Quantum computing for option pricing: three routes and when each is real
Three routes are on offer. Amplitude estimation prices an option on a quantum computer with quadratically fewer circuit calls than the samples Monte Carlo needs for the same error, a proven gain whose circuits have already run on small devices for toy contracts. Turning it into a faster price is the open part: every published resource estimate for a contract a bank would price needs a fault-tolerant machine far larger than any built so far. The quantum-inspired routes, tensor trains and tensor neural networks that compress a pricer, and tensor-network Fourier pricing for baskets, run on ordinary CPUs and GPUs today and have published speed-ups against classical reference solvers.
Updated
The three routes at a glance
- Amplitude estimation
- Runs on: a fault-tolerant quantum computer. Status: circuits demonstrated on small devices; the best resource estimates still need thousands of logical qubits[4][5]
- Tensor trains and tensor neural networks
- Runs on: CPUs and GPUs. Status: published speed-ups over finite-difference solvers and dense networks on Heston and Black-Scholes problems[8][10]
- Tensor-network Fourier pricing
- Runs on: a laptop. Status: a proof of concept on correlated baskets of European options[11]
- Which is quantum
- Only the first. The other two borrow maths from quantum physics and use no quantum hardware
- What this page does not do
- Compare amplitude estimation and Monte Carlo query by query. Here the question is which route a pricing desk could use, and when
What a pricing desk is choosing between
Pricing is rarely one contract once. A desk prices a book under many parameter sets for calibration, sensitivities and risk, and the expensive cases are path-dependent or multi-asset contracts that fall back on Monte Carlo. The three routes attack that cost in different places.
Amplitude estimation speeds up the sampling itself: it replaces Monte Carlo's average over paths with a quantum estimate whose error falls faster with the number of circuit calls. Tensor trains and tensor neural networks compress the pricing function, so that an expensive solver runs once, offline, and each later price is cheap. Tensor-network Fourier pricing compresses the high-dimensional integral of a basket so that it can be evaluated directly. Only the first needs a quantum computer; the others are quantum-inspired, which means they use tensor networks from many-body physics on classical hardware. The tensor networks page explains the structure they share.
Route 1: amplitude estimation, and the reality check
The idea is that the expected payoff of an option can be written as the probability that one qubit of a circuit ends in . Amplitude estimation, built on the same rotation as Grover's algorithm, estimates that probability with error from about circuit calls, where Monte Carlo needs about samples[2]. Rebentrost, Gupt and Bromley set out the pricing algorithm[1], and Stamatopoulos et al. built the circuits for vanilla, multi-asset and barrier options and ran a three-qubit European call on the IBM Q Tokyo device with error mitigation[3].
The reality check is the full cost of a contract a bank would actually price. Chakrabarti et al. estimated it for a basket autocallable and a target accrual redemption forward, at a target error of [4]. Even with their new re-parameterisation method, the autocallable needs 8k logical qubits[4] and a T-depth of 54 million[4]. For the quantum price to arrive in 1 second[4], which they take as the bar for advantage, the machine must execute T-gates at 10 MHz[4], against logical clock rates of about 10 kHz[4] in the estimates they cite. Classically, such an autocallable can take five to ten seconds[4] of Monte Carlo with at least forty thousand paths[4].
Later work lowered the qubit count without removing the bar. Stamatopoulos and Zeng encoded payoffs with quantum signal processing and arrived at 4.7k logical qubits[5] and T-gates[5] executed at 45 MHz[5], against 207 MHz[5] for the arithmetic method under their own updated estimate, so their clock rate does not compare one to one with the 10 MHz above. Babbush et al. argue more generally that quadratic speed-ups will not give an advantage on early fault-tolerant machines, because the overhead of error correction eats them[6]. For a buyer, amplitude estimation is a research position to hold, not a tool to procure.
What would bring amplitude estimation to a pricing desk
The estimates above name the conditions, and none of them is the query count, which already favours amplitude estimation. The first is the logical clock rate: the machine has to run T-gates fast enough to return a price inside the time a classical pricer takes, and the rate that requires is far above the logical rates in the estimates Chakrabarti et al. cite[4]. The second is the circuit: in Stamatopoulos and Zeng's own comparison, encoding the payoff with quantum signal processing needed fewer logical qubits and a lower clock rate than the arithmetic method[5], and further cuts of that kind move the bar closer. The third is the overhead of error correction, which Babbush et al. find large enough to cancel a quadratic gain on early fault-tolerant machines[6].
None of this can be tested at full size on today's hardware. Two parts can be tested now: the oracle calls amplitude estimation needs for a payoff of yours, counted on a simulator against your own Monte Carlo, and how far gate noise lets the Grover depth grow on a small circuit on a real QPU. The VaR and CVaR guide runs the first measurement for a tail probability, and the comparison with Monte Carlo shows how the query gap grows as the target tightens.
Run it on a real QPU with EigenOrbit RouterA small IQAE circuit on a QPU shows how noise limits the Grover depth. It does not price a real contract; that needs the fault-tolerant machines above.In eigenkit: pip install "eigenkit[router]", then provider=RouterConfig(device=...) (docs).
Route 2: tensor trains and tensor neural networks
Chebyshev interpolation turns parametric pricing into a two-phase job: offline, run the expensive solver at a grid of Chebyshev nodes in parameter space; online, evaluate a polynomial[7]. The grid grows exponentially with the number of parameters, which is where tensor trains come in. Glau, Kressner and Statti stored the grid of prices in tensor-train format and filled it by tensor completion from a sample of solver runs. For an American put in the Heston model with five varying parameters, the online phase was 75 times faster[8] than the finite-difference reference and at least as accurate, and the completed tensor was compressed by a factor of 115[8]. The offline completion took about 6 minutes[8]. For European baskets in Black-Scholes they treated up to 25 underlyings[8], with efficiency gains up to 4000[8] over a Monte Carlo reference with control variates, when only the initial prices vary. A related dynamic Chebyshev method prices American options faster than Longstaff-Schwartz least-squares Monte Carlo in the authors' tests[9].
A tensor neural network replaces dense layers with tensor-network layers (matrix product operators). Patel et al., from Multiverse Computing and Crédit Agricole, used it to solve the Heston pricing PDE through its stochastic formulation. On an Nvidia T4 GPU the tensor networks converged up to twelve times[10] faster than dense networks of the same parameter count[10]. Two caveats from the paper itself: the gain is in training time, not in the price, and a well-chosen dense network matched the tensor network's accuracy at the same parameter count, so the parameter saving depends on which dense design you compare against.
Route 3: tensor-network Fourier pricing
Fourier methods price a single-asset option almost exactly from the characteristic function of the model, but the integral becomes intractable for a basket because its grid grows exponentially with the number of assets. Kastoryano and Pancotti, at AWS, wrote both the characteristic function and the transformed payoff as matrix product states built by TT-cross, so the price becomes one cheap inner product. On a min-option over correlated geometric Brownian motions, run on a laptop, the method matched exact prices up to 4 assets[11] and matched Monte Carlo with a significant speed-up up to 15 assets[11], a speed-up of several orders of magnitude[11] over vanilla Monte Carlo in their proof of concept. A bond dimension of about 10 to 15[11] was often enough for their correlations.
It is the youngest of the three routes: one payoff, one model family, one correlation structure, and a comparison with vanilla rather than quasi-Monte Carlo. Models with path dependence or stochastic volatility are not covered.
Which route fits which workload
| Route | Best fit | Published evidence | What is missing |
|---|---|---|---|
| Amplitude estimation | Path-dependent contracts priced by Monte Carlo | Resource estimates only; small hardware runs[3][4] | A fault-tolerant machine that runs millions of T-gates per second |
| Tensor-train interpolation | Repricing one product across many parameter sets: calibration, risk, real-time quotes | Heston American put and Black-Scholes baskets[8] | Tests on production books; cost grows with the ranks the problem needs |
| Tensor neural networks | Deep PDE solvers where training time is the bottleneck | Heston PDE on one GPU[10] | Independent replication; a peer-reviewed version |
| Tensor-network Fourier | European baskets with a known characteristic function | Min-option on up to 15[11] correlated assets | Path dependence, other models, a quasi-Monte Carlo baseline |
How to read a quantum option pricing claim
Vendor and press figures on this topic often drop the context the papers give. Five questions recover it.
Quantum or quantum-inspired?
If the method runs on GPUs, it is a classical algorithm and should be compared with the best classical alternatives, not with a quantum roadmap.
Against which classical reference?
Vanilla Monte Carlo, quasi-Monte Carlo, finite differences and Fourier methods can differ by orders of magnitude on the same contract; for one European call, Kastoryano and Pancotti found vanilla Monte Carlo needed to samples[11] to reach a standard deviation of , where a Fourier grid ran in milliseconds. A speed-up over the slowest of them says little.
Is the offline phase counted?
Interpolation and trained networks move cost offline. That pays when the same product is repriced many times, and not for a one-off price.
Logical or physical qubits?
Resource estimates count error-corrected logical qubits; each needs many physical qubits. Today's devices offer physical qubits with no full error correction.
Is it a paper or a press release?
Use the figures in the paper, with its own caveats. A number that appears only in a talk or a press note has not been checked by anyone else.
Questions
Can quantum computers price options today?
Only toy instances. Amplitude estimation circuits for vanilla, basket and barrier options have been worked out and simulated, and a three-qubit European call has run on an IBM device, but a contract a bank would price needs a fault-tolerant machine with thousands of logical qubits. Classical Monte Carlo and PDE methods remain faster and more accurate on every real book.
How many qubits does quantum option pricing need?
Several thousand logical qubits in every published estimate for a contract of business interest: Chakrabarti et al. (2021) for an autocallable and a TARF, and Stamatopoulos and Zeng (2024) after encoding the payoff with quantum signal processing, which cut the count but not below thousands. These are error-corrected qubits, each made of many physical ones, and the machine must also execute millions of T-gates per second.
What is tensor network option pricing?
It is a family of classical methods that use tensor networks, such as tensor trains, to compress the high-dimensional objects of pricing: a grid of prices over many parameters, the weights of a neural network PDE solver, or the Fourier integral of a basket. They run on CPUs and GPUs and have published speed-ups against classical reference solvers.
Is quantum-inspired option pricing the same as quantum computing?
No. Quantum-inspired methods borrow mathematics from quantum physics, mostly tensor networks, and run on ordinary hardware. Their results stand or fall against other classical methods, and they need no quantum computer now or later.
Does amplitude estimation beat Monte Carlo for option pricing?
In the number of circuit calls, yes: its error falls as one over the number of calls, against one over the square root for Monte Carlo. In running time, not on any hardware that exists, because each call on an error-corrected machine is slow. The resource estimates on this page put the break-even far beyond current devices.
What does eigenkit offer for option pricing?
An IQAE solver that estimates an amplitude from a circuit you supply, with a confidence interval and an oracle-call count, and a portfolio VaR and CVaR estimator built on it. It has no option-pricing oracle, no Monte Carlo baseline and no tensor-train or tensor neural network pricer.
Sources
- P. Rebentrost, B. Gupt, T. R. Bromley. Quantum computational finance: Monte Carlo pricing of financial derivatives. Physical Review A 98, 022321, 2018. arXiv:1805.00109.
- A. Montanaro. Quantum speedup of Monte Carlo methods. Proceedings of the Royal Society A 471, 20150301, 2015. arXiv:1504.06987.
- N. Stamatopoulos, D. J. Egger, Y. Sun, C. Zoufal, R. Iten, N. Shen, S. Woerner. Option pricing using quantum computers. Quantum 4, 291, 2020. arXiv:1905.02666.
- S. Chakrabarti, R. Krishnakumar, G. Mazzola, N. Stamatopoulos, S. Woerner, W. J. Zeng. A threshold for quantum advantage in derivative pricing. Quantum 5, 463, 2021. arXiv:2012.03819.
- N. Stamatopoulos, W. J. Zeng. Derivative pricing using quantum signal processing. Quantum 8, 1322, 2024. arXiv:2307.14310.
- R. Babbush, J. R. McClean, M. Newman, C. Gidney, S. Boixo, H. Neven. Focus beyond quadratic speedups for error-corrected quantum advantage. PRX Quantum 2, 010103, 2021. arXiv:2011.04149.
- M. Gaß, K. Glau, M. Mahlstedt, M. Mair. Chebyshev interpolation for parametric option pricing. Finance and Stochastics 22(3), 2018.
- K. Glau, D. Kressner, F. Statti. Low-rank tensor approximation for Chebyshev interpolation in parametric option pricing. SIAM Journal on Financial Mathematics 11(3), 2020. arXiv:1902.04367.
- K. Glau, M. Mahlstedt, C. Pötz. A new approach for American option pricing: the dynamic Chebyshev method. SIAM Journal on Scientific Computing 41(1), 2019. arXiv:1806.05579.
- R. G. Patel, C.-W. Hsing, S. Sahin, S. Palmer, S. S. Jahromi et al. Quantum-inspired tensor neural networks for option pricing. arXiv preprint, 2022. arXiv:2212.14076.
- M. Kastoryano, N. Pancotti. A highly efficient tensor network algorithm for multi-asset Fourier options pricing. arXiv preprint, 2022. arXiv:2203.02804.
Run it yourself with eigenkit
eigenkit is EigenOrbit's open-source Python library: write a circuit once, run it on any simulator or QPU, and test it against a classical baseline on your own data before you invest in it or rule it out.
$ pip install eigenkit