A)}_\infty\simeq\operatorname{PsPerf}(B)/\operatorname{thick}_B(B^\vee) \simeq\mathcal{D}_{\mathrm{sg}}(B^{\mathrm{op}})^{\mathrm{op}}. \tag*{(57)}

If BB satisfies the Gorenstein condition (12), the first quotient is identified with Dsg(B)\mathcal{D}_{\mathrm{sg}}(B).

Proof. Since PP split-generates PsPerf⁡(B)\operatorname{PsPerf}(B), the functor Ψ\Psi is a quasi-equivalence with quasi-inverse −⊗AP-\otimes_A P. Tensor–Hom adjunction gives

Ψ(B∨)=Hom⁡B(P,B∨)≃P∨.\Psi(B^\vee) = \operatorname{Hom}_B(P,B^\vee) \simeq P^\vee.

Consequently, the identification of subcategories in (56) reduces to proving

PsPerf⁡(A)=thick⁡A(P∨).\operatorname{PsPerf}(A) = \operatorname{thick}_A(P^\vee).

We establish this equality and reflexivity of AA by verifying the hypotheses of Proposition 6.1.

Since BB is proper, B∨B^{\vee} belongs to PsPerf⁡(B)\operatorname{PsPerf}(B), so the preceding identification shows that P∨P^{\vee} is perfect as a right AA-module. The AA–BB-bimodule structure on PP gives P∨P^{\vee} a BB–AA-bimodule structure. Under Ψ(B∨)≃P∨\Psi(B^{\vee}) \simeq P^{\vee}, its left BB-action corresponds to the left action on B∨B^{\vee}, and its right AA-action corresponds to precomposition on Hom⁡B(P,B∨)\operatorname{Hom}_{B}(P,B^{\vee}). Linear duality and quasi-full faithfulness of Ψ\Psi therefore give quasi-isomorphisms of dg algebras

B→∼Hom⁡B(B∨,B∨)→∼Hom⁡A(P∨,P∨).B \xrightarrow{\sim} \operatorname{Hom}_{B}(B^{\vee},B^{\vee}) \xrightarrow{\sim} \operatorname{Hom}_{A}(P^{\vee},P^{\vee}).

Thus the generators AA of Perf⁡(A)\operatorname{Perf}(A) and P∨P^{\vee} of thick⁡A(P∨)\operatorname{thick}_{A}(P^{\vee}) have endomorphism algebras AA and BB, respectively. These are the algebra models used in Proposition 6.1, and the corresponding graph bimodule is

Hom⁡A(A,P∨)≃BPA∨.\operatorname{Hom}_{A}(A,P^{\vee}) \simeq{}_{B}P^{\vee}_{A}.

Since P∈PsPerf⁡(B)P \in\operatorname{PsPerf}(B), BPA∨{}_{B}P^{\vee}_{A} is also perfect over k\mathsf{k}. This verifies the pseudo-perfectness hypothesis on the graph bimodule.

The dual graph bimodule is (P∨)∨≃P(P^{\vee})^{\vee} \simeq P as an AA–BB-bimodule. Hence the comparison functor in Proposition 6.1 is −⊗AP-\otimes_{A}P, the quasi-inverse of Ψ\Psi already established above. Its action morphism (51) is the defining identification A=Hom⁡B(P,P)A=\operatorname{Hom}_{B}(P,P), and the density condition (6.3) is the assumption $\operatorname{PsPerf}(B)=\operatorname{thick}_{B}(P)$. Together with the assumed smoothness and degreewise finiteness of $A$, these verifications allow us to apply Proposition 6.1. It proves that $A$ is reflexive and supplies the Koszul duality hypotheses of {xref}Lemma 4.2. The generation statement {xref}(4.7)` then gives

PsPerf⁡(A)=thick⁡A(P∨)=Ψ(thick⁡B(B∨)),\operatorname{PsPerf}(A)=\operatorname{thick}_{A}(P^{\vee})=\Psi(\operatorname{thick}_{B}(B^{\vee})),

which proves (56).

