betaFeedback

WorksPhysics and Astronomy

Toward Quantum Monte Carlo for Sovereign Debt Risk: A proof of concept

Abstract

We present a proof-of-concept pipeline that applies quantum amplitude estimation to the probability that a sovereign debt-to-GDP ratio exceeds a critical threshold at a given horizon. The pipeline combines a reduced two-factor Gaussian debt model linearized around its mean, quantum state preparation on a 32 x 32 grid, a weighted-adder circuit, an integer comparator, and Maximum Likelihood Amplitude Estimation (MLAE). On a noiseless statevector simulator with 500 shots per Grover power, the circuit estimate (a = 0.1152) agrees with the circuit's exact target amplitude (0.1153), and the observed estimation error is consistent with the amplitude-estimation scaling in the number of oracle calls. The model used here is deliberately simple: it admits a closed-form solution, so the contribution is a demonstration of the encoding and estimation workflow, not a practical advantage over classical methods. We discuss the conditions under which such a pipeline could become relevant and the obstacles that remain.

Introduction

Sovereign debt sustainability assessment is a cornerstone of macroeconomic policy, yet it remains challenged by deep uncertainty in growth trajectories, interest rate dynamics, and geopolitical shocks [1, 2]. Monte Carlo simulation is widely used for this purpose, but its estimation error decreases only as O(1/N)\mathcal{O}(1/\sqrt{N}) in the number of samples NN. For rare-event probabilities, this scaling can impose high computational cost [3].

Quantum Amplitude Estimation (QAE), introduced by Brassard et al. [4], reduces the error scaling to O(1/N)\mathcal{O}(1/N) in the number of calls to the state-preparation oracle, a quadratic improvement in query complexity. This has motivated applications to financial risk analysis [5, 6] and option pricing [7]. The theoretical foundation is Grover’s search algorithm [8], with later variants that avoid quantum phase estimation, notably Maximum Likelihood Amplitude Estimation (MLAE) [9] and iterative approaches [10], which reduce circuit depth and improve robustness to finite sampling [11].

An important caveat accompanies these results. A quadratic query speedup does not by itself imply a practical advantage: the cost of implementing the oracle on quantum hardware, the need for fault tolerance at useful depths, and the strength of optimised classical baselines all matter. Resource analyses indicate that quadratic speedups alone may be insufficient for practical advantage without substantial hardware progress [12, 13]. We keep this caveat in view throughout.

In this paper we assemble and test a complete amplitude-estimation pipeline for a sovereign debt tail probability. Our contributions are modest and specific: (i) an explicit workflow from a two-factor linearized debt model to a quantum circuit with a weighted adder and integer comparator; (ii) an MLAE-based estimator evaluated on a statevector simulator with finite-shot sampling; and (iii) an honest accounting of query costs and of the gap between this toy problem and a setting where a quantum approach would be competitive. The model is simple enough to admit a closed-form answer, which we use as a validation reference.

Section 2 describes the methodology, Section 3 reports results, Section 4 discusses limitations and relevance, and Section 5 concludes.

Methodology

The pipeline has four stages: classical model reduction, quantum state preparation, quantum arithmetic, and amplitude estimation.

Classical Model Reduction and Linearisation

Debt dynamics.

Debt-to-GDP evolves over a horizon of T=20T = 20 years from d0=0.9d_0 = 0.9 as

λt=rt−gt1+gt(1)\lambda_t = \frac{r_t - g_t}{1 + g_t} \tag*{(1)}
pbt=a+b(dt−1−d∗)+pbpolicy,(2)pb_t = a + b(d_{t-1} - d^*) + pb_{\mathrm{policy}}, \tag*{(2)}
dt=(1+λt)dt−1−pbt,t=1,…,T,(3)d_t = (1 + \lambda_t)d_{t-1} - pb_t,\qquad t = 1,\ldots,T, \tag*{(3)}

where pbt>0pb_t > 0 denotes a primary surplus (ratio to GDP). In the full simulation model each of gtg_t, rtr_t and pbtpb_t carries its own yearly noise term; for the quantum pipeline we reduce the model to two aggregate shocks: the cumulative growth shock GG and the cumulative interest-rate shock RR, i.e. the sums over the horizon of the yearly shocks, assumed independent and Gaussian,

G∼N(0,σG2),R∼N(0,σR2),σG=σgT,σR=σrT,(4)G \sim\mathcal{N}(0,\sigma_G^2), \quad R \sim\mathcal{N}(0,\sigma_R^2), \quad\sigma_G = \sigma_g\sqrt{T}, \quad\sigma_R = \sigma_r\sqrt{T}, \tag*{(4)}

each spread evenly over the years:

gt=gˉ+GT,rt=rˉ+RT.(5)g_t = \bar{g} + \frac{G}{T}, \quad r_t = \bar{r} + \frac{R}{T}. \tag*{(5)}

This reduction drops the yearly noise on the primary balance and the persistence of growth. The event of interest is dT>τd_T > \tau with τ=0.95\tau= 0.95, and the target is a=Pr⁡(dT>τ)a = \Pr(d_T > \tau). Table 1 lists all symbols and values.

