MathOverflow will be down for maintenance for approximately 3 hours, starting Monday evening (06/24/2013) at approximately 9:00 PM Eastern time (UTC-4).

That said, one should always be careful about the word "duality" in this context -- often people just mean there's an a canonical isomorphism $H_G^k X \rightarrow H^G_{n-k}X$. (Which does exist using equivariant Borel-Moore homology, when $X$ is smooth.) The geometric meaning is the same as in the ordinary case: this isomorphism can be used to define the "class" of a codimension if $k$ V$is a$G$-invariant sub(manifold/variety/cycle/etc) of codimension$[V]$k$, it defines a fundamental class in $H_G^kX$.H^G_{n-k} X$, which is identified with$H_G^kX$by means of the isomorphism. Duality could also mean refer to a perfect pairing$H_G^* H_G^k X \otimes H_G^*X H_G^\ell X \rightarrow H_G^*(point)$H_G^{k+\ell-n}(point)$, given by the equivariant pushforward (integral), and coming from integral): $$\alpha\otimes\beta \mapsto \int \alpha\cdot\beta .$$ Using this pairing, dual bases for and assuming that $H_G^*X$ is free as a module over $H_G^*(point)$-modules. H_G^*(point)$, any$H_G^*(point)$-module basis for$H_G^*X$has a dual basis, in the usual sense of linear algebra. (Arguably, this duality pairing is what should really be called "Poincaré duality".) Again, the geometric meaning is similar to the ordinary case: if you're lucky enough to have invariant subvarieties whose classes form a basis for$H_G^*X$, the Poincaré dual basis is given by classes of subvarieties intersecting the original ones transversally (if they exist). However, the existence of geometrically defined dual bases is a stronger statement, because the equivariant integral is generally nonzero on classes of degree greater than$\dim X$. (I should say all this is prejudiced toward the equivariant cohomology rings that usually show up in algebraic geometry, e.g.,$H_G^*X$is a free module over$H_G^*(point)$, so that it makes sense to talk about dual bases.) PS: Defining equivariant Borel-Moore homology requires a little more care, since the spaces$EG\times^G X \rightarrow BG$are infinite-dimensional fiber bundles. But they have finite-dimensional approximations$EG_m \times^G X \to BG_m$, so it makes sense to define $$H^G_k X = H_{k+\dim BG_m}(EG_m\times^G X)$$for$m\gg0$. The equivariant Poincaré isomorphism is just the ordinary one for these approximation spaces. 1 There is a paper by Brion: "Poincaré duality and equivariant (co)homology," Michigan Math. J. 48 (2000). As mentioned in one of the comments, he uses Borel-Moore as the homology theory. That said, one should always be careful about the word "duality" in this context -- often people just mean there's an isomorphism$H_G^k X \rightarrow H^G_{n-k}X$. (Which does exist using equivariant Borel-Moore homology, when$X$is smooth.) The geometric meaning is the same as in the ordinary case: this isomorphism can be used to define the "class" of a codimension$kG$-invariant sub(manifold/variety/cycle/etc)$[V]$in$H_G^kX$. Duality could also mean a perfect pairing$H_G^* X \otimes H_G^*X \rightarrow H_G^*(point)$, given by the equivariant pushforward (integral), and coming from this, dual bases for$H_G^*X$as$H_G^*(point)$-modules. Again, the geometric meaning is similar to the ordinary case: if you're lucky enough to have invariant subvarieties whose classes form a basis for$H_G^*X$, the Poincaré dual basis is given by classes of subvarieties intersecting the original ones transversally (if they exist). However, the existence of dual bases is a stronger statement, because the equivariant integral is generally nonzero on classes of degree greater than$\dim X$. (I should say all this is prejudiced toward the equivariant cohomology rings that usually show up in algebraic geometry, e.g.,$H_G^*X$is a free over$H_G^*(point)\$, so that it makes sense to talk about dual bases.)