---
title: 'FOQCS-LCU: Efficient Block Encoding for Spin Models'
url: https://www.emergentmind.com/topics/foqcs-lcu
type: topic
---

# FOQCS-LCU: Efficient Block Encoding for Spin Models

FOQCS-LCU is a structured linear-combination-of-unitaries framework for block encoding operators expressed as Pauli-string sums. The acronym stands for **Fast One-Qubit Controlled Select Linear Combination of Unitaries**. Its defining move is to replace the standard binary-indexed, multiply controlled `SELECT` oracle of LCU by a check-matrix-style representation in which the \(X\)- and \(Z\)-support of each Pauli string are stored explicitly in two ancilla registers and applied through two parallel layers of two-qubit gates. Introduced as a compact LCU formulation for practical block encodings of structured spin Hamiltonians and then extended to products, powers, and matrix polynomials, FOQCS-LCU is characterized by a shallow `SELECT` oracle, structure-aware `PREPARE` circuits based on Dicke states, and a “trivially controlled” property that makes controlled block encodings unusually inexpensive [2507.20887, 2601.18767].

## 1. Origin and defining objective

FOQCS-LCU was introduced as a variant of standard LCU block encoding aimed at the practical bottleneck of explicit circuit synthesis. In ordinary prepare-select-unprepare constructions, one writes
\[
H=\sum_{m=0}^{M-1}\alpha_m U_m,
\]
prepares an ancilla superposition over the index \(m\), and applies
\[
\mathrm{SELECT}=\sum_{m=0}^{M-1}|m\rangle\langle m|\otimes U_m.
\]
The standard difficulty is that `SELECT` typically requires multiply controlled applications of the \(U_m\), which become deep and hardware-unfriendly. FOQCS-LCU was designed specifically to eliminate that depth bottleneck for Pauli-structured operators while preserving the usual LCU semantics of postselected block encoding [2507.20887].

The later matrix-polynomial work treats FOQCS-LCU not merely as a standalone Hamiltonian block encoding, but as the primitive from which one can assemble block encodings of products, powers, and
\[
p_d(H)=a_0 I+a_1H+\cdots+a_d H^d.
\]
In that extension, the practical claim is no longer only that a single block encoding is shallow, but that polynomial block encodings can be made additively shallow in \(n\) and \(d\), and that controlled versions remain similarly inexpensive [2601.18767].

A plausible summary is that FOQCS-LCU is best viewed as a depth-oriented re-engineering of LCU rather than a new abstract model of block encoding. Its central tradeoff is explicit throughout the literature: more ancillas, much less `SELECT` depth [2507.20887, 2601.18767].

## 2. Algebraic formulation and block-encoding structure

The FOQCS-LCU formulation begins from the usual Pauli-string LCU setting but rewrites the Hamiltonian in separated \(X/Z\) index form. For an \(n\)-qubit target matrix,
\[
H=\sum_{i=0}^{2^n-1}\sum_{j=0}^{2^n-1}\tilde\alpha_{ij}\bigotimes_{\ell=0}^{n-1} Z^{j_\ell}X^{i_\ell},
\]
where \(i_\ell,j_\ell\in\{0,1\}\). The coefficient normalization is
\[
\lambda=\sum_{i=0}^{2^n-1}\sum_{j=0}^{2^n-1} |\tilde\alpha_{ij}|.
\]
The right and left preparation oracles are
\[
P_R\ket{0}_{2n}= \frac{1}{\sqrt{\lambda}}\sum_{i,j} e^{i\arg(\tilde\alpha_{ij})}\sqrt{|\tilde\alpha_{ij}|}\ket{i}\ket{j},
\]
\[
P_L\ket{0}_{2n}= \frac{1}{\sqrt{\lambda}}\sum_{i,j}\sqrt{|\tilde\alpha_{ij}|}\ket{i}\ket{j}.
\]

The resulting FOQCS-LCU circuit has the form
\[
U_{\rm fox}=(P_L^\dagger\otimes I)\,\mathrm{select}\,(P_R\otimes I),
\]
acting on \(2n\) ancilla qubits and an \(n\)-qubit system register. Postselection of the ancillas onto \(\ket{0}_{2n}\) yields the normalized action of \(H\) on the system. The block-encoding notion used is the standard one: a unitary \(U\) is an \((\alpha,a,\varepsilon)\)-block encoding of \(A\) if
\[
\left\|A-\alpha(\bra{0^a}\otimes I)\,U\,(\ket{0^a}\otimes I)\right\|\le \varepsilon.
\]
In the idealized circuit formulas used in the matrix-polynomial construction, the constructions are exact and \(\varepsilon=0\) [2601.18767].

