Add ITensorNetworkOperator and QuadraticFormNetwork data structures - #195
Draft
jack-dunham wants to merge 18 commits into
Draft
jack-dunham wants to merge 18 commits into
jack-dunham wants to merge 18 commits into
Conversation
Member
|
Your PR no longer requires formatting changes. Thank you for your contribution! |
Codecov Report❌ Patch coverage is Additional details and impacted files@@ Coverage Diff @@
## main #195 +/- ##
==========================================
- Coverage 87.19% 82.86% -4.33%
==========================================
Files 16 16
Lines 687 782 +95
==========================================
+ Hits 599 648 +49
- Misses 88 134 +46
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
jack-dunham
commented
Sep 16, 2026
| struct ITensorNetworkOperator{T, V, I, P <: AbstractITensorNetwork{T, V}} <: | ||
| AbstractITensorNetwork{T, V} | ||
| parent::P | ||
| outputnames::Vector{I} |
Contributor
Author
There was a problem hiding this comment.
Consider fast lookup of this.
jack-dunham
commented
Sep 16, 2026
jack-dunham
force-pushed
the
jd/quadraticform
branch
2 times, most recently
from
September 28, 2026 19:38
39c3212 to
3457ed5
Compare
`QuadraticFormNetwork` represents `⟨tn|op|tn⟩` lazily, with an operator layer between the bra and ket layers. A ket index name is renamed in the bra layer when it is a link name or appears in the operator's `inputnames`, so a ket index the operator does not act on still contracts straight onto the bra. `operatortensor` renames each operator output name to the bra name of the input name it is paired with, leaving the input names to meet the ket. The constructor throws unless `vertices(operator)` and `vertices(ket)` are set-equal. `AbstractBilinearFormNetwork` is the new supertype of `NormNetwork` and `QuadraticFormNetwork`. A subtype implements `braname`, `kettensor` and, where it has an operator layer, `operatortensor`; `bratensor`, `conj_bratensor` and `indmap` are defined once in terms of those, so the two concrete files now share them instead of each carrying a copy. `AbstractBilinearFormNetworkView` does the same for `KetView`, `BraView` and the new `OperatorView`: each implements `Base.parent` and `get_vertex_data`, and takes its graph structure, assignment and mutability from the parent network. `test/test_quadraticformnetwork.jl` covers the name map, the three views, and three contractions: an identity operator layer reproduces `contract_network(NormNetwork(tn))`, a random on-site product operator matches a dense `⟨ψ|O|ψ⟩` reference, and a `2I` layer on three vertices scales the norm by 8. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The five files are byte-identical at their new paths; only the `include` paths in `src/ITensorNetworksNext.jl` change. `src/normnetworkview.jl` becomes `bilinearformnetworkview.jl`, since it now holds the views of both `NormNetwork` and `QuadraticFormNetwork`. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`OperatorView` now accepts any `AbstractBilinearFormNetwork` rather than only a `QuadraticFormNetwork`, and a second constructor throws an `ArgumentError` for a `NormNetwork`, which reports the reason rather than surfacing a `MethodError`. A `NormNetwork` maps every site index name to itself in its bra layer, so an identity operator between the layers would need its output name to equal its input name, and `ITensor` rejects that: "Dimension names must be distinct". Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`encoded_vertex`, `decoded_vertex` and `encoded_graph` move from a `for View in (:KetView, :BraView)` loop in `bilinearformnetworkview.jl` to `AbstractBilinearFormNetworkView` in `abstractbilinearformnetworkview.jl`, replacing the `vertex_positions`/`ordered_vertices`/`position_graph` methods that NamedGraphs 0.13 removed. `OperatorView` is now covered as well, as it was by the methods being replaced. The loop also `@eval`ed against `BraView` above that type's own definition, so including the file threw `UndefVarError: BraView`. `QuadraticFormNetwork` gets the same rename; it still defined the three removed names, so the package did not load. `test_quadraticformnetwork.jl` takes `incident_edges`, `named_grid` and `named_path_graph` from `NamedGraphs` itself, since the `GraphsExtensions` and `NamedGraphGenerators` submodules are gone. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`ITensorNetworkOperator` is the network equivalent of `ITensorBase.ITensorOperator`:
a tensor network together with a positional pairing of its dangling index names,
`outputnames[i]` with `inputnames[i]`. The constructor, reached through
`ITensorBase.operator(tn, out, in)`, throws an `ArgumentError` when the two name
lists differ in length and when a named index is not dangling, since
`dimnamevertices` must return exactly one vertex for an operator leg.
A pair may straddle two vertices, as it does for a swap or a translation, so
`getindex` wraps each vertex tensor as an `ITensorOperator` carrying only the pairs
whose two halves both sit on that vertex; the remaining legs are dangling on the
wrapper, which `NamedTensorOperator` permits for names in neither list. The wrapper
is built even when no pair is local, so `eltype` is `NamedTensorOperator{I, T}` for
every vertex.
