Allow rescaling flat networks with bp

Author: JoeyT1994

src/normalize.jl

@@ -1,32 +1,31 @@
 using LinearAlgebra
 
-function rescale(tn::AbstractITensorNetwork, c::Number, vs=collect(vertices(tn)))
+function rescale(tn::AbstractITensorNetwork; alg="exact", kwargs...)
+  return rescale(Algorithm(alg), tn; kwargs...)
+end
+
+function rescale(
+  alg::Algorithm"exact", tn::AbstractITensorNetwork, vs=collect(vertices(tn)); kwargs...
+)
+  logn = logscalar(alg, tn; kwargs...)
+  c = 1.0 / (exp(logn / length(vs)))
   tn = copy(tn)
   for v in vs
     tn[v] *= c
   end
   return tn
 end
 
-function LinearAlgebra.normalize(tn::AbstractITensorNetwork; alg="exact", kwargs...)
-  return normalize(Algorithm(alg), tn; kwargs...)
-end
-
-function LinearAlgebra.normalize(alg::Algorithm"exact", tn::AbstractITensorNetwork)
-  norm_tn = QuadraticFormNetwork(tn)
-  c = exp(logscalar(alg, norm_tn) / (2 * length(vertices(tn))))
-  return rescale(tn, 1 / c)
-end
-
-function LinearAlgebra.normalize(
+function rescale(
   alg::Algorithm"bp",
-  tn::AbstractITensorNetwork;
+  tn::AbstractITensorNetwork,
+  vs=collect(vertices(tn));
   (cache!)=nothing,
   update_cache=isnothing(cache!),
   cache_update_kwargs=default_cache_update_kwargs(cache!),
 )
   if isnothing(cache!)
-    cache! = Ref(BeliefPropagationCache(QuadraticFormNetwork(tn)))
+    cache! = Ref(BeliefPropagationCache(tn, group(v -> v, vertices(tn))))
   end
 
   if update_cache
@@ -35,19 +34,57 @@ function LinearAlgebra.normalize(
 
   tn = copy(tn)
   cache![] = normalize_messages(cache![])
-  norm_tn = tensornetwork(cache![])
-
   vertices_states = Dictionary()
-  for v in vertices(tn)
-    v_ket, v_bra = ket_vertex(norm_tn, v), bra_vertex(norm_tn, v)
-    pv = only(partitionvertices(cache![], [v_ket]))
+  for pv in partitionvertices(cache![])
+    pv_vs = filter(v -> v ∈ vs, vertices(cache![], pv))
+
+    isempty(pv_vs) && continue
+
     vn = region_scalar(cache![], pv)
-    norm_tn = rescale(norm_tn, 1 / sqrt(vn), [v_ket, v_bra])
-    set!(vertices_states, v_ket, norm_tn[v_ket])
-    set!(vertices_states, v_bra, norm_tn[v_bra])
+    if isreal(vn) && vn < 0
+      tn[first(pv_vs)] *= -1
+      vn = abs(vn)
+    end
+
+    vn = vn^(1 / length(pv_vs))
+    for v in pv_vs
+      tn[v] /= vn
+      set!(vertices_states, v, tn[v])
+    end
   end
 
   cache![] = update_factors(cache![], vertices_states)
+  return tn
+end
+
+function LinearAlgebra.normalize(tn::AbstractITensorNetwork; alg="exact", kwargs...)
+  return normalize(Algorithm(alg), tn; kwargs...)
+end
+
+function LinearAlgebra.normalize(
+  alg::Algorithm"exact", tn::AbstractITensorNetwork; kwargs...
+)
+  norm_tn = QuadraticFormNetwork(tn)
+  vs = filter(v -> v ∉ operator_vertices(norm_tn), collect(vertices(norm_tn)))
+  return ket_network(rescale(alg, norm_tn, vs; kwargs...))
+end
+
+function LinearAlgebra.normalize(
+  alg::Algorithm,
+  tn::AbstractITensorNetwork;
+  (cache!)=nothing,
+  cache_construction_function=tn ->
+    cache(alg, tn; default_cache_construction_kwargs(alg, tn)...),
+  update_cache=isnothing(cache!),
+  cache_update_kwargs=default_cache_update_kwargs(cache!),
+)
+  norm_tn = QuadraticFormNetwork(tn)
+  if isnothing(cache!)
+    cache! = Ref(cache_construction_function(norm_tn))
+  end
+
+  vs = filter(v -> v ∉ operator_vertices(norm_tn), collect(vertices(norm_tn)))
+  norm_tn = rescale(alg, norm_tn, vs; cache!, update_cache, cache_update_kwargs)
 
   return ket_network(norm_tn)
 end

test/test_normalize.jl

@@ -5,32 +5,32 @@ using ITensorNetworks:
   edge_scalars,
   norm_sqr_network,
   random_tensornetwork,
-  vertex_scalars
+  vertex_scalars,
+  rescale
 using ITensors: dag, inner, siteinds, scalar
 using Graphs: SimpleGraph, uniform_tree
 using LinearAlgebra: normalize
 using NamedGraphs: NamedGraph
-using NamedGraphs.NamedGraphGenerators: named_grid
+using NamedGraphs.NamedGraphGenerators: named_grid, named_comb_tree
 using StableRNGs: StableRNG
 using Test: @test, @testset
 @testset "Normalize" begin
 
-  #First lets do a tree 
-  L = 6
+  #First lets do a flat tree 
+  nx, ny = 2, 3
   χ = 2
   rng = StableRNG(1234)
 
-  g = NamedGraph(SimpleGraph(uniform_tree(L)))
-  s = siteinds("S=1/2", g)
-  x = random_tensornetwork(rng, s; link_space=χ)
+  g = named_comb_tree((nx, ny))
+  tn = random_tensornetwork(rng, g; link_space=χ)
 
-  ψ = normalize(x; alg="exact")
-  @test scalar(norm_sqr_network(ψ); alg="exact") ≈ 1.0
+  tn_r = rescale(tn; alg="exact")
+  @test scalar(tn_r; alg="exact") ≈ 1.0
 
-  ψ = normalize(x; alg="bp")
-  @test scalar(norm_sqr_network(ψ); alg="exact") ≈ 1.0
+  tn_r = rescale(tn; alg="bp")
+  @test scalar(tn_r; alg="exact") ≈ 1.0
 
-  #Now a loopy graph
+  #Now a state on a loopy graph
   Lx, Ly = 3, 2
   χ = 2
   rng = StableRNG(1234)