This decomposition can also be described in check-matrix language. For a single qubit,
\[
I \leftrightarrow (0,0),\quad Z \leftrightarrow (0,1),\quad X \leftrightarrow (1,0),\quad Y \leftrightarrow (1,1),
\]
with
\[
\sigma_{m_\ell}=(-i)^{i_\ell z_\ell} Z^{z_\ell}X^{i_\ell}.
\]
FOQCS-LCU stores the two binary strings \(i=[i_{n-1}\cdots i_0]\) and \(z=[z_{n-1}\cdots z_0]\) directly in ancilla registers, so the Pauli support pattern is encoded distributively rather than by a single binary term label [2507.20887].

## 3. `SELECT` compression and structure-aware `PREPARE`

The distinctive circuit innovation of FOQCS-LCU is the `SELECT` oracle. Instead of conditionally applying one entire \(U_m\) under a multi-qubit control register, FOQCS-LCU applies the \(X\)- and \(Z\)-patterns sitewise:
\[
\mathrm{SELECT}_{\rm FOQCS}
=
\left(\prod_{\ell=0}^{n-1}\mathrm{C}X_{i_\ell\to s_\ell}\right)
\left(\prod_{\ell=0}^{n-1}\mathrm{C}Z_{z_\ell\to s_\ell}\right),
\]
where \(s_\ell\) is the \(\ell\)-th system qubit. For a fixed ancilla basis state \(\ket{i}\ket{z}\),
\[
\ket{i}\ket{z}\ket{\varphi}\mapsto \ket{i}\ket{z}\left(\bigotimes_{\ell=0}^{n-1} Z^{z_\ell}X^{i_\ell}\right)\ket{\varphi}.
\]
Because the controlled-\(X\) gates act on distinct targets, they form one parallel layer; the controlled-\(Z\) gates form a second parallel layer. Accordingly, the logical `SELECT` depth is exactly two: one layer of \(n\) CNOTs and one layer of \(n\) CZs, with no Toffolis and no generic multiply controlled unitaries [2507.20887].

This is only half of the FOQCS-LCU story. Once `SELECT` becomes cheap, `PREPARE` becomes the dominant cost. The 2025 construction therefore develops tailored Dicke-state routines for coefficient loading. The relevant state families include the single-excitation Dicke state
\[
|D^n_1\rangle = \frac{1}{\sqrt{n}}\sum_{\ell=0}^{n-1}|2^\ell\rangle,
\]
its unbalanced version
\[
|D^n_1(\alpha)\rangle = \sum_{\ell=0}^{n-1}\alpha_\ell |2^\ell\rangle,\qquad \|\alpha\|_2=1,
\]
the constrained two-excitation states
\[
|D^n_{2,k_N}\rangle = \frac{1}{\sqrt{n-k_N}} \sum_{\ell=0}^{n-k_N-1}|2^\ell+2^{\ell+k_N}\rangle,
\]
and the correlated “double Dicke” states such as
\[
|D^{n}_{1,\mathrm{double}}\rangle = \frac{1}{\sqrt{n}}\sum_{\ell=0}^{n-1}|2^\ell\rangle|2^\ell\rangle.
\]
These states encode one-body terms, fixed-distance two-body terms, and the correlated \(X/Z\) support needed for \(Y\)-type terms [2507.20887].

The reported CNOT counts are explicit. The basic subroutine \(\mathrm{subr}(\theta)\) uses \(2\) CNOTs; \(\mathrm{staircase}(n)\) uses \(2n-2\); \(\mathrm{CChain}(k_N)\) uses \(n-k_N\); and \(\mathrm{copygate}\) uses \(n\). For the prepared states themselves,
\[
|D^n_1\rangle: 2n-2,\qquad
|D^n_{2,\mathrm{NN}}\rangle: 3n-5,\qquad
|D^n_{2,k_N}\rangle: 3n-3k_N-2,
\]
\[
|D^n_{1,\mathrm{double}}\rangle: 3n-2,\qquad
|D^n_{2,\mathrm{NN},\mathrm{double}}\rangle: 4n-5,\qquad
|D^n_{2,k_N,\mathrm{double}}\rangle: 4n-3k_N-2.
\]
The unbalanced versions have the same CNOT count as the balanced ones; only phase gates and rotation angles change [2507.20887].

## 4. Products, powers, and matrix polynomials

The 2026 extension uses FOQCS-LCU as an input primitive for matrix polynomial block encoding. If \(M_1\) and \(M_2\) each admit FOQCS-LCU encodings, then their product \(M_2M_1\) can be block encoded by stacking FOQCS gadgets with separate ancilla pairs. Repeating this \(k\) times yields a direct block encoding of \(H^k\) using \(k\) copies of \(P_R\), \(P_L^\dagger\), and \(k\) `select` layers. The construction is explicitly not a nested reuse of the original block encoding in the QSVT sense; it is assembled as a larger LCU-type gadget [2601.18767].

