add Tensor.rand_reduce

docs/changelog.md

@@ -9,6 +9,9 @@ Release notes for `quimb`.
 
 - expose [`qtn.edge_coloring`](quimb.tensor.tensor_arbgeom_tebd.edge_coloring) as top level function and allow layers to be returned grouped.
 - add docstring for [`tn.contract_compressed`](quimb.tensor.tensor_core.TensorNetwork.contract_compressed)
+- add [`Tensor.rand_reduce`](quimb.tensor.tensor_core.Tensor.rand_reduce) for randomly removing a tensor index by contracting a random vector into it. One can also supply the value `"r"` to `isel` selectors to use this.
+
+---
 
 
 (whats-new-1-8-4)=
@@ -19,6 +22,8 @@ Release notes for `quimb`.
 - fix for MPS sampling with fixed seed ({issue}`247` and {pull}`248`)
 - fix for `mps_gate_with_mpo_lazy` ({issue}`246`).
 
+---
+
 
 (whats-new-1-8-3)=
 ## v1.8.3 (2024-07-10)
@@ -31,6 +36,8 @@ Release notes for `quimb`.
 - add [`edges_1d_chain`](quimb.tensor.geometry.edges_1d_chain) for generating 1D chain edges
 - [operatorbuilder](quimb.experimental.operatorbuilder): better coefficient placement for long range MPO building
 
+---
+
 
 (whats-new-1-8-2)=
 ## v1.8.2 (2024-06-12)
@@ -48,6 +55,9 @@ Release notes for `quimb`.
 - [`approx_spectral_function`](quimb.linalg.approx_spectral.approx_spectral_function) add plotting and tracking
 - add dispatching to various tensor primitives to allow overriding
 
+---
+
+
 (whats-new-1-8-1)=
 ## v1.8.1 (2024-05-06)
 
@@ -76,6 +86,9 @@ Release notes for `quimb`.
 - fix for retrieving `opt_einsum.PathInfo` for single scalar contraction ({issue}`231`)
 
 
+---
+
+
 (whats-new-1-8-0)=
 ## v1.8.0 (2024-04-10)
 
@@ -160,6 +173,9 @@ Release notes for `quimb`.
 - fix autoblock bug where connected sectors were not being merged ({issue}`223`)
 
 
+---
+
+
 (whats-new-1-7-3)=
 ## v1.7.3 (2024-02-08)
 
@@ -173,6 +189,9 @@ Release notes for `quimb`.
 - restore fallback (to `scipy.linalg.svd` with driver='gesvd') behavior for truncated SVD with numpy backend.
 
 
+---
+
+
 (whats-new-1-7-2)=
 ## v1.7.2 (2024-01-30)
 
@@ -186,6 +205,9 @@ Release notes for `quimb`.
 - removed import of deprecated `numba.generated_jit` decorator.
 
 
+---
+
+
 (whats-new-1-7-1)=
 ## v1.7.1 (2024-01-30)
 
@@ -219,6 +241,9 @@ Release notes for `quimb`.
 - fix bug in quantum discord computation when the state was diagonal ({issue}`217`)
 
 
+---
+
+
 (whats-new-1-7-0)=
 ## v1.7.0 (2023-12-08)
 
@@ -269,6 +294,9 @@ Release notes for `quimb`.
   performing tensor network simplifications.
 
 
+---
+
+
 (whats-new-1-6-0)=
 ## v1.6.0 (2023-09-10)
 
@@ -293,6 +321,9 @@ Release notes for `quimb`.
 
 - fix gauge size check for some backends
 
+---
+
+
 (whats-new-1-5-1)=
 ## v1.5.1 (2023-07-28)
 
@@ -313,6 +344,9 @@ Release notes for `quimb`.
 - fix for {meth}`gate_with_auto_swap` for `i > j`.
 - fix bug where calling `tn.norm()` would mangle indices.
 
+---
+
+
 (whats-new-1-5-0)=
 ## v1.5.0 (2023-05-03)
 
@@ -377,6 +411,9 @@ Release notes for `quimb`.
 - fix {func}`~quimb.tensor.decomp.qr_stabilized` bug for strictly upper
   triangular R factors.
 
