sparse Matérn (SPDE) precision on Delaunay meshes, for the matrix-free interferometer route #41
Unanswered
HRSAstro
asked this question in
Ideas & Proposals
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Context
aa.reg.MaternSparse(#646) builds the ν = 1 Matérn precision from the 5-point Laplacian of a rectangular mesh, and raisesNotImplementedErrorfor anything else. Since the matrix-free route (#638,Settings(interferometer_route="matrix_free")) needs a sparse H, it is limited to rectangular meshes too. #638 notes under known limitations that Delaunay meshes could use the finite-element form of the same prior; I couldn't find an issue or PR for this, so I'm proposing it here.Adaptive Delaunay meshes are where large interferometric sources most need many source pixels, so this would let the matrix-free route cover them.
Proposal
Lindgren, Rue & Lindström (2011) define the SPDE prior on a triangulation in the first place; the rectangular version is the special case.
Basis. Piecewise-linear "tent" functions φ_i on the triangulation, f(x) = Σ_i s_i φ_i(x). This is the barycentric interpolation the Delaunay mapper already uses, so the mapping matrix, the data operator, the Powell preconditioner and the SLQ evidence are unchanged.
Finite-element matrices (sparse, from the vertices and
scipy.spatial.Delaunay):Precision (ν = 1), with K = κ²C̃ + G and κ = √2 / scale:
H = coefficient · W⁻¹ c K C̃⁻¹ K W⁻¹ = coefficient · RᵀR, R = √c C̃^(−½) K W⁻¹
H links each vertex to its neighbours and their neighbours. R gives the evidence probes as now, and ln det H follows exactly from one sparse factorisation of K: N ln c + 2 ln|det K| − ln det C̃ − 2 Σ ln w_i (plus N ln coefficient).
Normalisation. On a uniform grid one constant c gives unit prior variance everywhere. On an adaptive mesh the variance can depend on the local triangle size, particularly where triangles approach the correlation length. Rescaling each vertex to unit variance, H → V^(½) H V^(½) with v_i the prior variances (from a selected inverse, or a stochastic estimate once per mesh and scale), would match the dense kernels' C_ii = 1.
Edges. The finite-element form has the same reflecting (zero-flux) boundary as the rectangular version, which raises the prior variance near the mesh edge. Lindgren et al. extend the mesh with a ring of coarse triangles about one correlation length wide; these carry no data and add few vertices. Zeroed edge pixels (
zeroed_pixels) could stay as they are.API
Either
MaternSparsedispatches on the mesh type, or a separateMaternSparseDelaunaywith the same methods (regularization_matrix_sparse_from,regularization_factor_sparse_from,log_det_regularization_matrix_term_from). The matrix-free route would then accept Delaunay mappers.Tests
MaternSparse(up to the triangulation's own discretisation error).Out of scope
As for the rectangular case, only integer α = ν + 1 has a sparse precision, so ν = 1.5 stays dense-only.
I have a rough version of the maths worked through on the PyUVimage side and am happy to help test.
All reactions