Likelihood Geometry of the Squared Grassmannian

By the excision property, \chi(X_{n+1}) = \chi(X_{n + 1}^\circ) +\chi(X_{n+1} \backslash X_{n + 1}^\circ). We shall prove that \chi(X_{n+1} \backslash X_{n + 1}^\circ) = 0 by showing that the map below is a fibration on its image:

It suffices to show that the fibers are homeomorphic. Every nonempty fiber F is the complement in (\mathbb{C}^*)^2 of n - 1 lines through the origin: each minor p_{i(n + 1)} for i \geq 3 contributes one line and Q_{n + 1} degenerates into a double line since Q_n = 0.

The Euler characteristic of finitely many lines through the origin in (\mathbb{C}^*)^2 is zero because each line is a \mathbb{P}^1 with two points removed and the lines do not intersect. Because \chi((\mathbb{C}^*)^2) = 0, the Euler characteristic of the complement of n -1 lines through the origin in (\mathbb{C}^*)^2 is zero, too. By the fibration property of the Euler characteristic and because \chi(F) = 0, we conclude that \chi(X_{n + 1} \backslash X_{n + 1}^\circ) = 0.

To prove \pi_{n + 1} is a fibration, it suffices to show that the fibers of \pi_{n + 1} are homeomorphic on each open stratum S^\circ.

The fiber in \mathbb{C}^2 of a point M_n \in X_n is the complement of the lines p_{i(n + 1)} and the conic Q_{n + 1}. Since all the lines p_{i(n + 1)} pass through the origin, their intersection is the same for any M_n \in X_n. The variation in the fibers comes from the intersection with the conic. We claim that every line p_{i(n+1)} either intersects the conic Q_{n + 1} in two distinct points or not at all.

In the intersection locus of p_{i(n+1)} = 0 and Q_{n + 1} = 0,

If \sum_{j = 1}^n p_{ij}^2 \neq 0, then there are two distinct intersection points. If \sum_{j = 1}^n p_{ij}^2 = 0, then (?) becomes Q_n = 0, which is a contradiction, so the intersection of Q_{n + 1} and p_{i(n + 1)} is empty. Thus, in the fiber F_{X_n}, every line intersects the conic in two points and in the fiber F_{S_i}, every line intersects the conic in two points, except p_{i(n+1)} which does not intersect the conic.

It remains to be shown that there can be at most two lines p_{i(n+1)} which do not intersect Q_{n + 1}. When p_{i(n+1)} does not intersect Q_{n + 1} in \mathbb{C}^2, the line and conic intersect at a double point at infinity. In other words, p_{i(n+1)} is tangent to the projectivization of Q_{n + 1} at infinity. In projective space, p_{i(n + 1)} intersects Q_{n + 1} tangentially if and only if p_{i(n + 1)}^\vee lies in the dual conic Q_{n + 1}^\vee, i.e., \begin{pmatrix} 0 & -y_i & x_i \end{pmatrix} A_{n}^{-1} \begin{pmatrix} 0 & -y_i & x_i \end{pmatrix}^T =\sum_{j = 1}^n p_{ij}^2 = 0. Since at most two lines passing through the origin can be tangent to a conic, only two of the lines p_{i(n + 1)} can have an empty intersection with the conic Q_{n + 1} at once. Thus, in the fiber F_{S_i \cap S_j}, every line intersects the conic in two points, except p_{i(n + 1)} and p_{j(n + 1)} which do not intersect the conic.

In \mathbb{C}^2\backslash \{(0,0)\}, we have \chi(Q_{n + 1}) = 0, since the conic Q_{n + 1} has nonzero constant term Q_n and is therefore topologically equivalent to a \mathbb{P}^1 with two points at infinity removed. Similarly, in \mathbb{C}^2\backslash \{(0,0)\}, we have \chi(p_{i(n + 1)}) = 0, since the line p_{i(n + 1)} is topologically equivalent to a \mathbb{P}^1 with the origin and a point at infinity removed. Suppose that M_n \in X_n. Because \chi(\mathbb{C}^2\backslash \{(0,0)\}) = 0, by the excision property,

where the last equality holds because the intersection of a line with a conic is zero dimensional.