SymbolMeaningValue
TTHorizon20 years
d0d_0Initial debt/GDP0.90
d∗d^*Target debt ratio in the fiscal rule0.6
aaIntercept of the primary-balance (pb) rule0.0
bbFeedback of the pbpb on the debt gap0.03
pbpolicypb_{\mathrm{policy}}Constant policy effect on pbpb (%GDP/year)0.0
gˉ,rˉ\bar{g}, \bar{r}Trend growth, mean interest rate0.015, 0.025
σg,σr\sigma_g, \sigma_rStd. of the yearly growth / rate shocks0.01, 0.005
E[G],E[R]\mathbb{E}[G], \mathbb{E}[R]Means of the cumulative shocks0, 0
σG,σR\sigma_G, \sigma_RStd. of GG, RR (=σT=\sigma\sqrt{T})0.04472, 0.02236
λ\lambda (nominal)(rˉ−gˉ)/(1+gˉ)(\bar{r}-\bar{g})/(1+\bar{g})0.009852
pb1pb_1 (nominal)b(d0−d∗)b(d_0-d^*)0.0090
dTd_T (nominal, G=R=0G=R=0)Terminal debt ratio0.8978
τ\tauThreshold of the event dT>τd_T > \tau0.95
nnQubits per random factor (2n2^n grid points)5
mmQubits of the sum register6
{k}\{k\}Grover powers{0,1,2,4}\{0,1,2,4\}
nshotsn_{\mathrm{shots}}Shots per Grover power500

Table 1. Model and simulation parameters.

Linearisation and sign convention.

Growth and interest rate enter with opposite signs: ∂dT/∂G=−0.7422\partial d_T/\partial G = -0.7422 and ∂dT/∂R=+0.7349\partial d_T/\partial R = +0.7349 (central finite differences at G=R=0G=R=0 with a step of σ/2\sigma/2). To obtain two positive weights for the quantum adder we use the adverse variables u1=−Gu_1=-G (low growth is bad) and u2=+Ru_2=+R. This is equivalent to reversing the index of the growth register, and it is exact because the Gaussian is symmetric about zero, so the probabilities pip_i are unchanged. Each uiu_i is truncated at ±3.0σ\pm3.0\sigma and discretised on 2n=322^n = 32 regular points ui=ℓi+sixiu_i = \ell_i + s_i x_i, xi∈{0,…,2n−1}x_i \in\{0,\ldots,2^n-1\}, with step s1=0.00866s_1 = 0.00866 and s2=0.00433s_2 = 0.00433 (Table 2). The linearised map is

QuantityValue
Truncation range of GG[−0.13416,+0.13416][-0.13416,+0.13416] (= ±3σG\pm3\sigma_G)
Truncation range of RR[−0.06708,+0.06708][-0.06708,+0.06708] (= ±3σR\pm3\sigma_R)
Probability mass cut per variable0.0027
Grid step of GG / RR0.00866 / 0.00433
∂dT/∂G\partial d_T/\partial G, ∂dT/∂R\partial d_T/\partial R (per unit)−0.7422/+0.7349-0.7422 / +0.7349
Same, per standard deviation−0.03319/+0.01643-0.03319 / +0.01643
c1c_1, c2c_2 (increase of dTd_T per grid step of u1=−Gu_1=-G, u2=+Ru_2=+R)+0.006424+0.006424, +0.003180+0.003180
dconstd_{\mathrm{const}} (intercept at x1=x2=0x_1=x_2=0)0.748922
Exact dTd_T at best / worst corner0.7624 / 1.0620
Linear model at best / worst corner0.7489 / 1.0467
max⁡∣dTlin−dTexact∣\max\lvert d_T^{\mathrm{lin}}-d_T^{\mathrm{exact}}\rvert on the grid0.0154
max⁡∣dTcircuit−dTexact∣\max\lvert d_T^{\mathrm{circuit}}-d_T^{\mathrm{exact}}\rvert on the grid0.0163

Table 2. Truncation, discretisation and linearisation.

dT≈dconst+c1x1+c2x2,(6)d_T \approx d_{\mathrm{const}} + c_1 x_1 + c_2 x_2, \tag*{(6)}

where ci=(∂dT/∂ui)si>0c_i = (\partial d_T/\partial u_i)s_i > 0 is the increase of dTd_T per grid step, and dconst=dT(0,0)−(c1+c2)(2n−1)/2d_{\mathrm{const}} = d_T(0,0) - (c_1+c_2)(2^n-1)/2 is the intercept at x1=x2=0x_1=x_2=0, the most favourable corner of the grid (it is not the value at the mean, dT(0,0)=0.8978d_T(0,0)=0.8978).

