From 0d8b8cbf14040834755885f94b15d7b929c8f170 Mon Sep 17 00:00:00 2001
From: Manan Gupta <mnn@yogins.com>
Date: Sun, 9 Aug 2026 12:06:12 -0700
Subject: [PATCH] QSVD: skip the unused state-vector simulation in the loss
 inner loop

---
 pyqpanda-algorithm/pyqpanda_alg/QSVD/QSVD.py |  7 ++-
 test/QAlgBase/Test_QSVD_SVD.py               | 59 ++++++++++++++++++++
 2 files changed, 63 insertions(+), 3 deletions(-)

diff --git a/pyqpanda-algorithm/pyqpanda_alg/QSVD/QSVD.py b/pyqpanda-algorithm/pyqpanda_alg/QSVD/QSVD.py
index d43f868e..94961a6f 100644
--- a/pyqpanda-algorithm/pyqpanda_alg/QSVD/QSVD.py
+++ b/pyqpanda-algorithm/pyqpanda_alg/QSVD/QSVD.py
@@ -134,14 +134,15 @@ def loss(self, ls, return_type=True):
         machine.run(prog, 1000)
         re = machine.result().get_prob_list(qvec0 + qvec1)
         re = np.array(parse_quantum_result_list(re, qvec0+qvec1, select_max=-1))
-        stv = StateVector(self.q0 + self.q1)
-        phase = stv.evolve(cir).ndarray().real
-        phase = phase.reshape(2**self.q1, 2**self.q0)
         prob = np.diagonal(re.reshape(2**self.q1, 2**self.q0))
         same_p = np.sum(prob)
         if return_type:
             return 1-same_p
         else:
+            # state vector is only needed for the singular vectors, not for the loss
+            stv = StateVector(self.q0 + self.q1)
+            phase = stv.evolve(cir).ndarray().real
+            phase = phase.reshape(2**self.q1, 2**self.q0)
             return phase, np.argmax(abs(phase))
 
     def QSVD_min(self):
diff --git a/test/QAlgBase/Test_QSVD_SVD.py b/test/QAlgBase/Test_QSVD_SVD.py
index 2d6ad838..91a3d698 100644
--- a/test/QAlgBase/Test_QSVD_SVD.py
+++ b/test/QAlgBase/Test_QSVD_SVD.py
@@ -1,6 +1,9 @@
+import time
+
 import pytest
 import numpy as np
 from pyqpanda_alg.QSVD import SVD
+from pyqpanda_alg.QSVD import QSVD as qsvd_module
 import warnings
 
 
@@ -30,6 +33,62 @@ def test_orthogonality_properties(self):
         assert np.all(np.diff(q_sorted) <= 0), "QSVD奇异值应该降序排列"
         assert np.all(np.diff(np_sorted) <= 0), "NumPy奇异值应该降序排列"
 
+    def test_singular_values_against_numpy(self):
+        matrix = np.random.random(16).reshape([4, 4])
+
+        qsvd_instance = SVD(matrix_in=matrix)
+        para = qsvd_instance.QSVD_min()
+        q_singular = np.sort(np.diag(qsvd_instance.return_diag(para)))[::-1]
+
+        np_singular = np.linalg.svd(matrix, compute_uv=False)
+
+        rel_err = np.max(np.abs(q_singular - np_singular) / np_singular)
+        assert rel_err < 1e-2, f"singular value relative error too large: {rel_err}"
+
+    def test_singular_vectors_against_numpy(self):
+        matrix = np.random.random(32).reshape([4, 8])
+
+        qsvd_instance = SVD(matrix_in=matrix)
+        para = qsvd_instance.QSVD_min()
+        # loss() reports the index the dominant singular value ended up on
+        max_index = qsvd_instance.loss(para, return_type=False)[1]
+        left = qsvd_instance.max_eig('0', para, max_index)
+        right = qsvd_instance.max_eig('1', para, max_index)
+
+        u_np, s_np, v_np = np.linalg.svd(matrix)
+
+        cos_left = abs(np.dot(left, u_np[:, 0])) / (np.linalg.norm(left) * np.linalg.norm(u_np[:, 0]))
+        cos_right = abs(np.dot(right, v_np[0])) / (np.linalg.norm(right) * np.linalg.norm(v_np[0]))
+        assert cos_left > 0.99, f"left singular vector overlap too low: {cos_left}"
+        assert cos_right > 0.99, f"right singular vector overlap too low: {cos_right}"
+
+    def test_loss_skips_state_vector_on_scalar_path(self, monkeypatch):
+        # the scalar loss is the optimizer inner loop, it must not simulate the state vector
+        built = []
+        real_state_vector = qsvd_module.StateVector
+
+        def counting_state_vector(*args, **kwargs):
+            built.append(1)
+            return real_state_vector(*args, **kwargs)
+
+        monkeypatch.setattr(qsvd_module, 'StateVector', counting_state_vector)
+
+        qsvd_instance = SVD(matrix_in=np.random.random(16).reshape([4, 4]))
+        qsvd_instance.loss(qsvd_instance.parameter)
+        assert len(built) == 0, "scalar loss built a state vector"
+
+        qsvd_instance.loss(qsvd_instance.parameter, return_type=False)
+        assert len(built) == 1, "matrix loss did not build a state vector"
+
+    def test_qsvd_min_runtime(self):
+        matrix = np.random.random(64).reshape([8, 8])
+
+        start = time.perf_counter()
+        SVD(matrix_in=matrix).QSVD_min()
+        elapsed = time.perf_counter() - start
+
+        assert elapsed < 20, f"QSVD_min on an 8x8 matrix took {elapsed:.1f}s"
+
 
 if __name__ == "__main__":
     # 运行测试