A 6-parameter model beats a 481-parameter neural network on Merton HJB training loss, yet cannot produce the portfolio allocation. Wang, Lisser, Sreekumar and Toffano's smaller model is classically simulable by construction. Reproducing it requires no quantum computer. The authors acknowledge this in the fixed-volatility test, where the entangled circuit beats its simulable twin by a gap they describe as visible but relatively moderate. Their comparison tests how closely the models approximate a closed-form value function. They report neither a runtime saving nor a realized investment outcome.

A physics-informed neural network (PINN) learns without solution data. Training minimizes the squared partial differential equation (PDE) residual at sampled points, together with the squared mismatch at the boundaries. A QPINN (quantum physics-informed neural network) substitutes a parameterized quantum circuit for the network. The measured expectation of that circuit supplies the candidate solution.

The target PDE is the Hamilton-Jacobi-Bellman (HJB) equation for Merton's problem. Its solution v(t,x) gives the investor's maximum expected utility; derivatives of that solution determine the optimal risky fraction. The exercise involves no trading signal or market data. The authors rank solvers by final loss over 30 and 10 runs, measuring only their agreement with a known closed form.

The answer is known before training

The benchmark uses Merton's problem with power utility and constant market parameters: r = 0.02, µ = 0.0219, σ = 0.2, γ = 0.95, horizon T = 1. These assumptions make the value function a time term multiplied by x^0.95/0.95. They also fix the optimal risky fraction at 0.95. The investor keeps 95% in the risky asset and 5% in cash throughout the one-year horizon. No market prices enter either experiment. All the paper's figures come from synthetic targets with known answers, either a closed-form PDE solution or an explicit test polynomial.

The circuits use quantum signal processing to construct polynomials. A single qubit alternates the encoding rotation R_x(−2 arccos x) with trainable R_z phases, producing a polynomial in x whose degree is at most L. A Hadamard test mixes two such circuits through a control qubit. This removes a parity restriction, while limiting each factor to 1/2. An ancilla register sums R products of these univariate pieces. The authors name this a tensor-decomposed polynomial: a sum of R products of one-variable polynomials.

The construction requires 2D + ⌈log R⌉ + 1 qubits and O(RDL) parameters. For the test polynomial with D = 2, R = 4 and L = 3, the bounds give 7 qubits, with leading-order depth and parameter quantities of 48 and 24. Applying the bound for the general multivariate construction of Yu et al. at total degree 5 gives 10, 6400 and 1120. After training, the tensor model reaches a relative L2 error of 1.06×10⁻⁵.

The authors build two solvers from this construction. The QPINN adds a controlled entangling layer, which becomes the identity when its parameter is zero. The Quantum-inspired PINN holds that layer at the identity. Its output is then a sum of products of single-qubit expectations, or a single product when R = 1. A classical computer can evaluate this directly. Across 30 fixed-volatility runs, the final-loss ranking is QPINN (7 parameters), Quantum-inspired (6), fully connected PINN (481), then Counterpart PINN (6). The last of these is a baseline that trains polynomial coefficients directly.

Training settings are fully specified

The code is public, and the paper fixes every optimizer setting. All four models use LAMB with weight decay 0 and betas (0.0, 0.0), removing both moment averages. Training starts with a learning rate of 5×10⁻³, multiplied by 0.995 after each optimizer step. Runs last 2000 epochs, preceded by an 8-candidate, 70-epoch multistart. The fixed-volatility runs use seeds 51 + 100i and loss weights W_d = W_1 = 1, W_2 = 5. Each model receives an output scale of 5; the second experiment also adds an output shift of 1.

With decay of 0.995 per step, the rate at step 2000 is 0.995^2000 ≈ 4.4×10⁻⁵ of its starting value, assuming each epoch is one full-batch step. Training has effectively frozen well before the final epoch. By about step 920, the rate is already below 1% of its start.

Where does the setup leave ambiguity?