The shocks are discretised on a regular grid with 2nbits=322^{n_{\mathrm{bits}}}=32 points per dimension (nbits=5n_{\mathrm{bits}}=5 qubits per variable).

Error measures.

We use five reference probabilities. aconta_{\mathrm{cont}}: continuous Gaussians and the exact nonlinear model, by classical Monte Carlo with 2 000 000 samples (seed 123), standard error 0.0002. alin,conta_{\mathrm{lin,cont}}: continuous Gaussians and the linear model, which is Gaussian and has the closed form alin,cont=1−Φ((τ−μ)/σd)a_{\mathrm{lin,cont}} = 1-\Phi((\tau-\mu)/\sigma_d) with μ=dT(0,0)=0.8978\mu=d_T(0,0)=0.8978 and σd=0.03704\sigma_d=0.03704. atrunca_{\mathrm{trunc}}: Gaussians truncated at ±3σ\pm3\sigma, exact model. agrida_{\mathrm{grid}}: the 32×3232\times32 grid with the exact model. acirca_{\mathrm{circ}}: the grid with the linear model and integer arithmetic of the circuit, which is the amplitude the circuit encodes. The modelling error decomposes as

Emodel=acirc−acont=(atrunc−acont)⏟truncation+(agrid−atrunc)⏟discretisation+(acirc−agrid)⏟linearisation + integer,(7)E_{\mathrm{model}} = a_{\mathrm{circ}} - a_{\mathrm{cont}} = \underbrace{(a_{\mathrm{trunc}}-a_{\mathrm{cont}})}_{\text{truncation}} + \underbrace{(a_{\mathrm{grid}}-a_{\mathrm{trunc}})}_{\text{discretisation}} + \underbrace{(a_{\mathrm{circ}}-a_{\mathrm{grid}})}_{\text{linearisation + integer}}, \tag*{(7)}

and the estimation error is EQAE=a^−acircE_{\mathrm{QAE}}=\hat{a}-a_{\mathrm{circ}}. Table 2 reports both, consistently with the rest of the paper.

The value Elin,int=+0.0000E_{\mathrm{lin,int}}=+0.0000 on the grid is specific to this threshold and grid: it arises because the grid with exact and with linear integer arithmetic flags the same grid points. It must not be read as the linearisation error. The linearisation error in the continuum is acont−alin,cont=+0.0067a_{\mathrm{cont}}-a_{\mathrm{lin,cont}}=+0.0067, larger than the estimation error and in the opposite direction to the net modelling error, so that part of the modelling error cancels.

Quantum State Preparation

Each discretised Gaussian is loaded into a register of n=5n = 5 qubits with the amplitudes pi\sqrt{p_i} of the truncated, renormalised, discretised Gaussian (the density evaluated at the grid points and normalised), producing

∣ψ⟩=∑i=031∑j=031piqj ∣i⟩∣j⟩,(8)|\psi\rangle= \sum_{i=0}^{31}\sum_{j=0}^{31}\sqrt{p_i q_j}\,|i\rangle|j\rangle, \tag*{(8)}

where pi,qjp_i, q_j are the discretised probabilities of x1,x2x_1,x_2. The loading uses Qiskit’s generic StatePreparation (exact amplitude-vector synthesis); we did not use NormalDistribution or the Grover–Rudolph construction. For two registers of five qubits, this loading is inexpensive. We note that for general distributions on many qubits, exact loading can require resources exponential in the number of qubits, which is one of the known bottlenecks of this approach.

