QAE vs Monte Carlo for Risk EstimationQuadratic Speedup, Honestly Scoped
Where iterative quantum amplitude estimation actually outpaces classical sampling, and where Monte Carlo still wins today.
QAE vs Monte Carlo: How the Two Methods Compare
Classical Monte Carlo estimates an unknown quantity — here, a portfolio's probability of loss — by drawing N random samples from a modeled distribution and averaging. Sampling error shrinks only as 1/√N, so halving the error means roughly quadrupling the sample count. Quantum amplitude estimation (QAE) prepares that same probability distribution as a quantum state and estimates the amplitude of the "loss exceeds threshold" subspace directly: O(1/ε) oracle calls reach additive error ε, against O(1/ε²) samples for classical Monte Carlo to reach the same guarantee.
"Quadratic speedup" describes that exponent, not a fixed multiplier on run time. It means QAE's query count grows linearly with 1/ε while Monte Carlo's grows with 1/ε² — it does not mean QAE is faster in wall-clock terms today. Each QAE oracle call is a full quantum circuit — state preparation plus a scheduled run of Grover-operator repetitions — and on current simulators and NISQ devices that per-call cost outweighs the query-count advantage for most practical precisions. The two error curves are normalized to cross around ε ≈ 10⁻²: to the right of that point, Monte Carlo's simplicity wins; to the left, QAE's query count pulls ahead — asymptotically.
Canonical QAE reads out the amplitude via quantum phase estimation, which needs a wide, low-error ancilla register and coherence times today's devices don't reliably hold. Iterative Quantum Amplitude Estimation (IQAE) instead narrows a confidence interval on the amplitude step by step, applying the Grover operator a scheduled, limited number of times each round and resetting the circuit before decoherence and gate noise build up. That's the variant behind Superpositions Studio's quantum amplitude estimation page and its VaR/CVaR risk-estimation proof of concept, which is also the worked comparison further down this page.
None of this runs on production quantum hardware yet. The proof of concept below ran on a noiseless simulator — and simulating that circuit classically for verification was itself capped at 16 qubits by compute cost, a reminder that the "quantum" side of this comparison is still emulated on classical machines almost everywhere. A clear, industrial-scale advantage over Monte Carlo for realistic financial models most likely needs fault-tolerant hardware: resource-estimation studies for derivative pricing put that at thousands of logical qubits and roughly 10⁷–10⁹ logical gate operations, and public roadmaps from IBM, PsiQuantum, and others tentatively place that hardware generation in the 2030s. Until then, simulators are where IQAE gets exercised, refined, and — as below — checked against the classical answer it's supposed to match.
Side by Side: Error Scaling and Requirements
| Dimension | Classical Monte Carlo | QAE (IQAE) |
|---|---|---|
| Error scaling | Θ(log(1/α)/ε²) samples — error falls as 1/√N | O(log(2/α · log²(π/4ε))/ε) oracle calls — error falls close to 1/N |
| Where the curves cross | Cheaper to the right of ε ≈ 10⁻² (looser precision targets) | Asymptotically cheaper to the left of ε ≈ 10⁻² (tighter targets) |
| What has to be prepared | A stochastic model of the distribution (here, a Gaussian log-return fit) and a random-number generator | An amplitude-embedding operator U for the discretized distribution plus a loss oracle L, combined into A = LU |
| Error-reduction technique | Antithetic variates, importance sampling, control variates, quasi-Monte Carlo sequences — shrink the constant, not the exponent | The Grover operator built from A, applied a scheduled number of times per IQAE round — changes the exponent itself |
| Hardware today | Any classical CPU; embarrassingly parallel, no qubits required | Noiseless simulator or small NISQ device; the risk-estimation proof of concept used 11 qubits and 500–1,000 gates for 2 assets |
| Who wins today | Essentially always, once state-preparation and circuit-depth costs are counted | Only asymptotically, on paper, until fault-tolerant hardware closes the per-query cost gap |
When to Choose Which
Reach for Monte Carlo when
- You need a result today: it runs on any classical machine, with no circuit design and no qubit budget to plan around.
- The target precision is loose enough that O(1/ε²) sampling is still cheap — roughly ε ≳ 10⁻² in the risk-estimation setting below.
- The payoff or loss function is easy to sample directly but awkward to compile into a compact quantum oracle.
- You want mature variance-reduction tooling — antithetic variates, importance sampling, control variates — rather than a research-stage circuit.
Reach for QAE / IQAE when
- The target precision sits left of the ε ≈ 10⁻² crossover, where the query-count exponent starts to dominate the sample count.
