Assume now that f : X → B is a special Lagrangian fibration with topology and discriminant locus ∆ being an amoeba. Let b ∈ Int(∆), and set M = f^{−1}(b). Set M^{o} = M\{x_{1}, x_{2}}, where x_{1}, x_{2} are the two conical singularities of M. Suppose that the tangent cones to these two conical singularities, C_{1} and C_{2}, are both cones of the form M^{0}. Then the links of these cones, Σ_{1} and Σ_{2}, are T^{2}’s, and one expects that topologically these can be described as follows. Note that M^{o} ≅ (T^{2}\{y_{1}, y_{2}}) × S^{1} where y_{1}, y_{2} are two points in T^{2}. We assume that the link Σ_{i} takes the form γ_{i} × S^{1}, where γ_{i} is a simple loop around y_{i}. If these assumptions hold, then to see how M can be smoothed, we consider the restriction maps in cohomology

H^{1}(M^{o}, R) → H^{1}(Σ_{1}, R) ⊕ H^{1}(Σ_{2}, R)

The image of this map is two-dimensional. Indeed, if we write a basis e^{i}_{1}, e^{i}_{2} of H^{1}(Σ_{i}, R) where e^{i}_{1} is Poincaré dual to [γ_{i}] × pt and e^{i}_{2} is Poincaré dual to pt × S^{1}, it is not difficult to see the image of the restriction map is spanned by {(e^{1}_{1}, e^{2}_{1})} and {(e^{1}_{2}, −e^{2}_{2})}. Now this model of a topological fibration is not special Lagrangian, so in particular we don’t know exactly how the tangent cones to M at x_{1} and x_{2} are sitting inside C^{3}, and thus can’t be compared directly with an asymptotically conical smoothing. So to make a plausibility argument, choose new bases f^{i}_{1}, f^{i}_{2} of H^{1}(Σ_{i}, R) so that if M(a,0,0), M(0,a,0) and M(0,0,a) are the three possible smoothings of the two singular tangent cones at the singular points x_{1}, x_{2} of M. Then Y(M_{i}^{(a,0,0)}) = πaf^{i}_{1}, Y(M_{i}^{(0,a,0)}) = πaf^{i}_{2}, and Y(M_{i}(0,0,a)) = −πa(f^{i}_{1} + f^{i}_{2}).

Suppose that in this new basis, the image of the restriction map is spanned by the pairs (f^{1}_{1}, rf^{2}_{2}) and (rf^{1}_{2}, f^{2}_{1}) for r > 0, r ≠ 1. Then, there are two possible ways of smoothing M, either by gluing in M_{1}^{(a,0,0)} and M_{2}^{(0,ra,0)} at the singular points x_{1} and x_{2} respectively, or by gluing in M_{1}^{(0,ra,0)} and M_{2}^{(a,0,0)} at x_{1} and x_{2} respectively. This could correspond to deforming M to a fiber over a point on one side of the discriminant locus of f or the other side. This at least gives a plausibility argument for the existence of a special Lagrangian fibration of the topological type given by f. To date, no such fibrations have been constructed, however.

On giving a special Lagrangian fibration with codimension one discriminant and singular fibers with cone over T^{2} singularities, one is just forced to confront a codimension one discriminant locus in special Lagrangian fibrations. This leads inevitably to the conclusion that a “strong form” of the **Strominger-Yau-Zaslow conjecture** cannot hold. In particular, one is forced to conclude that if f : X → B and f’ : X’ → B are dual special Lagrangian fibrations, then their discriminant loci cannot coincide. Thus one cannot hope for a fiberwise definition of the dualizing process, and one needs to refine the concept of dualizing fibrations. Let us see why the discriminant locus must change under dualizing. The key lies in the behaviour of the positive and negative vertices, where in the positive case the critical locus of the local model of the fibration is a union of three holomorphic curves, while in the negative case the critical locus is a pair of pants. In a “generic” special Lagrangian fibration, we expect the critical locus to remain roughly the same, but its image in the base B will be fattened out. In the negative case, this image will be an amoeba. In the case of the positive vertex, the critical locus, at least locally, consists of a union of three holomorphic curves, so that we expect the discriminant locus to be the union of three different amoebas. The figure below shows the new discriminant locus for these two cases.

Now, under dualizing, positive and negative vertices are interchanged. Thus the discriminant locus must change. This is all quite speculative, of course, and underlying this is the assumption that the discriminant loci are just fattenings of the graphs. However, it is clear that a new notion of dualizing is necessary to cover this eventuality.