Quantum Arithmetic: Weighted Addition and Comparison

Let δ\delta be the debt resolution of one integer unit, chosen so that the full range of dTd_T on the grid spans the sum register, δ=(c1+c2)(2n−1)/(2m−1)=0.004726\delta= (c_1 + c_2)(2^n - 1)/(2^m - 1) = 0.004726. The weights of bit jj of each register are w1,j=round⁡(c12j/δ)w_{1,j}=\operatorname{round}(c_1 2^j/\delta) and w2,j=round⁡(c22j/δ)w_{2,j}=\operatorname{round}(c_2 2^j/\delta), so that the circuit computes

S=∑j=0n−1w1,j[x1]j+∑j=0n−1w2,j[x2]j,(9)S = \sum_{j=0}^{n-1} w_{1,j}[x_1]_j + \sum_{j=0}^{n-1} w_{2,j}[x_2]_j, \tag*{(9)}

with [x]j[x]_j the jj-th bit of xx, using a WeightedAdder; then dT≈dconst+δSd_T \approx d_{\mathrm{const}}+\delta S. For n=5n=5 the weights are w1=[1,3,5,11,22]w_1=[1,3,5,11,22] and w2=[1,1,3,5,11]w_2=[1,1,3,5,11]. The maximum is max⁡S=∑jw1,j+∑jw2,j=63\max S=\sum_j w_{1,j}+\sum_j w_{2,j}=63, and the sum register of m=6m=6 qubits holds up to 2m−1=632^m-1=63, so max⁡S≤2m−1\max S\leq2^m-1 and no overflow occurs (the slack is 0; the check is tight because δ\delta is chosen to use the full range).

An IntegerComparator flags the grid points with S≥τintS\geq\tau_{\mathrm{int}}, where τint=⌈(τ−dconst)/δ⌉=43\tau_{\mathrm{int}}=\lceil(\tau-d_{\mathrm{const}})/\delta\rceil=43. Because of this rounding the circuit’s effective threshold is dconst+τintδ=0.9521d_{\mathrm{const}}+\tau_{\mathrm{int}}\delta=0.9521 instead of τ=0.95\tau=0.95; its size depends on nn (Table 5). The comparator sets an objective qubit to ∣1⟩\lvert1\rangle on the flagged points, so that the probability of measuring ∣1⟩\lvert1\rangle equals acirca_{\mathrm{circ}}.

nnGridmmτint\tau_{\mathrm{int}}Effective τ\taumax⁡S/(2m−1)\max S/(2^{m} - 1)acirca_{\mathrm{circ}}acirc−aconta_{\mathrm{circ}} - a_{\mathrm{cont}}
38×88 \times84110.967315/150.0611−0.0250-0.0250
416×1616 \times165210.950631/310.0834−0.0027-0.0027
532×3232 \times326430.952163/630.0818−0.0043-0.0043

Table 5. Dependence on the grid size (m=n+1m = n+1 sum qubits, nshots=500n_{\mathrm{shots}} = 500; a^\hat{a} from one run, seed 1).

Amplitude Estimation via MLAE

We estimate aa using MLAE [9], which replaces the quantum phase estimation of canonical QAE by a set of Grover powers and classical maximum-likelihood inference. Let AA denote the state-preparation circuit (loading, adder, comparator). The Grover operator is Q=−AS0A†SfQ=-AS_0A^\dagger S_f, where S0S_0 reflects about the all-zero state and SfS_f flips the sign of states with the objective qubit in ∣1⟩\lvert1\rangle. For a schedule of powers {k}\{k\} we apply QkA∣0⟩Q^kA\lvert0\rangle and measure the objective qubit nshots=500n_{\mathrm{shots}}=500 times, recording the number yky_k of ∣1⟩\lvert1\rangle outcomes. Then

Pr⁡(objective=1∣k)=sin⁡2((2k+1)θ),sin⁡2θ=a.(10)\Pr(\mathrm{objective}=1\mid k)=\sin^2((2k+1)\theta),\qquad\sin^2\theta=a. \tag*{(10)}

The likelihood is

L(θ)=∏k(nshotsyk)[sin⁡2((2k+1)θ(a))]yk[cos⁡2((2k+1)θ(a))]nshots−yk,(11)\mathcal{L}(\theta)=\prod_k\binom{n_{\mathrm{shots}}}{y_k}\left[\sin^2((2k+1)\theta(a))\right]^{y_k}\left[\cos^2((2k+1)\theta(a))\right]^{n_{\mathrm{shots}}-y_k}, \tag*{(11)}

and a^=sin⁡2θ^\hat{a}=\sin^2\hat{\theta} with θ^\hat{\theta} maximising L\mathcal{L} on a grid of 200,001 points of (0,π/2)(0,\pi/2). A geometric schedule of powers is what gives MLAE its Heisenberg-type scaling [9]; our schedules (0,1,2,4,…)(0,1,2,4,\ldots) are of this type.