- The quantity of interest is naturally an amplitude — a probability, expected value, or definite integral — with a clean oracle to encode it.
- You're characterizing IQAE's behavior on a simulator against a known classical answer, ahead of any deployment on real hardware.
- You can work within today's constraints: small qubit counts, noiseless simulators, and gate counts in the hundreds for a multi-asset oracle.
The same question — asymptotic advantage versus today's practical cost — comes up for combinatorial problems too; see the quantum annealing vs QAOA comparison for how it plays out there.
The VaR/CVaR Proof of Concept: A Worked Comparison
9,909
Time steps, AAPL + MSFT
11
Qubits (5/asset + 1 ancilla)
0.001
VaR MSE vs. Monte Carlo
0.006
CVaR MSE vs. Monte Carlo
This isn't run on a synthetic toy distribution. The underlying risk-estimation use case draws on a universe of 8,049 instruments — 5,884 stocks and 2,165 ETFs, six numerical features each — but the IQAE-vs-Monte-Carlo proof of concept itself narrows that to two large-cap names, AAPL and MSFT, whose combined log-price time series (9,909 time steps) is converted to log-scale growth rates and fitted to a Gaussian process for the mean and variance.
That distribution is discretized into 2ⁿ probabilities spanning μ−5σ to μ+5σ and loaded into a quantum state through an amplitude-embedding operator U. A loss oracle L flips a shared ancilla qubit whenever the simulated loss exceeds a chosen threshold; the resulting amplitude a = sin²θₐ is the probability that the loss exceeds it. Sweeping the threshold traces out the loss boundary (VaR) and the mean loss beyond it (CVaR) — combined into a single operator A = LU whose amplitude IQAE then estimates iteratively, with the result compared back to a classical Monte Carlo estimate of the same VaR and CVaR values.
The proof of concept ran with 5 qubits per asset plus 1 shared ancilla — 11 qubits total for the 2-asset case — using 500–1,000 gates for the embedding and loss operators together, and 1,000 shots per amplitude-estimation run, on a noiseless quantum simulator (classical simulation of the circuit for verification was itself capped at 16 qubits by compute cost). A single amplitude-estimation run reported 202 ms; the full run, at 1,000 shots across 8 CPU cores, took 20.6 s end to end.
Measured against the classical Monte Carlo reference on the same distribution, IQAE's VaR estimate came in at 0.001 mean squared error and its CVaR estimate at 0.006 MSE. At a coarser setting — 5 qubits per asset with 100 Grover-operator iterations — the relative error was 0.007 for VaR and 0.018 for CVaR. Read those as what they are: agreement with Monte Carlo on a 2-asset simulator run, not a demonstrated advantage over it — the query-count argument in the section above is asymptotic, and 2 assets on a simulator sits nowhere near where that asymptote starts to bite. The full workflow, dataset preparation, and quantum-vs-classical comparison are detailed on the VaR/CVaR risk-estimation proof of concept page; readers looking past single-portfolio tail risk toward full portfolio construction should see quantum portfolio optimization.
Business Impact
Significant business value shows up once portfolios grow past a handful of assets — which is also where Monte Carlo's O(1/ε²) sampling starts to get expensive.
~700
Qubits for a 100+ asset portfolio
$30/run
Today's cost per 1,000-shot run
$10/run
Projected cost by 2027–2028
0.1%
Target precision behind that estimate
At a target precision of ε = 0.1%, the same quadratic-speedup argument that makes IQAE asymptotically cheaper than Monte Carlo implies that simulating a 100+ asset portfolio at this proof of concept's per-asset qubit budget would need roughly 700 qubits. Priced at roughly $30 per 1,000-shot run today, per AWS Braket's published pricing, that run costs about $30; anticipated hardware and pricing improvements could bring that to around $10 per run by 2027–2028 — at which point the computational cost becomes comparable to a high-dimensional Monte Carlo simulation on classical hardware, rather than a premium paid for the quantum method. None of this is available yet: it's a projection built on today's per-qubit pricing and tomorrow's expected qubit counts, not a benchmark that has been run.
The realistic timeline, per the same use case's own outlook, runs in three stages. Near-term, NISQ devices can only simulate toy cases at high precision, since limited coherence time rules out long chains of Grover-operator applications. Medium-term, early error-corrected hardware should support 10–20 assets at roughly 0.1% precision. Long-term, on fault-tolerant hardware, IQAE is positioned to become the standard approach for portfolios of hundreds of assets — the same regime where Monte Carlo alone becomes the practical bottleneck this comparison exists to address.
Quant risk teams
Deciding whether to prototype IQAE against an existing Monte Carlo VaR/CVaR pipeline before committing engineering time.
Applied quantum researchers
Benchmarking IQAE's query-count advantage against a real classical reference instead of a synthetic distribution.
Product and hardware-roadmap leads
Scoping when a quantum amplitude-estimation deployment could beat Monte Carlo on cost, given today's per-qubit pricing and coherence limits.
How It Works on the Platform
One distribution, two estimation methods, one reproducible comparison against the same reference.
Prepare
Fit the asset return distribution (a Gaussian log-return model) and discretize it into the probability table both methods estimate against
Run Monte Carlo
Draw samples from the modeled distribution and compute the classical VaR/CVaR reference values
Run IQAE
Encode the same distribution into an 11-qubit state, build the loss oracle, and estimate the loss-exceedance amplitude iteratively
Compare
Check IQAE's estimates against the Monte Carlo reference: MSE, relative error, run count, and shots per run
Export
Download the methods, both implementations, and a reproducible benchmark report
Related pages
Algorithms, applied use cases and benchmarks connected to this page
Algorithms
Quantum Amplitude Estimation (QAE)
Estimate probabilities, expectations, and VaR/CVaR with IQAE's quadratic speedup over Monte Carlo sampling, benchmarked with a real proof of concept.
Finance
VaR and CVaR Estimation with IQAE
Estimate Value at Risk and Conditional Value at Risk with Iterative Quantum Amplitude Estimation, Monte Carlo benchmarks, and reproducible reports.
Finance
Quantum Portfolio Optimization: QAOA, HHL and Annealing
Compare quantum portfolio optimization: QAOA with quantum walk mixers, HHL mean-variance allocation, and quantum annealing for cardinality-constrained QUBOs.
Comparisons
Quantum Annealing vs QAOA for QUBO Problems
Compare quantum annealing with QAOA on QUBO/Ising optimization: hardware, constraints, scaling, and two real portfolio-optimization benchmarks.
Frequently Asked Questions
Is quantum amplitude estimation faster than Monte Carlo?
In query count, yes: QAE needs O(1/ε) oracle calls to reach additive error ε, while classical Monte Carlo needs O(1/ε²) samples for the same guarantee — a quadratic reduction that widens as the target precision tightens. In wall-clock terms, not yet: each QAE oracle call is a full quantum circuit, and on today's simulators and NISQ devices that per-call cost outweighs the query-count advantage for most practical precisions.
How many qubits does QAE need for risk estimation?
It scales with how finely the distribution is discretized and how many assets are involved. The VaR/CVaR proof of concept behind this page used 5 qubits per asset plus 1 shared ancilla — 11 qubits total for 2 assets — and scaling that same resolution to a 100+ asset portfolio would need roughly 700 qubits.
Can QAE replace Monte Carlo for VaR today?
Not in production. The proof of concept ran on a noiseless simulator, needed 500–1,000 gates to encode the loss operator for just 2 assets, and classical simulation of that circuit for verification was itself capped at 16 qubits by compute cost. Monte Carlo remains the practical choice on today's hardware; a clear cost advantage for QAE most likely needs fault-tolerant qubits that public roadmaps place in the 2030s.
What is iterative quantum amplitude estimation (IQAE)?
IQAE is the NISQ-friendly variant of QAE. Instead of reading out the amplitude with quantum phase estimation, which needs a wide, low-error ancilla register, it narrows a confidence interval on the amplitude step by step, applying the Grover operator a scheduled number of times each round and resetting before decoherence and gate noise build up. It's the variant used in the risk-estimation proof of concept this page compares against Monte Carlo.
Does classical Monte Carlo have its own version of a speedup?
Techniques like antithetic variates, importance sampling, control variates, and quasi-Monte Carlo sequences all reduce the constant in front of Monte Carlo's 1/√N error, sometimes substantially, but none of them change the exponent. QAE's quadratic speedup changes the exponent itself, from 1/√N to close to 1/N, which is a fundamentally different kind of improvement.
Where do the two methods actually cross over?
The two error-scaling curves are normalized to intersect around ε ≈ 10⁻². Above that precision threshold, Monte Carlo's simplicity and hardware maturity win; below it, IQAE's query count pulls ahead asymptotically, though today's state-preparation and circuit-depth costs mean the practical crossover, in wall-clock terms, sits well past where the asymptotic one does.
Benchmark QAE Against Monte Carlo
Run the 11-qubit IQAE VaR/CVaR benchmark and review its 0.001 and 0.006 MSE against a classical Monte Carlo reference.
Try your first case for free