Abstract
Pose Graph Optimization (PGO) is the problem of estimating a set of poses from pairwise relative measurements. PGO is a nonconvex problem, and currently no known technique can guarantee the efficient computation of a global optimal solution. In this paper, we show that Lagrangian duality allows computing a globally optimal solution, under certain conditions that are satisfied in many practical cases. Our first contribution is to frame the PGO problem in the complex domain. This makes analysis easier and allows drawing connections with the recent literature on unit gain graphs. Exploiting this connection we prove nontrival results about the spectrum of the matrix underlying the problem. The second contribution is to formulate and analyze the properties of the Lagrangian dual problem in the complex domain. The dual problem is a semidefinite program (SDP). Our analysis shows that the duality gap is connected to the number of eigenvalues of the penalized pose graph matrix, which arises from the solution of the SDP. We prove that if this matrix has a single eigenvalue in zero, then (1) the duality gap is zero, (2) the primal PGO problem has a unique solution, and (3) the primal solution can be computed by scaling an eigenvector of the penalized pose graph matrix. The third contribution is algorithmic: we exploit the dual problem and propose an algorithm that computes a guaranteed optimal solution for PGO when the penalized pose graph matrix satisfies the Single Zero Eigenvalue Property (SZEP). We also propose a variant that deals with the case in which the SZEP is not satisfied. This variant, while possibly suboptimal, provides a very good estimate for PGO in practice. The fourth contribution is a numerical analysis. Empirical evidence shows that in the vast majority of cases (100 % of the tests under noise regimes of practical robotics applications) the penalized pose graph matrix does satisfy the SZEP, hence our approach allows computing the global optimal solution. Finally, we report simple counterexamples in which the duality gap is nonzero, and discuss open problems.
Access this chapter
Tax calculation will be finalised at checkout
Purchases are for personal use only
Notes
- 1.
We use the term “rotation subproblem” to denote the problem of associating a rotation to each node in the graph, using relative rotation measurements. This corresponds to disregarding the translation measurements in PGO.
- 2.
We use the somehow standard term “composition” to denote the group operation for SE(2). For two poses T 1 ≐(p 1, R 1) and T 2 ≐(p 2, R 2), the composition is T 1 ⋅ T 2 = (p 1 + R 1 p 2, R 1 R 2) [16]. Similarly, the identity element is (02, I 2).
- 3.
When composing measurements along the loop, edge direction is important: for two consecutive edges (i, k) and (k, j) along the loop, the composition is T ij = T ik ⋅ T kj , while if the second edge is in the form (j, k), the composition becomes T ij = T ik ⋅ T jk −1.
- 4.
\(\tilde{V }\) can be computed from singular value decomposition of \(\tilde{W}(\lambda ^{\star })\).
- 5.
This was not included in the first row of Fig. 7 as it does not provide a guess for the positions of the nodes.
References
Aragues, R., Carlone, L., Calafiore, G., Sagues, C.: Multi-agent localization from noisy relative pose measurements. In: IEEE International Conference on Robotics and Automation (ICRA), Shanghai, pp. 364–369 (2011)
Arie-Nachimson, M., Kovalsky, S.Z., Kemelmacher-Shlizerman, I., Singer, A., Basri, R.: Global motion estimation from point matches. In: 3DIMPVT, 2012
Bandeira, A.S., Singer, A., Spielman, D.A.: A Cheeger inequality for the graph connection Laplacian. SIAM. J. Matrix Anal. Appl. 34 (4), 1611–1630 (2013)
Barooah, P., Hespanha, J.P.: Estimation on graphs from relative measurements. Control Syst. Mag. 27 (4), 57–74 (2007)
Biswas, P., Lian, T., Wang, T., Ye, Y.: Semidefinite programming based algorithms for sensor network localization. ACM Trans. Sensor Netw. 2 (2), 188–220 (2006)
Borra, D., Carli, R., Lovisari, E., Fagnani, F., Zampieri, S.: Autonomous calibration algorithms for planar networks of cameras. In Proceedings of the 2012 American Control Conference, pp. 5126–5131 (2012)
Boyd, S., Vandenberghe, L.: Convex Optimization. Cambridge University Press, Cambridge (2004)
Calafiore, G.C., Carlone, L., Wei, M.: Distributed optimization techniques for range localization in networked systems. In: IEEE Conference on Decision and Control, pp. 2221–2226 (2010)
Carlevaris-Bianco, N., Eustice, R.M.: Generic factor-based node marginalization and edge sparsification for pose-graph SLAM. In: IEEE International Conference on Robotics and Automation (ICRA), pp. 5728–5735 (2013)
Carlone, L.: Convergence analysis of pose graph optimization via Gauss-Newton methods. In: IEEE International Conference on Robotics and Automation (ICRA), pp. 965–972 (2013)
Carlone, L., Censi, A.: From angular manifolds to the integer lattice: guaranteed orientation estimation with application to pose graph optimization. IEEE Trans. Robot. 30 (2), 475–492 (2014)
Carlone, L., Dellaert, F.: Duality-based verification techniques for 2D SLAM. In: International Conference on Robotics and Automation (ICRA), pp. 4589–4596 (2014)
Carlone, L., Aragues, R., Castellanos, J.A., Bona, B.: A linear approximation for graph-based simultaneous localization and mapping. In: Robotics: Science and Systems (RSS) (2011)
Carlone, L., Aragues, R., Castellanos, J.A., Bona, B.: A fast and accurate approximation for planar pose graph optimization. Intl. J. Robot. Res. 33 (7), 965–987 (2014)
Carlone, L., Tron, R., Daniilidis, K., Dellaert, F.: Initialization techniques for 3D SLAM: a survey on rotation estimation and its use in pose graph optimization. In: IEEE International Conference on Robotics and Automation (ICRA) (2015)
Chirikjian, G.S.: Stochastic Models, Information Theory, and Lie Groups, Volume 2: Analytic Methods and Modern Applications (Applied and Numerical Harmonic Analysis). Birkhauser, Basel (2012)
Chiuso, A., Picci, G., Soatto, S.: Wide-sense estimation on the special orthogonal group. Commun. Inf. Syst. 8, 185–200 (2008)
Chung, F.R.K.: Spectral Graph Theory. CBMS Regional Conference Series in Mathematics, vol. 92 American Mathematical Society, Providence, RI (1996)
Costa, J., Patwari, N., Hero, A.: Distributed weighted-multidimensional scaling for node localization in sensor networks. ACM Trans. Sensor Netw. 2 (1), 39–64 (2006)
Cucuringu, M., Lipman, Y., Singer, A.: Sensor network localization by eigenvector synchronization over the Euclidean group. ACM Trans. Sensor Netw. 8 (3), 19:1–19:42 (2012)
Cucuringu, M., Singer, A., Cowburn, D.: Eigenvector synchronization, graph rigidity and the molecule problem. Inf. Infer.: J. IMA 1 (1), 21–67 (2012)
Dellaert, F., Carlson, J., Ila, V., Ni, K., Thorpe, C.E.: Subgraph-preconditioned conjugate gradient for large scale slam. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (2010)
Doherty, L., Pister, K., El Ghaoui, L.: Convex position estimation in wireless sensor networks. In: IEEE INFOCOM, vol. 3, pp. 1655–1663 (2001)
Dubbelman, G., Browning, B.: Closed-form online pose-chain slam. In: IEEE International Conference on Robotics and Automation (ICRA) (2013)
Dubbelman, G., Esteban, I., Schutte, K.: Efficient trajectory bending with applications to loop closure. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1–7 (2010)
Dubbelman, G., Hansen, P., Browning, B., Dias, M.B.: Orientation only loop-closing with closed-form trajectory bending. In: IEEE International Conference on Robotics and Automation (ICRA) (2012)
Duckett, T., Marsland, S., Shapiro, J.: Fast, on-line learning of globally consistent maps. Auton. Robot. 12 (3), 287–300 (2002)
Eckart, C., Young, G.: The approximation of one matrix by another low rank. Psychometrika 1, 211–218 (1936)
Eren, T., Goldenberg, O.K., Whiteley, W., Yang, Y.R.: Rigidity, computation, and randomization in network localization. In: INFOCOM 2004. Twenty-third Annual Joint Conference of the IEEE Computer and Communications Societies, vol. 4, pp. 2673–2684. IEEE, Hong Kong (2004)