Query cost

A circuit with Grover power kk calls AA (or A†A^\dagger) 2k+12k+1 times. The total number of oracle calls for a schedule {km}\{k_m\} with nshotsn_{\mathrm{shots}} shots per power is therefore

Ncalls=nshots∑m(2km+1).(12)N_{\mathrm{calls}}=n_{\mathrm{shots}}\sum_m(2k_m+1). \tag*{(12)}

For the schedule (0,1,2,4)(0,1,2,4) used in the main experiment, ∑m(2km+1)=1+3+5+9=18\sum_m (2k_m+1)=1+3+5+9=18, giving Ncalls=9,000N_{\mathrm{calls}}=9{,}000, while the number of circuit executions is 4×500=2,0004\times500=2{,}000. We report oracle calls, not circuit executions, as the cost measure when comparing with classical sampling.

Implementation Parameters

We use nbits=5n_{\mathrm{bits}}=5 qubits per variable (32×32=102432\times32=1024 grid points), nsum=6n_{\mathrm{sum}}=6 sum qubits, Grover powers (0,1,2,4)(0,1,2,4), and 500 shots per power. All simulations use a noiseless statevector simulator with shot-based sampling to emulate finite-measurement noise. Table 4 presents the reproducibility of this work:

ItemValue
Python3.14.7
Numpy2.5.4
Qiskit version2.5.2
Aernot used
RNGnumpy.random.default_rng(PCG64)
State loadingqiskit.circuit.library.StatePreparation, amplitudes πi\sqrt{\pi_i} of the truncated, renormalised discretised Gaussian.
ArithmeticWeightedAdder + IntegerComparator (qiskit.circuit.library).
Amplitude estimationMLAE: AQkAQ^k, k∈[0,1,2,4]k\in[0,1,2,4] grid search over θ\theta (200,001 points).
Circuit simulationExact Statevector (no sampling noise); shots drawn with numpy binomial.
Reference Monte Carlo seed/size123 / 2,000,000
RMSE repetitions/seeds200/0..199

Table 4. Reproducibility

Results

Figure 1 presents four diagnostics. Figure 1a Illustrates the baseline precision across successive approximation layers for evaluating the tail event probability (P(debt/GDP>95%)P(\mathrm{debt}/\mathrm{GDP}>95\%) at year 20). The comparison highlights minimal discretization error between the continuous exact model (0.086) and the discretized representation (0.082), with the final MLAE ideal emulation yielding a highly accurate estimate of 0.081.

Results for the sovereign debt tail probability

Figure 1. (a) Comparison of probability estimates: continuous exact model, discretised exact model, circuit target (linear + integer arithmetic), and QAE estimate. (b) Distribution of dTd_T encoded in the circuit. Red bars indicate grid points above the 95% threshold; their total mass equals the target probability aa. (c) Probability of measuring the objective qubit in state ∣1⟩|1\rangle versus Grover power kk. Red markers: simulation; blue curve: theoretical sin⁡2((2k+1)θ)\sin^2((2k + 1)\theta). (d) RMSE versus number of oracle calls (log-log). Classical Monte Carlo (grey dashed) scales as O(1/N)O(1/\sqrt{N}); MLAE (orange) is shown for six schedules. Fig. 1: Results for the sovereign debt tail probability. (a) Comparison with reference values. (b) Encoded debt distribution and crisis tail. (c) Grover amplification. (d) Estimation error versus oracle calls.

Figure 1b displays the discrete probability distribution mapped across the sorted debt/GDP horizon grid points as encoded by the quantum circuit layout. The targeted regions exceeding the 95% critical risk threshold are explicitly highlighted in red, visually demonstrating the state preparation phase before amplification. Figure 1c traces the success probability of the objective qubit as a function of the Grover iteration count kk. The ideal emulation data points tightly fit the theoretical sin⁡2((2k+1)θ)\sin^{2}((2k + 1)\theta) oscillation pattern, validating the coherent amplification of the target state from the initial baseline amplitude of a=0.082a = 0.082. And the Figure 1d benchmarks the resource efficiency by plotting the RMSE of the probability estimate against the total computational cost (measured via circuit oracle calls or classical sample sizes). The MLAE ideal emulation demonstrates a visibly steeper scaling profile relative to the standard classical Monte Carlo rate (∼1/N\sim1/\sqrt{N}), confirming the anticipated quantum scaling advantage within the tested domain.

