To illustrate the guiding principle, let us look at the first expansion: from vector bundles to perfect complexes.
Pushforwards and the Riemann–Roch theorem
A crucial ingredient in Grothendieck’s relative version of the Riemann–Roch theorem [Borel and Serre (1958)] is that, for every proper map \(f\colon X \to Y\) of smooth quasi-projective varieties, there is a pushforward map \[ f_*\colon K(\Vect (X)) \to K(\Vect (Y)). \] At the level of vector bundles, however, it is not at all obvious where such a map should come from: the pushforward \(f_*\mathcal {E}\) of a vector bundle is in general neither a vector bundle nor concentrated in a single degree, and one is forced to use the derived pushforward together with resolutions. If instead we allow perfect complexes as input, the map simply comes from the functor \[ f_*\colon \Perf (X) \to \Perf (Y), \] and in this generality also the annoying assumption of quasi-projectivity disappears. In other words: K-theory admits proper pushforwards because the categories \(\Perf (-)\) admit proper pushforwards.
Further examples
The same pattern appears everywhere once one works with stable categories:
- Descent. The functor \(X \mapsto K(\Perf (X))\) satisfies Zariski (and even Nisnevich) descent by the work of Thomason–Trobaugh [Thomason and Trobaugh (1990)], which fails for \(X \mapsto K(\Vect (X))\) in general. The categorical reason is that for a quasi-compact open \(U \subseteq X\) with closed complement \(Z\), the restriction functor \(\Perf (X) \to \Perf (U)\) is a localization (up to idempotent completion) with kernel \(\Perf _Z(X)\), the perfect complexes supported on \(Z\). K-theory sends such sequences to fiber sequences, and descent follows formally.
- Additive inverses. The cofiber sequence \(x \to 0 \to \Sigma x\) gives \([x]+[\Sigma x]=0\) in \(K_0(C)\) after imposing the cofiber relations. At the spectrum level, additivity implies that suspension induces multiplication by \(-1\) on \(K(C)\). This does not make \(x\) and \(\Sigma x\) inverses under direct sum in the category itself.
- Projective bundle formulas. For a commutative ring \(R\), there is a semiorthogonal decomposition (in the sense of Section 2.4) \[ \Perf (\P ^1_R) \simeq \lra {\Perf (R) \otimes \Oo (-1),\; \Perf (R) \otimes \Oo }, \] and K-theory turns this into a splitting \(K(\P ^1_R) \simeq K(R) \oplus K(R)\).
In each of these cases, a property of K-theory that is hard to establish by hand becomes a formal consequence of a structural property of the input category, combined with the single fundamental property of K-theory: it is a localizing invariant, i.e. it sends short exact sequences of small stable categories to fiber sequences of spectra (Definition 2.30).