Passing to dg quotients and applying (39) gives the first quasi-equivalence in (57). Linear duality sends thick⁡B(B∨)\operatorname{thick}_{B}(B^{\vee}) to Perf⁡(Bop)op\operatorname{Perf}(B^{\mathrm{op}})^{\mathrm{op}}, so Lemma 5.4 gives the second. Finally, under the Gorenstein condition (12), one has thick⁡B(B∨)=Perf⁡(B)\operatorname{thick}_{B}(B^{\vee})=\operatorname{Perf}(B), and the first quotient is therefore D⁡sg(B)\operatorname{D}_{\mathrm{sg}}(B). □

Corollary 6.5. Let BB be a finite-dimensional positive A∞A_{\infty} algebra over RR with B1=0B^{1}=0. Then A=Hom⁡B(R,R)A=\operatorname{Hom}_{B}(R,R) has a nonpositive smooth dg model with H0(A)=RH^{0}(A)=R and finite-dimensional cohomology in every degree. The conclusions of Proposition 6.4 hold with P=RP=R.

Proof. Put V=(B>0[1])∨V=(B^{>0}[1])^{\vee}. The relative bar construction gives a dg model for AA whose underlying graded algebra is TR(V)T_{R}(V); compare [15 Section 2.2 and Lemma 5.3(b)]. Since BB is finite-dimensional and B1=0B^{1}=0, the space VV is finite-dimensional and concentrated in strictly negative degrees. Consequently, only finitely many tensor lengths contribute to each graded piece. In particular, this model is nonpositive and has finite-dimensional cohomology in every degree.

The augmentation excludes constant terms, and minimality of BB excludes linear terms in the differential of each generator. Thus, for a homogeneous generator v∈Vv\in V, the differential d(v)d(v) is a finite sum of words v1⋯vℓv_{1}\cdots v_{\ell} with ℓ≥2\ell\ge2. Each such word has degree ∣v∣+1|v|+1, and all its factors have negative degree. It follows that

∣vi∣=∣v∣+1−∑j≠i∣vj∣>∣v∣for every i.|v_i|=|v|+1-\sum_{j\ne i}|v_j|>|v| \qquad\text{for every }i.

Hence d(v)d(v) involves only generators of strictly greater degree than vv.

To deduce smoothness of AA, let δ:A→ΩA/R1\delta:A\to\Omega^{1}_{A/R} be the universal relative derivation. As a graded AA-bimodule, one has

ΩA/R1≃A⊗RV⊗RA,\Omega^{1}_{A/R}\simeq A\otimes_{R}V\otimes_{R}A,

with 1⊗v⊗11 \otimes v \otimes1 corresponding to δv\delta v. Its differential satisfies

d(δv)=δ(d(v)).d(\delta v)=\delta(d(v)).

By the Leibniz rule, the right-hand side is a sum of bimodule multiples of elements δvi\delta v_i with ∣vi∣>∣v∣|v_i|>|v|. Thus, when the generators δv\delta v are adjoined in decreasing order of degree, the differential of each new generator lies in the bimodule generated by those already adjoined. Since VV is finite-dimensional, this gives a finite filtration of ΩA/R1\Omega^1_{A/R} by dg sub-bimodules whose successive quotients are finite direct sums of shifts of Aei⊗kejAAe_i\otimes_k e_jA. These quotients are perfect, so ΩA/R1\Omega^1_{A/R} is a perfect AA-bimodule.

Since R=krR=k^r, the bimodule A⊗RAA\otimes_R A is also a finite direct sum of perfect bimodules Aei⊗keiAAe_i\otimes_k e_iA. The exact sequence

0⟶ΩA/R1⟶A⊗RA⟶A⟶00\longrightarrow\Omega^1_{A/R}\longrightarrow A\otimes_R A\longrightarrow A\longrightarrow0

therefore shows that the diagonal bimodule AA is perfect, hence smooth. Moreover, A0=RA^0=R, and d(A−1)=0d(A^{-1})=0 because the differential of a degree −1-1 generator has no constant term. Thus H0(A)=RH^0(A)=R.