Eren, T., Whiteley, W., Belhumeur, P.N.: Using angle of arrival (bearing) information in network localization. In: IEEE Conference on Decision and Control, pp. 4676–4681 (2006)
Eustice, R.M., Singh, H., Leonard, J.J., Walter, M.R.: Visually mapping the RMS Titanic: conservative covariance estimates for SLAM information filters. Int. J. Robot. Res. 25 (12), 1223–1242 (2006)
Fredriksson, J., Olsson, C.: Simultaneous multiple rotation averaging using Lagrangian duality. In: Asian Conference on Computer Vision (ACCV) (2012)
Frese, U., Larsson, P., Duckett, T.: A multilevel relaxation algorithm for simultaneous localisation and mapping. IEEE Trans. Robot. 21 (2), 196–207 (2005)
Govindu, V.M.: Combining two-view constraints for motion estimation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 218–225 (2001)
Govindu, V.M.: Lie-algebraic averaging for globally consistent motion estimation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2004)
Grisetti, G., Stachniss, C., Burgard, W.: Non-linear constraint network optimization for efficient map learning. Trans. Intell. Transport. Syst. 10 (3), 428–439 (2009)
Hartley, R., Trumpf, J., Dai, Y., Li, H.: Rotation averaging. IJCV 103 (3), 267–305 (2013)
Hatanaka, T., Fujita, M., Bullo, F.: Vision-based cooperative estimation via multi-agent optimization. In: IEEE Conference on Decision and Control (2010)
Hazewinkel, M. (ed.) Complex number. In: Encyclopedia of Mathematics. Springer, New York (2001)
Huang, S., Lai, Y., Frese, U., Dissanayake, G.: How far is SLAM from a linear least squares problem? In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 3011–3016 (2010)
Huang, S., Wang, H., Frese, U., Dissanayake, G.: On the number of local minima to the point feature based SLAM problem. In: IEEE International Conference on Robotics and Automation (ICRA), pp. 2074–2079 (2012)
Indelman, V., Nelson, E., Michael, N., Dellaert, F.: Multi-robot pose graph localization and data association from unknown initial relative poses via expectation maximization. In: IEEE International Conference on Robotics and Automation (ICRA) (2014)
Johannsson, H., Kaess, M., Fallon, M., Leonard, J.J.: Temporally scalable visual SLAM using a reduced pose graph. In: IEEE International Conference on Robotics and Automation (ICRA), pp. 54–61 (2013)
Kaess, M., Ranganathan, A., Dellaert, F.: iSAM: incremental smoothing and mapping. IEEE Trans. Robot. 24 (6), 1365–1378 (2008)
Kaess, M., Johannsson, H., Roberts, R., Ila, V., Leonard, J., Dellaert, F.: iSAM2: incremental smoothing and mapping using the Bayes tree. Int. J. Robot. Res. 31, 217–236 (2012)
Kim, B., Kaess, M., Fletcher, L., Leonard, J., Bachrach, A., Roy, N., Teller, S.: Multiple relative pose graphs for robust cooperative mapping. In: IEEE International Conference on Robotics and Automation (ICRA), Anchorage, Alaska, May 2010, pp. 3185–3192
Knuth, J., Barooah, P.: Collaborative 3D localization of robots from relative pose measurements using gradient descent on manifolds. In: IEEE International Conference on Robotics and Automation (ICRA), pp. 1101–1106 (2012)
Knuth, J., Barooah, P.: Collaborative localization with heterogeneous inter-robot measurements by Riemannian optimization. In: IEEE International Conference on Robotics and Automation (ICRA) (2013)
Knuth, J., Barooah, P.: Error growth in position estimation from noisy relative pose measurements. Robot. Auton. Syst. 61 (3), 229–224 (2013)
Konolige, K.: Large-scale map-making. In: Proceedings of the 21th AAAI National Conference on AI, San Jose, CA (2004)
Kümmerle, R., Grisetti, G., Strasdat, H., Konolige, K., Burgard, W.: g2o: A general framework for graph optimization. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), Shanghai, May 2011
Levenberg, K.: A method for the solution of certain nonlinear problems in least squares. Quart. Appl. Math 2 (2), 164–168 (1944)
Liu, M., Huang, S., Dissanayake, G., Wang, H.: A convex optimization based approach for pose SLAM problems. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 1898–1903 (2012)
Lu, F., Milios, E.: Globally consistent range scan alignment for environment mapping. Auton. Robots 4, 333–349 (1997)