To pass from powers to a polynomial,
\[
p_d(H)=\sum_{k=0}^d a_k H^k,
\]
the paper adds an outer LCU indexed by \(k\). The degree register is encoded in unary:
\[
\ket{0_u}=\ket{000\cdots0},\quad
\ket{1_u}=\ket{100\cdots0},\quad
\ket{2_u}=\ket{110\cdots0},\ \dots,\ 
\ket{d_u}=\ket{111\cdots1}.
\]
This avoids multi-controlled selection of the power \(H^k\): the first \(k\) ones in the unary string activate the first \(k\) FOQCS-LCU blocks.

The polynomial coefficients are folded into the outer LCU by
\[
w_k:=\sqrt{|a_k|\,\lambda^k},\qquad
\phi_k:=\arg(a_{k+1})-\arg(a_k),\qquad
W=\sum_{k=0}^d |a_k|\lambda^k.
\]
The outer preparation oracles are
\[
P_{{\rm poly},R}\ket{0}_d
=
\frac{1}{\sqrt{W}}
\sum_{k=0}^{d} e^{i\sum_{j=0}^{k-1}\phi_j} w_k \ket{k_u},
\]
\[
P_{{\rm poly},L}\ket{0}_d
=
\frac{1}{\sqrt{W}}
\sum_{k=0}^{d} w_k \ket{k_u}.
\]
Their unary state-preparation circuits are explicit CRY ladders with parallel phase gates, using
\[
\theta_k:=2\arccos\!\left( \frac{|w_k|}{\sqrt{1-\sum_{p=0}^{k-1}|w_p|^2}} \right),
\]
and each of \(P_{{\rm poly},R}\) and \(P_{{\rm poly},L}\) requires only \(2(d-1)\) CNOTs [2601.18767].

The main theorem states that the
\[
(d+(2d+1)n)\text{-qubit}
\]
circuit implements a block encoding of \(p_d(H)\). After postselection, the constructed action is exactly the desired polynomial up to a global phase:
\[
\sum_{k=0}^d e^{i\sum_{j=0}^{k-1}\phi_j}|a_k|H^k\ket{\varphi}
=
e^{-i\arg(a_0)}p_d(H)\ket{\varphi}.
\]

The main performance claim is a depth statement. Standard approaches such as QSVT give depth scaling
\[
\Theta\!\bigl(d\times \text{depth of block encoding of }H\bigr),
\]
so for spin chains with block-encoding depth linear in \(n\), total depth is \(O(dn)\). FOQCS-LCU changes this to an additive form: the extra depth overhead for going from \(H\) to \(p_d(H)\) is linear in \(d\) with no dependence on system size or on the cost of block encoding \(H\). For the spin models analyzed, the total CNOT depth has the form
\[
c_1 n + c_2 d,
\]
rather than \(O(dn)\) [2601.18767].

## 5. Trivial control and explicit resource estimates

A second defining property of FOQCS-LCU is that controlled block encodings are unusually cheap. The key lemma is structural: if a unitary decomposes as
\[
U=B_sA_s\cdots B_1A_1
\]
and the \(B_i\) share a common eigenstate \(\ket{\xi}\) with eigenvalue \(1\), then on input \(\ket{\xi}\) only the \(A_i\) need to be controlled. FOQCS-LCU satisfies this because
\[
\mathrm{select}\ket{0}_{2n}\ket{\varphi}=\ket{0}_{2n}\ket{\varphi}
\]
for any system state \(\ket{\varphi}\). Thus, when the ancillas start in \(\ket{0}_{2n}\), one controls only \(P_R\) and \(P_L^\dagger\), not the `select` layer [2601.18767].

The later paper adds an assumption that
\[
P_R=P_R^{\rm mod}\,P_R^{c},\qquad
P_L=P_L^{\rm mod}\,P_L^{c},
\]
where \(P_R^{\rm mod}\) and \(P_L^{\rm mod}\) have \(\ket{0}_{2n}\) as eigenstate with eigenvalue \(1\), while \(P_R^{c}\) and \(P_L^{c}\) contain only \(O(1)\) single-qubit gates. For the spin-model circuits considered there, \(P_R^{c}\) and \(P_L^{c\dagger}\) are each a single \(X\) gate. Consequently, the controlled FOQCS-LCU block encoding requires only two extra CNOTs. For matrix polynomials the simplification is even stronger: controlling the full polynomial block encoding requires controlling only the first \(R_y\) in \(P_{{\rm poly},R}\) and the last \(R_y\) in \(P_{{\rm poly},L}^\dagger\), giving at most four extra CNOTs total, independent of \(n\) and \(d\) [2601.18767].

