update problem 4 and fix problem 10

Author: refraction-ray

examples/challenge_suite/evaluate_10.py

@@ -125,7 +125,7 @@ def evaluate(solution_module, config):
 
     print("Challenge 10 evaluation")
     print(f"Solution module: {solution_module}")
-    print(f"Solution time: {solution_elapsed:.2f}s")
+    print(f"End-to-end solution time: {solution_elapsed:.2f}s")
     print(f"Exact reference time: {exact_elapsed:.2f}s")
     print(f"Qubits: {config['n_qubits']}")
     print(f"Selected CMZ qubits: {len(config['selected_qubits'])}")

examples/challenge_suite/problem_10.md

@@ -65,7 +65,13 @@ Required result keys:
 - `energy_history`: NumPy array with shape `(max_steps,)`.
 - `final_parameters`: final rotation tensor with shape `(n_layers, n_qubits, 3)`.
 
-Each history records one value per optimizer update, evaluated immediately before applying that update. The evaluator derives initial/final energy density from the first/last entries of `energy_history`. The largest gate size and requested step count are fixed by the evaluator configuration and are not returned by the solution.
+Each entry of `energy_history` must store the **energy density**
+
+```text
+<H> / n_qubits
+```
+
+evaluated immediately before the corresponding optimizer update, not the raw total energy `<H>`. The evaluator derives initial/final energy density from the first/last entries of `energy_history`. The largest gate size and requested step count are fixed by the evaluator configuration and are not returned by the solution.
 
 ## Evaluation Interface
 
@@ -75,7 +81,7 @@ The evaluator file is `evaluate_10.py`. It dynamically imports a solution module
 python evaluate_10.py --solution solution_10
 ```
 
-The evaluator consumes only the returned result dictionary. It prints solution time, exact-reference time, initial and final energy densities, exact ground-state energy density, VQE gap, energy-history length, largest gate size, returned keys, and pass/fail criteria. It does not save files or create plots by default.
+The evaluator consumes only the returned result dictionary. It prints end-to-end solution time, exact-reference time, initial and final energy densities, exact ground-state energy density, VQE gap, energy-history length, largest gate size, returned keys, and pass/fail criteria. It does not save files or create plots by default.
 
 ## Passing Criteria
 
@@ -99,7 +105,7 @@ python evaluate_10.py --solution solution_10
 
 Observed TensorCircuit-NG/JAX baseline in the current validation environment with the default 200-step configuration:
 
-- Solution time: `66.30s`.
+- End-to-end solution time: `66.30s`.
 - Exact reference time: `30.69s`.
 - Initial energy density: `0.9718770385`.
 - Final energy density: `-1.1766326427`.

examples/challenge_suite/problem_4.md

@@ -34,7 +34,7 @@ Here `p01` is the probability for `0 -> 1` and `p10` is the probability for `1 -
 
 Use four fixed 12-qubit inputs and the same noisy probe circuit for every input. The four inputs are a GHZ state `(|0...0> + |1...1>) / sqrt(2)`, Bell-like pair states `(|01> + |10>) / sqrt(2)` on pairs `(0,1), (2,3), ..., (10,11)`, `|0>^12`, and `|+>^12`.
 
-The shared probe circuit is one even-bond brickwork entangler layer. On every bond `(0,1), (2,3), ..., (10,11)`, apply `RXX(0.31)` with convention `RXX(theta) = exp(-i theta X_i X_j / 2)`, and then apply the asymmetric bit-flip channel independently to the two qubits in that bond.
+The shared probe circuit has two brickwork entangler sublayers. First apply one even-bond layer on bonds `(0,1), (2,3), ..., (10,11)`, where each bond receives `RXX(0.31)` with convention `RXX(theta) = exp(-i theta X_i X_j / 2)` followed by the asymmetric bit-flip channel independently on both qubits in that bond. Then apply one offset odd-bond layer on bonds `(1,2), (3,4), ..., (9,10)` with the same `RXX(0.31)` entangler and the same per-qubit asymmetric bit-flip channel after each odd-bond entangler.
 
 All probes therefore have identical noisy entangler structure and differ only in the initial state.
 
@@ -99,15 +99,15 @@ The TensorCircuit-NG solution in `solution_4.py` can be evaluated with:
 python evaluate_4.py --solution solution_4
 ```
 
-Observed TensorCircuit-NG/JAX baseline in the current validation environment with the entangled-probe default configuration:
+Observed TensorCircuit-NG/JAX baseline in the current validation environment with the two-sublayer entangled-probe default configuration:
 
-- End-to-end solution time: `47.45s`.
-- Initial loss: `6.70448504e-03`.
-- Final loss: `3.09971142e-08`.
-- Fitted `p01`: `0.03398004`.
-- Fitted `p10`: `0.01109863`.
+- End-to-end solution time: `32.91s`.
+- Initial loss: `7.07078446e-03`.
+- Final loss: `2.54623540e-08`.
+- Fitted `p01`: `0.03403885`.
+- Fitted `p10`: `0.01104033`.
 - Trace-preserving error: `1.11022302e-16`.
 
 ## Implementation Hint
 
-For a TensorCircuit-NG/JAX baseline, use `DMCircuit` with `apply_general_kraus` to insert the custom one-qubit channel after each entangling operation. The probes share one noisy entangler structure and differ only in their initial states, keeping the benchmark focused on the trainable channel. Keep the Kraus operators as differentiable tensor algebra, enforce positivity with sigmoid-parameterized probabilities, and verify trace preservation by contracting `sum_a K_a^\dagger K_a`.
+For a TensorCircuit-NG/JAX baseline, use `DMCircuit` with `apply_general_kraus` to insert the custom one-qubit channel after each entangling operation in both the even-bond and odd-bond sublayers. The probes share one noisy entangler structure and differ only in their initial states, keeping the benchmark focused on the trainable channel. Keep the Kraus operators as differentiable tensor algebra, enforce positivity with sigmoid-parameterized probabilities, and verify trace preservation by contracting `sum_a K_a^\dagger K_a`.

examples/challenge_suite/runtime_summary.md

@@ -7,7 +7,7 @@ Evaluator-reported `End-to-end solution time` for the current 12 problems on Mac
 | 1 | 27.22s |
 | 2 | 2.87s |
 | 3 | 2.46s |
-| 4 | 47.45s |
+| 4 | 36.11s |
 | 5 | 45.50s |
 | 6 | 26.83s |
 | 7 | 63.84s |

examples/challenge_suite/solution_4.py

@@ -71,9 +71,10 @@ def apply_channel(circuit, kraus, qubits):
 
 
 def apply_noisy_entangler_layer(circuit, kraus, config):
-    for i in range(0, config["n_qubits"] - 1, 2):
-        circuit.rxx(i, i + 1, theta=config["entangler_angle"])
-        apply_channel(circuit, kraus, (i, i + 1))
+    for start in (0, 1):
+        for i in range(start, config["n_qubits"] - 1, 2):
+            circuit.rxx(i, i + 1, theta=config["entangler_angle"])
+            apply_channel(circuit, kraus, (i, i + 1))
 
 
 def prepare_initial_state(circuit, probe_index, config):