We computed these intersections in Lemma 3.2. In the generic fiber, each of the n lines intersects the conic in two points, for a total of 2n intersection points. The intersections for S_i, S_i \cap S_j are the same, but with n - 1 and n - 2 lines, respectively.

It is sufficient to prove the claim for S_n by symmetry. We first remark that S_n \subset X_n^\circ.

We argue that S_n has Euler characteristic zero by showing that \pi_n|_{S_n} is a fibration whose fiber is the union of two lines through the origin in (\mathbb{C}^*)^2.

Since the fiber has Euler characteristic zero, the result follows from the fibration property.

Because \sum_{j = 1}^{n-1} p_{jn}^2 is a homogeneous polynomial in x_{n} and y_n, the fiber of a point in \pi_n(S_n) is a degenerate conic in the variables x_n, y_n, so each fiber is the union of two lines through the origin in the complement of the lines p_{jn} and conic Q_n in (\mathbb{C}^*)^2. Since the p_{jn} all pass through the origin, they do not intersect the degenerate conic in (\mathbb{C}^*)^2. The conic Q_n also does not intersect \sum_{j = 1}^{n-1} p_{jn}^2, because Q_n - \sum_{j = 1}^{n-1} p_{jn}^2 = Q_{n - 1} = 0 on the intersection, contradicting S_n \subset X_n^\circ. Since the conic \sum_{j = 1}^{n-1} p_{jn}^2 does not intersect the lines p_{jn} or the conic Q_n, the fiber is the union of two lines through the origin in (\mathbb{C}^*)^2.

By (?), the number of critical points is equal to the Euler characteristic \chi(X_n), so it suffices to show that \chi(X_n) = 2^{n-2} (n-1)!.

We proceed by induction on n. The base case is furnished by Example 1.4. By Lemmas 2.2, 3.1, and 3.2, we have

By Lemma 3.4, \chi(S_i) = 0, so the first sum is zero. By Lemma 3.3 and the values of the Möbius function in Figure ?, the second sum is zero:

By induction and Lemma 3.3, \chi(X_{n + 1}) = \chi(F_{X_n}) \chi(X_n) = 2n(2^{n-2}(n-1)!) = 2^{n-1}n!.

Since f_0 + \cdots + f_n > 0, the function L is smooth on \mathbb{R}^d.

Because the denominator of L has larger degree than the numerator, L approaches 0 at infinity and is therefore bounded on all regions. Because L is smooth and bounded, it attains a local maximum or minimum on each region, depending on the sign of L. This proves the first inequality and second statement. The second inequality follows from the definition of ML degree.

The result follows from Lemma 4.2 with f_0 = p_{1\cdots d}^2 = 1 and the fact that (?) is nonnegative on \mathbb{R}^{d(n-d)}.

We will prove that \# {\rm sgn}({\rm Gr}_\mathbb{R}(2, n)^\circ) = 2^{n-2}(n-1)!. By Proposition 3.5, the total number of critical points is 2^{n-2}(n-1)!, so the result follows from Corollary 4.3.

To count the sign vectors, we fix 2 \leq k \leq n and begin with a matrix

in X^\mathbb{R}_{2,n} such that x_3,\ldots, x_n, y_3, \ldots, y_n > 0. We take any permutation of the last n - 2 columns and flip the sign of any of the last n - 2 columns. This process yields 2^{n - 2} (n - 2)! different matrices. We argue that given a fixed M_n, each of these matrices has a distinct sign vector. Because every matrix in X_{2,n} can arise in this way, there are 2^{n - 2} (n - 2)! possible sign vectors for a fixed k. Hence, in total, we have 2^{n - 2} (n - 1)! total possible sign vectors.

Let p be the Plücker vector of M_n and q be the Plücker vector of a matrix produced from M_n by the permutation and sign changes described above. We identify the permutation and columns whose signs were flipped. The columns i with q_{1i} < 0 had their signs flipped. Assuming that q_{1i} > 0 for all i, we uniquely identify the permutation from its inversions: for i, j \geq 3 the signs of p_{ij} and q_{ij} agree if and only if the pair ij is not an inversion.