New Classes of Counterexamples to Hendrickson’s Global Rigidity Conjecture

Write the stress condition for each vertex v_i, and arrange them into a matrix such that every vector in the kernel is a stress. It is not difficult to show that this matrix is exactly df_G(p)^T.

If a_2, \ldots, a_{k-1} \geq d+1, then removing any d vertices leaves at least one vertex in each independent set, so the graph remains connected. If not, then for some i, 2 \leq i \leq k-1, a_i \leq d, so one can remove A_i, disconnecting the graph.

This k-chain is the subgraph of some complete bipartite graph. Both independent sets of this complete bipartite graph have more than d + 2 vertices. Since the complete bipartite graph has \binom{d+2}{2} vertices, by Theorem 5 it is GPR, and thus it is not GGR. So, the k-chain is the subgraph of a graph which is not GGR, and so is not GGR itself.

We use induction on n.

[Base Case \Rightarrow] For n = 1, first suppose G\setminus\{e_1\} is GLR. Since G(p) is locally rigid, \operatorname{rank} df_G(p) = \operatorname{rank} df_{G\setminus\{e_1\}}(p). Adding e_1 to G\setminus\{e_1\} does not increase the rank of the rigidity matrix, so it increases the dimension of \ker(df_G(p)^T) by 1. This means adding e_1 adds a new dimension of stresses, which is only possible if there is some stress with a non-zero value on e_1. By scaling this stress, we can achieve any prescribed value on e_1.

[Base Case \Leftarrow] Assume there is some stress with a non-zero value on e_1. Removing one edge can decrease the dimension of \ker (df_G(p)^T) by at most 1. Moreover, there is a stress with a non-zero value on e_1, and since this stress cannot exist without e_1, removing e_1 must decrease \dim\ker(df_G(p)^T) by exactly 1. But, the number of rows of df_G(p) also decreases by 1, so \operatorname{rank} df_{G}(p) stays the same. Therefore, \operatorname{rank} df_G(p) = \operatorname{rank} df_{G\setminus\{e_1\}}(p), and so G\setminus\{e_1\} is GLR.

[Inductive Step \Rightarrow] Assume that for some n, when G\setminus \{e_1, \ldots, e_n\} is GLR, there is a stress on G(p) with any a_1, \ldots, a_n on e_1, \ldots, e_n. Then assume that G\setminus \{e_1, \ldots, e_{n+1}\} is GLR. Let H = G\setminus \{e_1, \ldots, e_n\}. First, note that H is also GLR, so we can create a stress on a generic framework G(p) with values a_1, \ldots, a_n on e_1, \ldots, e_n. Call the stress we create \omega. Because H\setminus \{e_{n+1}\} is GLR, we can create some stress of H(p) with any value we like on e_{n+1} by Base Case \Rightarrow, and we can artificially extend it to a stress of G(p) with values of 0 on e_1, \ldots, e_n. We give this stress the value on e_{n+1} such that, when we compose it with \omega, we create a stress with values a_1, \ldots, a_{n+1} on e_1, \ldots, e_{n+1}.

[Inductive Step \Leftarrow] Assume that for some n, if we can find a stress with any value on e_1, \ldots, e_n, G\setminus \{e_1, \ldots, e_n\} is GLR. Then, suppose we can find some stress with any value we want on G\setminus \{e_1, \ldots, e_{n+1}\}. By the inductive hypothesis, H is GLR. If we set a_1, \ldots, a_{n} to all be zero, then we can find a stress on H with any value we wish on e_{n+1}. So, by Base Case \Leftarrow, G\setminus \{e_1, \ldots, e_{n+1}\} is GLR.

If there is a non-zero stress on every edge of G, then by scaling, we can find a stress of any value we want on any edge of G. Thus, by Proposition 14, each edge is redundant and G is GRR.

On the other hand, if the only stress on some edge is the zero stress, we cannot find a stress of any value on that edge. Thus, by Proposition 14, G is not GRR.

C_{a_1, a_2, \cdots, a_k} is a subgraph of a complete bipartite graph with the same vertices, which has already been proved to be GLR. Moreover, as shown above, we can put arbitrary stresses on all of the edges that must be removed to create C_{a_1, a_2, \cdots, a_k}. By Proposition 14, the k-chain is GLR.