`QuadraticFormNetwork` takes one as its operator layer. Its `acted` set is now
`Set{I}(inputnames(operator))` rather than a loop over per-vertex input names, and
`operatortensor` reads the network-level pairing: reading `qf.operator[vertex]`
would skip a crossing pair's output leg, which is dangling on that wrapper, and the
bra layer would not meet it. `operatortensor` therefore returns a plain tensor and
`get_vertex_data` drops its `state` call.
`supportof` gains a method for the new type. Its body moves to
`supportof_dimnames(tn, opnames)` in `abstracttensornetwork.jl` so both methods
share it; a `Union` signature is not possible there because
`ITensorNetworkOperator` is not yet defined at that point in the include order.
`replacedimnames` is not implemented for the new type: `ITensorNetwork` has no
`replacedimnames` method to delegate to.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…etwork` `ketnetwork` and `operatornetwork` return the stored ket network and `ITensorNetworkOperator`. `branetwork` defaults to a `BraView`, since a bilinear-form network does not store its bra layer. The belief propagation code for `NormNetwork` now reads its link names through these accessors. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ds on `BraView` `BraView` was the only remaining subtype, so it now subtypes `AbstractITensorNetwork` directly and carries the graph, encoding, assignment and mutability forwarding itself. The include order places `bilinearformnetworkview.jl` first, since the forwarding methods now name `BraView`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pure move: the contents of `bilinearformnetworkview.jl` and `abstractbilinearformnetworkview.jl` are appended verbatim, and their `using` lines are merged into the existing imports. `src/bilinearforms/` now holds the interface, `NormNetwork` and `QuadraticFormNetwork`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…in `contract_network` Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`contraction_order(tn; alg)` now splits Gramians before dispatching, matching `contract_network(alg::Exact, tn)`; previously `contraction_order(nn)` on a `NormNetwork` threw a `MethodError` calling `size` on a Gramian. Adds tests for that fix and for `split_gramians`' pass-through on inputs with no Gramian. Also extends the `braname` docstring to cover its `AbstractGramian` method and shortens the `operatortensor(g::QuadraticFormGramian)` comment. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Gramian's indices are the layer indices no other layer at that vertex shares, so the network-level `linkinds`, `linknames`, `linkaxes`, `siteinds`, `sitenames`, `siteaxes` and `dimnamevertices` work on bilinear-form networks again. The bra indices are read from `conj_bratensor`, so no tensor data is conjugated. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…operator pairs `ITensorNetworkOperator` now throws when a pair's output and input sit on different vertices, so each vertex operator carries its full pairing. `QuadraticFormGramian` stores that operator in place of the raw tensor and the network-wide name vectors. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…map` and the network-level accessors Graph forwarding, the ket part of `is_vertex_assigned` and `braname` are defined once on `AbstractBilinearFormNetwork` through `ketnetwork` and a new `branamemap` accessor, and both constructors build the name map with `select_branames`. `kettensor(bn, v)`, `bratensor(bn, v)`, `conj_bratensor(bn, v)` and `operatortensor(bn, v)` are removed in favour of calling the accessor on the Gramian `bn[v]`, and the unused `indmap` is removed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
jack-dunham
force-pushed
the
jd/quadraticform
branch
from
September 29, 2026 14:48
3457ed5 to
2903a18
Compare
The output and input name vectors are replaced by one map from each output name to its input name, so `getindex` looks up a vertex's pairs by name instead of scanning every pair. A repeated output name now throws an `ArgumentError`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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
2 participants
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.
Adds
QuadraticFormNetworkfor⟨ψ|O|ψ⟩, and changes the vertex data of the norm and quadratic-form networks from lazy tensor products to Gramian objects.NormNetworkandQuadraticFormNetworkshare a supertype,AbstractBilinearFormNetwork, insrc/bilinearforms/.ketnetwork,operatornetworkandbranetworkreturn the layers. The bra layer is aBraViewwhose tensors are built on access.ITensorNetworkOperatoris a tensor network with a pairing of output and input index names, made withoperator(tn, outputnames, inputnames). Each output has to sit on the same vertex as its input.nn[v]andqf[v]return aNormGramianorQuadraticFormGramian. It holds the ket tensor, the operator at that vertex for a quadratic form, and the ket-to-bra name map.kettensor,bratensorandoperatortensorread the layers, and the bra is only built when asked for. Gramians defineinds,namesandaxes, solinkinds,siteindsetc. work on these networks.contract_networkandcontraction_ordersplit each Gramian into its layer tensors, so the contraction order can absorb messages into the ket before joining it to the bra. This replacesfactor_tensors.Breaking:
eltypeofNormNetworkandQuadraticFormNetworkis now the Gramian type.factor_tensors,indmap,KetView,OperatorView, and the two-argument accessors such askettensor(bn, v). Usekettensor(bn[v])instead.subgraphdoesn't work on these networks yet.