Agreement with Reference Values

The single-run MLAE estimate a^=0.0809\hat{a} = 0.0809 agrees with the circuit target acirc=0.0818a_{\mathrm{circ}} = 0.0818 (Figure 1a); over 200 repetitions the estimator has bias −0.00000-0.00000 and standard deviation 0.001050.00105 (Table 3), so the single-run error EQAE=−0.0009E_{\mathrm{QAE}} = -0.0009 is typical. This confirms that the estimator recovers the amplitude encoded in the circuit. The circuit target in turn differs from the continuous exact model by Emodel=−0.0043E_{\mathrm{model}} = -0.0043, composed of truncation (−0.0015-0.0015), discretisation (−0.0027-0.0027) and linearisation with integer arithmetic on the grid (+0.0000+0.0000); the estimation error is thus 4.1 times smaller than the modelling error. The closed-form probability of the linear Gaussian model is alin,cont=0.0793a_{\mathrm{lin,cont}} = 0.0793, 0.0067 below aconta_{\mathrm{cont}}.

QuantityValue
aconta_{\mathrm{cont}}: continuous Gaussians, exact model0.0861 (±0.0002\pm0.0002)
alin,conta_{\mathrm{lin,cont}}: continuous Gaussians, linear model (closed form)0.0793
atrunca_{\mathrm{trunc}}: truncated at ±3σ\pm3\sigma, exact model0.0845
agrida_{\mathrm{grid}}: 32 ×\times 32 grid, exact model0.0818
acirca_{\mathrm{circ}}: grid + linear + integer (circuit target)0.0818
a^\hat{a}: QAE estimate (single run, seed 1)0.0809
Linearisation in the continuum, acont−alin,conta_{\mathrm{cont}} - a_{\mathrm{lin,cont}}+0.0067
Emodel=acirc−acontE_{\mathrm{model}} = a_{\mathrm{circ}} - a_{\mathrm{cont}}−0.0043-0.0043
truncation, atrunc−aconta_{\mathrm{trunc}} - a_{\mathrm{cont}}−0.0015-0.0015
discretisation, agrid−atrunca_{\mathrm{grid}} - a_{\mathrm{trunc}}−0.0027-0.0027
linearisation + integer on the grid, acirc−agrida_{\mathrm{circ}} - a_{\mathrm{grid}}+0.0000
EQAE=a^−acircE_{\mathrm{QAE}} = \hat{a} - a_{\mathrm{circ}} (single run)−0.0009-0.0009
Etotal=a^−acontE_{\mathrm{total}} = \hat{a} - a_{\mathrm{cont}}−0.0052-0.0052
RMSE of a^\hat{a} vs. acirca_{\mathrm{circ}} (200 repetitions, seeds 0–199)0.00105
Bias / std. of a^\hat{a} over the repetitions−0.00000-0.00000 / 0.00105
Oracle calls per run, nshots∑k(2k+1)n_{\mathrm{shots}}\sum_k(2k + 1)9 000 (= 500 ×\times 18)
Circuit executions per run2 000

Table 3. Probability a=Pr⁡(dT>τ)a = \Pr(d_T > \tau) with τ=0.95\tau= 0.95 and its errors (n=5n = 5, m=6m = 6, nshots=500n_{\mathrm{shots}} = 500).

Table 5 varies the grid size. The modelling error does not decrease monotonically with nn: the effective threshold fluctuates around τ\tau because of the integer rounding (0.9673, 0.9506, 0.9521 for n=3,4,5n = 3, 4, 5), and the linearisation error in the continuum does not vanish with nn. Increasing nn therefore improves the representation of the Gaussians but not the linearisation, and we do not claim convergence to aconta_{\mathrm{cont}}.

Encoded Distribution

Figure 1b shows the probability mass of the linearised debt as computed by the circuit. The total mass of the flagged region (red) equals acirc≈0.1153a_{\mathrm{circ}} \approx0.1153. Note that this is a probability of order 10−110^{-1}, i.e. not a rare event; the threshold is only about 5.2 percentage points above the nominal terminal ratio dT(0,0)=0.8978d_T(0,0) = 0.8978.

Grover Amplification

Figure 1c shows Pr⁡(objective=1)\Pr(\mathrm{objective}=1) against kk. The simulated frequencies follow the theoretical sin⁡2((2k+1)θ)\sin^2((2k+1)\theta) curve with sin⁡2θ=acircuit\sin^2\theta=a_{\mathrm{circuit}}, up to the statistical fluctuation expected for 500 shots. This validates the construction of the Grover operator, with the caveat that the theoretical curve uses θ\theta obtained from the circuit’s own k=0k=0 probability.

