Add lazy EMD solver with O(n) memory requirement

RELEASES.md

@@ -2,15 +2,18 @@
 
 ## 0.9.7.dev0
 
-This new release adds support for sparse cost matrices in the exact EMD solver. Users can now pass sparse cost matrices (e.g., k-NN graphs, sparse graphs) and receive sparse transport plans, significantly reducing memory footprint for large-scale problems. The implementation is backend-agnostic, automatically handling scipy.sparse for NumPy and torch.sparse for PyTorch, and preserves full gradient computation capabilities for automatic differentiation in PyTorch. This enables efficient solving of OT problems on graphs with millions of nodes where only a sparse subset of edges have finite costs.
+This new release adds support for sparse cost matrices and a new lazy EMD solver that computes distances on-the-fly from coordinates, reducing memory usage from O(n×m) to O(n+m). Both implementations are backend-agnostic and preserve gradient computation for automatic differentiation.
 
 #### New features 
+- Add lazy EMD solver with on-the-fly distance computation from coordinates (PR #788)
 - Migrate backend from deprecated `scipy.sparse.coo_matrix` to modern `scipy.sparse.coo_array` (PR #782)
-- Geomloss function now handles both scalar and slice indices for i and j. Using backend agnostic reshaping. Allows to do plan[i,:] and plan[:,j] (PR #785)
+- Geomloss function now handles both scalar and slice indices for i and j (PR #785)
 - Add support for sparse cost matrices in EMD solver (PR #778, Issue #397)
 
 #### Closed issues
-- Fix O(n³) performance bottleneck in sparse bipartite graph arc iteration  (PR #785)
+- Fix NumPy 2.x compatibility in Brenier potential bounds (PR #788)
+- Fix MSVC Windows build by removing __restrict__ keyword (PR #788)
+- Fix O(n³) performance bottleneck in sparse bipartite graph arc iteration (PR #785)
 - Fix deprecated JAX function in `ot.backend.JaxBackend` (PR #771, Issue #770)
 - Add test for build from source (PR #772, Issue #764)
 - Fix device for batch Ot solver in `ot.batch` (PR #784, Issue #783)

ot/__init__.py

@@ -41,6 +41,7 @@
 from .lp import (
     emd,
     emd2,
+    emd2_lazy,
     emd_1d,
     emd2_1d,
     wasserstein_1d,
@@ -82,6 +83,7 @@
 __all__ = [
     "emd",
     "emd2",
+    "emd2_lazy",
     "emd_1d",
     "sinkhorn",
     "sinkhorn2",

ot/gromov/_estimators.py

@@ -122,8 +122,8 @@ def GW_distance_estimation(
 
     for i in range(nb_samples_p):
         if nx.issparse(T):
-            T_indexi = nx.reshape(nx.todense(T[index_i[i], :]), (-1,))
-            T_indexj = nx.reshape(nx.todense(T[index_j[i], :]), (-1,))
+            T_indexi = nx.reshape(nx.todense(T[[index_i[i]], :]), (-1,))
+            T_indexj = nx.reshape(nx.todense(T[[index_j[i]], :]), (-1,))
         else:
             T_indexi = T[index_i[i], :]
             T_indexj = T[index_j[i], :]
@@ -243,16 +243,18 @@ def pointwise_gromov_wasserstein(
     index = np.zeros(2, dtype=int)
 
     # Initialize with default marginal
-    index[0] = generator.choice(len_p, size=1, p=nx.to_numpy(p))
-    index[1] = generator.choice(len_q, size=1, p=nx.to_numpy(q))
+    index[0] = int(generator.choice(len_p, size=1, p=nx.to_numpy(p)).item())
+    index[1] = int(generator.choice(len_q, size=1, p=nx.to_numpy(q)).item())
     T = nx.tocsr(emd_1d(C1[index[0]], C2[index[1]], a=p, b=q, dense=False))
 
     best_gw_dist_estimated = np.inf
     for cpt in range(max_iter):
-        index[0] = generator.choice(len_p, size=1, p=nx.to_numpy(p))
-        T_index0 = nx.reshape(nx.todense(T[index[0], :]), (-1,))
-        index[1] = generator.choice(
-            len_q, size=1, p=nx.to_numpy(T_index0 / nx.sum(T_index0))
+        index[0] = int(generator.choice(len_p, size=1, p=nx.to_numpy(p)).item())
+        T_index0 = nx.reshape(nx.todense(T[[index[0]], :]), (-1,))
+        index[1] = int(
+            generator.choice(
+                len_q, size=1, p=nx.to_numpy(T_index0 / nx.sum(T_index0))
+            ).item()
         )
 
         if alpha == 1:
@@ -404,10 +406,15 @@ def sampled_gromov_wasserstein(
         )
         Lik = 0
         for i, index0_i in enumerate(index0):
+            T_row = (
+                nx.reshape(nx.todense(T[[index0_i], :]), (-1,))
+                if nx.issparse(T)
+                else T[index0_i, :]
+            )
             index1 = generator.choice(
                 len_q,
                 size=nb_samples_grad_q,
-                p=nx.to_numpy(T[index0_i, :] / nx.sum(T[index0_i, :])),
+                p=nx.to_numpy(T_row / nx.sum(T_row)),
                 replace=False,
             )
             # If the matrices C are not symmetric, the gradient has 2 terms, thus the term is chosen randomly.

