nLab Kan lift

Idea

The notion of Kan lift in a 2-category or in a bicategory is dual to the notion of Kan extension, by a process of turning the 1-cells around, much in the same way as a homotopy lifting property is dual to a homotopy extension property (in model category theory, for instance).

Informally, a Kan lift is a best approximate solution to the problem of finding a lift $\widetilde{f}: A \to B$ of an arrow (morphism) $f: A \to C$ through an arrow $p: B \to C$, as in the diagram

Of course, lifts typically don’t literally exist in the sense of an equation $p \circ \widetilde{f} = f$ or an isomorphism $p \circ \widetilde{f} \cong f$. But in good situations, one may have the next best thing: a 2-cell $p \circ \widetilde{f} \Rightarrow f$ which is universal among 2-cells of this form. This gives the notion of right Kan lift. The notion of left Kan lift is similar, but with 2-cells in the opposite direction.

Terminology

Usually, when working within a 2-categorical context, Kan lifts are simply refered to as lifts/liftings, just as what happens with Kan extensions.

Definition

As in the discussion at Kan extension, there is a local notion of Kan lift and a global notion. Expanding on the informal description above, we give the local notion first, followed by the (conceptually more perspicuous) global notion in terms of adjoint functors.

Local description

Given 1-cells $p: B \to C$, $f: A \to C$ in a bicategory, a right Kan lift of $f$ through $p$, denoted $Rift_p f$, is a 1-cell $\widetilde{f}: A \to B$ equipped with a 2-cell

$$\varepsilon: p \circ \widetilde{f} \Rightarrow f$$

satisfying the universal property: given any pair $(g: A \to B, \eta: p \circ g \Rightarrow f)$, there exists a unique 2-cell

$$\zeta: g \Rightarrow \widetilde{f}$$

such that $\varepsilon \bullet (p \circ \zeta) = \eta$. (Here $\circ$ refers to composition across a 0-cell, and $\bullet$ to composition across a 1-cell.) As with any universal description, the pair $(Rift_p f, \varepsilon)$ is unique up to unique 2-cell isomorphism.

A left Kan lift of $f$ through $p$, denoted $Lift_p f$, is a 1-cell $\widetilde{f}: A \to B$ equipped with a 2-cell

$$\eta: f \Rightarrow p \circ \widetilde{f}$$

such that given any pair $(g: A \to B, \theta: f \Rightarrow p \circ g)$, there exists a unique 2-cell

$$\zeta: \widetilde{f} \Rightarrow g$$

such that $(p \circ \zeta) \bullet \eta = \theta$.

Global description

Given $p: B \to C$ and a 0-cell $A$ in a bicategory, if the right Kan lift $Rift_p f$ exists for any $f: A \to C$, then we speak of a global Kan lift. When this is the case, we may define a functor between hom-categories

$$Rift_p: [A, C] \to [A, B]$$

which at the object level is of course $f \mapsto Rift_p f$. At the morphism level, given a 2-cell

$$\eta: f \Rightarrow g: A \to C$$

there is an induced 2-cell $Rift_p f \Rightarrow Rift_p g$, the one which corresponds (by the universal property of $Rift_p g$) to the composite

$$p \circ Rift_p f \overset{\varepsilon}{\Rightarrow} f \overset{\eta}{\Rightarrow} g.$$

A standard universality argument shows that $Rift_p$ thus defined is functorial.

Proposition: $Rift_p$ is right adjoint to the functor $[A, p]: [A, B] \to [A, C]$ obtained by postcomposing with $p$, i.e., the functor $f \mapsto p \circ f$.

This is the easier, more conceptual way to remember right Kan lifts: in a global sense, they are right adjoint to postcomposition. Similarly, global left Kan lifts are left adjoint to postcomposition.

• In a post at the nCafé, David Corfield put the question: why are Kan extensions more frequently mentioned than Kan lifts? One proposed answer is that in many cases a left or right Kan lift will exist for a trivial reason: namely that $p$ itself has a left or right adjoint. Other examples will be given below.

