---
title: Linear Combination of Unitaries (LCU)
url: https://www.emergentmind.com/topics/linear-combination-of-unitaries-lcu
type: topic
---

# Linear Combination of Unitaries (LCU)

Linear Combination of Unitaries (LCU) is a quantum-algorithmic framework in which a target operator is expressed as a weighted sum of unitary operators, typically in the discrete form \(A=\sum_j a_j U_j\), and more generally in continuous forms such as \(F(A)=\int_V f(\mathbf t)\,G(A,\mathbf t)\,d\mathbf t\). In contemporary usage, LCU denotes both the algebraic decomposition itself and the circuit constructions that turn such decompositions into block encodings, postselected transformations, quasi-probability samplers, or classical-post-processing schemes for nonunitary functions. It is a standard input model for block-encoding-based algorithms, including qubitization and QSVT, and it appears in Hamiltonian simulation, quantum chemistry, state preparation, optimization, and variational circuit design [2507.20887], [2407.06571], [2509.14451].

## 1. Algebraic formulation and coherent realization

In the standard coherent formulation, one starts from a decomposition
\[
A=\sum_{j=0}^{L-1}\alpha_j U_j,
\]
with unitary \(U_j\) and coefficients whose phases can be absorbed into the unitaries. Writing
\[
\alpha=\sum_{j=0}^{L-1}|\alpha_j|,
\]
LCU constructs a block encoding of \(A/\alpha\) by combining a coefficient-preparation oracle with a multiplexed unitary. A canonical PREPARE map acts on an ancilla register as
\[
P(\alpha)\ket{\bar 0}=\sum_{\ell=0}^{L-1}\sqrt{\frac{\alpha_\ell}{|\alpha|_1}}\ket{\ell},
\]
while SELECT applies
\[
S(\Upsilon)=\sum_{\ell=0}^{L-1}\ket{\ell}\!\bra{\ell}\otimes \mathcal P_\ell.
\]
The PREPARE–SELECT–UNPREPARE pattern then yields a larger unitary whose \(\ket{\bar 0}\)-ancilla block is \(\Upsilon/|\alpha|_1\) [2501.18515].

This construction is equivalently described in block-encoding language. If \(V\) and \(W\) are ancilla unitaries whose first columns encode the magnitudes and phases of the coefficients, then
\[
\ket{0}^{\otimes n_a}\ket{\psi}\mapsto \ket{0}^{\otimes n_a}\frac{\sum_j \kappa_j U_j}{\alpha}\ket{\psi}+\dots
\]
realizes an \((\alpha,n_a,\epsilon)\)-block encoding of the target operator [2601.18024]. In standard postselected use, the success probability on input \(\ket{\psi}\) is
\[
p=\frac{1}{|\alpha|_1^2}\,\langle\psi|\Upsilon^\dagger \Upsilon|\psi\rangle,
\]
and oblivious amplitude amplification is the conventional mechanism for boosting this probability [2501.18515].

The same formalism also accommodates parametrized constructions. In variational settings, one writes
\[
V(\boldsymbol\theta,\vec c)=\sum_j c_j U_j(\boldsymbol\theta^{(j)}),
\]
and the output is naturally sub-normalized. In that setting, the expectation value \(\mathrm{tr}(\rho O)\) already incorporates the LCU postselection probability rather than separating it as a preprocessing overhead [2506.22310].

## 2. Resource measures, optimality, and decomposition quality

The central cost parameter of an LCU decomposition is its coefficient \(1\)-norm. For a Hamiltonian written as
\[
\hat H=\gamma \hat 1+\sum_k u_k \hat U_k,
\]
the quantity
\[
\lambda=\sum_k |u_k|
\]
sets the normalization in the block encoding and governs the query complexity of Hamiltonian simulation. In particular, truncated Dyson-series simulation has complexity
\[
\mathcal O\!\left(\lambda t\,\mathrm{polylog}\frac{\lambda t}{\epsilon}\right),
\]
while qubitization has complexity
\[
\mathcal O\!\left(\lambda t+\mathrm{polylog}\frac{1}{\epsilon}\right)
\]
in oracle uses [2208.08272].

A general lower bound links LCU quality to spectral geometry. For any bounded Hermitian operator, every LCU decomposition obeys
\[
\lambda\ge \frac{\Delta E}{2},
\]
where \(\Delta E=E_{\max}-E_{\min}\) is the spectral range. The same work shows that \(\lambda_{\min}=\Delta E/2\) is achievable in principle by using unitaries built from
\[
\hat U_\pm=\hat H_{sr}\pm i\sqrt{\hat 1-\hat H_{sr}^2},
\]
after shifting and rescaling the spectrum to \([-1,1]\). The practical limitation is that the square-root term is generally highly nonlocal, so the optimal-\(\lambda\) construction is “practically unattainable for general Hamiltonians” [2208.08272].

This lower-bound perspective is especially important in quantum chemistry, where different fermionic and qubit LCUs induce large differences in \(\lambda\). The “Majorana Tensor Decomposition” provides a unifying description in which existing Pauli, anti-commuting, single-factorization, double-factorization, CSA, THC, and newer low-rank constructions are all instances of a common tensor-factorized Majorana LCU framework [2407.06571]. A complementary line of work, the Block Invariant Symmetry Shift method, reduces the spectral range and the resulting LCU \(1\)-norm by modifying the Hamiltonian only outside the target electron-number block; the 2024 linear-programming reformulation guarantees global optimality within the BLISS ansatz and scales to active spaces of up to 76 orbitals [2409.18277].

A recurring misconception is that “LCU cost” is determined only by the number of terms. The chemistry literature instead shows that coefficient structure, symmetry shifts, factorization rank, and fragment spectral ranges can matter as much as raw term count. This suggests that decomposition design is often the dominant classical preprocessing problem in practical LCU pipelines.

## 3. Continuous, Fourier, and ancilla-free generalizations

LCU is not limited to finite sums. In the LCU via classical post-processing paradigm, one assumes an integral representation
\[
F(A)=\int_V f(\mathbf t)\,G(A,\mathbf t)\,d\mathbf t,
\]
where each \(G(A,\mathbf t)\) is proportional to a unitary operator. A quantum device estimates
\[
\mathrm{Re}\bigl[\mathrm{Tr}(G(A,\mathbf t)\rho)\bigr]
\]
with the Hadamard test, and the nonunitary quantity \(\mathrm{Tr}[F(A)\rho]\) is reconstructed by classical integration. Within this setting, Monte Carlo, trapezoid, and quasi-Monte Carlo integration induce different asymptotic numerical-integration errors:
\[
\mathcal O(K^{-1/2}),\qquad \mathcal O(K^{-2/d}),\qquad \mathcal O((\log K)^d/K),
\]
respectively, while all retain Hadamard-test statistical error of order \(\mathcal O(M^{-1/2}K^{-1/2})\). In ground-state property estimation and Green’s-function estimation, quasi-Monte Carlo was found to yield the lowest overall errors for practically relevant shot counts per unitary [2509.14451].

A second generalization replaces coherent LCU gadgets by quasi-probability or sampling constructions. In ancilla-free Fourier-based LCU for optimization, a target unitary is written as
\[
U=\sum_{j=0}^m c_j V_j,
\]
and the induced channel is decomposed into a quasi-probability combination with cost
\[
\Gamma=\|c\|_1^2.
\]
Dropping the signed interference terms yields an ancilla-free sampler with output probabilities \(\tilde p_x\) satisfying
\[
\tilde p_x\ge p_x/\Gamma.
\]
For diagonal unitaries generated by \(g(x)\in\{0,\ldots,m\}\), the Fourier construction gives \(\Gamma\le m+1\); for arbitrary permutation-invariant unitaries expanded over collective \(SU(2)\) rotations, the cost satisfies \(\Gamma=O(n^3)\) [2605.18985]. The same framework makes explicit the connection between Fourier penalties in QAOA and classical Lagrangian relaxation.

A third direction studies general nonunitary operators through Fourier extension. There, one approximates the identity function on a buffered interval by a smooth sine series with exponentially decaying coefficients, applies it separately to the Hermitian and anti-Hermitian parts of \(A\), and rewrites the sines as complex exponentials. The resulting LCU achieves exponential convergence in the truncation order, while the block-encoding subnormalization scales as
\[
\alpha=O\!\left(\|A\|_2\log\log\frac{1}{\epsilon}\right),
\]
which is substantially better than the polynomial dependence on \(1/\epsilon\) in earlier general-purpose constructions [2601.18024].

## 4. Hamiltonian simulation and error-compensated synthesis

Hamiltonian simulation has been the most prominent application of LCU. In truncated-Taylor simulation, one expands \(e^{-iHt}\) as a linear combination of products of Hamiltonian terms and then block-encodes the truncated series. For electronic-structure Hamiltonians with highly nonuniform coefficients, it is not optimal to keep all terms at a fixed Taylor order. An adaptive truncation scheme that iteratively retains larger-magnitude terms at each order improves the simulation accuracy typically by an order of magnitude for a given circuit depth across a range of molecular instances [2007.11624].

A distinct strategy uses LCU to compensate the error of product formulas rather than replacing them. For a \(K\)-th order Trotter formula \(S_K(x)\), define the multiplicative remainder
\[
V_K(x)=U(x)S_K(x)^\dagger.
\]
The “Paired Taylor-Series Compensation” construction builds an LCU only for \(V_K(x)\) and pairs the leading anti-Hermitian error terms with the identity so that the per-segment LCU norm remains close to one. The resulting algorithm improves the time scaling of \(K\)-th order Trotter from \(t^{1+1/K}\) to \(t^{1+1/(2K+1)}\) and yields logarithmic dependence on \(1/\epsilon\). For local lattice Hamiltonians, the “Nested-Commutator Compensation” variant uses explicit nested-commutator expansions of the Trotter remainder and achieves
\[
O\Big(n^{1+\frac{2}{2K+1}}\,t^{1+\frac{1}{2K+1}}\,\epsilon^{-\frac{1}{2K+1}}\Big)
\]
gate complexity, thereby quadratically improving the accuracy exponent of \(K\)-th order Trotter while retaining almost linear system-size dependence [2212.04566].

LCU has also been demonstrated on hardware for Hamiltonian dynamics. In an ion-trap realization for squared overlaps of time-evolved states, the Taylor-expanded propagator was collapsed into one large LCU, then reduced by pre-selecting only those Pauli terms contributing to the target observable. Together with a multiplexor-based SELECT implementation, this yielded a two-qubit gate count
\[
2^{\lceil \log_2(L)\rceil}(2n+1)-n-2
\]
for an \(L\)-term \(n\)-qubit LCU. In the Rabi–Hubbard example discussed there, a 400–500-term expansion was reduced to at most 11 terms in the observable-relevant subspace, enabling a practical implementation on an ion-trap device [2501.18515].

These developments show that LCU is no longer used only as a monolithic asymptotic primitive. In practice it is frequently combined with truncation heuristics, factorization structure, or product formulas to rebalance ancilla cost, circuit depth, and classical preprocessing.

## 5. Structured block encodings and application-specific constructions

A major challenge in practical LCU is the SELECT oracle. In conventional implementations it is realized by multi-controlled application of the indexed unitary, so two-qubit and \(T\)-gate costs are often dominated by control logic rather than by the target operator itself. The FOQCS-LCU construction addresses this for Pauli-sum Hamiltonians by rewriting the sum in check-matrix form and using two \(n\)-qubit ancilla registers that separately encode the \(X\) and \(Z\) patterns of each Pauli string. SELECT then reduces to one parallel layer of CNOTs and one parallel layer of CZs, giving a constant-depth implementation with a linear number of singly controlled Pauli gates and ancillae. For Heisenberg and spin-glass models, the resulting block encodings show an order-of-magnitude improvement in CNOT count over conventional LCU [2507.20887].

The same paper exploits generalized Dicke-state preparation to reduce PREPARE overhead. Because many structured Hamiltonians require superpositions over single-excitation, nearest-neighbor double-excitation, or paired-register excitation patterns, PREPARE can be compiled into explicit linear-CNOT Dicke routines rather than treated as a generic amplitude-loading problem [2507.20887]. This is an example of a broader theme in LCU: once coefficient structure is known, PREPARE can often be specialized far beyond the generic \(\lceil\log_2 M\rceil\)-qubit model.

LCU has also been adapted to black-box quantum state preparation. Given an oracle that returns \(n\)-bit values \(x_j\), the amplitude \(x_j\) can be transduced into a quantum amplitude by interpreting the binary expansion
\[
x_j=\sum_{i=0}^{n-1}x_{j,i}2^{-(i+1)}
\]
as an LCU over fixed dyadic coefficients. Two concrete constructions were given: a standard LCU version requiring \(2\log(n)-1\) additional qubits and \(3n-4\) Toffoli gates for the amplitude-transduction core, and a modified LCU version using \(n+2\) ancillas and \(n\) Toffoli gates. With logical-AND optimization, the standard variant’s Toffoli count can also be reduced to \(n\) [2105.06230].

Application areas have expanded accordingly. In optimization, Fourier LCU replaces fully connected penalty terms and XY-mixer evolutions by ensembles of simple single-qubit layers or collective rotations [2605.18985]. In chemistry, tensor-factorized fermionic LCUs turn four-index integral tensors into orbital-rotation-conjugated reflection operators [2407.06571]. In both cases, LCU functions less as a single algorithm than as a compilation language for structured operators.

## 6. Variational behavior, ancilla data, and parallel execution

The growth of parametrized and hybrid algorithms has made trainability a central question for LCU. For a coherent LCU of \(k\) parametrized circuits with Haar-model assumptions on the constituent families, the variance of the unnormalized expectation satisfies
\[
\mathrm{Var}[m]\ge \sum_{j=1}^k E[c_j^2]^2\,\mathrm{Var}[m_j]\ge \frac{1}{k^3}\min_j \mathrm{Var}[m_j].
\]
Accordingly, if each constituent family is trainable at scale, the coherent LCU remains trainable up to a polynomial penalty in \(k\). For incoherent superpositions, the bound improves to
\[
\mathrm{Var}(\tilde m)\ge \frac{1}{k}\min_j \mathrm{Var}[m_j],
\]
and for uniform Dirichlet weights with identical moment structure one obtains
\[
\mathrm{Var}(\tilde m)=\frac{2}{k+1}\,\mathrm{Var}(m_i).
\]
The analysis therefore rules out LCU itself as a generic barren-plateau mechanism and supports the use of LCU as a controlled expressivity increase in variational ansätze [2506.22310].

Another recent development reinterprets the “junk” ancilla outcomes of LCU as useful data. In an alternative circuit with Hadamards on an index register and a single rotation qubit, every ancilla outcome \((i,r)\) projects the system onto a distinct linear combination of the \(U_t\ket{\psi}\). Collecting these outcome states into a matrix \(\Phi\) yields the exact factorization
\[
\Phi=CX,
\]
with \(\operatorname{rank}(\Phi)\le K\). This low-rank structure enables matrix completion from partial observations and classical recovery of the target branch, and it also motivates candidate cryptographic applications based on hiding the coefficient matrix \(C\) as a secret key [2605.02986].

Depth reduction has likewise become a major theme. Parallelization schemes based on partitioning SELECT into commuting groups, together with adaptive Clifford circuits, teleportation, and parallel QROM techniques, reduce logical depth with only an \(O(\log n)\) qubit overhead. Prior work cited there suggests \(O(n)\) depth savings for molecular Hamiltonians, and numerics indicate a depth reduction by a factor approximately \(n/2\). Because the \(T\)-depth decreases by the same factor while the \(T\)-count is unchanged, the approach can reduce overall space-time volume in fault-tolerant LCU implementations [2312.00696].

These strands indicate a broadening of the subject. A plausible implication is that LCU is increasingly best understood not as a single postselection gadget, but as a family of operator-encoding techniques spanning coherent block encodings, randomized samplers, continuous integral representations, low-rank data structures, and hardware-aware compilation strategies.

Source: https://www.emergentmind.com/topics/linear-combination-of-unitaries-lcu