Finally, Proposition 6.3(ii) gives PsPerf⁡(B)=thick⁡B(R)\operatorname{PsPerf}(B)=\operatorname{thick}_B(R). Since BB is proper and AA is smooth with finite-dimensional cohomology in every degree, Proposition 6.4 applies with P=RP=R. □

Remark 6.6. Degreewise finiteness can also hold when B1≠0B^1\ne0. For a finite-dimensional positive BB, the reduced bar construction is

Bar⁡(B)=⨁ℓ≥0(B>0[1])⊗Rℓ,Hom⁡B(R,R)≃(Bar⁡(B)∨)op,\operatorname{Bar}(B)=\bigoplus_{\ell\geq0}(B^{>0}[1])^{\otimes_R\ell},\qquad\operatorname{Hom}_B(R,R)\simeq(\operatorname{Bar}(B)^\vee)^{\mathrm{op}},

where the opposite is taken for the forward convolution product. If the degree-one arrows, indexed by bases of ejB1eie_jB^1e_i from ii to jj, form an acyclic quiver, blocks of degree-zero bar letters have bounded length; see [ASS06, Section II.1 and Lemma II.1.4(c)]. Each graded piece of Bar⁡(B)\operatorname{Bar}(B) is then finite-dimensional, giving the finiteness required in Proposition 6.1.

Geometric examples

For a finite-type Weinstein manifold X2nX^{2n}, let W(X)\mathcal{W}(X) and F(X)\mathcal{F}(X) denote its wrapped and compact Fukaya categories of exact graded branes with finite-rank local systems. W(X)\mathcal{W}(X) is smooth with a smooth nn-Calabi–Yau structure and is generated by cocores, while Poincaré duality gives a proper nn-Calabi–Yau structure on F(X)\mathcal{F}(X); see [Gan13; GPS24a; CDRGG24] and [Gan23, Theorem 2].

The following examples illustrate the algebraic criteria and the scheme comparison (46), giving

Perf⁡^(W(X))∞≃Dsg(F(X)).(58)\widehat{\operatorname{Perf}}(\mathcal{W}(X))_\infty\simeq\mathcal{D}_{\mathrm{sg}}(\mathcal{F}(X)). \tag*{(58)}

By [GGV22, Theorem 1.1], this also identifies the Rabinowitz Fukaya category up to Morita equivalence. In the first two examples, we first apply the criteria to the full subcategory on the chosen compact cores and then verify that these cores split-generate Perf⁡(F(X))\operatorname{Perf}(\mathcal{F}(X)).

(1) Cotangent bundles. Let QQ be a closed simply connected spin manifold, with coefficients in any field and zero background class. A cotangent fibre GG and the zero section KK give

A≃C−∗(ΩQ;k),B≃C∗(Q;k),M≃k;A\simeq C_{-*}(\Omega Q;k),\qquad B\simeq C^*(Q;k),\qquad M\simeq k;

see [Abol1, Theorem 1.1] and [Ekl23, Introduction] for the compatible Yoneda and Koszul actions. A finite minimal model of BB is positive with B1=0B^1=0, so Corollary 6.5 applies. Since AA is nonpositive with H0(A)=kH^0(A)=k, Proposition 6.3(i) places every compact object’s fibre Yoneda module in thick⁡A(k)\operatorname{thick}_A(k). Thus the zero section split-generates Perf⁡(F(T∗Q))\operatorname{Perf}(\mathcal{F}(T^*Q)), giving (58) for the full compact category. (2) Tree plumbings. Let XX be a finite tree plumbing of cotangent bundles of closed simply connected spin nn-manifolds, n≥3n \ge3, with the same coefficient convention. Take GG to be the sum of the cocores and KK the sum of the cores. Grade the cores so that opposite intersection generators have degrees 1,n−11,n-1. Then BB is proper and admits a positive minimal model with acyclic degree-one quiver, and the summands of MM are the augmentation modules RiR_i. The action map (51) is a quasi-isomorphism by [EkL23, Theorems 66, 68, and 83, Section 6.1.4]. Proposition 6.3 and Remark 6.6 verify (52). Moreover, AA is nonpositive with finite-dimensional H0(A)H^0(A), and the cocore Yoneda modules of the cores represent all simple H0(A)H^0(A)-modules. The same truncation argument therefore proves compact generation and yields (58).

(3) Punctured tori. Over a perfect field, the standard graded torus TrT_r with r≥1r \ge1 punctures has a quasi-equivalence of pairs

(Perf⁡(W(Tr)),Perf⁡(F(Tr)))≃(Coh⁡(Cr),Perf⁡(Cr))\left(\operatorname{Perf}(\mathcal{W}(T_r)),\operatorname{Perf}(\mathcal{F}(T_r))\right) \simeq\left(\operatorname{Coh}(C_r),\operatorname{Perf}(C_r)\right)

where C1C_1 is the split nodal cubic and CrC_r is the cycle of rr projective lines for r>1r>1; see [LekPer15, Theorem 9.1] and [LekPo17, Corollary 3.4.1 and Theorem 3.7.2]. Since its dualizing complex is OCr[1]\mathcal{O}_{C_r}[1], formula (46) gives (58), with target Dsg(Cr)\mathcal{D}_{\mathrm{sg}}(C_r). Calculations for double-bubble plumbings appear in [21 Proposition 5.2 and Corollaries 5.7–5.8], and for Milnor fibres in [LekUe21, Section 3] and [LekUe22, Theorem 6.11]. For the description of double suspensions by cluster categories, see [LekUe26, Theorems 2.2 and 4.1].

Hochschild cohomology

We deduce multiplicative comparisons of Hochschild cohomology from Theorem A, then relate them to singular Hochschild cohomology under Keller’s hypotheses. All module Hom and tensor products use the dg models fixed in § 2.

The boundary comparison

For a small dg category C\mathcal{C} and a C\mathcal{C}-bimodule MM, write

C∗(C;M)=Hom⁡Ce(CΔ,M),HHm(C,M)=HmC∗(C;M),\mathcal{C}^{*}(\mathcal{C};M)=\operatorname{Hom}_{\mathcal{C}^{e}}(\mathcal{C}_{\Delta},M), \qquad HH^{m}(\mathcal{C},M)=H^{m}\mathcal{C}^{*}(\mathcal{C};M),

and omit MM for diagonal coefficients; see [5 Section 5.4] for the Hochschild bar model and cup product. Let q:C→Dq:\mathcal{C}\to\mathcal{D} be a dg localization, meaning that restriction q∗:Mod⁡(D)→Mod⁡(C)q^{*}:\operatorname{Mod}(\mathcal{D})\to\operatorname{Mod}(\mathcal{C}) is quasi-fully faithful; dg quotients have this property [6 Section 3.9]. Set

∂qC=(qe)∗DΔ,(∂qC)(X,Y)=D(qY,qX),qe=qop⊗q.(59)\partial_q\mathcal{C}=(q^{e})^{*}\mathcal{D}_{\Delta}, \qquad(\partial_q\mathcal{C})(X,Y)=\mathcal{D}(qY,qX), \qquad q^{e}=q^{\mathrm{op}}\otimes q. \tag*{(59)}

Composition in D\mathcal{D} and the map η:CΔ→∂qC\eta:\mathcal{C}_{\Delta}\to\partial_q\mathcal{C} induced by qq give this bimodule a multiplication and unit, hence a cup product on HH∗(C,∂qC)HH^{*}(\mathcal{C},\partial_q\mathcal{C}).

Proposition 7.1. Restriction gives a quasi-isomorphism

q∗:C∗(D)⟶C∗(C;∂qC)(60)q^{*}:\mathcal{C}^{*}(\mathcal{D})\longrightarrow\mathcal{C}^{*}(\mathcal{C};\partial_q\mathcal{C}) \tag*{(60)}

and an isomorphism of unital graded algebras on cohomology.