Properties

Dual to the situation in Kan extensions, one is interested in whether a Kan lift is respected by a 1-cell with codomain the domain of the lift. This is defined as follows:

• given $(\widetilde{f}, \varepsilon) = \mathop{Rift}_p f$, $g$ is said to respect this right lift if $(\widetilde{f}g, \varepsilon \bullet g) = \mathop{Rift}_p(f g)$

and analogously for left Kan lifts. A Kan lift $\widehat{f}$ is absolute if it is respected by any 1-cell into $\mathop{dom}(\widehat{f})$. Absolute Kan lifts subsume adjunctions and relative adjunctions, and are prominently present in the axioms of a Yoneda structure; for more see the examples below.

Examples

• Let $Mod$ denote the framed bicategory of small categories $C$, bimodules $r: C \to D$ (i.e., functors $r: D^{op} \times C \to Set$), and transformations between such bimodules. Then global right Kan lifts exist (as do global right Kan extensions). These may be computed as weighted limits or via end formulas, e.g.,
  $$(Rift_r s)(b, c) = \int_{d: D} hom(r(d, c), s(d, b))$$

Examples of this construction abound in mathematics, especially when generalized to the enriched category theory context. For example, in the bicategory $Rel$, which corresponds to enrichment in $\mathbf{2}$, the right Kan lift is essentially a universally quantified predicate of the form

$$\forall_{d: D} r(d, c) \Rightarrow s(d, b)$$

(“for all $d$ satisfying condition $r$, we impose condition $s$”).

• More generally, a biclosed bicategory is precisely a bicategory where global right Kan extensions and right Kan lifts exist for every 1-cell $p$. Monoidal bicategories provide instances of this.

• adjunctions in a 2-category can be defined in terms of Kan lifts: a 1-cell $u\colon A \to B$ has a left adjoint iff $\mathop{Lift}_u 1_B$ exists and is absolute; in this case putting $(f,\iota) = \mathop{Lift}_u 1_B$ we have $f \dashv u$ with unit $\iota \colon 1_B \Rightarrow u f$. The universal property of the left Kan lift plus absoluteness are enough to construct the counit and to verify the triangular equations. There’s of course a dual definition in terms of absolute right Kan lifts.

• relative adjoints in $\mathbf{Cat}$ can also be expressed as absolute kan lifts; see relative adjoint for a precise statement.

• representably fully faithful 1-cells, meaning those for which $B(X,f)$ is fully faithful in $\mathbf{Cat}$ for every object $X\colon B$, are those for which $(1_A, 1_f) = \mathop{Lift}_f f$, and this lifting is absolute.

• In $Cat$, if $A$ is small and $B$ is locally small, and if $F: A \to B$ is a functor, then we have a Yoneda embedding $y: A \to P A = Set^{A^{op}}$ and a functor $B(F-, -): B \to P A$, and there is a canonical map

  $$y_A \to B(F-, -) \circ f$$

  (essentially, $hom(a, b) \to hom(F a, F b)$ taking $f: a \to b$ to $F f: F a \to F b$). This arrow exhibits $F$ as a left Kan lift of $y$ through $B(F-, -)$, which is moreover absolute. This example is important in the theory of Yoneda structures, due to Street and Walters; see Mark Weber’s updated development in the context of 2-topos theory.

References

Kan lifts were mentioned in passing in Definition I.12.6 of:

• John Gray, Formal category theory: adjointness for $2$-categories, Lecture Notes in Mathematics, 391, Springer (1974) [doi:10.1007/BFb0061280]

where it is mentioned the concept does not have a name since they had never been considered before in Cat.

Kan lifts were called “liftings” in:

• Ross Street, Bob Walters, Yoneda structures on 2-categories, JPAA 50 (1978) 350-379 [doi:10.1016/0021-8693(78)90160-6]

The terminology “Kan lifting” is used in:

• Charles Grellois, Algebraic theories, monads, and arities (2011) [arXiv:1110.3294, hal:00670841]

The terminology “Kan lift” seems to first appear in:

• Weng Kin Ho, Characterising E-projectives via Co-monads, Electronic Notes in Theoretical Computer Science 301 (2014) 61-77 [doi:10.1016/j.entcs.2014.01.006]