Example 5.30 (Efimov). Let \(k\) be a field, and consider the category \[ D := \Big \{(V,W,\phi ) \;\Big |\; V,W \in \D (k),\ \phi \colon \bigoplus _{\N }V \to \bigoplus _{\N }W\Big \}. \] The functors \(W \mapsto (0,W,0)\) and \((V,W,\phi ) \mapsto V\) exhibit \(D\) as an extension of \(\D (k)\) by \(\D (k)\), both of which are compactly generated. Nevertheless, \(D\) is not dualizable.
Proof (following [Efimov (2024), Proposition 6.8]). Here \(\Phi (W) = \prod _{\N }\bigoplus _{\N }W\), since \(\hom (\bigoplus _{\N }V,\bigoplus _{\N }W) \simeq \hom (V,\prod _{\N }\bigoplus _{\N }W)\). By Proposition 5.28, it suffices to show that \(\Psi (k) \notin \mathrm {Tate}(\Perf (k))\), where \(\Psi (k)\) is the pro-object corepresenting \(W \mapsto \hom (k,\Phi (W)) = \prod _{\N }\bigoplus _{\N }W\). Using (AB6) in \(\D (k)\), we may rewrite this functor as \[ \prod _{\N }\bigoplus _{\N }W \simeq \colim _{f\colon \N \to \N }\ \prod _{n \in \N }\bigoplus _{i=0}^{f(n)-1}W \simeq \colim _{f\colon \N \to \N }\hom (X_f,W), \qquad X_f := \bigoplus _{n \in \N }k^{f(n)}, \] where \(\N ^{\N }\) is ordered pointwise, and for \(f \leq g\) the transition map is induced by the projection \(X_g \to X_f\). Hence \(\Psi (k) \simeq \text {``}\!\lim \nolimits _f\!\text {} X_f\), a pro-system in which all transition maps are split epimorphisms.
Let \(q\colon \D (k) \to \Calk (\Perf (k))\) be the projection to \(\Calk (\Perf (k)) := \bigcup _{\kappa }\Calk _{\kappa }(\Perf (k))\) (the transition functors are fully faithful by Remark 2.27); this is an idempotent complete stable category, and \(q(x) = 0\) if and only if \(x \in \Perf (k)\). Suppose that \(\Psi (k) \in \mathrm {Tate}(\Perf (k))\). Then \(\Psi (k) \oplus \Sigma \Psi (k) \in \mathrm {Tate}^{\mathrm {el}}(\Perf (k))\), so there is a cofiber sequence \(u \to \Psi (k) \oplus \Sigma \Psi (k) \to v\) with \(u \in \Pro (\Perf (k))\) and \(v \in \D (k)\). Applying \(\Pro (q)\) kills \(u\), so \(\Pro (q)(\Psi (k)) = \text {``}\!\lim \nolimits _f\!\text {} q(X_f)\) is a retract of the constant pro-object \(q(v)\), and hence is itself isomorphic to a constant pro-object \(c\).
We now use the following elementary observation: if a pro-system \((Y_f)_f\) indexed by a directed poset, with split epimorphisms as transition maps \(t_{gf}\colon Y_g \to Y_f\), is isomorphic to a constant pro-object \(c\), then there is \(f_0\) such that \(t_{gf}\) is an isomorphism for all \(f_0 \leq f \leq g\). Indeed, the isomorphism is given by maps \(p_f\colon Y_f \to c\) for \(f \geq f_0\) (compatible with the \(t_{gf}\)) and a compatible family \(s_f\colon c \to Y_f\), such that \(p_fs_f = \id _c\) for \(f \geq f_0\), and for every \(f\) there is \(g \geq f,f_0\) with \(s_fp_g = t_{gf}\). If \(\sigma \) is a section of \(t_{gf}\), then \(s_f(p_g\sigma ) = \id \). So for \(f \geq f_0\), the map \(s_f\) is both a split mono and a split epi, hence an isomorphism, and then \(t_{gf} = s_fs_g^{-1}\) is an isomorphism for \(f_0 \leq f \leq g\).
Applied to \(Y_f = q(X_f)\), this gives a contradiction: for every \(f\), the fiber of \(X_{f+1} \to X_f\) is \(\bigoplus _{\N }k \notin \Perf (k)\), so \(q(X_{f+1}) \to q(X_f)\) is not an isomorphism.