A straight line nearly satisfies the Merton HJB

The paper multiplies the HJB through by ∂²v/∂x². Only the third term lacks that derivative, and its coefficient is ½(0.0019/0.2)² ≈ 4.5×10⁻⁵. A function linear in x makes the other two terms vanish. The remaining residual is roughly 4.5×10⁻⁵ times the squared slope.

In the fixed-volatility experiment, the Quantum-inspired PINN has L = 1. Its hypothesis takes the form p1(x)p2(t), with degree-1 factors. The second x-derivative is identically zero, making its PDE loss near zero by construction. The loss ranking therefore mostly measures boundary fit to x^0.95, itself already close to a line.

The allocation is undefined for this model.

The paper's control formula is α̂ = −(µ−r)/σ² · v_x/(x v_xx). With v_xx ≡ 0, the ratio is undefined, so the model cannot deliver the allocation a trader would use. The paper says α̂ "can be directly derived" from the learned functions. We did not find a reported α̂ error for any model.

The fixed-volatility Counterpart PINN uses degree-2 factors. That gives it a richer class with actual curvature, yet it comes last in both experiments, behind the 481- and 491-parameter networks. This ranking suggests that parameterization and optimizer behaviour matter more here than expressivity.

How much does entanglement contribute?

The abstract bases "quantum-induced improvement" on both quantum models beating the Counterpart PINN, a classical model with a similar tensor-decomposed bias. The fixed-volatility comparison already gives reason to question that attribution. The Counterpart's degree-2 product class contains the Quantum-inspired PINN's degree-1 class. The loss gap thus reflects parameterization or optimizer choices. And the Quantum-inspired model achieves its win without any quantum resource.

An entangling layer accounts for the QPINN's margin over the Quantum-inspired model. It contributes 1 parameter in the first experiment and 4 in the second (34 against 30). Discussing the first experiment, the authors write that "the added entangling layer improves performance in this case, while the main gain appears to come from using a hypothesis function matched to the separable structure of the solution." Their own account assigns little of the gain there to entanglement.

The conclusion makes a stronger claim: the entangling layer "yields a QPINN that is no longer efficiently simulable on classical hardware." Section 3.4.1 is narrower. It says the model can no longer be simulated efficiently by the factorized procedure used for the Quantum-inspired PINN. The second QPINN has 8-qubit circuitry, with a state vector of 2^8 = 256 amplitudes. The authors ran noiseless simulations in PyTorch on an M2 Pro CPU, differentiating exact expectation values automatically. Neither shot noise nor parameter-shift cost enters those runs. Losing the factorized evaluation route falls short of classical intractability, as the circuit's size illustrates.

The authors defend entanglement by pointing to the wider QPINN lead when volatility becomes an input. In that experiment, t and σ couple through k(σ)(T−t). They interpret the larger gap as evidence that an entangling design tailored to the problem can add useful expressivity. They also say that this term in the analytical solution is why the layer couples only the t and σ registers. Alongside that choice, they fix the per-variable degrees at L_t = 1, L_x = 2 and L_σ = 3. The answer supplied the architecture.

We did not find a classical model supplied with the same four-parameter t and σ cross term. That comparison would isolate the entanglement. The evidence is also thinner in this experiment: 10 runs rather than 30. Meanwhile, the network's parameter multiple over the Quantum-inspired PINN drops from more than 80 times to about 15 times.

Limits of our replication

We have no quantum execution. Anything we build is therefore the classically simulable Quantum-inspired PINN evaluated against the closed form. Such a test would assess the quantum-inspired implementation; it would leave the QPINN's resource and advantage claims untested. We also could not backtest an allocation. The method requires a traded risk-free leg, and we hold no bond price data. Turning the constant-parameter HJB into a rolling calibration would also create a different strategy. We encountered the same obstacle with another Merton benchmark tested only against its own closed form.

The clearest result in the paper remains the 24-versus-1120 leading-order parameter count from the two resource bounds. To change our view of the PDE results, we would want an α̂ error against the known 0.95. That calculation needs a model with nonzero curvature in x and a settled sign for k.