Table of Contents
This article is part of the Quantum Snake Oil Dictionary — a series examining terms used in quantum technology marketing. The series is divided into Red Flag Terms (terms with no established technical meaning that almost always signal hype or fraud) and Misused Terms (legitimate concepts routinely stripped of context in marketing). This entry is a Misused Term.
“Deterministic Quantum Computing”
A note before we begin. This article examines “deterministic quantum computing” as a sales claim: the promise that a quantum computer, or software running on one, returns exact answers that come out identical on every run without any sampling. I am not referring to any specific company, product, or individual. A vendor using the term may have built useful classical software and described it loosely; the physics is the same either way, and so are the questions at the end of this entry.
Researchers use “deterministic” in several precise senses, which I describe below, and I have no quarrel with any of them. My concern is vendors who apply the word to the final output of a computation. (Chaos-based ciphers sold as deterministic are a separate problem, covered in Deterministic Chaos Encryption.) If anyone ever demonstrates exact, sampling-free answers from quantum hardware in a peer-reviewed paper, I will update this entry.
The Pitch to Risk Desks
Vendors now make this pitch to banks, opening with a problem every quant recognizes. Banks price derivatives and measure risk with Monte Carlo simulation: generate a large set of possible market paths, value the position along each one, and average the results.
Monte Carlo error falls only as one over the square root of the number of paths, as Shouvanik Chakrabarti and colleagues note in their study of quantum derivative pricing, so each additional decimal digit costs about a hundred times more computation. Two runs started from different random seeds also disagree in their last digits, which annoys anyone who has to reconcile the numbers.
The vendor’s answer is to stop sampling altogether.
In the version I reviewed most recently, a quantum processor computes the expectation directly, every reported price matches its analytic reference to the limit of double-precision arithmetic, and a rerun reproduces the result byte for byte. The vendor describes this as fault-tolerant quantum computation on hardware any customer can rent today (a claim I examined separately in Software-Defined Fault Tolerance) and sells access through an API. Other vendors make an older, cruder pitch and simply promise certainty from a technology the buyer has heard is probabilistic.
Vendors in both camps use “deterministic” to promise two separate things. One is reproducibility: rerunning the job gives back the first answer. The other is exactness: the output equals the true value to the last digit. For the answers these vendors sell (a price, a risk figure, a molecular energy, any real number read out of a quantum state), each promise has a cost in repeated runs or circuit depth, and the cost grows with every digit. Inside quantum computing, the word means something narrower.
What “Deterministic” Means Inside Quantum Computing
Quantum computing researchers use “deterministic” in at least four precise ways. The most common concerns fault-tolerant protocols, many of which check for errors with extra “flag” qubits and discard any run in which a flag fires. Discarding runs is post-selection. It makes a protocol non-deterministic, since nobody can say in advance whether a given attempt will be kept. A deterministic protocol handles a flagged run with additional correction circuitry instead of throwing it away.
Ludwig Schmid and colleagues built their 2025 method for deterministic fault-tolerant state preparation on exactly this distinction, adding a correction circuit that removes the need for post-selection. Sascha Heußen and colleagues compared the two kinds of circuit directly in a 2023 trapped-ion study. Friederike Butt and colleagues at RWTH Aachen University designed deterministic code-switching protocols for logical gates and non-deterministic ones that can prepare magic states.
By the field’s own definition, a pipeline that removes the runs failing a quality check and reports only the runs that pass is the non-deterministic kind. Calling such a pipeline deterministic reverses the technical meaning of the word.
The second use concerns gates. In photonic and networked architectures, some entangling operations succeed only part of the time, and researchers call those gates non-deterministic. In 2010 Ying Li, Sean Barrett, Thomas Stace, and Simon Benjamin showed that fault-tolerant computation still works with such gates even when more than 90 percent of attempts fail, as long as each failure is announced and other errors stay rare. The label refers to whether a gate fires. A computation built entirely from deterministic gates still ends in a probabilistic measurement.
In error detection, “deterministic” has a third use. Surface-code experiments measure parity checks in repeated rounds, and certain products of those outcomes are predictable whenever no error has occurred. Craig Gidney’s Stim simulator calls these combinations detectors. A decoder treats any detector that fires as evidence of an error. On noisy hardware the products stop being predictable, which is why engineers build decoders.
The fourth use is the name of a model of computation. In 1998 Emanuel Knill and Raymond Laflamme introduced deterministic quantum computation with one quantum bit (DQC1), in which a single clean qubit and a register of maximally mixed qubits estimate the normalized trace of a unitary operator. I wrote about the model in more detail in my DQC1 explainer.
In DQC1, the answer comes from repetition. In the procedure as Animesh Datta, Steven Flammia, and Carlton Caves describe it, the experimenter executes the circuit many times, and after L runs the accuracy improves as one over the square root of L. That is the sampling curve the vendors claim to have escaped, inside the one model with “deterministic” in its name.
None describes a certain answer.
Quantum Outputs Are Samples, Even After Error Correction
A quantum computation ends in a measurement, which returns one outcome drawn from a probability distribution. Max Born stated that rule in 1926. Every quantum processor built since behaves accordingly. When the quantity you want is a real number, such as an option price encoded as an expectation value, the processor never prints it. You run the circuit many times, tally the outcomes, and estimate the number from the tallies. The statistical error shrinks only as more runs come in.
Engineers who build fault-tolerant machines cut the error rate of every logical operation. The measurement at the end still returns a random outcome. The threshold theorem behind quantum error correction states that once physical error rates fall below a critical value, logical error rates drop exponentially as the code grows. They approach zero without reaching it. On its Willow processor, Google Quantum AI ran a distance-7 surface code with a logical error rate of 0.143% per cycle of error correction, a result I analyzed when it appeared.
That works out to roughly one failure every 700 cycles.
Universal fault-tolerant computation adds a second source of randomness. The T gates that complete a universal gate set consume magic states, which dedicated factories produce through probabilistic distillation. In a May 2026 analysis of non-determinism in fault-tolerant execution, Aditi Awasthi, Jonathan Baker, and colleagues model both the production and the consumption of those states as stochastic; in their model, a magic-state injection needs a corrective fix-up half the time. A fault-tolerant computer with negligible logical error would still return samples from its circuit’s ideal output distribution. It would return the right samples, and the estimates built from them would still have error bars.
What Each Extra Digit Costs
Quantum computers can estimate an average faster than classical sampling, and the size of the advantage is well established. Gilles Brassard, Peter Høyer, Michele Mosca, and Alain Tapp described quantum amplitude estimation in 2000. Ashley Montanaro showed in 2015 that it delivers a near-quadratic speedup over classical Monte Carlo methods for estimating the mean of any quantity with bounded variance. Classical sampling needs on the order of 1/ε² samples to reach accuracy ε, while amplitude estimation needs on the order of 1/ε applications of the underlying circuit. That scaling is the Heisenberg limit, and no algorithm beats it by more than a constant factor. In an August 2026 analysis of amplitude estimation, Jona Erle and Bálint Koczor restate the matching lower bound, which they credit to Zalka and Burchard: circuit depth squared, multiplied by the number of repetitions, must grow at least as fast as 1/ε².
An election poll shows what those exponents mean. A pollster who interviews 1,000 voters reports a margin of error of about three points; cutting it to a point and a half requires polling four times as many people. Amplitude estimation behaves like a better polling method, one that halves the margin with twice the interviews. Either way, the pollster still reports a margin. A vendor promising exact, zero-variance answers with no sampling is describing a poll with no margin of error, conducted without interviewing anyone.
The arithmetic for a twelve-digit answer follows from the bound. At the Heisenberg limit, an accuracy of 10⁻¹² takes on the order of 10¹² sequential applications of the circuit inside one coherent computation; even at one microsecond per application, that is more than 11 days without an uncorrected error. Shorter circuits make the problem worse, because depth squared times repetitions must still reach about 10²⁴, which raises the total work. In the same Willow experiment, Google’s distance-7 logical qubit lasted about 291 microseconds.
Shouvanik Chakrabarti and his colleagues at Goldman Sachs and IBM published the first complete resource estimate for useful quantum derivative pricing in 2021. Pricing the simpler of their two benchmark derivatives would take about 8,000 logical qubits and a T-depth of 54 million, and the computation would have to finish in about a second to beat classical methods. Each of those logical qubits is built from many physical ones. The advantage the authors target is a faster estimate with a stated error bar.
A claim of twelve-digit, zero-variance prices from today’s noisy cloud processors conflicts with a proven lower bound and with the field’s own engineering estimates. The bound has no exemptions.
Where Identical Reruns Come From
If no quantum algorithm can produce a twelve-digit price for free, identical outputs have to come from somewhere else. Each run on a noisy processor returns a different set of bitstrings, because quantum measurement is random and the device’s calibration drifts over time. A reported number that comes out byte for byte identical across those runs cannot depend sensitively on the measurement record.
My assessment is that its digits come from a deterministic classical stage downstream. Engineers can produce this behavior with three kinds of stage: an optimizer that converges to the same fixed point from any reasonable input, a projection onto an answer computed in advance, and a closed-form formula evaluated in floating point.
I documented the first pattern in Engineered Signature as Evidence, where identical outputs across different processors came from a deterministic post-processing step, and the second in Circular Validation / Answer-Smuggling. The third is the easiest to spot. A vendor that validates its prices against the Black-Scholes-Merton formula and matches it to the last digit a double-precision number can hold has most likely evaluated the same formula classically. A quantum estimate from any feasible number of runs on current hardware would scatter around the reference by many orders of magnitude more.
Reproducibility also has an ordinary classical solution. A Monte Carlo engine started from a fixed random seed draws the same sequence of random numbers on every run of the same software stack, and so returns the same answer every time. From a byte-identical rerun, a buyer learns only that a pipeline is deterministic, and still needs separate evidence that the answer is correct and that a quantum processor computed it.
Ask which stage produced the digits.
What Legitimate Practice Looks Like
Quantum finance is a serious research field, and its published results read very differently from the pitch. Patrick Rebentrost, Brajesh Gupt, and Thomas Bromley set out quantum Monte Carlo pricing of financial derivatives in 2018 as a quadratic speedup in the number of steps needed to reach a given error. Nikitas Stamatopoulos and colleagues worked out option pricing with amplitude estimation in 2020 for vanilla, multi-asset, and path-dependent contracts. They also ran a three-qubit European call option on a 20-qubit IBM device and reported how far the noisy results fell from the exact price.
The Chakrabarti estimate came next. In a 2024 follow-up, Stamatopoulos and William Zeng cut the requirement to about 4,700 logical qubits executing a billion T-gates at 45 MHz. Both teams spell out the resources needed for an advantage and how far current hardware is from providing them.
Credible fault-tolerance teams are just as explicit about their numbers. Google’s surface-code papers report the logical error per cycle, with uncertainties, at each code distance, and show it falling as the code grows. Careful groups that use post-selection report the fraction of runs they kept. With that number, a reader can work out how many runs a larger circuit would need. I describe the cases where groups leave it out in Post-Selection Laundering. Credible teams also submit their results to peer review and publish enough detail for another group to rerun them.
Questions to Ask a Vendor
“What happens when you replace the measurement data with random bitstrings?” Run the identical pipeline, but feed it random bits in place of the processor’s output. If the reported answer does not change, the quantum step contributed nothing to it. Craig Gidney of Google Quantum AI has applied this randomized-data control to claimed quantum results. A vendor confident in its method can run it in an afternoon.
“How many shots did each result use, and what fraction did any filter discard?” A quantum result is an estimate built from repeated runs. Its statistical error depends on the number of runs. If any quality check or parity filter removed data, ask for the total rejection rate across every stage. Vendors sometimes quote a rate for one stage and omit the others. (I once asked a vendor this question in public and never received a number.)
“What is the error bar, and how does it shrink as you buy more machine time?” Every estimate read out of a quantum state has one. A vendor who reports zero error at any budget is making a claim that conflicts with the Heisenberg limit. The follow-up is to ask which part of the system is classical.
“Can you show the raw measurement counts from several reruns next to the final answer?” Raw counts from a noisy processor vary from run to run. If the counts vary and the final answer does not, the digits came from the classical stage. The vendor should be able to say which one.
“Who outside your company has reproduced the result?” Look for peer review, published code, and circuits that an independent group has rerun. A license that forbids reuse of the method, or evidence offered only under NDA, blocks exactly that check. If the vendor compares itself with an audited industry benchmark, ask for the audit report on its own run; a comparison against someone else’s audited record is a self-reported number. When the evidence is missing, vendors usually answer with one of the deflection patterns I have cataloged.
The Bottom Line
Quantum computers return samples. With error correction those samples become trustworthy, and with amplitude estimation they cost quadratically less than classical ones. Each extra digit of precision still has a proven minimum cost. Nobody has published a method that reads an exact real number out of a quantum state.
In quantum error correction, “deterministic” names a protocol that repairs flagged runs instead of discarding them, close to the opposite of what the sales pitch implies. When a vendor offers deterministic quantum computing with exact, zero-variance answers on hardware you can rent today, ask which stage produced the digits. In the case I examined most closely, a classical fixed-point optimizer produced them.
