Quantum Subspace Expansion

The vastness of a quantum system's Hilbert space represents one of the greatest challenges in quantum computation and simulation. Finding the properties of a molecule or material by exploring this entire space is computationally intractable. Furthermore, today's quantum hardware operates in a noisy environment, where errors corrupt an algorithm's intended results. This article introduces Quantum Subspace Expansion (QSE), an elegant and powerful method that addresses both of these problems simultaneously. QSE provides a practical framework for extracting highly accurate information from a quantum system by focusing on a small, intelligently chosen corner of its Hilbert space.

This article is structured to provide a comprehensive understanding of this vital technique. In the first chapter, ​​Principles and Mechanisms​​, we will delve into the core idea of QSE, drawing parallels to familiar concepts in quantum mechanics and numerical linear algebra to explain how solving a small, classical generalized eigenvalue problem can yield precise quantum information. The second chapter, ​​Applications and Interdisciplinary Connections​​, will then explore how this principle is applied in the real world to mitigate errors on noisy quantum processors, calculate molecular energies for quantum chemistry, and even diagnose the specific flaws in quantum hardware.

Imagine you are trying to find the lowest point in a vast, fog-covered mountain range. The full landscape—the entire Hilbert space of a quantum system—is impossibly large to map out. You have a rough starting point, perhaps a location given by a noisy GPS, which is our imperfect state prepared on a quantum computer. How can you find a better, lower-altitude position without exploring the entire range? You could look around your immediate vicinity, checking the altitude in a few specific directions—north, east, and perhaps along the direction of steepest descent. By analyzing just this small, local patch of land, you can find a much better estimate for the lowest point nearby.

This is the central philosophy behind ​​Quantum Subspace Expansion (QSE)​​. It is a wonderfully elegant and powerful idea that says we can learn a great deal about a complex quantum system not by tackling its full infinite complexity, but by asking the right questions within a tiny, intelligently chosen corner of its world.

If you have studied quantum mechanics, you have already met the core mathematical machinery of QSE in a different guise: ​​degenerate perturbation theory​​. Remember what happens when you have a set of states, say, $n$ of them, which all share the exact same energy $E_0$? These states form a degenerate subspace. Now, if you introduce a small perturbation, $\hat{V}$, to the system, this "accidental" equality is often broken. The energy level splits into $n$ distinct new levels.

How do we find these new energies? We can't simply calculate the energy shift for each state individually. The perturbation mixes the original states. The new, correct states are linear combinations of the old ones. To find the right combinations and their corresponding energy shifts, we must construct an $n \times n$ matrix of the perturbation, with elements $M_{\alpha\beta} = \langle \phi_\alpha | \hat{V} | \phi_\beta \rangle$, and find its eigenvalues. Diagonalizing this small matrix tells us everything we need to know to first order. We have solved the problem not in the infinite-dimensional Hilbert space, but within the small, relevant subspace of degenerate states.

QSE takes this idea and runs with it. It asks: why reserve this powerful technique only for subspaces that happen to be degenerate? What if we could design a small subspace of our own choosing, a subspace that we believe is physically relevant, and then apply the same logic?

This is precisely what QSE does. We start with a reference state, let’s call it $| \psi_0 \rangle$. This could be the state prepared by a variational algorithm like VQE, which is our best, albeit noisy, guess for the ground state. Then, we generate a handful of new states by acting on $| \psi_0 \rangle$ with a chosen set of operators, $\{ \hat{O}_i \}$. Our subspace is now the collection of all possible linear combinations of these basis states:

According to the ​​Rayleigh-Ritz variational principle​​, the best possible approximations for the eigenenergies of the true Hamiltonian, $\hat{H}$, that can be found within this subspace are the stationary points of the energy expectation value. The search for these stationary points transforms into a familiar problem: finding the eigenvalues of the Hamiltonian projected onto our subspace.

There is, however, one crucial twist. The basis states in degenerate perturbation theory, $\{|\phi_\alpha\rangle\}$, are typically chosen to be orthonormal. Our custom-built basis states, $\{ \hat{O}_i | \psi_0 \rangle \}$, are almost certainly not. They can overlap, and they are not necessarily normalized to one. For example, the states in problems and are constructed from a noisy root state and are clearly not orthogonal.

