Is there some criterion for whether a space has the homotopy type of a closed manifold (smooth or topological)? Poincare duality is an obvious necessary condition, but it's almost certainly not sufficient. Are there any other special homotopical properties of manifolds?
In surgery theory (which is basically a whole field of mathematics which tries to answer questions as the above), the next obstruction to the existence of a manifold in the homotopy type is that every finite complex with Poincaré duality is the base space of a certain distinguished fibration (Spivak normal fibration) whose fibre is homotopy equivalent to a sphere. (In order to get a unique such fibration, identify two fibrations if they are fiber homotopy equivalent or if one is obtained from the other by fiberwise suspension.) For manifolds, this fibration is the sphere bundle of the normal bundle, so the Spivak normal fibration comes from a vector bundle. This is invariant under homotopy equivalence. Thus the next obstruction is: the Spivak normal fibration must come from a vector bundle. If I remember right, then it was Novikov who first proved that for simplyconnected spaces of odd dimension at least 5, this is the only further obstruction. In general, there is a further obstruction with values in a group L_n(pi_1,w) which depends on the fundamental group, first StiefelWhitney class and the dimension. See Lück's notes on surgery theory at http://wwwmath.unimuenster.de/u/lueck/publ/lueck/ictp.pdf . 


The main result of the BrowderNovikovSullivanWall surgery theory (19621969) is that for n>4 a space X is homotopy equivalent to a compact ndimensional topological (resp. differentiable) manifold if and only if X is homotopy equivalent to a finite CW complex M with ndimensional Poincaré duality, and there is a topological (resp. vector) bundle over M for which the corresponding normal map (f,b):N> M from an ndimensional manifold N has zero surgery obstruction in the Wall Lgroup of quadratic forms over Z[π_{1}(X)]. Thus there are two obstructions, a primary topological Ktheory obstruction to the existence of a bundle, and (depending on the vanishing of the primary one, and a choice of reason) a secondary obstruction in algebraic Ltheory. The original theory was for differentiable manifolds: the extension to topological manifolds due to Kirby and Siebenmann (1970) remains a major success of surgery theory. All this is explained (at some length) in Wall's own book Surgery on compact manifolds (1970/1998) and also in my own books Algebraic Ltheory and topological manifolds (1992) and Algebraic and geometric surgery (2002), as well as many other references (such as Wolfgang Lück's notes listed in a previous post). I have made available a large number of surgeryrelated resources on my website. 


Sean: this gives Poincare space which is not homotopy equivalent to a closed manifold. the idea is that the Spivak fibration of the 5 dimensional Poincare space doesn't lift to a stable vector bundle. One can prove this as follows: let X^5 be as in Madsen and Milgram. Then X fibers over S^3 with fiber S^2. Call this fibration xi, and let the projection X > S^3 be p. Note that p has a section, call it s. It's not hard to prove that the Spivak fibration of X in this case is just p*\xi. One then needs to verify that p*\xi doesn't lift to a stable vector bundle. But if it did, then so would xi = (p o s)*\xi = s*p* \xi. But it's very easy to see that xi doesn't lift (xi is given by the nontrivial element of pi_2(F) = Z_2, where F classifies spherical fibrations with section, whereas pi_2(O) is trivial). 


Jacob Lurie gave a talk last week at Peter May's birthday conference on noncommutative Poincare duality. The idea is to take an nmanifold M and a (n1)connected space X. Then, he showed that the compact mapping space Map_c(M,X) is isomorphic to a certain homotopy colimit over a certain category of open subsets of M. This is equivalent to the usual commutative Poincare duality. However, it is not clear (to me) what the natural generalization of the statement is to nonmanifolds. So, I am not sure how to use it as a test. However, if you could use it as a test of being a manifold, it seems feasible that if the noncommutative statement held for your test space M and all (n1)connected spaces X for some X, then it would seem reasonable ask whether your test space is the homotopy type of an nmanifold. The category of open sets over which Lurie takes the colimit is the category of disjoint balls (homeomorphic to R^n) in M. Thus, a guess might be something like: if M is a space, and if U is a category of open sets of M that cover M, and if Map_c(M,X) is equivalent to the homotopy colimit of Map_c(U_i,X) for all U_i in U for all (n1)connected X for some n, then M has the homotopy type of an nmanifold. I have no idea if this is true, and even if it is true, it is not clear if it would be useful. 


Sure, lots of necessary conditions. For example, there are relations between Steenrod operations and those place some restrictions on the cohomology of a smooth manifold. 


I'm relying on memory here. A good example, which is discussed in Madsen and Milgram's book on surgery and classifying spaces for topological, PL and smooth manifolds is the set of 1connected Poincare duality spaces of dimension 5 with 4skeleton h. e. to the 4skeleton of (S^2) x (S^3), which is (S^2) v (S^3). ::::: sorry, M. O. won't let me use image tags :( :::::::: a, (S^2) x (S^3), which is obviously a manifold b. a manifold homeomorphic to SU(3)/SO(3) (though this fact isn't mentioned in Madsen & Milgram). c. a Poincare duality space whose Spivak fibration cannot be reduced to a smooth vector bundle. This is proved using secondary cohomology operations based on Steenrod squares. This was first proven by Gitler and Spivak(?). 