To find the dimension of the space of stresses of the 3-chain cover, use the inclusion-exclusion principle. The overlap among the stresses stems from the 2-chains shared by adjacent 3-chains. So, using Corollary 13 and some simple algebra, the dimension of the space of stresses of the 3-chain cover is:

If v = \binom{d+2}{2}, the reader can verify that this is also e - vd + \binom{d+1}{2}. Since the stresses of the 3-chain cover constitute a subspace of the total space of stresses with equal dimension, they account for the entire space of stresses of C_{a_1, a_2, \cdots, a_k}.

First, note that by Lemma 17, C_{a_1, a_2, \cdots, a_k} is GLR. We will apply Corollary 15 directly in the rest of this proof, so we only have to determine whether the stresses of the graphs are non-zero on every edge.

[\Rightarrow] Suppose either condition 1 or condition 2 does not hold. If a_2 < d+2, then by Corollary 13, the 3-chain C_{a_1,a_2,a_3} (or bipartite graph K_{a_1+a_3,a_2}) will have a zero-dimensional space of stresses, and this is the only 3-chain which includes the edges connecting A_1 and A_2. Any stress on the k-chain will have entries of 0 on these edges, meaning that they are not redundant and as a consequence, the graph is not GRR. The same argument applies to a_{k-1}.

Moreover, suppose that for some i, a_i = a_{i+1} = d + 1. Then by Corollary 13, both C_{a_{i-1}, a_i, a_{i + 1}} and C_{a_i, a_{i+1}, a_{i+2}} have only the zero stress. These are the only two 3-chains that cover the edges between A_i and A_{i+1}, so by Lemma 18, any stress on the k-chain will have entries of 0 on these edges.

[\Leftarrow] Assume that a_2, a_{k-1} \geq d + 2 and there is no i such that a_i = a_{i+1} = d+1. Firstly, there is a non-zero stress covering the edges between A_1 and A_2. To see this, consider C_{a_1,a_2,a_3}, where each of a_1+a_3 and a_2 is at least d+2, so this 3-chain or bipartite graph has a non-zero space of stresses. Hence, each edge between A_1 and A_2 has a non-zero stress covering it. The same argument can be applied to the edges between A_{k-1} and A_k.

For the other edges, there are two cases to consider. In the first case, for all 3 \leq i \leq k-2, a_i \geq d+2. In this case, there is obviously a stress with non-zero values everywhere. Otherwise, there exists some 3 \leq i \leq k-2 such that a_i = d+1. Then we know that a_{i-1}, a_{i+1} \geq d+2, so using Corollary 13 on C_{a_{i-2},a_{i-1},a_i} and C_{a_i,a_{i+1},a_{i+2}}, we know that the edges between A_{i-1} and A_i and the edges between A_i and A_{i+1} have non-zero stresses covering them. Thus, the k-chain is GRR.

We have not yet found a conceptual proof of this fact. However, using the algorithm for testing generic local rigidity described earlier, we have found one locally rigid realization, and the algorithm cannot return a false positive for a graph being GLR, since the rank of the rigidity matrix can only decrease due to non-generic realizations and special primes. This proves that the graph is GLR.

We will prove the lemma by induction. For the base case, by Proposition 21, K_6 ~ {\cup}_{2,4} ~ C_{2,3,5,4} is generically locally rigid in \mathbb{R}^5.

For the inductive step, assume K_n ~ {\cup}_{2,4} ~ C_{2,3,5,4} is generically locally rigid in \mathbb{R}^5. Use a Hennenberg operation to create a new graph, choosing v_i and v_j to be any vertices in K_n and connecting v' to 4 other vertices in K_n. This new graph is GLR in \mathbb{R}^5 by the above theorem. Finally, we can add the edge between v_i and v_j, and connect v' to the rest of the vertices in K_n. This constructs the graph K_{n+1} ~ {\cup}_{2,4} ~ C_{2,3,5,4}, which is GLR because it is constructed by adding edges to a GLR graph.

There are two cases to be examined. We can either remove an edge from G or from C_{2,3,5,4}. We want to show that each of the resulting graphs is GLR.

