Interface suggestion: make TNR schemes iterable to access intermediate tensors

Currently, the run! interface runs the TNR scheme to the end of the RG process, with only a limited information of the intermediate RG steps only accessible at the end from the returned value of Finalizer. This also negatively affects the tests (especially in test/schemes/schemes.jl): sometimes TNR are repeatedly run for different number of steps to examine different properties (free energy or CFT data). For example:

TNRKit.jl/test/schemes/schemes.jl

Lines 20 to 31 in 24f658e

    
           @testset "TRG - Ising Model" begin 
        
               @info "TRG ising free energy" 
        
               scheme = TRG(T) 
        
               data = run!(scheme, truncrank(24), maxiter(25)) 
        
               @test free_energy(data, ising_βc) ≈ f_onsager rtol = 2.0e-6 
        
               @info "TRG ising CFT data" 
        
               scheme = TRG(T) 
        
               run!(scheme, truncrank(24), maxiter(10)) 
        
               cft = sort(CFTData(scheme; shape = [1, 1, 0]).scaling_dimensions[2:end]; by = abs) .|> real

A nicer interface in my opinion is to make the TNR schemes iterable (similar to the TimeEvolver in PEPSKit or IterativeSolver in MPSKit). Each iteration returns the tensors at the current step, and the normalization factor (for free energy calculation, etc). The TRG test above then may be written as

@info "BTRG ising free energy"
scheme = TRG(T)
data = scalartype(T)[]
# may make `scheme` callable; or use other approaches
rg_runner = scheme(truncrank(24), maxiter(25))
for (n, (T, nrm)) in enumerate(rg_runner)
    push!(data, nrm)
    (n != 10) && continue
    # examine CFT data at step 10
    cft = sort(CFTData(scheme; shape = [1, 1, 0]).scaling_dimensions[2:end]; by = abs) .|> real
    # CFT tests
end
@test free_energy(data, ising_βc) ≈ f_onsager rtol = 6.0e-8

Then, no need to run TRG twice.

I can make a PR next week on TRG to illustrate how this works. If you like it, we can gradually roll out the new interface to all TNR schemes.

I'm quite curious to see this in action! We could keep supporting the run! functionality but at the same time offer an "expert mode" which people can use by using the schemes as an iterable.

If do not consider the backward compatibility, I would dream to have a structure called

struct TensorBlock{horizontal, vertical}
  Ts::Vector{AbstractTensorMap}
end

and a struct called

struct RG
  ...
end

And define a function

evaluate(scheme::RG, tensorblock::TensorBlock)

which returns a tensorblock.

Navigation Menu

Repository navigation

Interface suggestion: make TNR schemes iterable to access intermediate tensors #191

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	@testset "TRG - Ising Model" begin
	@info "TRG ising free energy"
	scheme = TRG(T)
	data = run!(scheme, truncrank(24), maxiter(25))

	@test free_energy(data, ising_βc) ≈ f_onsager rtol = 2.0e-6

	@info "TRG ising CFT data"
	scheme = TRG(T)
	run!(scheme, truncrank(24), maxiter(10))

	cft = sort(CFTData(scheme; shape = [1, 1, 0]).scaling_dimensions[2:end]; by = abs) .\|> real

Interface suggestion: make TNR schemes iterable to access intermediate tensors #191

Description

Activity

Yue-Zhengyuan commented on Jun 12, 2026

VictorVanthilt commented on Jun 16, 2026

Chenqitrg commented on Jul 7, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions