Add reverse Index map to ITensorNetwork; drop preserve_graph and fix_edges!#365
Merged
Conversation
…ges!
Adds an `Index → Set{vertex}` reverse map as a field on `ITensorNetwork`
and uses it to maintain the graph-edge ↔ shared-`Index` invariant on
every vertex write in O(deg(v) + |inds(value)|), eliminating the need
for the `@preserve_graph` bypass and the O(n) `fix_edges!` sweep.
Removed:
- `@preserve_graph` macro, `setindex_preserve_graph!`, `fix_edges!`
- `map_vertex_data_preserve_graph`, `map_vertices_preserve_graph!`
- `map_vertex_data` (replaced by `Base.map` / `Base.map!`)
- `scale_tensors` / `scale_tensors!` (inlined at the three callsites
in `normalize.jl` and `caches/abstractbeliefpropagationcache.jl`)
- `Base.setindex!(::AbstractBeliefPropagationCache, ::ITensor, vertex)
= not_implemented()` stubs (BPC writes now go through the underlying
`ITensorNetwork`'s `set_vertex_data!`)
Each wrapper type (`TreeTensorNetwork`, `BeliefPropagationCache`,
`BilinearFormNetwork`, `LinearFormNetwork`, `QuadraticFormNetwork`)
defines its own `set_vertex_data!` that forwards to its wrapped network;
there is no abstract `tensornetwork(::AbstractITensorNetwork)` interface
method.
Simplified `Graphs.add_edge!(::AbstractITensorNetwork, edge)`: under the
invariant the "fill missing half" branch collapses to a single `qr!`.
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## main #365 +/- ##
==========================================
+ Coverage 76.05% 76.46% +0.41%
==========================================
Files 57 57
Lines 2710 2639 -71
==========================================
- Hits 2061 2018 -43
+ Misses 649 621 -28
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
Storage is now `graph::NamedGraph{V}`, `vertex_data::Dict{V, ITensor}`, and
`ind_to_vertices::Dict{Index{S}, Set{V}}`, with the constructor surface
narrowed to `ITensorNetwork(tensors)` / `{V}` / `{V, S}` plus a graph-only
empty-network form for `similar_graph`-style scaffolding. The 2-arg
`(tensors, graph)` ctor is gone.
Drop `data_graph` / `data_graph_type` from the `AbstractITensorNetwork`
interface; subtypes (TreeTensorNetwork, AbstractBeliefPropagationCache,
AbstractFormNetwork, QuadraticFormNetwork) now forward via
`underlying_graph` + `vertex_data` directly. Add `Base.values(tn)` so
callers don't depend on whether `vertex_data` is `Dict`- or
`Dictionary`-shaped. Inline the old reverse-map / edge-reconciliation
helpers into `set_vertex_data!` and `rem_vertex!`.
Refactor `opsum_to_ttn` and the TTN `directsum` paths that previously
relied on the 2-arg ctor to build an empty TTN-shaped scaffold: both now
buffer into a plain `Dict{V, ITensor}` and wrap as
`TreeTensorNetwork(ITensorNetwork(...))` once the tensor data is filled,
so the tree invariant only has to hold on a fully-populated network.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- TreeTensorNetwork now holds graph, vertex_data, ind_to_vertices, and ortho_region directly, matching the ITensorNetwork field layout instead of wrapping an inner ITensorNetwork. - Extract _set_vertex_data! and _rem_vertex! helpers shared by both ITensorNetwork and TreeTensorNetwork. - Drop similar_graph and the empty-then-fill construction paths from the form networks and belief-propagation cache; networks are now always constructed with their full vertex data. - Rewrite induced_subgraph_from_vertices to feed tensors through the standard ITensorNetwork(tensors) constructor. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- TreeTensorNetwork now uses the auto-generated all-fields constructor
plus a single outer ctor that performs the is_tree check, matching
the ITensorNetwork constructor design.
- Add ITensorNetwork{V}() empty ctor used as the seed for the
tensor-collection constructor, replacing the explicit three-field
call.
- Extract _unregister_inds! so _set_vertex_data! and _rem_vertex!
share their reverse-map cleanup, without altering the in-place
vertex_data update path.
- Build dictionaries via map(Indices(...)) in test/utils.jl.
- Fix the TreeTensorNetwork jldoctest to use a properly connected
two-vertex example.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Drop ITensorNetwork{V}() empty-arg ctor; inline the three-field call
at the single seed site inside ITensorNetwork{V}(tensors). External
callers that want an empty network can pass an empty tensor
collection (Dict{V, ITensor}() etc.).
- Drop the TreeTensorNetwork(::ITensorNetwork; ortho_region=...) and
TreeTensorNetwork{V}(::ITensorNetwork) overloads. All construction
now routes through TreeTensorNetwork(tensors; ortho_region=nothing),
which builds an ITensorNetwork first and then performs the is_tree
check.
- Define Base.keytype on AbstractITensorNetwork so an AbstractITN can
be passed as a tensor collection to ITensorNetwork(tensors).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mirror the `Vector()` / `Dictionary()` convention by providing
parameterless and `{V}`-only constructors that yield an empty network:
- `ITensorNetwork()` / `ITensorNetwork{V}()`
- `TreeTensorNetwork()` / `TreeTensorNetwork{V}()`
The default vertex type is `Any`. The body of `ITensorNetwork{V}(tensors)`
now seeds from `ITensorNetwork{V}()` and writes each tensor through
`setindex!`, removing the explicit field call and the prior
`set_vertex_data!` call at this layer.
Also drop the redundant `ITensorNetwork(::TreeTensorNetwork)` and the
`ITensorNetwork(::AbstractTTN)` "not implemented" placeholder — the
generic `ITensorNetwork(tensors)` now handles a `TreeTensorNetwork`
via `keytype` and `setindex!`. Update the docs to drop the dangling
references.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`Dict{Any, ITensor}` has `keytype = Any`, so `TreeTensorNetwork(tensors)`
already produces a TTN with `V = Any` — the explicit
`ITensorNetwork{Any}(...)` wrap is redundant.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The intermediate `ITensorNetwork` wrap was redundant — `TreeTensorNetwork` accepts a tensor collection directly. Use a literal vertex label (`[1]`) in `ortho_region` instead of looking it up via `vertices(itn)`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…data! The `ortho_region` keyword on `TreeTensorNetwork(tensors)` was not used by any caller and let users assert a gauge without performing the associated transformation. Removing it nudges callers toward `orthogonalize` (which does the QR sweep) and `set_ortho_region` (the low-level metadata update used internally). Also drop our `set_vertices_data!` override on `AbstractITensorNetwork` — `DataGraphs` already provides exactly the same loop as a default. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The two-argument `ITensorNetwork(tensors, graph)` ctor was removed
earlier in this PR but lingered in three docs/src/ pages and one
prose reference. Migrate them to the single-arg form, which infers
edges from shared `Index`es.
Also fix `truncate(tn::AbstractTTN, edge)` to call the generic
`TreeTensorNetwork(...)` ctor instead of the now-removed
`TreeTensorNetwork{V}(::ITensorNetwork)` specialization, and simplify
the empty `TreeTensorNetwork{V}()` to build the all-fields struct
directly instead of indirecting through `ITensorNetwork{V}()`.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
mtfishman
added a commit
that referenced
this pull request
May 13, 2026
…tindex!(bpc, ...) `data_graph_type` was an accessor in the pre-#365 DataGraph-based storage layout. After #365 inlined storage into NamedGraph + Dictionary fields, no implementation of `data_graph_type` remains in src/, but the docs still listed it for `AbstractFormNetwork`, `QuadraticFormNetwork`, `AbstractBeliefPropagationCache`, `AbstractTTN` (as a prose label), and `TreeTensorNetwork`. Drop all of those entries. `edge_data_eltype(::Type{<:AbstractIndsNetwork})` was a misnamed reference — the actual function is `DataGraphs.edge_data_type`. Fix the name. `setindex!(bpc, factor, vertex)` is documented for both `AbstractBeliefPropagationCache` and `BeliefPropagationCache`, but neither type has a corresponding method in src/. Drop both entries.
mtfishman
added a commit
that referenced
this pull request
May 13, 2026
## Summary Doc-only cleanup: remove references in `docs/src/experimental_methods.md` and `docs/src/developer_methods.md` to methods that no longer exist in src/. - **`visualize(is::IndsNetwork, args...; kwargs...)`** — removed from `src/indsnetwork.jl` in [#356](#356) along with the `IndsNetwork`-based `ITensorNetwork` constructor family. The doc entry was missed. - **`data_graph_type`** — was an accessor for the pre-[#365](#365) `DataGraph`-based storage layout. After #365 inlined storage into `NamedGraph + Dictionary` fields, no implementation remains in src/, but the docs still listed it for `AbstractFormNetwork`, `QuadraticFormNetwork`, `AbstractBeliefPropagationCache`, `AbstractTTN` (as a prose label), and `TreeTensorNetwork`. Drop all of those entries. - **`edge_data_eltype(::Type{<:AbstractIndsNetwork})`** — misnamed; the actual function is `DataGraphs.edge_data_type`. Fix the name. - **`setindex!(bpc, factor, vertex)`** — documented for both `AbstractBeliefPropagationCache` and `BeliefPropagationCache`, but neither type has a corresponding method in src/. Drop both entries. No version bump — substantive change accumulates against the existing `0.22.0-DEV` prerelease.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
ITensorNetwork'sDataGraphstorage with explicit fields(
graph,vertex_data, and a newIndex → Set{vertex}reverse map)so plain
tn[v] = ...reconciles edges inO(deg(v) + |inds(value)|)instead of the oldO(n)fix_edges!sweep.
preserve_graphmachinery (macro,setindex_preserve_graph!,fix_edges!, etc.) now that vertex writesare cheap.
TreeTensorNetworkstorage to mirrorITensorNetworkinsteadof wrapping one, and narrows the constructor surface on both types to
()/{V}()/(tensors).🤖 Generated with Claude Code