The first step is to reformulate K-theory of small categories in terms of large categories. Every small stable category \(C_0\) gives rise to a compactly generated presentable stable category \(\Ind (C_0)\), and if \(C_0\) is idempotent complete, we can recover it as \(C_0 \simeq \Ind (C_0)^{\omega }\). (Our nonconnective K-theory is extended to a small stable category by first taking its idempotent completion. In contrast, connective K-theory need not be invariant under idempotent completion: the map on \(K_0\) can fail to be surjective, although the positive K-groups are unchanged.) This gives an equivalence of categories \[ \Cat _{\perf } \iso \PrL _{\omega ,\st } = \left \{ \begin {array}{c} \text {compactly generated categories in $\PrLst $,} \\ \text {strongly continuous functors} \end {array} \right \}, \qquad C_0 \mapsto \Ind (C_0), \] where a functor \(F\colon C \to D\) in \(\PrLst \) is called strongly continuous if its right adjoint again preserves colimits. Between compactly generated stable categories, this is equivalent to \(F\) preserving compact objects. So K-theory may equally well be regarded as a functor on compactly generated categories and strongly continuous functors.

From this perspective, the obvious next question is: which presentable stable categories, beyond the compactly generated ones, admit a reasonable K-theory? The answer is given by the following theorem, which provides many equivalent characterizations of a single class of categories.

Theorem 1.2 (Lurie, [Lurie (2018), Proposition D.7.3.1]). Consider \(C \in \PrLst \). The following are equivalent:

(1)

There exists a localization \(\phi \colon D \to E\) in \(\PrL _{\omega ,\st }\) such that \(C = \ker (\phi )\);

(2)

\(C\) is dualizable in \((\PrLst ,\otimes )\);

(3)

The functor \(\colim \colon \Ind (C) \to C\) has a left adjoint (“compactly assembled”);

(4)

\(C\) is a retract in \(\PrLst \) of a compactly generated category;

(5)

(Efimov, [Efimov (2024), Theorem 5.2]) \(C \otimes (-)\colon \PrLst \to \PrLst \) preserves fully faithful functors (“flat”);

(6)

(Clausen–Scholze, see [Efimov (2024), Proposition 2.51]) arbitrary products distribute over filtered colimits (“AB6”).

The equivalence of (1)–(4) will be proved in Theorem 3.16, the axiom (AB6) is discussed in Chapter 4, and flatness in Chapter 10. Characterization (1) is the one most relevant for K-theory: a dualizable category is precisely something that can be described as the “difference” of two compactly generated categories. Characterization (3) is the one that is easiest to work with in practice. In a compactly generated category, the presentation of an object \(x\) as the formal colimit of all compact objects mapping to it is the finest presentation of \(x\): it maps uniquely into every ind-object \(X \in \Ind (C)\) with \(\colim X \simeq x\), and as a consequence maps out of \(x\) into filtered colimits can be computed at the level of ind-objects. Characterization (3) says precisely that every object \(x\) of \(C\) still admits such a finest presentation \(\hat {y}(x) \in \Ind (C)\), whose terms are however no longer required to be compact. This heuristic is explained in more detail at the beginning of Chapter 3.

In particular, compactly generated categories are dualizable, with dual \[ \Ind (C_0)^{\vee } = \Ind (C_0\catop ). \] But there are many more examples.

Example 1.3. The following categories are dualizable but in general not compactly generated:

  • Keller’s example of a localization of a compactly generated category with a non-compactly generated kernel (Example 2.20);
  • the category \(\Shv (X;\Sp )\) of sheaves of spectra on a locally compact Hausdorff space \(X\) (see Chapter 9); for instance \(\Shv (\R ^n;\Sp )\) with \(n \geq 1\) has no nonzero compact objects at all (for sheaves of abelian groups on a non-compact connected manifold this is due to Neeman [Neeman (2001), Theorem 0.1]; for sheaves of spectra it follows from Efimov’s description of the compact objects [Efimov (2024), Proposition 10.17]);
  • the categories of nuclear modules arising in analytic geometry: \(\Nuc (\Z _p^{\square })\), \(\Nuc (\R ^{p\text {-}\mathrm {liq}})\), \(\Nuc (\C ^{\mathrm {gas}})\) [Scholze (2026); Clausen and Scholze (2026)].

The nuclear examples are self-dual because they are rigid. Sheaf categories on compact Hausdorff spaces are rigid as well. For general locally compact Hausdorff spaces they are locally rigid, rather than rigid; their self-duality is provided by categorical Verdier duality, see [Krause et al. (2024), Section 4.4] and Example 6.13.

Definition 1.4 (cf. [Gaitsgory and Rozenblyum (2017), Chapter 1, Definition 9.1.2]). A rigid category is \(C \in \CAlg (\PrLst )\) such that for every finite set \(I\), the multiplication functor \(\mu \colon C^{\otimes I} \to C\) admits a right adjoint in \(\Mod _{C^{\otimes I}}(\PrLst )\).

Example 1.5. Let \(C_0\) be a small idempotent complete symmetric monoidal stable category in which every object is dualizable. Then \(\Ind (C_0)\) is rigid. For instance, \(\D (R) = \Ind (\Perf (R))\) is rigid for every commutative ring \(R\).

Proposition 1.6 ([Gaitsgory and Rozenblyum (2017), Chapter 1, Section 9.2.1]). If \(C\) is rigid, then \(C\) is self-dual via \[ C \otimes C \xrightarrow {\mu } C \xrightarrow {\Gamma } \Sp , \] where \(\Gamma = \hom _C(\unit ,-)\).

Rigid categories, trace-class morphisms and nuclear modules are discussed in detail in Section 6.1, Section 6.2.

Remark 1.7. There is a useful dictionary between the compactly generated and the dualizable world, which will be developed in Chapter 4: compact objects are replaced by compact morphisms. A dualizable category is generated under colimits by sequential colimits \(\colim (x_0 \to x_1 \to x_2 \to \cdots )\) along compact morphisms, and in fact its \(\omega _1\)-compact objects are precisely of this form. In a symmetric monoidal setting, the analogy continues: just as a compactly generated category with compact unit is rigid if its compact objects are precisely its dualizable objects, a dualizable symmetric monoidal category is rigid if its compact morphisms are precisely its trace-class morphisms.

The morphisms between dualizable categories which we will consider are the strongly continuous functors.

Definition 1.8. We denote by \[ \Pr ^{\dual } \subseteq \PrLst \] the subcategory consisting of the dualizable categories and the strongly continuous functors, i.e. those functors whose right adjoint again preserves colimits (see Definition 3.11).

By the above discussion, Ind-completion defines a fully faithful functor \[ \Ind \colon \Cat _{\perf } \hookrightarrow \Pr ^{\dual }. \] The category \(\Pr ^{\dual }\) has excellent formal properties: it is closed under tensor products in \(\PrLst \), the inclusion into \(\PrLst \) creates colimits, and \(\Pr ^{\dual }\) admits all limits and internal homs, although these are more subtle (see Chapter 7).