Proof. By [6 Proposition 3.10(a)], extension of scalars along qeq^{e} sends CΔ\mathcal{C}_{\Delta} to DΔ\mathcal{D}_{\Delta}. Adjunction therefore gives the quasi-isomorphism (60). On bar cochains, its arity-mm component is

(q∗f)r(am,…,a1)=fm(q(am),…,q(a1))(q^{*}f)_r(a_m,\ldots,a_1)=f_m(q(a_m),\ldots,q(a_1))

for composable morphisms a1,…,ama_1,\ldots,a_m in C\mathcal{C}. The cup product uses deconcatenation of these inputs and composition in D\mathcal{D}, so this formula preserves products and sends the diagonal unit to η\eta. ∎

Under the hypotheses of Theorem A, apply this proposition to

q:C=Perf⁡(A)⟶D=Perf⁡(A)/PsPerf⁡(A).q : \mathcal{C} = \operatorname{Perf}(\mathcal{A}) \longrightarrow\mathcal{D} = \operatorname{Perf}(\mathcal{A})/\operatorname{PsPerf}(\mathcal{A}).

For the Yoneda functor i:A⟶Ci : \mathcal{A} \longrightarrow\mathcal{C}, define the boundary coefficient bimodule by

∂∞A=(ie)∗∂qC.(61)\partial_{\infty}\mathcal{A} = (i^{\mathrm{e}})^{*}\partial_q\mathcal{C}. \tag*{(61)}

Proposition 5.2 identifies it with the pullback of the diagonal of A^∞\widehat{\mathcal{A}}_{\infty} along (37).

Corollary 7.2. Under the hypotheses of Theorem A, there are natural isomorphisms of unital graded algebras

HH∗(A,∂∞A)≅HH∗(A^∞)≅HH∗(Perf⁡(A)^∞).(62)HH^{*}(\mathcal{A},\partial_{\infty}\mathcal{A}) \cong HH^{*}(\widehat{\mathcal{A}}_{\infty}) \cong HH^{*}(\widehat{\operatorname{Perf}(\mathcal{A})}_{\infty}). \tag*{(62)}

If A\mathcal{A} is reflexive, these algebras are isomorphic to HH∗(Dsg(Bop)op)HH^{*}(\mathcal{D}_{\mathrm{sg}}(\mathcal{B}^{\mathrm{op}})^{\mathrm{op}}). If A\mathcal{A} is reflexive and B\mathcal{B} satisfies the Gorenstein condition (12), the functor (4) induces an isomorphism of unital graded algebras

HH∗(Dsg(B))→∼HH∗(A,∂∞A).(63)HH^{*}(\mathcal{D}_{\mathrm{sg}}(\mathcal{B})) \xrightarrow{\sim} HH^{*}(\mathcal{A},\partial_{\infty}\mathcal{A}). \tag*{(63)}

Proof. The Yoneda Morita equivalence preserves Hochschild cohomology with corresponding coefficients and their cup products by [6 Proposition 3.10(e)]. Together with Proposition 7.1, it identifies HH∗(A,∂∞A)HH^{*}(\mathcal{A},\partial_{\infty}\mathcal{A}) with HH∗(D)HH^{*}(\mathcal{D}). Proposition 5.2 gives (62), and Theorem A gives the comparisons with the singularity categories. □

After translating the opposite-category and grading conventions, this is the cup product of [19 Definition 4.2 and Remark 4.3]; the boundary restriction map is the unital map of [Gao25, Proposition 3.10]. The same argument applies to the algebraic pairs (56) and the geometric equivalences (58).