For the 1D XYZ Heisenberg model,
\[
H = g\sum_{i=0}^{n-1} Z_i+\sum_{i=0}^{n-2} J_xX_iX_{i+1}+J_yY_iY_{i+1}+J_zZ_iZ_{i+1},
\]
the non-asymptotic resource counts in the polynomial paper are:

| Construction | CNOT depth (all-to-all) | Qubits |
|---|---:|---:|
| FOQCS-LCU | \(8n+10\) | \(3n+1\) |
| Controlled FOQCS-LCU | \(8n+12\) | \(3n+1\) |
| Polynomial \(p_d(H)\) | \(8n+6d+6\) | \(2dn+d+n+1\) |
| Controlled polynomial \(p_d(H)\) | \(8n+6d+10\) | \(2dn+d+n+1\) |

On square-grid hardware the corresponding XYZ depths are \(8n+20\), \(8n+22\), \(8n+10d+12\), and \(8n+10d+16\), respectively. The same additive pattern appears for the XXZ and Ising examples:
\[
\text{depth}_{\rm XXZ}(p_d(H))=4n+6d+10\quad\text{(all-to-all)},
\]
\[
\text{depth}_{\rm XXZ}(p_d(H))=4n+10d+12\quad\text{(square grid)},
\]
\[
\text{depth}_{\rm Ising}(p_d(H))=4n+6d-2\quad\text{(all-to-all)},
\]
\[
\text{depth}_{\rm Ising}(p_d(H))=4n+10d+8\quad\text{(square grid)}.
\]
The practical point is explicit in the formulas: total gate count still grows like \(dn\), but depth is additive in \(n\) and \(d\), not multiplicative [2601.18767].

## 6. Spin-model scope, limitations, and place within LCU research

FOQCS-LCU has been worked out explicitly for spin Hamiltonians such as the 1D XYZ Heisenberg, 1D XXZ Heisenberg, and 1D Ising models, with square-grid nearest-neighbor mappings and detailed non-asymptotic gate counts [2601.18767]. In the original construction, representative applications also included Heisenberg and spin-glass Hamiltonians, where the check-matrix representation and Dicke-state preparation expose repeated support patterns that standard structure-agnostic LCU would not exploit efficiently. For those examples, the paper reports an order-of-magnitude improvement in CNOT count over conventional LCU, and states that FOQCS-LCU scales as \(\mathcal O(n)\) for Heisenberg and \(\mathcal O(n^2)\) for spin glass [2507.20887].

The framework is nonetheless limited by its central tradeoff. The base block encoding uses \(2n\) ancillas for the \(X/Z\) support registers, and the polynomial construction uses
\[
d+(2d+1)n
\]
qubits in the general theorem, with explicit tables giving
\[
2dn+d+n+1
\quad\text{or}\quad
2dn+2d+n+1
\]
depending on architecture [2601.18767]. This suggests that FOQCS-LCU is depth-efficient rather than qubit-frugal. A second limitation is dependence on efficient implementations of \(P_R\) and \(P_L\). For arbitrary matrices, naive state preparation would be exponentially costly; the practical claims in the FOQCS-LCU papers rely on structured Hamiltonians whose coefficients admit compact Dicke-state-based or model-specific preparations [2507.20887, 2601.18767].

A third limitation is the usual normalization and postselection cost of LCU. FOQCS-LCU uses the same \(\ell_1\)-type normalization as ordinary Pauli-string LCU:
\[
\lambda=\sum_{i,j}|\tilde\alpha_{ij}|,
\]
and for matrix polynomials the outer normalization is
\[
W=\sum_{k=0}^d |a_k|\lambda^k.
\]
The papers note that \(\lambda\) is not worse than standard Pauli-string LCU normalization, but postselection cost remains inherent [2601.18767].

Within the broader LCU landscape, FOQCS-LCU occupies a specific niche. It is a deterministic, Pauli-structured, ancilla-rich method that makes coherent block encodings shallow and easily controllable. Other recent LCU directions instead trade coherence for sampling overhead through randomization or grouped virtual/coherent hybrids, using Hadamard-test-based estimators or reduction-factor tradeoffs rather than constant-depth `SELECT` oracles [2506.15658, 2512.06260]. A plausible implication is that FOQCS-LCU is complementary to those approaches: it addresses the circuit-depth bottleneck of coherent block encoding, whereas randomized and hybrid LCU frameworks address the depth–sampling tradeoff by weakening the requirement of exact coherent realization.

In that sense, FOQCS-LCU is not simply “an LCU method.” It is a particular answer to the question of how one should compile structured Pauli LCUs when shallow depth and inexpensive control are the primary constraints: store support explicitly, parallelize `SELECT`, encode coefficients with structured Dicke-state routines, and exploit the trivial eigenstate of `select` to make controlled versions nearly free [2507.20887, 2601.18767].

Source: https://www.emergentmind.com/topics/foqcs-lcu