Mao, G., Fidan, B., Anderson, B.: Wireless sensor network localization techniques. Comput. Networks 51 (10), 2529–2553 (2007)
Martinec, D., Pajdla, T.: Robust rotation and translation estimation in multiview reconstruction. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–8 (2007)
Olfati-Saber, R.: Swarms on sphere: a programmable swarm with synchronous behaviors like oscillator networks. In: IEEE Conference on Decision and Control, pp. 5060–5066 (2006)
Olson, E., Leonard, J., Teller, S.: Fast iterative alignment of pose graphs with poor initial estimates. In: IEEE International Conference on Robotics and Automation (ICRA), May 2006, pp. 2262–2269
Peters, J.R., Borra, D., Paden, B., Bullo, F.: Sensor network localization on the group of 3D displacements. SIAM J. Control Optim. (2014, submitted)
Piovan, G., Shames, I., Fidan, B., Bullo, F., Anderson, B.: On frame and orientation localization for relative sensing networks. Automatica 49 (1), 206–213 (2013)
Reff, N.: Spectral properties of complex unit graphs (2011). arXiv 1110.4554
Rosen, D.M., Kaess, M., Leonard, J.J.: An incremental trust-region method for robust online sparse least-squares estimation. In: IEEE International Conference on Robotics and Automation (ICRA), St. Paul, MN, May 2012, pp. 1262–1269
Rosen, D.M., Kaess, M., Leonard, J.J.: RISE: an incremental trust-region method for robust online sparse least-squares estimation. IEEE Trans. Robot. 30 (5), 1091–1108 (2014)
Russell, W.J., Klein, D.J., Hespanha, J.P.: Optimal estimation on the graph cycle space. IEEE Trans. Signal Process. 59 (6), 2834–2846 (2011)
Sarlette, A., Sepulchre, R.: Consensus optimization on manifolds. SIAM J. Control Optim. 48 (1), 56–76 (2009)
Saunderson, J., Parrilo, P.A., Willsky, A.: Semidefinite descriptions of the convex hull of rotation matrices (2014). arXiv preprint: http://arxiv.org/abs/1403.4914
Saunderson, J., Parrilo, P.A., Willsky, A.: Semidefinite relaxations for optimization problems over rotation matrices. In: IEEE Conference on Decision and Control, May 2014
Schrijver, A.: Theory of Linear and Integer Programming. Wiley, New York (1998)
Sharp, G.C., Lee, S.W., Wehe, D.K.: Multiview registration of 3D scenes by minimizing error between coordinate frames. IEEE Trans. Pattern Anal. Mach. Intell. 26 (8), 1037–1050 (2004)
Singer, A.: A remark on global positioning from local distances. Proc. Natl. Acad. Sci. 105 (28), 9507–9511 (2008)
Singer, A.: Angular synchronization by eigenvectors and semidefinite programming. Appl. Comput. Harmon. Anal. 30, 20–36 (2010)
Singer, A., Shkolnisky, Y.: Three-dimensional structure determination from common lines in Cryo-EM by eigenvectors and semidefinite programming. SIAM J. Imag. Sci. 4 (2), 543–572 (2011)
Stanfield, R.: Statistical theory of DF finding. J. IEE 94 (5), 762–770 (1947)
Thunberg, J., Montijano, E., Hu, X.: Distributed attitude synchronization control. In: IEEE Conference on Decision and Control (2011)
Tron, R., Afsari, B., Vidal, R.: Intrinsic consensus on SO(3) with almost global convergence. In: IEEE Conference on Decision and Control (2012)
Tron, R., Afsari, B., Vidal, R.: Riemannian consensus for manifolds with bounded curvature. IEEE Trans. Autom. Control 58 (4), 921–934 (2012)
Tron, R., Carlone, L., Dellaert, F., Daniilidis, K.: Rigid components identification and rigidity enforcement in bearing-only localization using the graph cycle basis. In: American Control Conference (2015)
Wang, H., Hu, G., Huang, S., Dissanayake, G.: On the structure of nonlinearities in pose graph SLAM. In: Robotics: Science and Systems (RSS) (2012)
Wang, L., Singer, A.: Exact and stable recovery of rotations for robust synchronization. Inf. Infer.: J. IMA pp. 1–53 (2013) doi: 10.1093/imaiai/drn000
Zhang, S.: Quadratic maximization and semidefinite relaxation. Math. Programm. Ser. A 87, 453–465 (2000)
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Appendix
Appendix
Proof of Proposition 1: Zero Cost in Trees
We prove Proposition 1 by inspection, providing a procedure to build an estimate that annihilates every summand in (8). The procedure is as follows:
-
1.