ot/lp/EMD.h

@@ -51,5 +51,21 @@ int EMD_wrap_sparse(
     uint64_t maxIter             // Maximum iterations for solver
 );
 
+int EMD_wrap_lazy(
+    int n1,                      // Number of source points
+    int n2,                      // Number of target points
+    double *X,                   // Source weights (n1)
+    double *Y,                   // Target weights (n2)
+    double *coords_a,            // Source coordinates (n1 x dim)
+    double *coords_b,            // Target coordinates (n2 x dim)
+    int dim,                     // Dimension of coordinates
+    int metric,                  // Distance metric: 0=sqeuclidean, 1=euclidean, 2=cityblock
+    double *G,                   // Output: transport plan (n1 x n2)
+    double *alpha,               // Output: dual variables for sources (n1)
+    double *beta,                // Output: dual variables for targets (n2)
+    double *cost,                // Output: total transportation cost
+    uint64_t maxIter             // Maximum iterations for solver
+);
+
 
 #endif

ot/lp/EMD_wrapper.cpp

@@ -370,4 +370,108 @@ int EMD_wrap_sparse(
         }
     }
     return ret;
-}
+}
+
+int EMD_wrap_lazy(int n1, int n2, double *X, double *Y, double *coords_a, double *coords_b, 
+                  int dim, int metric, double *G, double *alpha, double *beta, 
+                  double *cost, uint64_t maxIter) {
+    using namespace lemon;
+    typedef FullBipartiteDigraph Digraph;
+    DIGRAPH_TYPEDEFS(Digraph);
+
+    // Filter source nodes with non-zero weights
+    std::vector<int> idx_a;
+    std::vector<double> weights_a_filtered;
+    std::vector<double> coords_a_filtered;
+
+    // Reserve space to avoid reallocations
+    idx_a.reserve(n1);
+    weights_a_filtered.reserve(n1);
+    coords_a_filtered.reserve(n1 * dim);
+
+    for (int i = 0; i < n1; i++) {
+        if (X[i] > 0) {
+            idx_a.push_back(i);
+            weights_a_filtered.push_back(X[i]);
+            for (int d = 0; d < dim; d++) {
+                coords_a_filtered.push_back(coords_a[i * dim + d]);
+            }
+        }
+    }
+    int n = idx_a.size();
+
+    // Filter target nodes with non-zero weights
+    std::vector<int> idx_b;
+    std::vector<double> weights_b_filtered;
+    std::vector<double> coords_b_filtered;
+
+    // Reserve space to avoid reallocations
+    idx_b.reserve(n2);
+    weights_b_filtered.reserve(n2);
+    coords_b_filtered.reserve(n2 * dim);
+
+    for (int j = 0; j < n2; j++) {
+        if (Y[j] > 0) {
+            idx_b.push_back(j);
+            weights_b_filtered.push_back(-Y[j]);  // Demand is negative supply
+            for (int d = 0; d < dim; d++) {
+                coords_b_filtered.push_back(coords_b[j * dim + d]);
+            }
+        }
+    }
+    int m = idx_b.size();
+
+    if (n == 0 || m == 0) {
+        *cost = 0.0;
+        return 0;
+    }
+
+    // Create full bipartite graph
+    Digraph di(n, m);
+
+    NetworkSimplexSimple<Digraph, double, double, node_id_type> net(
+        di, true, (int)(n + m), (uint64_t)(n) * (uint64_t)(m), maxIter
+    );
+
+    // Set supplies
+    net.supplyMap(&weights_a_filtered[0], n, &weights_b_filtered[0], m);
+
+    // Enable lazy cost computation - costs will be computed on-the-fly
+    net.setLazyCost(&coords_a_filtered[0], &coords_b_filtered[0], dim, metric, n, m);
+
+    // Run solver
+    int ret = net.run();
+
+    if (ret == (int)net.OPTIMAL || ret == (int)net.MAX_ITER_REACHED) {
+        *cost = 0;
+
+        // Initialize output arrays
+        for (int i = 0; i < n1 * n2; i++) G[i] = 0.0;
+        for (int i = 0; i < n1; i++) alpha[i] = 0.0;
+        for (int i = 0; i < n2; i++) beta[i] = 0.0;
+
+        // Extract solution
+        Arc a;
+        di.first(a);
+        for (; a != INVALID; di.next(a)) {
+            int i = di.source(a);
+            int j = di.target(a) - n;
+
+            int orig_i = idx_a[i];
+            int orig_j = idx_b[j];
+
+            double flow = net.flow(a);
+            G[orig_i * n2 + orig_j] = flow;
+
+            alpha[orig_i] = -net.potential(i);
+            beta[orig_j] = net.potential(j + n);
+
+            if (flow > 0) {
+                double c = net.computeLazyCost(i, j);
+                *cost += flow * c;
+            }
+        }
+    }
+
+    return ret;
+}

ot/lp/__init__.py

@@ -9,7 +9,7 @@
 # License: MIT License
 
 from .dmmot import dmmot_monge_1dgrid_loss, dmmot_monge_1dgrid_optimize
-from ._network_simplex import emd, emd2
+from ._network_simplex import emd, emd2, emd2_lazy
 from ._barycenter_solvers import (
     barycenter,
     free_support_barycenter,
@@ -35,6 +35,7 @@
 __all__ = [
     "emd",
     "emd2",
+    "emd2_lazy",
     "barycenter",
     "free_support_barycenter",
     "cvx",