Estimation Error versus Oracle Calls

Figure 1d and Table 6 compare the RMSE of the MLAE estimate with the classical Monte Carlo error a(1−a)/N\sqrt{a(1-a)/N}, as a function of the number of oracle calls NcallsN_{\mathrm{calls}} from Eq. (12), for six schedules, with 200 independent repetitions per schedule (seeds 0–199) and 500 shots per power. The standard error of each RMSE is approximated by RMSE/2(nrep−1)\mathrm{RMSE}/\sqrt{2(n_{\mathrm{rep}}-1)}

ScheduleNcallsN_{\mathrm{calls}}MLAE RMSE (±\pm s.e.)Classical RMSE at N=NcallsN=N_{\mathrm{calls}}
(0)5000.01286 (±\pm0.00064)0.01226
(0, 1)2 0000.00410 (±\pm0.00021)0.00613
(0, 1, 2)4 5000.00205 (±\pm0.00010)0.00409
(0, 1, 2, 4)9 0000.00105 (±\pm0.00005)0.00289
(0, 1, 2, 4, 8)17 5000.00071 (±\pm0.00004)0.00207
(0, 1, 2, 4, 8, 16)34 0000.00033 (±\pm0.00002)0.00149

Table 6. RMSE versus oracle calls (acirc=0.0818a_{\mathrm{circ}}=0.0818).

For the schedule (0), MLAE reduces to plain sampling and the two errors coincide (0.01286 versus 0.01226), which is a consistency check. At the main schedule (Ncalls=9 000N_{\mathrm{calls}}=9\,000) the MLAE RMSE is 0.00105 against 0.00289 for classical sampling at the same number of calls; at the largest schedule (Ncalls=34 000N_{\mathrm{calls}}=34\,000) it is 0.00033 against 0.00149, a factor of 4.5. A least-squares fit on the six points gives a log-log slope of −0.85±0.02-0.85\pm0.02 for MLAE, versus −0.5-0.5 for classical sampling.

We emphasise the scope of this observation. With a fixed number of shots per Grover power and powers up to k=16k=16, the experiment covers about 1.8 decades of budget. It cannot, by itself, establish an asymptotic O(1/N)O(1/N) slope, and the fitted slope (−0.85-0.85) lies between the classical and the quantum rate. We therefore regard Figure 1d as consistent with, but not a demonstration of, the quadratic scaling. The comparison is in oracle calls only: one call of A\mathcal{A} is a circuit, whereas one classical sample costs a few arithmetic operations, and the MLAE circuits are simulated without hardware noise.

For reference, the theoretical query complexity of canonical QAE to reach additive error ε\varepsilon is O(1/ε)O(1/\varepsilon), versus O(a(1−a)/ε2)O(a(1-a)/\varepsilon^2) for classical sampling. For ε=0.005\varepsilon=0.005 and a≈0.0818a\approx0.0818, the classical requirement is about 3 004 samples, whereas the QAE bound is of order π/(4ε)≈157\pi/(4\varepsilon)\approx157 up to constants and logarithmic factors; in our data the schedule (0, 1) with 2 000 calls already gives an RMSE of 0.00410.

Discussion and Limitations

The test problem admits a closed form.

Under independent Gaussian shocks and a linear map, dTd_T is itself Gaussian, and Pr⁡(dT>τ)=1−Φ((τ−μ)/σd)\Pr(d_T>\tau)=1-\Phi((\tau-\mu)/\sigma_d) is available analytically (Table 3). Quantum (or classical) Monte Carlo is therefore unnecessary for this linearised problem; the closed form serves here as the validation reference. The contribution of the paper is the encoding and estimation workflow, not a computational gain on this instance.

The event is not rare.

The estimated probability is 0.0818. The motivating use case, tail events with small probability, is not exercised by the present experiment.

Classical baselines.

We compared against plain Monte Carlo only. Competitive classical alternatives, such as quasi-Monte Carlo, importance sampling, and deterministic quadrature, would likely perform much better on this low-dimensional problem and must be included in any claim of practical advantage.

Oracle cost and hardware.

The quadratic speedup is in the number of oracle calls. The oracle here (loading, adder, comparator) is cheap to simulate but, on hardware, each call has a non-negligible circuit cost, and Grover powers multiply the depth. Whether this circuit is executable on near-term noisy devices has not been tested: our simulations are noiseless and we have not transpiled to a hardware topology. Results on fault-tolerant resource requirements for similar financial problems suggest that useful advantage may require large logical-qubit and gate budgets [12, 13].

Model limitations.

The first-order linearisation is least accurate in the tails, which is exactly where the target probability is defined; the continuous linearisation error is +0.0067 in probability, and the linear model deviates from the exact one by up to 0.0154 in dTd_T on the grid. Its assessment is limited to the present parameters. We assume Gaussian shocks and independence; real debt dynamics involve fat tails, correlation, regime changes and policy feedback. The integer resolution of the adder (δ=0.004726\delta= 0.004726 of GDP) and the rounding of the threshold shift the effective threshold to 0.9521. Extending to nonlinear, multi-period, higher-dimensional models is where an oracle-based approach could in principle become relevant, but it is also where oracle construction becomes the main difficulty.

Estimator schedule.

The Grover schedule (0, 1, 2, 4) was chosen heuristically, and we did not compare MLAE against canonical QAE or iterative QAE on the same problem.

Conclusion and Outlook

We presented a proof-of-concept pipeline that encodes a two-factor linearised sovereign debt model into a quantum circuit and estimates a tail probability with MLAE. On a noiseless simulator, the estimator recovers the amplitude encoded in the circuit, and the Grover amplification behaves as theory predicts. We have been explicit that the linearised problem has a closed-form solution, that the event is not rare, that only a naive classical baseline was considered, and that the experiment does not by itself establish an asymptotic quantum advantage.

Natural next steps are: (i) non-Gaussian, nonlinear, multi-period debt dynamics with more risk factors, where closed forms are unavailable; (ii) comparison against variance-reduced classical estimators and against other amplitude-estimation variants; (iii) noise models and hardware experiments; and (iv) fault-tolerant resource estimates.

As a longer-term perspective, an amplitude-estimation routine of this kind could be exposed as a tool to software agents operating within digital-twin or simulation environments, such as the geopolitical sandbox of Dessureault et al. [14]. We have not implemented or evaluated such an integration, and we do not claim it as a contribution of this work.

Code and data availability

The source code and computational pipeline required to replicate the findings of this study are publicly available on GitHub. An archival version of the repository, including the exact package versions and random seeds specified in Table 4.

References

  1. [1]Fund, I.M.: Fiscal monitor: Fiscal policy in the great tribulation. Technical report, International Monetary Fund (2023)
  2. [2]International Settlements, B.: The global debt database: methodology and sources. Technical report, Bank for International Settlements (2022)
  3. [3]Glasserman, P.: Monte Carlo Methods in Financial Engineering vol. 53. Springer, ??? (2004)DOI
  4. [4]Brassard, G., Høyer, P., Mosca, M., Tapp, A.: Quantum amplitude amplification and estimation. Contemporary Mathematics 305, 53–74 (2002)DOI
  5. [5]Woerner, S., Egger, D.J.: Quantum risk analysis. npj Quantum Information 5(1), 1–9 (2019)DOI
  6. [6]Egger, D.J., Gámez, P.R., Sun, Y., Zoufal, C., Iten, R., Shen, S., Woerner, S.: Credit risk analysis using quantum computers. npj Quantum Information 7(1), 1–8 (2021)arxiv.org/abs/1907.03044
  7. [7]Stamatopoulos, N., Egger, D., Sun, Y., Zoufal, C., Iten, R., Shen, S., Woerner, S.: Option pricing using quantum computers. Quantum 4, 291 (2020)
  8. [8]Grover, L.K.: A fast quantum mechanical algorithm for database search. Proceedings of the twenty-eighth annual ACM symposium on Theory of computing, 212–219 (1996)arxiv.org/abs/quant-ph/9605043
  9. [9]Suzuki, Y., Uno, S., Raymond, R., Tanaka, T., Onodera, T., Yamamoto, N.: Amplitude estimation without phase estimation. npj Quantum Information 6(1), 1–9 (2020)DOI
  10. [10]Grinko, D., Gacon, J., Zoufal, C., Woerner, S.: Iterative quantum amplitude estimation. npj Quantum Information 7(1), 52 (2021)
  11. [11]Montanaro, A.: Quantum amplitude estimation: variants and beyond. Physical Review A 102(2), 022403 (2020)
  12. [12]Babbush, R., et al.: Focused quantum algorithms for monte carlo methods. arXiv preprint arXiv:2109.03011 (2021)
  13. [13]Chakrabarti, R., Krishnakumar, E., Woerner, S., et al.: A survey of quantum computing for finance. ACM Computing Surveys 54(8), 1–37 (2021)
  14. [14]Dessureault, J.-S., Manzi, A.T.I., Alaoui Ismaili, S., Mijoro Barisoa, A., Yankam Djoké, D., Bergeron, A., Yared-Amare, A., Laliancette, M., Bélanger, É.: Evaluating agentic ai through a geopolitical simulation sandbox. In: Computational Intelligence, pp. 194–213. Springer, ??? (2026). https://doi.org/10.1007/978-3-032-15632-7_11

Paper details

vlnc:W_cdvzaapx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents