↵ select ↓ ↑ navigate esc close

Paper Feed: Elucidating Reaction Mechanisms on Quantum Computers

Paper Feed ·

Listen: https://serve.buildtall.systems/b54e27081ac312ffc136057591280469f500d2c1bcdbbdea3487e5d16b8ba844.mp3

Source: https://serve.buildtall.systems/34343e00ffd4ea98f9b120e99b71586785d8ae4af984f3630b1133681508e1d8.pdf


This episode covers Elucidating Reaction Mechanisms on Quantum Computers, a paper that asks whether a quantum computer could map the steps of a chemical reaction in the enzyme that fixes nitrogen from the air. The authors work through one concrete case and estimate the machine it would take, rather than reporting a working device.

First, a note from the citation index: OpenAlex, the online index of scholarly works, has no record that matches this paper's title and authors as of October 22, 2026, so this episode cannot report a citation count or a citation trend. The paper is a theoretical feasibility study, not a laboratory experiment. Its question is how a quantum computer would be used in practice to work out a reaction mechanism, the sequence of short-lived intermediates and transition structures a reaction passes through, and how large and how fast such a machine would need to be. For a test case the authors choose biological nitrogen fixation by the enzyme nitrogenase, which turns dinitrogen into ammonia at room temperature and pressure, in contrast to the high-temperature, high-pressure Haber-Bosch process used to make fertilizer. The enzyme's active site, a metal cluster called FeMoco, has a mechanism the authors say is still not known, and they argue that classical methods lack the accuracy needed to settle it. The data are computed rather than measured. The authors optimize structures of a model of the FeMoco cofactor with density functional theory, a widely used approximate method, in several charge and spin states, and they compute the electron interaction inputs for two active spaces, one of 54 electrons in 54 spatial orbitals and one of 65 electrons in 57. They then describe a quantum workflow built on quantum phase estimation, a second-order Trotter-Suzuki approximation of the time evolution, and the surface code for error correction. They count the costly T gates, the runtime and the number of logical and physical qubits for three ways of running the rotations, serial, nested and parallel, against an energy error target of 0.1 millihartree, a thousandth of a hartree, the usual unit of energy in chemistry. The analysis is explicit about its assumptions, which is a strength: it sets an error budget, divides it across three error sources, and reports ranges rather than single figures. Several of its numbers are pessimistic by the authors' own account, since their rigorous bounds are described as extremely pessimistic and their simpler gate counts are said to probably exceed what is actually needed. The evidence is thinner elsewhere. The estimates for the large molecule are extrapolated from small molecules that could be computed exactly, and the authors say the data show little clear trend to support extrapolation to a molecule of FeMoco's size. Their check on how well simple starting states overlap with the true ground state is also limited to small systems. The workflow assumes a good starting state is already available, and the paper does not estimate what preparing one would cost. The material on hand contains no funding or data availability statement and does not name a journal or other venue.

The paper opens by explaining what a reaction mechanism is. A chemical reaction mechanism is a network of molecular structures, some short-lived and some long-lived intermediates, connected by transition structures. The authors explain that this network sits on what is called the Born–Oppenheimer surface, which maps the energy of the electrons as the positions of the nuclei change, and which can be refined with corrections for nuclear motion.

The relative energies of the stable structures set how thermodynamically stable each one is. The energy gaps between those stable structures and the transition structures that link them set how fast the reaction proceeds. Those rates come from rate constants that enter exponential functions, so the authors conclude that "very accurate energies are required for the reliable evaluation of the rate constants."

The paper argues that understanding complex mechanisms, such as transition-metal catalyzed transformations, therefore requires highly accurate electronic structure methods. It calls the electron correlation problem, the difficulty of describing how electrons influence one another's motion, "one of the most vexing problems in quantum chemistry," one that remains despite decades of progress. Density functional theory, a widely used approximate approach, has been applied to many kinds of electronic structures. The authors say its accuracy "is too low for truly reliable quantitative predictions." On classical computers, they note, molecules with much less than a hundred strongly correlated electrons are already out of reach for systematically improvable ab initio methods, which are first-principles calculations that could reach the required accuracy.

The authors recall that Richard Feynman proposed quantum computers because of this intractability. Lloyd and Zalka first investigated the exponential speedups that quantum simulation might offer, and Lidar, Aspuru-Guzik and others applied the idea directly to quantum chemistry. The field has stayed active, with increasingly sophisticated methods aimed at lowering the cost of quantum chemistry simulation.

The paper then raises an open question: how exactly to use a quantum computer to solve an important chemistry problem. The authors write that

the inability to point to a clear use case complete with resource and cost estimates is a major drawback.

An exponential speedup alone may not produce a useful algorithm if a typical application would need so much time and memory that it stays out of reach even for a quantum computer.

To address this, the authors set out to show, for one prototypical chemical system, how a quantum computer would be used in practice, and how large and how fast such a machine would need to be to finish the calculation in a reasonable time. They say this sets a target for the kind of device that existing research should aim to produce, and that it gives confidence that quantum simulation can answer problems that matter both scientifically and economically.

The chemical process they choose is biological nitrogen fixation by the enzyme nitrogenase. This enzyme turns a dinitrogen molecule into two ammonia molecules and one dihydrogen molecule "under ambient conditions." The industrial Haber-Bosch process needs high temperatures and pressures and is therefore energy intensive. By contrast, the active site of the molybdenum-dependent nitrogenase, the iron molybdenum cofactor, known as FeMoco, "can split the dinitrogen triple bond at room temperature and standard pressure." The enzyme has two parts: the Fe protein, made of two identical copies of one subunit, and the MoFe protein, a four-subunit assembly made of two copies each of two different subunits.

The paper's first figure shows the nitrogenase MoFe protein from the bacterium Clostridium pasteurianum, taken from the protein data base, along with the close environment of its active site. That active site is a cluster called FeMoco, which sits buried inside the protein, and the authors build a structural model of it for their calculations. Nitrogenase is the enzyme that performs nitrogen fixation, turning nitrogen from the air into a form that plants can use. The authors stress that this process matters for fertilizer production, yet they state plainly that "the mechanism of nitrogen fixation at FeMoco is not known."

They explain why the question remains open. Experiment has not yet revealed enough detail about the chemical steps, and theory is limited by what the paper calls "intrinsic methodological limitations of traditional quantum chemical methods."

The problem shows up in synthetic chemistry too. Chemists have found only three homogeneous catalysts, meaning catalysts that work in the same phase as the reactants, that fix nitrogen under ambient conditions, meaning ordinary temperature and pressure. The paper notes that all three decompose under reaction conditions, and their turnover number, the count of reactions a catalyst completes before it stops working, is less than a dozen.

Every chemical mechanism rests on a simple idea, and the paper starts there. The authors write that the work of elucidating a mechanism "requires the identification of all relevant stable intermediates and transition states" and the calculation of their properties. Those properties matter for two reasons. They allow experimental verification, and the energies attached to these structures reveal thermodynamic stability and kinetic accessibility, which together describe how readily a reaction can proceed.

The scale of the task is large. The paper notes that "a multitude of charge and spin states need to be explicitly calculated" in search of the relevant ones that make the whole chemical process viable. Such a search can involve thousands of elementary reaction steps, and the reaction energy of each must be calculated reliably.

For nitrogenase, the enzyme that fixes nitrogen from the air, the authors describe numerous protonated intermediates, meaning species carrying extra hydrogen ions, of the iron-molybdenum cofactor known as FeMoco, the cofactor that binds dinitrogen. Reduced intermediates in different charge and spin states are also feasible, and each must be assessed for its energy relative to the others. The paper points to kinetic modelling as the stage with the tightest demands, stating that "kinetic modelling poses tight limits on the accuracy of activation energies entering the argument of exponentials in rate expressions." An activation energy is the energy barrier a reaction must cross. Because it sits inside an exponential, a small error in it produces a large error in the predicted rate.

The authors also observe that most chemical processes are local. Usually only one or two bonds are broken or formed at a time, so the reacting core involves only a small number of atoms. That core is still shaped by its surroundings, such as a protein matrix or a solvent. Many embedding methods exist to account for these environment effects. In the paper's words, they "incorporate environment terms, ranging from classical electrostatics to full quantum descriptions, into the Hamiltonian to be diagonalized." The Hamiltonian is the mathematical operator whose energy levels the calculation finds. For nitrogenase, the authors point to an electrostatic quantum-mechanical and molecular-mechanical model, which treats the reactive core quantum mechanically and the surrounding protein with classical physics. They say such a model "can properly account for the energy modulation of the chemical transformation steps at FeMoco," meaning the way the protein pocket tunes the energy of each step.

The calculations use a structural model of the active site, the part of the enzyme where the chemistry happens. This model carries only simplified versions of the groups that anchor the cofactor to the protein, and it serves as the quantum-mechanical part of the calculation. The authors argue that studying this bare model is no limitation for their feasibility analysis, because electrostatic embedding does not change the number of orbitals considered when building the wavefunction, the mathematical description of the electrons' state. They ran full molecular structure optimizations with density functional theory, a widely used approximate quantum method, on this model in several charge and spin states, so that their analysis does not rest on a single electronic structure. The model describes the resting state of the cofactor. The authors add that binding a small molecule such as dinitrogen, dihydrogen, diazene, or ammonia "will not decisively increase or reduce the complexity of its electronic structure." The paper's appendix on the FeMoco active site gives further detail.

The paper then lays out a workflow, described in the caption of its flowchart, that divides the labour between classical and quantum computers. On the classical side, researchers generate and optimize structures of the active species within a suitable environment, optimize the molecular orbitals for a chosen active space, transform four-index integrals from the atomic-orbital basis to the molecular basis, and add temperature and entropic corrections before kinetic modelling. The quantum computer supplies the core energy, which the caption describes as a "quantum full configuration interaction (QFCI) energy" in a complete active orbital space. Once the quantum computer returns the ground-state energy of the Hamiltonian, nuclear-motion corrections turn it into enthalpic and entropic quantities at a given temperature. The resulting energy differences between stable intermediates and transition structures then enter the rate expressions used in kinetic modelling. The caption also notes that for complicated mechanisms, "this modeling might point to the exploration of additional structures."

The authors begin by setting out how chemists think about a reaction. They write that "Understanding a complex chemical mechanism requires the exploration of a potential energy landscape", a map of how a molecule's energy changes as its atoms move into different shapes. That map rests on the Born–Oppenheimer approximation, which the authors describe as one that "assigns an electronic energy to every molecular structure." In plain terms, the nuclei are treated as fixed for each arrangement, the electrons' energy is worked out for that arrangement, and every possible shape of the molecule receives a single energy value. The authors call the accurate calculation of this energy "the pivotal challenge", and it is the problem they go on to take on with quantum computing. On this landscape, the structures that matter are found by optimization. Local minima correspond to stable intermediates, the points where a molecule can sit for a while. First-order saddle points correspond to transition structures, the highest point on the route between two minima. The text says that the energy difference between two minima connected through a transition structure enters the expressions for rate constants, by way of Eyring's absolute rate theory. The authors describe this theory as a detailed quantum- and statistical-mechanical interpretation of the older rate expression given by Arrhenius. They note that more information about the energy surface, along with dynamic and quantum effects, can be taken into account. They also state that the theory "is accurate even for large molecular systems such as enzymes." Finally, the rate constants for all the elementary steps feed a kinetic description that, in the authors' account, "ultimately provide a complete picture of the chemical mechanism under consideration."

The paper now turns to exact diagonalization methods in chemistry, and to why they fall short for the molecules it cares about. It starts with electron correlation, the way the electrons in a molecule respond to one another. The paper notes that if the frontier orbital region around the Fermi level is dense, as it is in pi-conjugated molecules and open-shell transition metal complexes, then "strong static electron correlation plays a decisive role already in the ground state." Frontier orbitals are the highest-energy occupied and lowest-energy empty orbitals, the ones most involved in chemical reactions and in absorbing light. Pi-conjugated molecules have chains of alternating single and double bonds that let electrons spread out, and open-shell transition metal complexes contain metal centers with unpaired electrons. The authors add that the effect is even stronger for electronically excited states, which matter in photophysical and photochemical processes such as light harvesting for clean energy applications.

The paper says such situations call for multiconfigurational methods, which describe a molecule as a mix of several electronic arrangements at once. It singles out the complete-active-space self-consistent-field approach, usually shortened to CASSCF, as "a well-defined model" that also forms the basis for more advanced approaches. In CASSCF, a small set of orbitals is chosen as the active space, usually from those near the Fermi energy. The authors admit that choosing this set is hard. They write that

This CAS selection is an art that, however, may be automatized.

The paper also separates two kinds of correlation. CAS-type methods handle static correlation well, but the rest, which it calls dynamic correlation, is what "is decisive for quantitative results." Dynamic correlation is the many small adjustments electrons make to avoid each other, and it is needed for numbers that match experiment. The authors count it as "a remaining major drawback" of exact diagonalization schemes that they must still account for the contribution of all the neglected virtual orbitals, meaning the unoccupied orbitals left outside the active space.

The scaling problem comes next. The paper says CASSCF is traditionally carried out by exact diagonalization, which limits it to about 18 electrons in 18 spatial orbitals, because of the "steep scaling of many-electron basis states" with the number of electrons and orbitals. A method called the density matrix renormalization group, or DMRG, scales polynomially and can reach about 100 spatial orbitals. The paper points out a cost, though. DMRG is an iterative procedure, and because it represents the electronic wave function as a matrix product state, a chain of matrices that compresses the wave function, its convergence for strongly correlated molecules is "neither easy to achieve nor is it guaranteed."

Finally, the paper says the ground-state energies from quantum computer simulations "are not affected by these issues," but they bring their own accuracy questions, which the authors quantify later in the paper.

The section turns from classical chemistry to the quantum computer itself, and it begins by setting out where the quantum approach would fit alongside existing methods. The authors note that while molecular structure optimizations cannot yet be carried out efficiently on a quantum computer, they can be handled with standard density functional theory, or DFT, an established classical method. DFT-optimized structures are described as generally reliable, even though the energies they produce can carry large errors that cannot be controlled. The energy problem could then be taken on by a quantum computer running a multi-configurational wave function model, specifically a CAS-CI calculation, which the authors say gives access to truly large active orbital spaces. In plain terms, this is a way of describing the electrons that matter for a reaction across a much wider set of orbitals than classical methods usually handle.

The authors add that such a calculation does not need carefully optimized orbitals, since natural orbitals can be taken from an unrestricted Hartree-Fock calculation or from a small CASSCF calculation. They also state that the missing dynamic correlation, the part of electron behaviour that these models leave out, still has to be added. They describe two routes for this, a perturb-then-diagonalize approach done before the quantum computation starts, and a diagonalize-then-perturb approach in which the quantum computer produces the higher-order reduced density matrices needed. The authors judge the first route to be considerably more advantageous, because no wave-function-derived quantities then need to be calculated. One option they mention is to treat dynamic correlation through DFT using range separation, which avoids double counting and has already been studied for CASSCF and DMRG.

The authors state that a flowchart in their figure lays out the steps of a quantum-computer-assisted exploration of chemical mechanisms. They also argue that quantum computer results could validate and improve parametrized approaches such as DFT, which would then serve for the large-scale pre-screening of structures and energies.

The section then sets out the method the authors focus on. Among several ways to find the energies of ground states on a quantum computer, they single out quantum phase estimation, or QPE. The idea rests on the time evolution of an energy eigenstate, which only multiplies it by a phase that depends on its energy. The task of QPE is to learn that phase, which the authors compare to the way a Mach-Zehnder interferometer reads out a difference in path lengths. If the starting trial state is a superposition of eigenvectors, QPE collapses it to one of them, and the authors say it will land on the ground state with high probability when the trial state overlaps strongly with the ground state. Preparing that trial state is a separate topic that they defer to an appendix.

In practice, the time-evolution operator has to be broken down into elementary operations. The authors note that several methods exist for this and say they focus on the Trotter-Suzuki decompositions, specifically the second-order Trotter-Suzuki formula. For the second-quantized Coulomb Hamiltonian, which describes the electrons and their interactions, the formula splits the evolution into a product of pieces, one for each term, arranged symmetrically in time. The authors state that the error of this approximation scales with the cube of the time step divided by the square of the number of steps, and that the approximation becomes exact in the limit of infinitely many steps. Each piece can be implemented with a known sequence of elementary gates and single-qubit rotations. The authors state that these rotations dominate the cost of the quantum simulation, which is why they matter so much to the resource estimates that follow.

Earlier estimates of the cost of quantum simulation treated the rotations in the Trotter formula as if they were basic operations that a quantum computer performs directly. The paper argues that this approach "belies the cost of performing the computation fault tolerantly." To get reliable results, the algorithm has to run fault tolerantly. That means each logical qubit, the unit of quantum information the algorithm uses, is encoded across many physical qubits with a quantum error-correcting code such as the surface code. The redundancy protects the logical qubit against decoherence, the loss of quantum information to the surrounding environment, and against other experimental imperfections.

The paper notes a limit on this protection:

Quantum error correction cannot directly protect any arbitrary quantum operation, such as arbitrary rotations.

What error correction can protect is a discrete set of gates, and any continuous operation can be approximated from those gates to within an arbitrarily small error. The approximation takes two steps. First, the exponentials in the Trotter formula are decomposed into single-qubit rotations and Clifford gates, which the surface code can carry out fault tolerantly. Second, the single-qubit rotations are approximated using Clifford operations plus at least one non-Clifford operation, usually the T gate, which is a rotation by pi over eight about the z-axis.

Each T gate requires a procedure called magic state distillation. It consumes many noisy quantum states and outputs a single accurate magic state, which is then used to teleport a T gate into the computation. The paper says that distillation, in space and time, makes it "by far the most costly aspect of quantum error correction," adding a large multiplicative overhead to every single-qubit rotation. For that reason,

The number of T gates therefore typically dominates the cost when implementing a fault tolerant algorithm.

The paper then turns to resource estimates. Its claim is that the resources needed to elucidate reaction mechanisms on quantum computers "not only scale polynomially, but also that the calculations are feasible on a relatively small quantum computer in reasonable time." These estimates are extrapolated from the costs found in simulations of smaller molecules, and the paper points to its appendix for loose upper bounds. The estimates focus on two prototypical structures of FeMoco, the iron-molybdenum cofactor at the active site of the nitrogenase enzyme, which nitrogen-fixing bacteria use to turn atmospheric nitrogen into a usable form.

The paper next works through what it calls the resource requirements with quantum error correction. Its runtime estimates start from a simple assumption: a quantum computer can perform a logical T gate every 10 nanoseconds, while Clifford gates take negligible time. A T gate is one of the operations that is hardest to build fault-tolerantly, and a logical qubit is an error-protected qubit built from many physical ones. The authors also assume that a good trial state for the ground state is already available, so the estimate covers the simulation itself rather than the search for a starting point.

The paper sets an accuracy target of a total error of at most 0.1 millihartree, adding up every source of systematic and statistical error. A millihartree is a thousandth of a hartree, the usual energy unit in chemistry. Three sources of error are named: the error in phase estimation, which is the procedure used to read out an energy from the quantum computer; the error from splitting the evolution of the molecule into Trotter-Suzuki steps; and the error from approximating those steps with the gates the machine actually provides. The authors then optimize over these three costs to find the most efficient way to divide the error budget among them. They also consider a looser target of 1 millihartree, which they describe as "a larger error that would still put our approach into the accuracy range of standard state-of-the-art quantum chemical methods" for simple transition metal complexes.

Three implementations of the algorithm are compared. In the first, called Serial, the single-qubit rotations must run one after another. In the second, called Nesting, the terms of the Hamiltonian that act on disjoint sets of spin orbitals run in parallel, which substantially cuts the runtime. In the third, programmable ancilla rotations, or PAR, rotations are stored in qubits and teleported into the circuit when they are needed.

The summary table gives, for two structures of FeMoco, the active site of the nitrogenase enzyme that fixes nitrogen from the air, the number of T gates, the number of Clifford gates, the runtime and the number of logical qubits. Structure 1 has spin state zero and a charge of plus three, with 54 electrons in 54 spatial orbitals. Structure 2 has spin one-half and charge zero, with 65 electrons in 57 spatial orbitals. For structure 1 at the 0.1 millihartree target, the serial approach needs about 1.1 times ten to the fifteenth T gates, runs for about 130 days and uses 111 logical qubits. Nesting needs about 3.5 times ten to the fifteenth T gates, runs in about 15 days and uses 135 logical qubits. PAR needs about 3.1 times ten to the sixteenth T gates, runs in about 110 hours and uses 1982 logical qubits. For structure 2 at the same target, the serial run takes about 240 days, nesting about 27 days and PAR about 204 hours, with logical qubit counts of 117, 142 and 2024.

The authors draw three conclusions from these figures. First,

If all gates are executed in series then we estimate that the simulation will complete in under a year and use a small number of logical qubits.

Second,

PAR can reduce the time required to several days at the price of requiring nearly 20 times as many logical qubits.

Third,

Nesting gives a reasonable tradeoff between the two and may be the preferred approach.

They conclude that "The costs at a logical level" are "within reach of a small scale quantum computer." For the looser 1 millihartree target, which the authors say is enough to learn qualitatively accurate reaction rates, the costs drop by roughly a factor of 10 in every case. The table carries its own caveat:

These runtimes and gate counts are likely to exceed the actual requirements.

The paper then adds the overheads needed to run the simulation fault-tolerantly, using the surface code, a widely studied error-correcting scheme. It calculates the costs for structure 1 under three physical error rates: ten to the minus three, a near-term figure; ten to the minus six, which the authors say might one day be feasible in superconducting qubits and is expected early on for topological qubits; and ten to the minus nine, which they say may be achievable for topological quantum computers. Rather than counting only physical qubits, the authors sketch a modular architecture. A classical supercomputer would interface with a quantum computer made of a main quantum processor, plus separate T factories that produce T gates and rotation factories that synthesize rotations. Only the main processor is general purpose; the factories are special-purpose hardware that the authors expect to be easier to build. The fault-tolerant tally begins with the number of logical qubits in that main processor.

The discussion turns from the numbers to what the authors think those numbers mean. The architecture they sketch has a classical control front end, a main quantum processor, and auxiliary units that build single-qubit rotations from pi-over-eight rotations made in T factories. The T factories are specialized units that produce the particular quantum states an error-corrected processor needs for some operations.

Their central claim is that a quantitative understanding of complex open-shell species such as FeMoco, the cofactor at the heart of biological nitrogen fixation, "remains beyond the capability of classical-computer simulations" today. The paper argues that a quantum computer used as an accelerator alongside a classical supercomputer could still help. In that role, the quantum computer would "obtain, validate, or correct the energies of intermediates and transition states and thus give accurate activation energies for various transitions." Activation energies are the energy barriers a reaction has to cross, and they largely determine how fast a proposed mechanism can run. The authors write that their work "shows that quantum computers used as accelerators to classical computers could be used to elucidate this mechanism using a manageable amount of memory and time."

To put the scale in context, they compare the requirements to Shor's factoring algorithm, a quantum method for breaking large numbers into their prime factors that is widely cited in discussions of encryption. The paper says the resources needed here are "comparable to that needed for Shor’s factoring algorithm for interesting 4096 bit numbers," both in the number of gates and in the number of physical qubits. From this they conclude that

the complexity of these simulations is thus typical of that required for other targets for quantum computing.

The estimates depend heavily on how error-prone the quantum gates are. For that reason the authors suggest that robust qubits, such as topologically protected qubits, "may be invaluable" for using a quantum computer to work out reaction mechanisms like the mode of action of nitrogenase.

On hardware, the paper says the serial calculation needs millions of physical qubits, which it calls "challenging but not out of reach for early devices." Most of those qubits sit in the T factories, and each factory needs fewer physical qubits than the main processor. The serial calculation would need about 202 such factories at a physical error rate of one in a thousand, and fewer with better qubits. Once the factory qubits are counted, the authors find that between one hundred thousand and one million physical qubits are enough for error rates of one in a million and one in a billion.

The parallel approach, called PAR, changes the picture. The number of T factories rises by roughly a factor of a thousand, while the qubits per factory stay about the same. The authors attribute this to PAR performing about a thousand times as many rotations as the serial code. A nested approach to parallelism is cheaper:

The costs required by the nesting approach, on the other hand, are much more modest.

It brings about an order of magnitude more qubits, with a comparable reduction in runtime.

At low error rates, the serial and nested totals are both of the order of a million qubits, which the paper says "could even be envisioned as a single quantum processor." The one-in-a-thousand error case and the PAR approach need hundreds of millions to nearly a trillion physical qubits, which the authors concede seems more daunting. Their answer is a modular design in which each factory uses about a million physical qubits. They judge this "certainly within the realm of possibility for future generations of quantum devices," and they add that the economic benefits of solving chemical catalysis problems of industrial importance "may offset the cost."

The authors also stress that nitrogen fixation is only one case. Reactions involving strongly correlated species, which are hard to describe with traditional multiconfiguration approaches that build the electronic state from many configurations, are common. The paper lists catalysts that break carbon-hydrogen bonds, catalysts for hydrogen and oxygen production, carbon dioxide fixation, and photochemical processes. They point to the economic and social weight of such chemistry, from fertilizer production to polymerization catalysis and clean energy. In their view, a versatile, reliable and fast quantum chemical approach would be valuable across these fields. The discussion closes on one requirement: "Parallelising the quantum computation of the energy landscape will be crucial" to getting answers quickly enough to be useful.

The appendix on the basics of quantum computing closes with resource tables for simulating nitrogenase's cofactor, FeMoco, and the caption explains what they mean. The authors assume physical gates running at 100 megahertz and an error target of 0.1 millihartree, which is a very small energy tolerance. They compare three ways of building the rotations the simulation needs, labelled serial, PAR and nested, at physical error rates of one in a thousand, one in a million and one in a billion. Across these cases the total physical qubit counts range from tens of thousands to several million. The caption concedes that the total "seems large", but defends it by saying the machine "is composed of many small and weakly coupled devices that can be mass produced", so the scale comes from manufacturing many modest units rather than one enormous chip. The same argument continues in the text that follows. The authors describe a time frame of several days rather than several years, and they conclude that quantum computers "must be designed such that the components can be mass produced and clusters of quantum computers can be built" so that the many molecular structures needed to estimate reaction rates accurately can be scanned.

The appendix opens with a brief review of quantum computing, written for readers who are not experts in the field. It sets out the basic concepts and notation, introduces the quantum algorithms used to simulate chemistry and the ideas of quantum error correction, and points readers to a longer textbook treatment for more detail. The authors note that these concepts return later, in their discussion of phase estimation.

They begin with the qubit, the quantum counterpart of the classical bit. A classical bit holds one value, zero or one. The paper describes a qubit as "a linear superposition of states": its state is a combination of the two basis states, zero and one, each weighted by a complex number called an amplitude. The squared sizes of the two amplitudes must add up to one. Those squared sizes set the probabilities: measuring the qubit yields zero with probability equal to the squared size of the zero amplitude, and one with probability equal to the squared size of the one amplitude.

The authors use Dirac notation, the bra-ket shorthand, throughout, because it captures the tensor product structure of multi-qubit states compactly. Their example is four qubits all in the zero state, which is the tensor product of four single-qubit zero states. A general state of n qubits is described by a list of 2 to the power n amplitudes, one for each basis state. The paper stresses that "the ability to represent a superposition over exponentially many states with only a linear number of qubits is one of the essential ingredients of a quantum algorithm", which it calls an innate massive parallelism.

Computation proceeds by unitary transformations, which carry one quantum state into another. There are infinitely many possible unitary transformations on n qubits, so a real machine cannot carry out every one exactly. Instead, gate-model quantum computers apply a discrete set of gates, each a unitary matrix, to approximate the continuous transformations they need. The paper calls a gate set universal if "any unitary transformation can be expressed, within arbitrarily small error, as a sequence of such gates."

The appendix then lists the gates used in the work. The X gate is the quantum version of the classical NOT, swapping zero and one. The Hadamard gate takes a basis state to an equal mix of zero and one. The Z gate flips the sign of the one state, and the T gate, which the paper calls the fourth root of Z, is used with Hadamards because the relation HZH = X lets Z and X be converted into each other. The Y gate swaps the basis states with phase factors, and the identity gate leaves a state unchanged. The two-qubit controlled-NOT, or CNOT, flips its second qubit when its first qubit is one, mapping the input x and y to x and x XOR y.

Measurement is the exception to unitary evolution. It collapses the state to the observed outcome and erases the information held in the amplitudes. The paper draws the consequence plainly: "extracting information from a quantum system irreversibly damages a state." For that reason the parallelism described earlier is limited in practice, since most computations must be repeated many times to build up the statistics needed to read out the result.

The paper next turns to how an abstract quantum operation is converted into the discrete gates a real machine can run. Its chosen gate set is the Clifford plus T library introduced earlier, and the authors note that "compilation into other gate sets is also possible," though they do not pursue those here.

The simplest case is a single qubit. Its operation is a two-by-two unitary matrix, which is a reversible transformation of that qubit's state. The paper uses an Euler angle decomposition, a standard way of writing any single-qubit operation as three successive rotations about two axes, up to a global phase with no physical effect. That reduces the whole problem to performing one kind of rotation, and the paper states the consequence directly:

Thus implementing single qubit rotations can be seen as an atomic operation on which compilation process for U is based.

Larger operations are built from these single-qubit pieces, interleaved with entangling gates such as the controlled-NOT, which links two qubits.

The paper then surveys the decomposition methods.

There are a host of methods known for decomposing rotations into Clifford and T gates.

The earliest is the Solovay\u2013Kitaev algorithm, which the paper describes as an efficient method based on Lie-algebraic techniques, a branch of mathematics about continuous symmetries. More recent methods rest on number theory. The paper says they require far fewer operations than Solovay\u2013Kitaev, and that they are "useful, if not necessary, for reducing the gate counts of the simulation algorithm to a palatable level."

The paper gives cost figures for approximating a rotation to within a chosen error, written here as epsilon. Without extra helper qubits, the known bounds for the worst possible input angle grow with the logarithm of one over epsilon, with a coefficient of four, plus or minus fixed constants. Later work gives an average case of roughly three quarters of that worst-case bound.

A further class of methods uses an ancilla, which is an extra helper qubit, to cut the number of T gates needed. The paper estimates that this approach brings the average T-gate cost to about 1.15 times that logarithm, plus about 9.2. It also mentions repeat-until-success synthesis with fallback, which is asymptotically better, but the paper argues that it does not outperform the ancilla method in the range of error the simulation needs to reach chemical accuracy, the precision at which predictions about a chemical reaction become useful.

The paper also notes that PAR rotations and gearbox circuits can substantially reduce the T-depth, the number of sequential layers of T gates, at the price of running more operations in parallel. It sets gearbox synthesis aside and examines the costs and benefits of PAR rotations.

Parallel execution creates a problem for the cost estimates. When several rotations run at once, the total cost is set by the longest gate sequence in each block. Repeat-until-success synthesis is probabilistic and cuts the expected cost by roughly a factor of four from the worst case. But the paper argues that "it is clear that when performing many of them in parallel it is very likely that at least one of them will saturate the bound." For this reason, the authors say, "we use upper bounds on deterministic synthesis," which have better worst-case scaling, so their resource estimates rest on the worst case rather than the average.

The authors say their approach to quantum chemistry "closely follows the strategy proposed" in an earlier study. They start by writing the molecule's energy function, called the Hamiltonian, in second quantized form. This keeps track of where the electrons are by recording which slots in a discrete set of spin-orbitals are occupied. A spin-orbital is a basis state that an electron can occupy. Because a spin-orbital can hold only one electron, the paper notes that such states are "very natural to express in a quantum computer." Each spin-orbital gets its own qubit, with the state one meaning the orbital is occupied and the state zero meaning it is empty.

The paper describes the electron-adding and electron-removing steps as creation and annihilation operators. These must respect the symmetries that electrons require, which are the symmetries appropriate for Fermions, particles such as electrons that cannot share a state. The authors call "the most important property" the anti-commutation rule. It is tempting to map these operators directly onto the Pauli operators X and Y, which are the basic single-qubit building blocks of quantum circuits. The paper warns that such a mapping "does not satisfy the anti–commutation relation for Fermionic operators." Instead, the operators are converted into Pauli operators using known encodings called the Jordan–Wigner or Bravyi–Kitaev transformations.

To simulate how the system changes over time, the computer must carry out the evolution described by the Hamiltonian as a sequence of gate operations. The Hamiltonian has two kinds of terms: one for electrons hopping between orbitals and one for electrons interacting with each other. Once those terms are translated into Pauli operators, the evolution can be broken into operations a quantum computer could perform. The paper points out that well-known circuits exist for this, built from the Trotter–Suzuki formulas, which approximate the continuous evolution by slicing time into many small steps.

The most direct method for estimating the ground-state energy, which is the lowest energy the molecule can have, is phase estimation. It uses quantum interference between controlled runs of the time evolution at different time steps to infer the eigenvalues, the characteristic energy values, of that evolution. The paper notes that if the time step is smaller than pi divided by the size of the Hamiltonian, the method also gives the eigenvalues of the Hamiltonian directly. The authors highlight a cost advantage: phase estimation needs a number of applications that scales with one over the target error to learn the energy within that error with high probability. In their words, this "is quadratically better than bounds on the variance" that optimal classical sampling methods using an unbiased estimator would produce.

Phase estimation only works well if it starts from a state with large overlap with the ground state, meaning the starting state closely resembles the lowest-energy state. The paper explains that if the starting state is a mix of the ground state and a state orthogonal to it, phase estimation returns the ground-state energy with a probability equal to the squared size of the ground-state part. The simplest starting point is the Hartree–Fock state, an approximate description in which electrons fill the lowest available orbitals. Preparing it needs only NOT gates, which flip qubits from zero to one. For systems with strong correlations, however, the authors note that configuration interaction with high excitations, which mixes in configurations where electrons occupy other orbitals, may be needed to reach high overlap.

The authors do not estimate the cost of preparing such a state in this work. They argue that the cost of preparing an accurate approximation to the ground state of FeMoco, the active site of nitrogenase, "is difficult to determine in absentia of large scale quantum computers." A later part of the appendix returns to this question in its discussion of state preparation.

The paper turns to quantum error correction, the set of techniques that protects a quantum computation from the errors its hardware makes. The authors start with a blunt comparison:

Quantum hardware is far less robust to errors as classical hardware.

The physical parts of the machine, the qubits themselves, the gates that act on them, and the measurements that read them out, all err at some rate. Error correction is meant to reduce those errors without giving up the quantum behavior that makes the computer useful. The paper explains that these codes only work when the physical error rates stay below a threshold. That threshold depends on the particular code and on the type of noise the hardware suffers. Below the threshold, the errors in logical gates and logical qubits, meaning the protected versions of the computation's operations and bits, can be made arbitrarily small, at a cost that grows only polylogarithmically, which is to say slowly, with the size of the problem. The authors single out the surface code as the most popular option:

The surface code is currently the most popular code due in part to its relatively high threshold of 1%.

They note that enthusiasm has grown because of evidence that existing superconducting quantum computers may already have error rates near that threshold. In the authors' words, this "raises the hope that a fault-tolerant, scalable quantum computer may be just over the horizon." They then add a caution. Error correction can in principle support arbitrarily long computations on a noisy device, but "the resources required to execute such a fault-tolerant computation can be large if the system operates close to the threshold." The authors begin to locate where most of that cost falls.

A large share of the cost of quantum error correction comes from one need: performing a universal set of quantum gates. The paper explains that the two kinds of gate differ sharply in price. Clifford gates, a family of simpler operations, need very little overhead in the surface code, the error-correcting scheme the authors use. A non-Clifford gate, such as a T gate or a Toffoli gate, is another matter, and the paper states that "protecting a non-Clifford gate, such as a T or Toffoli gate, requires substantial resources."

Several techniques for building fault-tolerant non-Clifford gates exist, the authors note, but they focus on one: magic state distillation, used together with the surface code. Distillation is a way of producing a clean resource state from dirty ones. It takes a set of noisy input states and outputs a cleaner one. Because the surface code cannot implement the T state directly, the T state has to be distilled before it can be used.

The authors consider the fifteen-to-one distillation scheme of Bravyi and Kitaev. In it, fifteen noisy input states, each with error rate p, produce a single magic resource state. The paper gives the output error rate as roughly thirty-five times the cube of the input rate, so the output is far cleaner than the inputs, provided the inputs are already fairly good. The paper adds that the distillation step itself is cheap to run on the surface code:

Magic state distillation consists only of Clifford operations, which can be implemented easily within the surface code.

The appendix then turns to phase estimation. The authors explain why this matters for their later cost figures: phase estimation is "the outer most loop of the quantum simulation" and so is "a major driver of the cost of the simulation." How it is implemented therefore sets much of the total resource bill for simulating the nitrogenase active site, FeMoco, which the rest of the paper estimates. The material that follows covers phase estimation itself and then controlled rotations.

The appendix on phase estimation begins by explaining why this step matters so much. Phase estimation is one of the most critical components of the quantum simulation algorithm. Without it, the simulation time needed to estimate the ground-state energy would grow quadratically with the required precision. Because the target precision is on the order of 0.1 millihartree, the phase for one time step of the evolution operator must be estimated to within 0.1 divided by r millihartree, where r is the number of Trotter steps required. The authors note that if statistical sampling were used instead of phase estimation, on the order of 10 to the 12th experiments would be needed to shrink the variance of the estimate far enough. That would be impractical, which is why the authors call phase estimation "crucial for most large scale applications in quantum chemistry."

The appendix then describes the standard quantum phase estimation algorithm, which it attributes to a prior reference. According to the text, that algorithm can learn an eigenvalue within error epsilon with probability at least one half, using a number of applications of the unitary circuit bounded above by a term of order 16 pi divided by epsilon. The authors call this "far from optimal" in its time scaling, which they put at roughly pi divided by epsilon. The standard algorithm has the advantage of needing neither measurements of the quantum system nor a classical computer to infer the most likely eigenvalue. The authors say they use this result because it gives an upper bound on the cost of the optimal phase estimation algorithm.

The text then turns to iterative phase estimation, which it presents as an alternative. Instead of storing the phase in a quantum register, iterative phase estimation uses a classical inference algorithm to learn the eigenphase from measurements of an auxiliary probe qubit. That qubit is measured, re-entangled with the system, and measured again over successive rounds. The text credits Kitaev with proposing the first variant of this approach, and a circuit for the process appears in the paper's figure four, which is not reproduced here.

The appendix gives the ultimate limit on this kind of estimation, in terms of how many times the unitary must be applied for a given error tolerance, as approximately pi divided by epsilon. It also notes that Gaussian strategies, such as one called RFPE, produce Bayesian Cramer-Rao bounds that come close to saturating that limit, at about 3.3 divided by epsilon. The authors argue that these bounds are often saturated for likelihood functions of this form. They also point out that the Gaussian assumption makes the user discard substantial information from the higher moments of the posterior distribution. On that basis they conclude that the limit of pi over epsilon is the ultimate limit for phase estimation. They expect that optimized measurement policies from previous studies will come close to reaching it.

For their own estimates, the authors use a surrogate cost of pi divided by two epsilon as the expected performance of these optimized approaches. The factor of two, they explain, comes from an optimization that applies to quantum simulation algorithms, which the appendix goes on to discuss in detail.

Phase estimation needs more than the single-qubit rotation covered so far. It needs controlled Z-rotations, in which a rotation is applied only when a control qubit is in a chosen state. The paper notes that well-known methods exist for this, built from a pair of rotations and two CNOT gates, where a CNOT is a controlled-NOT operation that flips a target qubit when the control is set. The authors then offer a better approach, "previously shown in unpublished work by Tsuyoshi Ito in 2012", and prove in a lemma that its circuit does what it claims. The lemma states that the circuit implements a controlled rotation by twice the angle, with the top qubit acting as the control. The proof checks the two cases in which the control is zero or one, tracks how the rotations and the CNOT gate shift the state, and extends the result to any input by linearity.

The authors then explain why this form is still not the best one for their purpose. They write that it is "not optimal here because we can implement a variant of a controlled rotation gate to in effect double the impact that the eigenphase has on the measurement probabilities." The eigenphase is the number the algorithm is trying to learn, and the variant lets each rotation push the measurement outcomes further, so fewer rotations are needed. A second lemma describes a protocol that uses a given number of rotations and then measures a single ancilla qubit, a helper qubit used only for the readout. The probability that the ancilla reads zero follows a cosine term whose argument is twice the number of rotations multiplied by the difference between the unknown phase and the chosen angle. The proof says it is built on earlier informal arguments. It replaces each controlled rotation in the standard phase estimation circuit with a controlled operation that applies the evolution on one branch and its inverse on the other. The inverse is obtained by flipping the signs of the rotation angles.

The numerical experiments for the RFPE method, a robust phase estimation approach, show that the eigenphase can be learned to a target error with probability one half. The authors state the result that matters for cost: the number of rotations needed to perform phase estimation is a quarter of the value expected if the earlier circuits were used, or half the cost in settings where operations run in parallel. They also compare the new bounds with older ones. The ultimate lower limit from previous studies would become pi over two, measured against the target precision. The Bayesian Cramer-Rao bound gives a lower limit of about 1.6 over the precision, under the Gaussian priors assumed in RFPE. The upper bounds on traditional phase estimation from an earlier study would fall from sixteen pi to eight pi over the precision if these circuits were used.

The paper then turns to the Trotter-Suzuki formulas, which approximate the evolution under the molecule's Hamiltonian, the operator describing its energy, by a sequence of simpler evolutions. The inversion trick described above depends on the formula being run backward by reversing the sign of the time step. The authors state that this assumption holds for the (2k+1)th order formulas for all k of at least one, but that it "does not apply to the second order formula unless further assumptions are made." The reason is that the forward and backward products of the second-order formula differ at second order in the time step, so the inversion is not exact. For the real-valued Hamiltonians used in quantum chemistry, and when only the ground-state energy is wanted, the error from flipping the signs is of order the time step cubed. The authors call this "by no means fatal", but it could add to the Trotter error. For a third or higher order formula, they write, "no such danger exists." Together with the stronger proven error bounds for the third-order formula, this is the reason the paper prefers it over the second-order formula, which has the same asymptotic cost.

Appendix C turns to a technical question: how the approximations used in quantum simulation distort the energies the algorithm reports. The main approximation is the Trotter formula. The authors state the core problem plainly:

A major source of error in most quantum simulation algorithms arises from the use of Trotter formula expansions.

These errors can be made as small as one likes, but the authors note that pushing them below chemical accuracy, the precision chemists treat as the target, can substantially change how long a simulation takes.

The underlying cause is that a molecule's Hamiltonian, the operator that gives its energy, is made of many pieces that do not commute. Commuting means the order of two operations does not matter. Because the pieces do not commute, evolving the system under each piece in turn is not exactly the same as evolving it under the whole. In the paper's words,

Trotter errors arise from the fact that the terms in the Hamiltonian used in the expansion do not commute.

The authors first consider the Zassenhaus formula, which expands the exact evolution in terms of nested commutators. They say that in principle it gives everything needed to understand how these errors scale. To leading order, the error in the second-order splitting, known as the Strang splitting, can be estimated by the ground-state expectation value of a double commutator sum. The ground state is the lowest-energy state, and the expectation value is the average result one would measure in it. The authors point out that the number of terms grows roughly as the fourth power of system size. From this, the error in the Trotter formula can grow at most roughly as the tenth power of system size, multiplied by the square of the simulation time. It does not reach the twelfth power, because the commutator structure restricts which two orbitals the two outer terms can act on. Even though this estimate can be computed in polynomial time, the authors say there are too many terms to estimate it reliably for molecules on the scale of nitrogenase, even with Monte Carlo sampling.

A second route uses the Baker-Campbell-Hausdorff formula, which the authors describe as the dual of the Zassenhaus formula and which is suited to estimating errors in eigenvalues, the energy levels the algorithm is actually after. Here the Hamiltonian is split into its terms, and the Strang splitting applies each term's evolution for half a time step, runs them in reverse order, and so produces the evolution under an effective Hamiltonian. The gap between that effective Hamiltonian and the true one sets the error.

The authors then explain why they do not rely on a crude bound. The triangle inequality applied to the double commutator sum gives an upper bound on the asymptotic scaling, but it is not useful for their purposes, because "we do not know apriori how large t must be for the leading order term to be dominant." Instead they use a bound that they state is provably an upper bound on the error in the energy of the evolution. It is a weighted sum over triples of terms, scaled by the sizes of the terms, and it includes an indicator that equals one only when the corresponding double commutator is nonzero. The authors say this is slightly tighter than a bound used in earlier work.

To keep the sum manageable, the authors list rules under which a double commutator is known to vanish. First, if two of the terms act on disjoint sets of qubits, the commutator is zero. Second, a term acting on a set of qubits disjoint from the set acted on by the other two also gives zero. Third, pairs of terms of certain types, the authors name them as PP or PQQP terms, vanish, and terms of the PR and PQQR form with the same P and R likewise drop out. Fourth, by the Jacobi identity, only one of three related nested commutators can be nonzero. The authors acknowledge that other symmetries could eliminate still more terms, but they do not use them. Their estimate therefore overcounts slightly, and it remains an upper bound.

Even with these rules, the bound is still computationally heavy, because the double commutator sum contains on the order of N to the tenth power terms. The authors note that the Cauchy-Schwarz inequality can reduce the work to on the order of N to the fourth power operations. The cost is accuracy: the inequality can overrepresent the influence of large terms in the Hamiltonian, so the resulting bound is looser.

The authors turn next to how large the Trotter error really is in practice. Computing the matrix elements of the error operator is computationally challenging, so instead they rely on Monte Carlo sampling to estimate the upper bound. The paper explains why sampling works better than summing the series directly:

Monte Carlo sampling is much more effective here because the use of the triangle inequality removes the alternating sign that we see when summing the original series.

In practice, the sampler draws random combinations of Hamiltonian terms. It throws out any combination that contributes nothing, and for each remaining one it multiplies the sizes of the three terms involved. Averaging these products gives an error estimate that grows with the square of the time step and with the cube of the number of Hamiltonian terms. The authors show that, for a fixed target accuracy, the time step can be chosen as the square root of the target divided by a constant that depends on the molecule. The variance of the estimate shrinks as the number of samples grows. Using Chebyshev's inequality, the authors conclude that the sampled error falls below the target with probability greater than 75 percent once the sample count is large enough.

The paper also points out a weakness of plain uniform sampling. One-body terms in the Hamiltonian tend to be much larger than two-body terms, but the two-body terms are far more numerous. As the paper puts it, "uniform sampling can underestimate the contributions of such terms because of their relative scarcity in the sample space." To correct this, the authors sample separately within each class of nested commutators, drawing each type of term uniformly, and then add the class estimates together. They use at least a hundred million samples per class, which they say keeps the sample standard deviation under one percent.

The results are compared with the Trotter number, meaning the number of slices the simulated time evolution is split into. More slices mean smaller steps and a deeper circuit. The comparison covers small molecules, ranging from simple hydrides and diatomic molecules up to FeMoco, the iron-molybdenum cofactor, which has more than a hundred spin orbitals in the largest cases. The Trotter number needed to reach an accuracy of 0.1 millihartree is computed assuming no other sources of error. The rough scaling of the upper bound is about the number of spin orbitals raised to the power 2.5, a figure that earlier numerical studies had also noted. The authors add that "this scaling should not be seen as definitive," because the molecules differ widely in their chemistry and the data scatter accordingly.

The upper bounds turn out to be far too loose for small molecules, by roughly a factor of ten thousand. For that reason the paper offers three extrapolations of the scaling. The pessimistic one rescales the upper bound so that all the data lie beneath the curve. The middle one rescales the upper bound by the average gap between the bound and the computed examples. The optimistic one is a polynomial fit to the computed numbers alone, ignoring the bound. The authors say they "expect the middle curve to be the most realistic estimates," and they carry all three curves, as well as the raw upper bound, into their resource estimates.

The paper then proves a set of basic results to show how errors in the simulation carry through to the energy being computed. It cautions that these proofs are likely to give "substantial overestimates of the error," because they do not account for the natural cancellations that tend to occur in practical eigenvalue estimation.

The first result says that if two Hermitian operators differ by at most some small amount, their lowest eigenvalues, which are the ground-state energies, differ by at most that same amount. The proof relies on the variational property, under which the lowest energy of one operator is no larger than the expectation value of that operator in the ground state of the other.

The second result concerns the time evolutions generated by two similar Hamiltonians. If the difference between the two evolutions is bounded by the time multiplied by some non-decreasing function of time, then the Hamiltonians themselves differ by at most that function's value. The authors note that this bound is tight as time goes to zero, because the error in that limit is exactly the Hamiltonian difference multiplied by time.

The main theorem applies these results to a Hamiltonian written as a sum of M terms, approximated by a second-order Trotter product formula, also known as Strang splitting. If each term of the approximating sum differs from the true term by at most a small amount delta, and the product formula has its own error, then the difference in ground-state energies between the exact and effective Hamiltonians is at most the product formula's error plus the number of terms minus one, times delta. The authors say these bounds hold for generic quantum simulation, though they apply particularly well to quantum chemistry.

A final lemma shows that any unitary operation that approximates the exact time evolution to within a given accuracy corresponds to some Hermitian operator close to the original Hamiltonian. The proof takes the principal matrix logarithm of the unitary. The paper then connects this to circuit synthesis. Each single-qubit rotation in the circuit carries a small error, and because there is one rotation for each Hamiltonian term, the error in each term of the effective Hamiltonian is at most delta. The total energy shift from circuit synthesis is therefore at most M times delta. The authors stress that this error arises from approximating the single-qubit rotations even when there is no decoherence.

The appendix on cost estimates for nitrogenase opens with an error result that the later estimates depend on. The paper takes a molecule's energy operator, the Hamiltonian, written as a weighted sum of M simpler terms called Pauli operators. Each short evolution step is approximated by a circuit built from gates. The corollary states that the error in the computed ground-state energy is at most the Trotter error plus (2M minus 1) times the synthesis error, divided by the timestep. The Trotter decomposition approximates continuous evolution by many short steps, and circuit synthesis is the process of building a gate sequence that carries out each small rotation. The authors draw a practical consequence:

This shows that it suffices to choose synthesis error that shrinks linearly with the timestep used in the Trotter decomposition.

They note that synthesis cost grows only logarithmically as that error shrinks, so the extra precision is affordable.

The appendix then sets out the basic accounting. It assumes the user can prepare an exact copy of the ground state at negligible cost, so the cost lies in the simulation itself. Two factors drive it. The first is the cost of the Trotter circuit. The second is how many times that circuit must be repeated inside phase estimation, the algorithm that reads out an energy from repeated runs. The authors answer an obvious objection, that the number of time steps also drives cost. They grant that the Trotter decomposition is a major driver, but say it enters only indirectly, through phase estimation. Phase estimation learns the eigenphases of a single Trotter step. Shorter steps give a smaller Trotter error, but they require more repetitions, and because phase estimation cost scales inversely with the desired uncertainty, the total cost scales inversely with the time step. In the paper's words,

Thus the cost of the simulation can be thought of as arising from only two sources, the cost of each depending on the error tolerances allowed for all three contributions to the error.

The three error contributions are the phase estimation error, the Trotter error, and the synthesis error. By the triangle inequality, the total error in the ground-state energy stays at or below a tenth of a millihartree, a hartree being the standard atomic unit of energy, when the three contributions add to less than ten to the minus four hartree. The authors call the tenth-of-a-millihartree level appropriate for quantitative calculations of reaction rates. Cost is counted in T gates, the gates that are expensive to perform on a fault-tolerant quantum computer, and the paper gives a formula for that count. It depends on the phase estimation scaling constant, the number of Trotter steps needed to reach the target accuracy, and two constants of the synthesis method. The exact best split of the error budget is awkward to find, because of a logarithm in the formula. So the authors optimize a simplified version that calculus solves directly. They argue that its value at those settings is an upper bound on the true minimum, so the estimates stay upper bounds, provided the assumptions about Trotter error hold.

The inputs come from earlier circuits for the molecule. The number of terms is 6.1 million for the nitrogenase active site in a 54-orbital basis, and 8.2 million in a 57-orbital basis. These correspond to the two structures, which need 108 and 114 spin orbitals. The paper then gives four cases. The rigorous case uses only proven upper bounds on cost, and the authors say these "lead to estimates that are clearly extremely pessimistic." They add that even with optimistic assumptions about the target computer, the worst-case bounds "suggest between millions and tens of thousands of years depending on whether parallelism is used." The pessimistic case uses empirical scalings for synthesis and phase estimation, with the worst scaling the Trotter bounds support, rescaled by the ratio the authors observed between actual and predicted Trotter numbers. The rescaled case divides the rigorous bound by the average ratio between exact Trotter numbers and their bounds for molecules small enough to compute exactly. The authors write,

We suspect these rescaled estimates may provide the most realistic estimate of the Trotter number required to simulate nitrogenase.

The optimistic case extrapolates the average scaling observed across those small molecules, scaled so that all the data sits beneath the curve. The authors caution that this case is optimistic, since "there is little evidence for a clear trend in the empirical data," and they say the range of molecules is "insufficient to meaningfully extrapolate out to 108 spin orbitals" or more. They offer it as the best scaling the data could reasonably support.

The numbers in the tables show the spread. For the first structure, the pessimistic case needs about 8 times 10 to the 15th T gates, around 30 months on a machine running 100 million T gates per second. With parallel factories it falls to about 31 days. The rescaled case needs about 2 times 10 to the 15th T gates and about 250 days, or 8.8 days with parallelism. The optimistic case needs about 1.6 times 10 to the 14th T gates and about 19 days, or 17 hours with parallelism. The authors say the optimization yields "modest reductions in the costs relative to the worst case bounds," and the tables give those reductions for both structures.

The error budget for the simulation is split across three sources. The authors require the total error to stay within 0.1 millihartree, a thousandth of a hartree, the energy unit chemists use. The three sources are the systematic error in the Trotter-Suzuki decomposition, which approximates the molecule's energy operator by a sequence of simpler steps; the statistical error tolerance for phase estimation, the algorithm used to read an energy value out of the quantum computer; and the statistical error in synthesizing rotations from Clifford and T gates. Clifford gates are a restricted family that error-corrected machines can perform cheaply, while T gates are the costly ones, so counting them is a standard way of measuring the cost of a simulation.

The paper says the three uncertainties are "chosen such that the number of T gates required for the simulation is minimized given the target accuracy." The earlier error analysis can be reused inside this budget, the authors note, provided the errors in phase estimation and synthesis are not adversarial, meaning they do not all line up in the worst possible direction. The main analysis took this approach, optimizing the three dominant costs against one another to reach the 0.1 millihartree target with the fewest resources.

The second topic is parallel circuits, which try to run several rotations at once to shorten the computation. The first approach the authors describe, called nesting, groups terms in the Hamiltonian that commute with each other so they can run simultaneously. In principle this can cut the circuit depth substantially, meaning the number of sequential steps. The authors say that in practice it is "difficult to assess the performance of these schemes here because of the size of the molecule," and because they keep the terms in their lexicographic order. To estimate what parallelization could do for this calculation, they need a method that reduces the T-depth without changing the order of terms in the Trotter-Suzuki decomposition.

The PAR method, which the authors draw from earlier work, is offered as that method. It works by teleporting a rotation into a state with probability one half, using only Clifford operations and a preparation step. Teleportation here means moving the effect of a gate into a different quantum state through entanglement and measurement, and the one-half success probability means the procedure is probabilistic rather than guaranteed.

The correction for a failed attempt is where the scheme gets clever. If an attempt fails, the circuit applies the rotation with the wrong sign, and that error can be repaired by teleporting in a rotation of twice the angle. If that correction also fails, the authors fix it with a rotation of four times the angle, and so on up the chain. Each step fails about half the time, so the number of pre-cached qubits needed for a single rotation follows a geometric distribution. The rotation angles are known in advance, so these states can be prepared offline and in parallel. The authors conclude that "a logarithmic multiplicative overhead in space is needed to guarantee that enough ancilla qubits are prepared to perform the rotations with such high probability that it is unlikely that the cache of ancillae will ever be depleted."

To bound how many ancillae are needed to parallelize a block of M rotations with high probability, the authors describe a protocol. The terms of the Trotter expansion, which is the standard way of splitting a simulated molecular energy operator into a sequence of simpler steps, are grouped into blocks of M consecutive terms. For each rotation angle in a block, and for each power of two up to a fixed value n, the states are prepared in parallel. Each rotation is then attempted with programmable ancilla rotations, up to n times. If all n attempts fail, that counts as a failure. The protocol then applies the correct rotation and moves on to the next pre-cached rotation.

The second theorem puts a number on this protocol. It says that the expected number of rotations performed before a failure occurs is 2 to the n, multiplied by one minus the quantity one minus 2 to the minus n, raised to the power M. The proof rests on the fact that each attempt succeeds with probability one half and that the success of each programmable circuit is independent of the others. The chance of a failure across a cache is 2 to the minus n, so the chance of getting through all M rotations without one is one minus 2 to the minus n, raised to the power M. Because the cache holds only M rotations, any run of successes longer than M yields only M rotations, and the mean accounts for that cap.

The practical lesson is that caching corrections for n failures can cut the T-depth, meaning the number of sequential layers of costly T gates, by that same factor, without adding a prohibitive number of rotations. As n grows, the expression approaches M, which is the most the cache can supply. The authors note the trade-off directly: "this comes at the price of requiring more T gates and hence increases the overheads of quantum error correction."

The authors then turn to how these rotations should be costed. A major challenge is that, in a fault-tolerant machine, all the non-Clifford operations can be prepared at once and offline. Clifford gates are the easier family that error correction handles cheaply, while non-Clifford gates such as T gates are expensive. If only T gates are counted, the authors say, "we come to the absurd conclusion that the costs of all quantum simulation algorithms can be reduced to that of synthesizing a single rotation." To prevent that, they assume a delay equal to one T gate is included for the measurement and feed-forward step that the programmable rotation needs. Feed-forward means using the results of a measurement to choose the operations that follow. This fixed cost means that even if every T gate is precached, the rest of the simulation never takes zero time.

With that assumption in place, the authors follow an alternative to the large offline cache, both in these cost estimates and in the main body of the paper. As they put it,

rather than constructing a large cache of rotations offline, it makes sense to produce the rotations just in time.

If synthesizing one rotation costs C T gates, the scheme keeps n times C factories running constantly, each producing rotation states of the form used by the protocol. The factories are staggered so that one set finishes its n states at least one cycle before the next rotation is needed, and a set that finishes moves on to the rotations not already in production. If a failure occurs, the factory approach halts, as the traditional approach does, and waits for a rotation to be synthesized that applies the correct rotation online.

To begin the analysis, the authors imagine the programmable circuit succeeding C consecutive times. Each attempt takes the time of one T gate, and synthesis costs C T gates, so the timing of the first set of factories is worked out against that run of attempts.

The factories are staggered, so each set is expected to have finished producing its states before the last set applies its own. The paper argues that with C such factories, rotation states can keep being generated even in what it calls the "worst case scenario", where each PAR circuit succeeds on the first attempt. The authors write:

This is why in this setting it does not make sense to use more than nC rotation factories given the assumption that the cost of each PAR attempt is 1 T gate and that the rotations for at most n failures are to be pre-cached.

The paper then states a theorem about the average time needed per rotation under the factory-based PAR strategy. The assumptions are that each PAR application takes at most one T gate, the costly operation in fault-tolerant quantum computing, and that all remaining Clifford operations, the comparatively cheap gates, are free. The result is an expression in the number of factories and the number of pre-cached failures.

The proof starts from the fact that each PAR attempt succeeds with probability one half. The expected number of T gates applied before a solution is found is therefore a finite sum, which the proof reduces to a geometric series. The second term covers failure. A failure occurs with probability one over two to the n, makes n PAR attempts, and incurs a cost of C T gates to apply the fallback rotation. Summing the series gives the result. If the costs of feed forwarding, the bookkeeping that routes measurement results to the corrections that depend on them, are neglected, the online cost of a PAR rotation is simply C over two to the n T gates.

The paper then returns to nesting, which runs several terms of the simulation at once. It says that in both the serial and PAR cases, the costs come from building the circuit for the TS formula and counting its gates. The nesting figures in the main body are, in the paper's words, "upper bounds based on empirical estimates of the number of terms that can be simultaneously executed." To estimate them, the authors take FeMoco, the iron-molybdenum active site of the nitrogenase enzyme, and greedily group terms that act on distinct sets of spin orbitals. They report that one grouping can run 26.43 terms at once for one structure and 27.83 for the second. They say these values roughly match the optimal scaling of N over four that the strategy can reach.

The time is then found by dividing the T count by these factors, on the assumption that the quantum computer can run each commuting group at once. For the Clifford gates, though, the paper does not execute the grouping. It uses a simple upper bound instead, because the authors write:

We do this because the grouping strategy breaks the lexicographic ordering that leads to dramatic cancelation of the CNOT strings that arise from the Jordan–Wigner decomposition.

The Jordan–Wigner decomposition is a standard way of mapping electrons onto qubits. For the nested data, the number of timesteps is taken to be the same as in the other cases, which the paper considers reasonable, since operator ordering tends not to have a dramatic impact on the error. As the paper puts it,

Subsequent work will investigate the precise interplay that operator ordering has in nesting.

The appendix then turns to state preparation, which it describes as "a major unanswered question that impacts the cost of quantum simulations." The paper says it does not discuss those costs in detail in the rest of the paper. Instead it looks at the problems that arise with elementary state preparation methods built on coupled cluster, configuration interaction, or Hartree–Fock states, which are standard approximations of a molecule's electronic state. It also examines adiabatic state preparation, where a system is slowly moved from an easy starting problem to the target problem so that it stays in its lowest-energy state. The appendix promises numerical results on how the overlap between the ground state and Hartree–Fock states scales for small molecules, along with a discussion of the costs of adiabatic state preparation.

The paper turns next to simpler starting states, the trial wavefunctions a quantum computer would prepare before running a phase estimation. It begins with a caveat: "we cannot rigorously prove that elementary state preparation methods ... will suffice for preparing a state with large overlap with the ground state." The candidates are Hartree–Fock states, unitary coupled cluster states, and truncated configuration interaction states such as CISD and difference dedicated configuration interaction, known as DDCI. Overlap here means how much of the prepared state lies along the true ground state, expressed as a percentage, so a higher number means the simple starting point is closer to the answer. The authors want to know how well these simple ansatzes perform on molecules small enough to check numerically.

For small molecules, the paper reports substantial overlap between the Hartree–Fock state and the true ground state, with the smallest overlap in its set at 89%. It notes that other studies of chains of hydrogen atoms stretched near the point where they come apart have found very small overlaps with Hartree–Fock states. From this, the authors conclude that small molecules can be constructed that are not well described by such ansatzes. They find only weak evidence for a downward trend in the overlap as the system grows. Roughly half of their data points fit a curve of about 107.75 percent times e to the power of minus 0.0076 times the number of spin orbitals, which are the single-electron slots in the calculation. Extrapolated to a molecule the scale of FeMoco, the nitrogenase cofactor discussed later, that curve gives an overlap of roughly 43 percent. The paper says it cannot yet tell whether that molecule is typical of the set it studied. To set up a route beyond these starting states, the paper also describes a Hamiltonian that changes over a dimensionless time from 0 to 1. Its ground state at the start is the Hartree–Fock state, and at the end it is the full configuration interaction ground state. The paper states that for Hamiltonians that are smooth enough, a sufficiently slow evolution will carry the Hartree–Fock state into the full configuration interaction ground state.

The paper expects stronger correlations in molecules containing atoms with d-electrons than in the ones it examined. Those molecules are often beyond classical simulation, so finding a suitable benchmark set remains an open problem. The authors point to the success of a method called CI dynamically extended active space in density matrix renormalization group calculations as a sign that truncated configuration interaction states may still suffice for quantum simulation. On scaling with the largest nuclear charge among a molecule's atoms, they find no compelling evidence in this set. Beryllium hydride has better fidelity with the ground state than beryllium alone does, and hydrogen fluoride and hydrogen chloride have nearly identical overlaps even though chlorine has nearly twice the nuclear charge of fluorine. The paper concludes that

more study is needed in order to understand how these overlaps scale for large molecules.

It also states that it is strongly suspected that multi-reference states will be needed to model highly correlated ground states accurately.

Adiabatic state preparation is the second route the paper considers for getting a quantum computer into the state it needs. The idea is to replace the Hamiltonian, the operator that encodes a molecule's energy, with a time-dependent one. The starting version has a ground state, the lowest-energy state, that is easy to prepare. By the end of the evolution, the ground state matches the target state. If the change is slow enough, the system stays in its lowest-energy state throughout and arrives at the answer. The speed limit is set by the eigenvalue gap, the energy distance between the ground state and the next level up. The smaller that gap, the slower the change has to be.

The paper bounds how close the result comes to the exact ground state, which it calls the full configuration interaction ground state, the exact answer within a chosen set of basis functions. The bound depends on the largest rate of change of the Hamiltonian, divided by the square of the smallest gap, and on the evolution time. Choosing that time so the chance of success is of order one, the authors find it must scale as N to the fourth power over the squared gap, where N sets the size of the problem.

The authors then turn to the size of the derivative of the Hamiltonian. Applying the triangle inequality gives a norm that scales as N to the fourth power. They note that on physical grounds the norm should be closer to N squared, because "the potential energy scales quadratically with the number of constituent particles."

To run the evolution on a gate-based machine, a standard method is Trotter-Suzuki simulation, which slices the evolution into small time steps and approximates each step with simple gates. The paper cites the best known bounds for this method. As written, they give a circuit size that grows steeply with N and falls off with the squared gap. Taken at face value, the authors say, the cost of adiabatic state preparation "may be prohibitively expensive even if the eigenvalue gap is constant."

The authors add an important qualification about the upper bound on the derivative, which the cost estimate depends on directly.

The document turns from its method to the chemistry it aims to model. It sets out to give background on why biological dinitrogen fixation matters, and on the model of nitrogenase's active site, prepared in different charge and spin states, that its feasibility analysis relies on.

The authors describe nitrogen fixation as a long-standing goal: "a Holy Grail in chemistry has been the catalytic fixation of molecular nitrogen under ambient conditions." Fewer than half a dozen synthetic catalysts have been developed for that purpose, and the document notes that "after decades of fruitless efforts" they all suffer from low turnover numbers, meaning they complete only a limited number of reaction cycles before they stop working. On that record, the authors conclude that "the synthetic dinitrogen-fixation problem under ambient conditions can thus be considered largely unsolved."

The stakes are practical. Fertilizers are made from ammonia, the end product of dinitrogen fixation. The document points out that ammonia is produced in the very efficient Haber–Bosch process, which needs high temperature and pressure and consumes up to 2 percent of annual energy production.

The authors doubt that a sophisticated synthetic homogeneous catalyst will replace that process, since such catalysts are likely to be less stable and more expensive to produce. They see more promise in a mono- or polynuclear iron-based catalyst, one built around one or several iron centers, that works under ambient conditions and draws on an easily accessible source of hydrogen and nitrogen from air. Such a catalyst, they suggest, "could become important for local small-scale fertilizer-production concepts."

The core difficulty is chemical. In the document's words, "nitrogen fixation represents a tremendous chemical challenge to activate and break the strong triple bond in dinitrogen at room temperature and pressure." The authors then note that

Nature found an efficient way to achieve this goal.

The enzyme nitrogenase does it, and its active site is the iron–molybdenum cofactor, FeMoco.

The paper turns to the method that would make such calculations possible. The electronic energy of a molecule comes from solving the electronic Schrödinger equation, and the energy and wave function from that solution determine how a molecule behaves in a reaction. Expanding the many-electron wave function over a complete set of basis states gives the exact answer, which chemists call full configuration interaction and physicists call exact diagonalization. The authors state plainly that "this approach is unfeasible for molecules of more than a few atoms." The number of many-electron basis states grows exponentially with the number of orbitals, so the method hits a hard ceiling. As the paper puts it,

Exact diagonalization is therefore limited to about 18 electrons distributed among 18 spatial orbitals.

The standard workaround is to restrict the orbital space. The complete-active-space self-consistent-field approach, known as CASSCF, applies exact diagonalization inside a small active space of orbitals near the Fermi energy, the level that separates occupied from empty states. It then relaxes the orbitals self-consistently to compensate for leaving the rest out. The authors note that the approach is "limited by the 18-orbital wall", and that it ignores most of the virtual orbitals, the unoccupied ones, when building the wave function. A molecule of a hundred atoms needs far more than a thousand one-electron basis functions, so most of the orbitals are dropped. The density matrix renormalization group, or DMRG, can be seen as a polynomially scaling version of the same idea, but the authors observe that even it "can push this wall only to about a hundred spatial orbitals."

What a CASSCF-type wave function captures is one particular kind of electron correlation. The paper says such a wave function "solves only the so-called static electron correlation problem", which is the difficulty that arises when several orbitals have nearly the same energy and must be treated together. For that reason it is "particularly well suited for molecular structures with near-degenerate orbitals." The paper concedes that

The resulting electronic structure is only qualitatively well described.

It argues, however, that this qualitative picture holds along a whole reaction coordinate, which is what makes the approach attractive as a general tool. Quantitative accuracy needs more. The authors note that

the many virtual orbitals not considered for the CAS make a nonnegligible contribution to electronic-energy differences.

This is dynamic correlation, and it is usually added with a multi-reference perturbation theory calculation to second order. The authors caution that such perturbation theories "require elements of the three- and four-electron reduced density matrices, which are difficult to calculate and to store."

The authors argue that the chemistry that will benefit most is chemistry dominated by strong static correlation, rather than by the weak dispersion interactions of dynamic correlation. The electronic energy at zero Kelvin, with vibrational, temperature and entropy corrections left out, should dominate the energy change of a chemical process. In their words, transition-metal catalysis is "a field that presents such situations." Along a reaction mechanism, the states "often represent typical static electron correlation problems as bonds are formed and broken on the way to stable products." Reliable prediction usually means studying more than one elementary step, especially when reactive intermediates and unwanted side reactions are involved, so "the total number of molecular structures whose electronic energy is required for an understanding of a reaction mechanism is, in general, very large."

Late 3d transition metals fit this profile, and the authors single out iron, which is "cheap and ubiquitous, often yields nontoxic compounds, and therefore represents an ideal catalytic center." Every intermediate and transition state must be optimized with an efficient quantum chemical method. The paper says density functional theory, or DFT, has a split record here: the accuracy of energies from present-day DFT "is often not satisfactory for predictive purposes", yet molecular structures "can be reproduced with remarkable accuracy." The authors therefore optimize structures of a model of the iron-molybdenum cofactor, the FeMoco active site, in its resting state across several charge and spin states, to create the electronic situations their feasibility analysis has to handle. Settings for the small-active-space orbital optimization are listed for these structures, giving total spin, charge, and the numbers of active electrons and orbitals. The appendix also gives the atomic coordinates of two of the structures, in angstroms.

The paper also looks at ways to prepare the problem before diagonalizing. The authors describe "perturb, then diagonalize" ideas, which produce a dressed many-electron Hamiltonian that is better conditioned for a CASSCF-type approach. A newer development they highlight combines DFT, which describes dynamic correlation well, with CASSCF through range separation, an idea introduced by Savin. The electron-electron interaction is split by distance: DFT handles the short range, while the full flexibility of a CASSCF-type wave function is kept at long range. The authors say this approach "is very efficient and does not compromise the efficiency of a CASSCF-type approach", unlike perturbation theory. Combined with DMRG, it "has delivered very promising results" for benchmark reactions involving transition metals.

The authors conclude that "a chemically sensible exact diagonalization technique is a CASSCF-type approach that captures dynamic correlation effects in the one-electron states." They argue that such an approach "could be directly implemented on a quantum computer with the specific exact diagonalization technology developed for such a machine." That is the bridge the paper proposes between classical chemistry and the quantum hardware it studies.

The paper argues that "CASSCF-type methods will mostly be required for the validation of electronic energies of DFT-optimized molecular structures," which it calls "the essential piece of information for establishing a reaction mechanism." CASSCF, short for complete active space self-consistent field, is a method that treats a chosen set of electrons in a chosen set of orbitals with high accuracy. DFT, density functional theory, is a cheaper approach often used to find the shapes of molecules.

The methods note that follows explains how the inputs for the paper's calculations were built, for the iron-molybdenum cofactor, the metal cluster at the heart of nitrogenase, the enzyme that turns nitrogen gas into ammonia. The authors first optimized the structure of a resting-state model of that cofactor. The model includes "those residues of the protein backbone" that hold the metal cluster in place inside the enzyme. They then varied the spin and charge states during optimization. Spin describes the magnetic orientation of the electrons. The authors explain that this was done "in order to obtain a variety of electronic structures for the assessment of a solution algorithm on a quantum computer." They also state that "These spin and charge states do not necessarily match the one of the resting state of nitrogenase," so the two structures serve as test cases rather than as the enzyme's actual state.

The first structure carries three positive excess charges and an equal number of alpha and beta spins, the two orientations an electron can take. The second structure is uncharged and has one unpaired alpha spin. Both were optimized with unrestricted density functional theory. In this approach the spin symmetry is broken, and the paper notes that "only Sz remains as a good quantum number." In practice, only the spin component along one chosen axis stays a clean label. The calculations used the Turbomole program, version 6.4, with the B3LYP density functional and the def2-TZVP triple-zeta basis set, with polarization functions on all atoms. Only the molybdenum atom received an effective core potential, a simplified treatment of its inner electrons. The authors say this potential "also takes care of all scalar-relativistic effects on this heavy atom."

Next, the researchers generated molecular orbital integrals, the numerical inputs describing how electrons interact, for a complete active space. For the first structure this covered 54 electrons in 54 spatial orbitals, which is 108 spin orbitals, taken from a singlet CASSCF run with 24 electrons in 16 orbitals. Singlet means the electron spins pair up overall. For the second structure the active space was 65 electrons in 57 spatial orbitals, or 114 spin orbitals, taken from a quartet run with 21 electrons in 12 orbitals. Quartet refers to a total spin multiplicity of four.

The choice of these large active spaces followed Pulay's UNO-CAS criterion, in which, as the authors put it, "the occupation number of the natural orbitals serves as a selection criterion." Natural orbitals are a set of orbitals describing the electron distribution, and their occupation numbers show how fully each one is filled. The authors departed from the usual starting point of unrestricted Hartree–Fock natural orbitals. Instead, they picked natural orbitals from the smaller CASSCF runs whose occupation numbers fell between 1.98 and 0.02 for the first structure, and between 1.99 and 0.01 for the second. Those are orbitals that are neither nearly full nor nearly empty. The integrals were computed with the Molcas program, and the authors say that all settings for the calculations are summarized in a table.

The paper's main takeaway is that quantum computers used as accelerators alongside classical supercomputers could, in the authors' view, supply accurate energies for the intermediates and transition structures of a nitrogenase mechanism with a manageable amount of memory and time. They set the scale of the job against Shor's factoring algorithm, finding comparable gate counts and physical qubit numbers for the cases considered, and they conclude that this makes the simulation typical of the demands of quantum computing targets. They also argue that the economic value of solving catalysis problems of industrial importance may offset the hardware cost. The open questions they leave are substantial. The cost of preparing an accurate starting state for FeMoco is not estimated and may be large. The paper finds only weak evidence of a downward trend in the overlap of simple starting states as molecules grow, and its extrapolation to FeMoco gives roughly 43 percent, which it cannot yet confirm is typical. Whether such simple states will suffice for strongly correlated molecules remains to be studied, as does how the ordering of operators affects nested circuits. On hardware, the serial implementation needs millions of physical qubits, and the authors concede that the parallel implementation may need hundreds of millions to nearly a trillion, which looks more daunting. They turn to modular designs and to more robust qubits, which they say may be invaluable. Their adiabatic route to a good starting state also looks potentially prohibitively expensive under the worst-case bounds they cite.

  1. Elucidating Reaction Mechanisms on Quantum Computers