Singular Hochschild cohomology comparison

For proper B\mathcal{B}, its diagonal is pseudo-perfect over Be\mathcal{B}^{\mathrm{e}} by Definition 2.3. Denote its image in Dsg(Be)\mathcal{D}_{\mathrm{sg}}(\mathcal{B}^{\mathrm{e}}) by B‾Δ\overline{\mathcal{B}}_{\Delta}, and define

HHsgm(B)=HmHom⁡Dsg(Be)(B‾Δ,B‾Δ),HH^{m}_{\mathrm{sg}}(\mathcal{B}) = H^{m}\operatorname{Hom}_{\mathcal{D}_{\mathrm{sg}}(\mathcal{B}^{\mathrm{e}})}(\overline{\mathcal{B}}_{\Delta},\overline{\mathcal{B}}_{\Delta}),

with product induced by composition; for ordinary finite-dimensional algebras, this is singular Hochschild cohomology as in [7 Section 1].

For a finite-dimensional ordinary algebra BB with B/rad⁡(B)B/\operatorname{rad}(B) separable over k\mathbb{k}, Keller’s canonical comparison is an isomorphism of unital graded algebras

HHsg∗(B)→∼HH∗(Dsg(B)).(64)HH^{*}_{\mathrm{sg}}(B) \xrightarrow{\sim} HH^{*}(\mathcal{D}_{\mathrm{sg}}(B)). \tag*{(64)}

Indeed, [ELS20, Theorem 3.7] gives smoothness of PsPerf⁡(B)\operatorname{PsPerf}(B), which is the hypothesis required by [7 Theorem 1.1]. Consequently, when B=B\mathcal{B}=B in Corollary 7.2, A\mathcal{A} is reflexive, and B\mathcal{B} satisfies (12), composing (64) with (63) identifies HHsg∗(B)HH^{*}_{\mathrm{sg}}(B) with HH∗(A,∂∞A)HH^{*}(\mathcal{A},\partial_{\infty}\mathcal{A}) as unital graded algebras. If BB is split, meaning B/rad⁡(B)≅∏iMdi(k)B/\operatorname{rad}(B)\cong\prod_i M_{d_i}(\mathbb{k}), [XZ26, Theorem 1.1], applied to BopB^{\mathrm{op}} for our right-module convention, lifts (64) to the homotopy category of B∞B_{\infty}-algebras.

The complete multiplicative solution to Conjecture 1.1 requires further structure on A\mathcal{A}. For example, A\mathcal{A} also carries a nondegenerate 4-truncated nn-pre-Calabi–Yau structure, in the sense of [19 Definitions 2.9 and 2.11], its induced Calabi–Yau duality identifies

HHr(A,∂∞A)≅HHn−r(A,∂∞A).HH^{r}(\mathcal{A},\partial_{\infty}\mathcal{A}) \cong HH_{n-r}(\mathcal{A},\partial_{\infty}\mathcal{A}).

With the Koszul signs for this regrading, [19 Proposition 5.8 and Theorem 5.11] identify the cup product with their product of homological degree −n-n.

References

School of Mathematics, Nanjing University, Nanjing, Jiangsu, 210093, China Email address: [email protected]

Institut für Mathematik, Universität Augsburg, Germany Email address: [email protected]