Select a root node, say the first node (p i , r i ), with i = 1, and set it to the origin, i.e., p i = 02, r i = [1 0]⊤ (compare with (7) for θ i = 0);
-
2.
For each neighbor j of the root i, if j is an outgoing neighbor, set r j = R ij r i , and p j = p i + D ij r i , otherwise set r j = R ji ⊤ r i , and p j = p i + D ji r j ;
-
3.
Repeat point 2 for the unknown neighbors of every node that has been computed so far, and continue until all poses have been computed.
Let us now show that this procedure produces a set of poses that annihilates the objective in (8). According to the procedure, we set the first node to the origin: p 1 = 02, r 1 = [1 0]⊤; then, before moving to the second step of the procedure, we rearrange the terms in (8): we separate the edges into two sets \(\mathcal{E} = \mathcal{E}_{1} \cup \bar{\mathcal{E}}_{1}\), where \(\mathcal{E}_{1}\) is the set of edges incident on node 1 (the root), and \(\bar{\mathcal{E}}_{1}\) are the remaining edges. Then the cost can be written as:
We can further split the set \(\mathcal{E}_{1}\) into edges that have node 1 as a tail (i.e., edges in the form (1, j)) and edges that have node 1 as head (i.e., (j, 1)):
Now, we set each node j in the first two summands as prescribed in step 2 of the procedure. By inspection one can verify that this choice annihilates the first two summands and the cost becomes:
Now we select a node k that has been computed at the previous step, but has some neighbor that is still unknown. As done previously, we split the set \(\bar{\mathcal{E}}_{1}\) into two disjoint subsets: \(\bar{\mathcal{E}}_{1} = \mathcal{E}_{k} \cup \bar{\mathcal{E}}_{k}\), where the set \(\mathcal{E}_{k}\) contains the edges in \(\bar{\mathcal{E}}_{1}\) that are incident on k, and \(\bar{\mathcal{E}}_{k}\) contains the remaining edges:
Again, setting neighbors j as prescribed in step 2 of the procedure, annihilates the first two summands in (51). Repeating the same reasoning for all nodes that have been computed, but still have unknown neighbors, we can easily show that all terms in (51) become zero (the assumption of graph connectivity ensures that we can reach all nodes), proving the claim.
Proof of Proposition 2: Zero Cost in Balanced Graphs
Similarly to section “Proof of Proposition 1: Zero Cost in Trees” in this Appendix, we prove Proposition 2 by showing that in balanced graphs one can always build a solution that attains zero cost.
For the assumption of connectivity, we can find a spanning tree \(\mathcal{T}\) of the graph, and split the terms in the cost function accordingly:
where \(\bar{\mathcal{T}}\doteq\mathcal{E}\setminus \mathcal{T}\) are the chords of the graph w.r.t. \(\mathcal{T}\).
Then, using the procedure in section “Proof of Proposition 1: Zero Cost in Trees” in this Appendix we construct a solution {r i ⋆, p i ⋆} that attains zero cost for the measurements in the spanning tree \(\mathcal{T}\). Therefore, our claim only requires to demonstrate that the solution built from the spanning tree also annihilates the terms in \(\bar{\mathcal{T}}\):
To prove the claim, we consider one of the chords in \(\bar{\mathcal{T}}\) and we show that the cost at {r i ⋆, p i ⋆} is zero. The cost associated to a chord \((i,j) \in \bar{\mathcal{T}}\) is:
Now consider the unique path \(\mathcal{P}_{ij}\) in the spanning tree \(\mathcal{T}\) that connects i to j, and number the nodes along this path as i, i + 1, …, j − 1, j.
Let us start by analyzing the second summand in (54), which corresponds to the rotation measurements. According to the procedure in section “Proof of Proposition 1: Zero Cost in Trees” in this Appendix to build the solution for \(\mathcal{T}\), we propagate the estimate from the root of the tree. Then it is easy to see that:
where R ii+1 is the rotation associated to the edge (i, i + 1), or its transpose if the edge is in the form (i + 1, i) (i.e., it is traversed backwards along \(\mathcal{P}_{ij}\)). Now we notice that the assumption of balanced graph implies that the measurements compose to the identity along every cycle in the graph. Since the chord (i, j) and the path \(\mathcal{P}_{ij}\) form a cycle in the graph, it holds:
Substituting (56) back into (55) we get:
which can be easily seen to annihilate the second summand in (54).
Now we only need to demonstrate that also the first summand in (54) is zero. The procedure in section “Proof of Proposition 1: Zero Cost in Trees” in this Appendix leads to the following estimate for the position of node j:
The assumption of balanced graph implies that position measurements compose to zero along every cycle, hence:
or equivalently:
Substituting (60) back into (58) we obtain:
which annihilates the first summand in (54), concluding the proof.
Proof of Proposition 4: Properties of \(\mathcal{W}\)
Let us prove that \(\mathcal{W}\) has (at least) two eigenvalues in zero. We already observed that the top-left block of \(\mathcal{W}\) is \(\bar{\mathcal{L}} = \mathcal{L}\otimes I_{2}\), where \(\mathcal{L}\) is the Laplacian matrix of the graph underlying the PGO problem. The Laplacian \(\mathcal{L}\) of a connected graph has a single eigenvalue in zero, and the corresponding eigenvector is 1 n (see, e.g., [18, Sects. 1.2–1.3]), i.e., \(\mathcal{L}\cdot 1_{n} = 0\). Using this property, it is easy to show that the matrix N ≐[0 n ⊤ 1 n ⊤]⊤⊗ I 2 is in the nullspace of \(\mathcal{W}\), i.e., \(\mathcal{W}N = 0\). Since N has rank 2, this implies that the nullspace of \(\mathcal{W}\) has at least dimension 2, which proves the first claim.
Let us now prove that the matrix \(\mathcal{W}\) is composed by 2 × 2 blocks \([\mathcal{W}]_{ij}\), with \([\mathcal{W}]_{ij} \in \alpha SO(2)\), \(\forall i,j = 1,\ldots,2n\), and \([\mathcal{W}]_{ii} =\alpha _{ii}I_{2}\) with α ii ≥ 0. We prove this by direct inspection of the blocks of \(\mathcal{W}\). Given the structure of \(\mathcal{W}\) in (14), the claim reduces to proving that the matrices \(\bar{\mathcal{L}}\), \(\bar{Q}\), and \(\bar{A}^{\top }\bar{D}\) are composed by 2 × 2 blocks in α S O(2), and the diagonal blocks of \(\bar{\mathcal{L}}\) and \(\bar{Q}\) are multiples of the identity matrix. To this end, we start by observing that \(\bar{\mathcal{L}} = \mathcal{L}\otimes I_{2}\), hence all blocks in \(\bar{\mathcal{L}}\) are multiples of the 2 × 2 identity matrix, which also implies that they belong to α S O(2). Consider next the matrix \(\bar{Q}\doteq\bar{D}^{\top }\bar{D} +\bar{ U}^{\top }\bar{U}\). From the definition of \(\bar{D}\) it follows that \(\bar{D}^{\top }\bar{D}\) is zero everywhere, except the 2 × 2 diagonal blocks:
Similarly, from simple matrix manipulation we obtain the following block structure of \(\bar{U}^{\top }\bar{U}\):
where d i is the degree (number of neighbours) of node i. Combining (61) and (62) we get the following structure for \(\bar{Q}\):
where we defined \(\beta _{i}\doteq d_{i} +\sum _{j\in \mathcal{N}_{ i}^{\text{out}}}\|\varDelta _{ij}\|_{2}^{2}\). Clearly, \(\bar{Q}\) has blocks in α S O(2) and the diagonal blocks are nonnegative multiples of I 2.
Now, it only remains to inspect the structure of \(\bar{A}^{\top }\bar{D}\). The matrix \(\bar{A}^{\top }\bar{D}\) has the following structure:
Note that \(\sum _{j\in \mathcal{N}_{ i}^{\text{out}}}D_{ij}\) is the sum of matrices in α S O(2), hence it also belongs to α S O(2). Therefore, also all blocks of \(\bar{A}^{\top }\bar{D}\) are in α S O(2), thus concluding the proof.
Proof of Proposition 5: Cost in the Complex Domain
Let us prove the equivalence between the complex cost and its real counterpart, as stated in Proposition 5.
We first observe that the dot product between two 2-vectors \(x_{1},x_{2} \in \mathbb{R}^{2}\), can be written in terms of their complex representation \(\tilde{x}_{1}\doteq x_{1}^{\vee }\), and \(\tilde{x}_{2}\doteq x_{2}^{\vee }\), as follows:
Moreover, we know that the action of a matrix Z ∈ α S O(2) can be written as the product of complex numbers, see (18).
Combining (65) and (18) we get:
where \(\tilde{z} = Z^{\vee }\). Furthermore, when Z is multiple of the identity matrix, it easy to see that z = Z ∨ is actually a real number, and Eq. (66) becomes:
With the machinery introduced so far, we are ready to rewrite the cost x ⊤ Wx in complex form. Since W is symmetric, the product becomes:
Using the fact that [W] ii is a multiple of the identity matrix, \(\tilde{W}_{ii}\doteq[W]_{ii}^{\vee }\in \mathbb{R}^{}\), and using (67) we conclude \(x_{i}^{\top }[W]_{ii}x_{i} =\tilde{ x}_{i}^{{\ast}}\tilde{W}_{ii}\tilde{x}_{i}\). Moreover, defining \(\tilde{W}_{ij}\doteq[W]_{ij}^{\vee }\) (these will be complex numbers, in general), and using (66), Eq. (68) becomes:
where we completed the lower triangular part of \(\tilde{W}\) as \(\tilde{W}_{ji} =\tilde{ W}_{ij}^{{\ast}}\).
Proof of Proposition 6: Zero Eigenvalues in \(\tilde{W}\)
Let us denote with N 0 the number of zero eigenvalues of the pose graph matrix \(\tilde{W}\). N 0 can be written in terms of the dimension of the matrix (\(\tilde{W} \in \mathbb{C}^{(2n-1)\times (2n-1)}\)) and the rank of the matrix:
Now, recalling the factorization of \(\tilde{W}\) given in (25), we note that:
where the second relation follows from the upper triangular structure of the matrix. Now, we know from [68, Sect. 19.3] that the anchored incidence matrix A, obtained by removing a row from the the incidence matrix of a connected graph, is full rank:
Therefore:
Now, since we recognized that \(\tilde{U}\) is the complex incidence matrix of a unit gain graph (Lemma 1), we can use the result of Lemma 2.3 in [61], which says that:
where b is the number of connected components in the graph that are balanced. Since we are working on a connected graph (Assumption 1), b can be either one (balanced graph or tree), or zero otherwise. Using (73) and (74), we obtain N 0 = b, which implies that N 0 = 1 for balanced graphs or trees, or N 0 = 0, otherwise.
Proof of Proposition 7: Spectrum of Complex and Real Pose Graph Matrices
Recall that any Hermitian matrix has real eigenvalues, and possibly complex eigenvectors. Let \(\mu \in \mathbb{R}^{}\) be an eigenvalue of \(\tilde{W}\), associated with an eigenvector \(\tilde{v} \in \mathbb{C}^{2n-1}\), i.e.,
From Eq. (75) we have, for i = 1, …, 2n − 1,
where v i is such that \(v_{i}^{\vee } =\tilde{ v}_{i}\). Since Eq. (76) holds for all i = 1, …, 2n − 1, it can be written in compact form as:
hence v is an eigenvector of the real anchored pose graph matrix W, associated with the eigenvalue μ. This proves that any eigenvalue of \(\tilde{W}\) is also an eigenvalue of W.
To prove that the eigenvalue μ is actually repeated twice in W, consider now Eq. (75) and multiply both members by the complex number \(\mathrm{e}^{\text{j} \frac{\pi }{ 2} }\):
For i = 1, …, 2n − 1, we have:
where w i is such that \(w_{i}^{\vee } =\tilde{ v}_{j}\mathrm{e}^{\text{j} \frac{\pi } { 2} }\). Since Eq. (79) holds for all i = 1, …, 2n − 1, it can be written in compact form as:
hence also w is an eigenvector of W associated with the eigenvalue μ.
Now it only remains to demonstrate that v and w are linearly independent. One can readily check that, if \(\tilde{v}_{i}\) is in the form \(\tilde{v}_{i} =\eta _{i}\mathrm{e}^{\text{j}\theta _{i}}\), then
Moreover, observing that \(\tilde{v}_{j}\mathrm{e}^{\text{j} \frac{\pi } { 2} } =\eta _{i}\mathrm{e}^{\text{j}(\theta _{i}+\pi /2)}\), then
From (81) and (82) is it easy to see that v ⊤ w = 0, thus v, w are orthogonal, hence independent. To each eigenvalue μ of \(\tilde{W}\) there thus correspond an identical eigenvalue of W, of geometric multiplicity at least two. Since \(\tilde{W}\) has 2n − 1 eigenvalues and W has 2(2n − 1) eigenvalues, we conclude that to each eigenvalue μ of \(\tilde{W}\) there correspond exactly two eigenvalues of W in μ. The previous proof also shows how the set of orthogonal eigenvectors of W is related to the set of eigenvectors of \(\tilde{W}\).
Proof of Theorem 1: Primal-dual Optimal Pairs
We prove that, given \(\lambda \in \mathbb{R}^{n}\), if an \(\tilde{x}_{\lambda } \in \mathcal{X}(\lambda )\) is primal feasible, then \(\tilde{x}_{\lambda }\) is primal optimal; moreover, λ is dual optimal, and the duality gap is zero.
By weak duality we know that for any λ:
However, if x λ is primal feasible, by optimality of f ⋆, it must also hold
Now we observe that for a feasible x λ , the terms in the Lagrangian associated to the constraints disappear and \(\mathcal{L}(x_{\lambda },\lambda ) = f(x_{\lambda })\). Using the latter equality and the inequalities (83) and (84) we get:
which implies f(x λ ) = f ⋆, i.e., x λ is primal optimal.
Further, we have that
which, combined with weak duality (d ⋆ ≤ f ⋆), implies that d ⋆ = f ⋆ and that λ attains the dual optimal value.
Numerical Data for the Toy Examples in Sect. 6
Ground truth nodes poses, written as x i = [p i ⊤, θ i ]:
Relative measurements, for each edge (i, j), written as (i, j): [Δ ij ⊤, θ ij ]:
Rights and permissions
Copyright information
© 2016 Springer International Publishing Switzerland
About this chapter
Cite this chapter
Calafiore, G.C., Carlone, L., Dellaert, F. (2016). Lagrangian Duality in Complex Pose Graph Optimization. In: Goldengorin, B. (eds) Optimization and Its Applications in Control and Data Sciences. Springer Optimization and Its Applications, vol 115. Springer, Cham. https://doi.org/10.1007/978-3-319-42056-1_5
Download citation
DOI: https://doi.org/10.1007/978-3-319-42056-1_5
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-319-42054-7
Online ISBN: 978-3-319-42056-1
eBook Packages: Mathematics and StatisticsMathematics and Statistics (R0)