+---
+
+
 (whats-new-1-4-2)=
 ## v1.4.2 (2022-11-28)
 
@@ -385,6 +422,9 @@ Release notes for `quimb`.
 - move from versioneer to to
   [setuptools_scm](https://pypi.org/project/setuptools-scm/) for versioning
 
+---
+
+
 (whats-new-1-4-1)=
 ## v1.4.1 (2022-11-28)
 
@@ -418,6 +458,9 @@ Release notes for `quimb`.
 > - fix force atlas 2 and `weight_attr` bug ({issue}`126`)
 > - allow unpickling of `PTensor` objects ({issue}`128`, {pull}`131`)
 
+---
+
+
 (whats-new-1-4-0)=
 ## v1.4.0 (2022-06-14)
 
@@ -437,6 +480,9 @@ Release notes for `quimb`.
 - Various memory and performance improvements
 - Various graph generators and TN builders
 
+---
+
+
 (whats-new-1-3-0)=
 ## v1.3.0 (2020-02-18)
 
@@ -473,6 +519,9 @@ Release notes for `quimb`.
 - Add environment variable `QUIMB_NUMBA_PAR` to set whether numba should use automatic parallelization - mainly to fix travis segfaults.
 - Make cache import and initilization of `petsc4py` and `slepc4py` more robust.
 
+---
+
+
 (whats-new-1-2-0)=
 ## v1.2.0 (2019-06-06)
 

quimb/tensor/tensor_core.py

@@ -1626,8 +1626,15 @@ def isel(self, selectors, inplace=False):
 
         Parameters
         ----------
-        selectors : dict[str, int], dict[str, slice]
-            Mapping of index(es) to which value to take.
+        selectors : dict[str, int or slice or "r"]
+            Mapping of index(es) to which value to take. The values can be:
+
+            - int: select a specific value for that index.
+            - slice: select a range of values for that index.
+            - "r": contract a random vector in.
+
+            The mapping can contain indices that don't appear on this tensor,
+            in which case they are ignored.
         inplace : bool, optional
             Whether to select inplace or not.
 
@@ -1643,18 +1650,29 @@ def isel(self, selectors, inplace=False):
 
         See Also
         --------
-        TensorNetwork.isel
+        TensorNetwork.isel, Tensor.rand_reduce
         """
         T = self if inplace else self.copy()
 
-        new_inds = tuple(
-            ix
-            for ix in self.inds
-            if (ix not in selectors) or isinstance(selectors[ix], slice)
-        )
+        new_inds = []
+        data_loc = []
 
-        data_loc = tuple(selectors.get(ix, slice(None)) for ix in self.inds)
-        T.modify(apply=lambda x: x[data_loc], inds=new_inds, left_inds=None)
+        for ix in T.inds:
+            sel = selectors.get(ix, slice(None))
+            if isinstance(sel, slice):
+                # index will be kept (including a partial slice of entries)
+                new_inds.append(ix)
+                data_loc.append(sel)
+            elif sel == "r":
+                # eagerly remove any 'random' selections
+                T.rand_reduce_(ix)
+            else:
+                # index will be removed by selecting a specific index
+                data_loc.append(sel)
+
+        T.modify(
+            apply=lambda x: x[tuple(data_loc)], inds=new_inds, left_inds=None
+        )
         return T
 
     isel_ = functools.partialmethod(isel, inplace=True)
@@ -2200,6 +2218,22 @@ def vector_reduce(self, ind, v, inplace=False):
 
     vector_reduce_ = functools.partialmethod(vector_reduce, inplace=True)
 
+    def rand_reduce(self, ind, dtype=None, inplace=False, **kwargs):
+        """Contract the index ``ind`` of this tensor with a random vector,
+        removing it.
+
+        Parameters
+        ----------
+        """
+        if dtype is None:
+            dtype = self.dtype
+
+        v = randn(self.ind_size(ind), dtype=self.dtype, **kwargs)
+
+        return self.vector_reduce(ind, v, inplace=inplace)
+
+    rand_reduce_ = functools.partialmethod(rand_reduce, inplace=True)
+
     def collapse_repeated(self, inplace=False):
         """Take the diagonals of any repeated indices, such that each index
         only appears once.
@@ -7983,8 +8017,13 @@ def isel(self, selectors, inplace=False):
 
         Parameters
         ----------
-        selectors : dict[str, int]
-            Mapping of index(es) to which value to take.
+        selectors : dict[str, int or slice or "r"]
+            Mapping of index(es) to which value to take. The values can be:
+
+            - int: select a specific value for that index.
+            - slice: select a range of values for that index.
+            - "r": contract a random vector in.
+
         inplace : bool, optional
             Whether to select inplace or not.
 

tests/test_tensor/test_tensor_core.py

@@ -1909,6 +1909,15 @@ def test_gen_inds_connected(self):
         patches = tuple(tn.gen_inds_connected(2))
         assert len(patches) == 34
 
+    def test_tn_isel_rand(self):
+        mps = qtn.MPS_rand_state(6, 7)
+        ramp = mps.isel({mps.site_ind(i): "r" for i in mps.sites})
+        assert ramp.outer_inds() == ()
+        # check we haven't selected an computation basis amplitude
+        rx = ramp.contract()
+        xs = mps.to_dense().ravel()
+        assert not any(np.allclose(rx, x) for x in xs)
+
 
 class TestTensorNetworkSimplifications:
     def test_rank_simplify(self):