What does this mean for our eigenvalue problem? It means we need to account for the geometry of our "messy" basis. We must solve not the standard eigenvalue equation, but a ​​generalized eigenvalue problem​​:

Here, $\mathbf{c}$ is the vector of coefficients that defines the eigenvector in our basis, and $E$ is the corresponding energy eigenvalue. The matrix $H$ is just what you'd expect: its elements are $H_{ij} = \langle \psi_0 | \hat{O}_i^\dagger \hat{H} \hat{O}_j | \psi_0 \rangle$. But what is $S$?

$S$ is the ​​overlap matrix​​. Its elements, $S_{ij} = \langle \psi_0 | \hat{O}_i^\dagger \hat{O}_j | \psi_0 \rangle$, are simply the inner products between all of our basis vectors. This matrix encodes the full geometry of our subspace—all the lengths and all the angles. It is a "metric" that tells the Hamiltonian how to operate correctly in a space where the basis vectors are not neatly orthonormal. Solving this generalized eigenvalue problem on a classical computer gives us a set of improved energy estimates for the ground and excited states—the best estimates possible, given our chosen subspace.

The power of QSE hinges on the choice of the expansion operators, $\{ \hat{O}_i \}$. A well-chosen set can dramatically improve our energy estimates and even correct for errors introduced by a noisy quantum device.

Correcting Errors with the Hamiltonian Itself

A natural and powerful choice is to use powers of the Hamiltonian itself, generating a basis like $\{|\psi_0\rangle, \hat{H}|\psi_0\rangle, \hat{H}^2|\psi_0\rangle, \dots \}$. This constructs what is known as a ​​Krylov subspace​​, the very same structure used in famous classical algorithms like the Lanczos and Arnoldi methods for finding eigenvalues. This connection reveals a beautiful unity: QSE can be seen as a quantum implementation of one of the most successful ideas in numerical linear algebra. This same principle of using the Hamiltonian's action to find a "better" direction is also at the heart of advanced classical simulation methods like the Density Matrix Renormalization Group (DMRG).

This choice has a remarkable consequence for error mitigation. Suppose our quantum computer has a systematic error, so it operates with a faulty Hamiltonian $\tilde{H}$ instead of the ideal one, $H$. We prepare a state $| \psi_{prep} \rangle$, which is the ground state of $\tilde{H}$. We now want to find the true ground state energy of the ideal Hamiltonian, $H$. We can do this by using the basis $\{ | \psi_{prep} \rangle, \tilde{H} | \psi_{prep} \rangle \}$ and then diagonalizing the ideal Hamiltonian $H$ within this subspace. As demonstrated in a hypothetical scenario, this procedure can sometimes project out the influence of the error entirely, recovering the exact ground state energy of the ideal Hamiltonian. It's like having a distorted lens but being able to mathematically reconstruct the original, clear image by understanding the nature of the distortion.

Using Physical Intuition

In quantum chemistry or materials science, we can use our physical intuition to choose the operators. If $| \psi_0 \rangle$ is an approximation to the ground state, then low-energy excited states often correspond to moving a particle from an occupied orbital to an unoccupied one. We can choose our operators $\{ \hat{O}_i \}$ to be precisely these physical excitation operators (e.g., Pauli operators like $\sigma_x$ and $\sigma_y$ in a qubit representation, which "flip" a qubit). This populates our subspace with states that are physically likely to be important components of the low-energy spectrum.

QSE is a projection method, and its power is ultimately limited by the subspace it projects onto. It is not a magic bullet. Imagine our true ground state is a mix of state A and state B. If our subspace contains only directions related to state A, we will never find the component from state B.

A compelling thought experiment illustrates this perfectly. Suppose a two-qubit system has a true Hamiltonian containing a two-qubit error term, like an unwanted $\epsilon X_1 X_2$ crosstalk. If we try to correct for this using only single-qubit operators in our expansion set (like $X_1$ and $I \otimes X_2$), we find that we cannot fully remove the error. The QSE procedure gives a better answer than our initial noisy state, but a residual error remains. The reason is simple: our basis of single-qubit excitations does not contain the "direction" corresponding to the two-qubit error. To correct it, we would need to include an operator like $X_1 X_2$ in our expansion set.

This teaches us a crucial lesson: the success of QSE depends critically on whether the chosen subspace has a significant overlap with the true eigenvectors we are seeking. If we use expansion operators that are blind to the essential physics or errors of the problem, our results will be incomplete.

From correcting the energies of degenerate states to mitigating errors on near-term quantum processors, Quantum Subspace Expansion embodies a single, elegant principle: complex problems can often be simplified by focusing on a small, well-chosen pocket of a much larger space. It is the quantum analogue of a politician focusing on a few swing states, or an engineer strengthening a bridge by reinforcing a few critical joints. By solving a small generalized eigenvalue problem—a task easily handled by a classical computer—we can leverage a few precious measurements from a quantum device to obtain remarkably accurate results. This beautiful synthesis of quantum measurement and classical linear algebra represents a powerful and practical pathway toward extracting meaningful results from the noisy quantum computers of today and tomorrow.

Now that we have grappled with the principles of Quantum Subspace Expansion, you might be asking a perfectly reasonable question: What is it all for? It is one thing to learn the rules of a new game, and quite another to see why that game is worth playing. The truth is, the elegant idea of projecting a problem onto a small, manageable subspace is not just a mathematical curiosity. It is a powerful, versatile tool that physicists and chemists are actively using to push the boundaries of what is possible with today's quantum technology. It is our lens for finding clarity in the quantum fog.

In this chapter, we will take a journey through the surprisingly diverse applications of this technique. We will see how it helps us clean up the messy results from real quantum computers, how it connects to the grand challenge of designing new molecules and materials, and how its underlying philosophy extends to the subtle arts of ultra-precise measurement and understanding the very nature of quantum noise.

Imagine you have a powerful but slightly faulty quantum computer. You ask it to prepare a very specific, delicate quantum state — say, the famous Bell state $|\psi_{\text{ideal}}\rangle = \frac{1}{\sqrt{2}}(|00\rangle + |11\rangle)$. You run your program, but because of unavoidable noise, the machine produces a slightly corrupted version, $|\psi_{\text{noisy}}\rangle$. What can you do? Is the result simply garbage?

This is where Quantum Subspace Expansion (QSE) enters as a rather brilliant error mitigation strategy. Instead of throwing the result away, we make an educated guess about the nature of the noise. We might not know the exact error process, but we might suspect that the most likely error is, for example, an accidental flip of one of the qubits, represented by a Pauli $X$ operator.

With this guess, we construct a "search space," or a subspace, spanned by our ideal, intended state $|\psi_{\text{ideal}}\rangle$ and the most likely erroneous state, $X|\psi_{\text{ideal}}\rangle$. The QSE procedure then takes our actual noisy output, $|\psi_{\text{noisy}}\rangle$, and finds the state within this small search space that is its closest cousin. This new state, $|\psi_{\text{corr}}\rangle$, is our corrected answer. The magic is that even if our initial guess about the error was not perfectly accurate, this projection often yields a state remarkably closer to the ideal one, dramatically improving the fidelity of our computation.

Of course, there is a practical wrinkle. To perform this correction, we need to calculate matrix elements like $\langle\psi_k|\hat{H}|\psi_l\rangle$. But how do we measure these quantities? On the very same noisy quantum computer we are trying to correct! This might seem like a circular problem, but it is not. The mathematics of QSE is designed to work with these noisy inputs. When we analyze the expectation values we measure, we find that the noise parameters, like an error probability $p$, are baked directly into our results. QSE provides a systematic way to process these noisy matrix elements to produce a final, purified result.

This wonderful clean-up service does not come for free, however. To get a high-quality result from a noisy measurement, we need to take many, many samples. Think of it like taking a photograph in low light; you need a long exposure time to gather enough photons for a clear image. Error mitigation techniques invariably increase the number of "shots," or circuit executions, required. When QSE is nested with other methods, such as Probabilistic Error Cancellation (PEC), this "sampling overhead" can grow very quickly. The total cost is a function of the size of our subspace and the error rates of the machine's gates. This is a crucial, real-world trade-off: we can buy higher accuracy, but the currency is a greater number of quantum computations.

One of the most anticipated applications of quantum computers is in revolutionizing quantum chemistry and materials science. The ability to accurately calculate the energy levels of electrons in molecules could unlock the secrets to designing new medicines, more efficient solar cells, and novel catalysts. This grand challenge is, at its heart, about solving the Schrödinger equation for a complex, many-electron system.

Here, QSE emerges not just as an error-correction tool, but as a quantum-native version of powerful classical simulation techniques. Consider the problem of finding the energy of a molecule that has an "open shell" of electrons, a common situation in chemistry. These electrons have a quantum property called spin, and the overall state of the molecule must respect the fundamental symmetries of spin physics. Many important chemical processes, like the absorption of light, involve an electron "flipping" its spin.

QSE provides a beautiful framework for studying these phenomena. We can start with a simple reference state and build our expansion subspace using physically motivated operators—specifically, operators that create these very spin-flip excitations. By then diagonalizing the molecular Hamiltonian within this chemically relevant subspace, we can find the energies of low-lying excited states. Crucially, because this procedure can be designed to respect the system's fundamental spin symmetries, it guarantees that our results are physically meaningful. It is a direct and powerful method for exploring the photochemistry of molecules on a quantum computer.

This idea extends naturally to the physics of materials. Models like the Fermi-Hubbard model are cornerstones for understanding phenomena like conductivity and magnetism. When simulating such a model on a quantum computer, errors might occur that violate fundamental physical laws, like the conservation of the number of particles. QSE can be used as a keen-eyed detective in this scenario. By constructing a basis that includes states with different particle numbers (e.g., the vacuum and a two-particle state), we can calculate the Hamiltonian matrix elements between them. Ideally, these should be zero if the Hamiltonian conserves particle number. If a coherent error introduces a term $\hat{H}_{\text{err}}$ that violates this symmetry, QSE will find a non-zero matrix element that is directly proportional to the strength of the error. This not only allows us to mitigate the error but also serves as a precise diagnostic tool for characterizing the flaws in our quantum hardware.

So far, we have mostly assumed that we have a good physical intuition for choosing our subspace operators. But can we do better? Can we learn the best subspace directly from the device itself? The answer is a resounding yes, and it turns QSE into a sophisticated, data-driven diagnostic tool.

By repeatedly preparing a state and measuring how it decays over time—a process called quantum process tomography—we can learn the "fingerprint" of the noise affecting our system. This fingerprint is mathematically described by a superoperator known as the Liouvillian. The dominant error channels correspond to the eigenoperators of this Liouvillian. We can then use these very operators, learned directly from data, to construct a bespoke QSE basis that is maximally effective at correcting the specific noise of that device. The method is even robust enough to handle complex, "non-Markovian" noise that has a memory of its past.

Furthermore, the utility of QSE extends far beyond just calculating energies. At its core, it is a general-purpose tool for quantum signal processing. Consider the field of quantum metrology, the science of making measurements with the highest possible precision. A quantum sensor works by encoding a parameter to be measured, say a magnetic field $\theta$, into a quantum state $\rho(\theta)$. Noise degrades this encoding, reducing the achievable precision, which is quantified by a value called the Quantum Fisher Information (QFI). By treating the noisy sensor state as our input, we can use QSE to project it onto a "clean" operator subspace. This process effectively filters the noise out of the quantum signal, resulting in a mitigated state with a higher QFI, thereby restoring the sensor's precision.

The power of QSE is further amplified when combined with fundamental physical principles. If we know that the true state of our system must respect a certain symmetry (for example, parity), we can enforce this explicitly. We can construct a projection operator $P_s$ that filters out any part of our basis that does not conform to the required symmetry $s$. By working entirely within this symmetry-verified subspace, our calculation becomes both more efficient and more robust, as it's guaranteed to produce physically consistent results.

As we zoom out, a unifying picture emerges. The core idea of QSE—identifying a small, relevant subspace and solving a simplified problem within it—is a profoundly general principle. It applies not just to quantum states and Hamiltonians, but to the very dynamics of quantum systems.

The evolution of any open quantum system, subject to both its own Hamiltonian and environmental noise, is governed by a Liouvillian superoperator, $\mathcal{L}$. The eigenvalues of this superoperator tell us everything about the system's characteristic timescales and decay rates. In a fascinating application of the QSE philosophy, we can study these decay rates by treating the noise part of the evolution, $\mathcal{L}_D$, as a "perturbation." We then diagonalize this perturbation within the subspace of operators that are left invariant by the ideal, coherent evolution. This procedure yields the system's dominant decay rates as the eigenvalues, providing deep insight into the system's open dynamics.

From correcting gate errors in a circuit, to finding the color of a molecule, to sharpening the precision of a quantum sensor, Quantum Subspace Expansion is revealed to be more than a single algorithm. It is a philosophy for taming complexity. In the face of the infinite, sprawling Hilbert space, and the relentless chaos of noise, it gives us a strategy: find the small corner where the important physics is happening, and focus your attention there. It is this principle of economy and physical insight that makes it one of the most beautiful and practical ideas in our quest to build a functioning quantum world.