In the first case, we remove an edge from G to get G', which is still GLR. Suppose G' has v vertices. G' is a subgraph of K_v, so by Proposition 14 we can assign stresses for K_v with any values on the edges of K_v \setminus G'. We know that K_v ~ {\cup}_{2,4} ~ C_{2,3,5,4} is GLR. By assigning zero stresses to the edges of C_{2,3,5,4}, it is possible to assign stresses for K_v ~ {\cup}_{2,4} ~ C_{2,3,5,4} with any values on K_v \setminus G'. Thus, by Proposition 14, G' ~ {\cup}_{2,4} ~ C_{2,3,5,4} is GLR.

In the second case, we remove an edge from C_{2,3,5,4} and denote the resulting graph C'_{2,3,5,4}. Now suppose that G has v vertices. The graph K_v ~ {\cup}_{2,4} ~ C_{2,3,5,4} is GLR, and each of the edges of the subgraph C_{2,3,5,4} is redundant. Therefore, by Proposition 14, K_v ~ {\cup}_{2,4} ~ C'_{2,3,5,4} is GLR. Finally, using the same argument as before, it is possible to create stresses for K_v with any values on K_v \setminus G. These stresses will still exist on K_v ~ {\cup}_{2,4} ~ C'_{2,3,5,4}, so by Proposition 14, G ~ {\cup}_{2,4} ~ C'_{2,3,5,4} is GLR. Consequently, G ~ {\cup}_{2,4} ~ C_{2,3,5,4} is GRR.

Since G is not GGR, we can find a generic framework G(p) and a non-equivalent framework G(q), up to Euclidean motions, in \mathbb{R}^d. Let p_1, \ldots, p_v,\allowbreak q_1, \ldots, q_v represent the locations of vertices in H, and p_{v+1}, \ldots,\allowbreak p_w,\allowbreak q_{v+1}, \ldots, q_w represent the locations of vertices in G \setminus H.

It is possible to transform G(q) into an equivalent framework G(q') with p_i = q'_i for i = 1, \ldots, v. First, through translations, make q'_1 = p_1. Since H is GGR, any other realization of vertices in H must be equivalent up to Euclidean motions. Hence, one can reflect and rotate the entire framework to ensure q'_i = p_i for i = 2, \ldots, n. Finally, since G(p) and G(q) are non-equivalent frameworks, p_i and q'_i are not all the same for i = v+1, \ldots, w. We have shown that for any generic framework G(p), there exists a non-equivalent framework with the location of the vertices in H the same.

Now connect all the edges of H for both frameworks, which has the same effect as replacing H with K_v in both frameworks. Name this new graph G'. Since no edges are added between G \setminus H and H, both p and q' preserve the edge lengths of G'. Any generic realization of G' is also a generic realization of G. Therefore, for any generic framework G'(p), there is another non-equivalent framework G'(q'), implying that G' is not GGR.

Next, replacing H with any K_i for i \geq v results in a graph that is not GGR. The case i = v is already proved. The graph obtained by replacing H with K_i, which we shall denote as G'', contains G' as a subgraph. Starting with G', p and q' as previously described, do the following to both frameworks: add i-v vertices, connect them to all the vertices in K_v only, project them onto the same set of locations in \mathbb{R}^d for both realizations, and finally, ensure that p is still generic. The new graph formed is precisely G''. No edges are added between any of the points in K_i and G'' \setminus K_i, so G''(q') is a non-equivalent realization of G''(p). Finally, all generic realizations of G'' must have the points in the subgraph G' be generic as well, so a non-equivalent realization can be found for any generic realization G''(p). In this way, G'' is not GGR for any i \geq v.

Finally, the replacement of H with H' is a subgraph of the graph obtained by replacing H with K_i for some i, so replacing H with H' results in a graph that is not generically globally rigid.

We examine K_6 ~ {\cup}_{2,4} ~ C_{2,3,5,4}. As discussed before, the only stresses of this graph come from the subgraph K_{7,7}. Consequently, K_6 ~ {\cup}_{2,4} ~ C_{2,3,5,4} has the same stress matrix, up to scale, as K_{7,7} for equivalent realizations. Both stress matrices have the same nullity and as a consequence of Theorem 3, K_6 ~ {\cup}_{2,4} ~ C_{2,3,5,4} is not GGR.

On the other hand, K_6 ~ {\cup}_{2,4} ~ C_{2,3,5,4} contains a subgraph K_6 which is GGR. Replacing K_6 with any G with at least 6 vertices forms the graph G ~ {\cup}_{2,4} ~ C_{2,3,5,4}, and by Lemma 27, this graph is not GGR.