References

  1. [1]Isambard Goodbody. Reflecting perfection for finite-dimensional differential graded algebras. In: Bulletin of the London Mathematical Society 56.12 (2024), pp. 3689–3707. DOI: 10.1112/blms.13160.DOI
  2. [2]Isambard Goodbody. Reflexivity and Hochschild cohomology. In: Journal of Algebra 705 (2026), pp. 32–60. DOI: 10.1016/j.jalgebra.2026.04.050.DOI
  3. [3]Norihiro Hanihara and Junyang Liu. Calabi–Yau structures on derived and singularity categories of symmetric orders. 2026. arXiv: 2512.03836v2 [math.RT]. URL: https://arxiv.org/abs/2512.03836v2.
  4. [4]Bernhard Keller. A-infinity algebras, modules and functor categories. In: Trends in Representation Theory of Algebras and Related Topics. Vol. 406. Contemporary Mathematics. Providence, RI: American Mathematical Society, 2006, pp. 67–93. DOI: 10.1090/conm/406/07654.DOI
  5. [5]Bernhard Keller. On differential graded categories. In: International Congress of Mathematicians, Vol. II. Zürich: European Mathematical Society, 2006, pp. 151–190. DOI: 10.4171/022-2/8. arXiv: math/0601185v5 [math.KT]. URL: https://arxiv.org/abs/math/0601185v5.
  6. [6]Bernhard Keller. Deformed Calabi–Yau completions. In: Journal für die reine und angewandte Mathematik 654 (2011). With an appendix by Michel Van den Bergh, pp. 125–180. DOI: 10.1515/crelle.2011.031. See also the erratum (2018).DOI
  7. [7]Bernhard Keller. Singular Hochschild cohomology via the singularity category. 2020. arXiv: 1809.05121v10 [math.RT]. URL: https://arxiv.org/abs/1809.05121v10.
  8. [8]Henning Krause. Localization theory for triangulated categories. In: Triangulated Categories. Ed. by Thorsten Holm, Peter Jørgensen, and Raphaël Rouquier. Vol. 375. London Mathematical Society Lecture Note Series. Cambridge: Cambridge University Press, 2010, pp. 161–235. DOI: 10.1017/CBO9781139107075.005. arXiv: 0806.1324v2 [math.CT]. URL: https://arxiv.org/abs/0806.1324v2.
  9. [15]Di Ming Lu, John H. Palmieri, Quan Shui Wu, and James J. Zhang. Koszul equivalences in A∞-algebras. In: New York Journal of Mathematics 14 (2008), pp. 325–378.
  10. [16]Volodymyr Lyubashenko and Oleksandr Manzyuk. Quotients of unital A∞-categories. In: Theory and Applications of Categories 20.13 (2008), pp. 405–496. DOI: 10.70930/tac/z2jgorg4.DOI
  11. [17]Volodymyr Lyubashenko and Sergiy Ovsienko. A construction of quotient A∞-categories. In: Homology, Homotopy and Applications 8.2 (2006), pp. 157–203. DOI: 10.4310/HHA.2006.v8.n2.a9.DOI
  12. [18]Dmitri Orlov. Triangulated categories of singularities and D-branes in Landau–Ginzburg models. In: Proceedings of the Steklov Institute of Mathematics 246 (2004), pp. 227–248.arxiv.org/abs/math/0302304
  13. [19]Manuel Rivera, Alex Takeda, and Zhengfang Wang. Algebraic string topology from the neighborhood of infinity. 2024. arXiv: 2308.09684v3 [math.AT]. URL: https://arxiv.org/abs/2308.09684v3.
  14. [20]Paul Seidel. Fukaya Categories and Picard-Lefschetz Theory. Zürich Lectures in Advanced Mathematics. Zürich: European Mathematical Society, 2008. ISBN: 978-3-03719-063-0. DOI: 10.4171/063.DOI
  15. [21]Ivan Smith and Michael Wemyss. Double bubble plumbings and two-curve flops. In: Selecta Mathematica (New Series) 29.2, 29 (2023). DOI: 10.1007/s00029-023-00828-z.DOI
  16. [22]Hao Su and Dong Yang. From simple-minded collections to silting objects via Koszul duality. In: Algebras and Representation Theory 22.1 (2019), pp. 219–238. DOI: 10.1007/s10468-018-9763-y.DOI
  17. [23]Alex Takeda. Efimov’s categorical formal punctured neighborhood of infinity, Rabinowitz Fukaya category, CY and pre-CY structures. In: Mini-Workshop: Flavors of Rabinowitz Floer and Tate Homology. Vol. 20. Oberwolfach Reports 4. Report 56/2023. EMS Press, 2023, pp. 3164–3166. DOI: 10.4171/OWR/2023/56.
  18. [24]Bertrand Toën. The homotopy theory of dg-categories and derived Morita theory. In: Invent. Math 167.3 (2007), pp. 615–667. DOI: 10.1007/s00222-006-0025-y.DOI

Paper details

vlnc:W_vtykgrg6
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents