---
title: 'LogQ Parameterization: Phase-Based QUBO Encoding'
url: https://www.emergentmind.com/topics/logq-parameterization
type: topic
---

# LogQ Parameterization: Phase-Based QUBO Encoding

LogQ parameterization is a phase-based encoding of Quadratic Unconstrained Binary Optimization (QUBO) variables in which each discrete spin is represented not by a binary register value or a real interval variable, but by a unit-modulus complex phase
\[
f(\theta_i)=e^{-i\pi R(\theta_i)},
\]
with \(\theta_i\in\mathbb R\) continuous and \(R(\theta_i)\) driven toward \(0\) or \(1\) so that \(f(\theta_i)\to \pm1\). In the formulation developed for QUBO, this encoding was introduced within a logarithmic-qubit quantum ansatz and later shown to admit an exact classical reformulation as a non-linear continuous relaxation on the complex unit circle. The term “LogQ” refers to logarithmic qubit scaling rather than logarithms in the objective, and the construction is explicitly described as neither a probability parameterization, nor a simplex relaxation, nor a standard logit parameterization [2604.12925].

## 1. Discrete optimization problem and motivation

LogQ is formulated for the spin form of QUBO. Instead of binary variables \(x_i\in\{0,1\}\), it uses spins \(s_i\in\{-1,1\}\), related by
\[
s_i=2x_i-1.
\]
The target optimization problem is
\[
s^*=\underset{s \in \{-1,1\}^{n}}{\arg\min}\; \left(-\frac{1}{2}\sum_{i,j=0}^{n-1} s_i Q_{ij} s_j\right).
\]
This is the baseline discrete problem that LogQ parameterizes [2604.12925].

A conventional continuous relaxation would replace each spin by a real variable in \([-1,1]\),
\[
s_i\rightarrow s_i^{(\mathrm{lin})}\in[-1,1],
\]
and then threshold the optimized value at zero. The LogQ motivation is that such a linear relaxation is weak because it does not enforce \(|s_i|\approx 1\). LogQ instead relaxes the variables onto the complex unit circle, where the modulus is fixed to \(1\) and the binary values \(\pm1\) appear as special points. This suggests that the method is best understood as a relaxation that preserves a strong analogue of the discrete spin norm while allowing continuous optimization over latent phase coordinates [2604.12925].

From a modeling standpoint, the decisive substitution is not
\[
s_i\mapsto s_i^{(\mathrm{lin})},
\]
but
\[
s_i\mapsto f(\theta_i)=e^{-i\pi R(\theta_i)}.
\]
The resulting scheme is therefore unlike standard LP relaxations, unlike SDP or vector relaxations, and unlike variational quantum encodings that use one qubit per logical variable.

## 2. Encoding map, latent variables, and decoding

The LogQ parameterization has three distinct levels. The actual optimization variables are the continuous real parameters \(\theta_0,\dots,\theta_{n-1}\). The intermediate coordinate is \(R(\theta_i)\in[0,1]\), which serves as a relaxed phase coordinate. The embedding used by the algorithm is the unit-modulus complex number
\[
f(\theta_i)=e^{-i\pi R(\theta_i)}.
\]
At convergence, the intended condition is
\[
R(\theta_i^*)=0 \text{ or } 1,
\]
so that
\[
f(\theta_i^*)\in\{1,-1\},
\]
and the recovered spin is
\[
s_i^*=\begin{cases}
-1 & \text{if } R(\theta_i^*)=0,\\
+1 & \text{if } R(\theta_i^*)=1.
\end{cases}
\]
Thus the \(\theta_i\) are latent continuous controls rather than relaxed spins themselves, \(R(\theta_i)\) is the relaxed coordinate, and \(f(\theta_i)\) is the actual phase/amplitude embedding [2604.12925].

The paper emphasizes that the only genuine modeling freedom in LogQ is the choice of \(R(\cdot)\). The stated sufficient conditions for a useful parameterization are:
\[
R(\theta_i)\in[0,1]\ \text{and differentiable almost everywhere,}
\]
\[
R'(\theta_i)=0 \Rightarrow R(\theta_i)\in\{0,1\},
\]
together with the existence of large intervals such that \(|R'(\theta_i)|\) is not small, and the condition that only “small barrier” local minima exist in \(C(\theta)\). These conditions are motivated entirely by trainability. Earlier LogQ variants exhibited vanishing gradients and therefore required evolutionary or global optimization; the revised design criterion is to preserve the binary endpoint structure without destroying useful gradient signal [2604.12925].

Two common misconceptions are explicitly excluded by the source formulation. First, LogQ is not a probability parameterization and does not encode spins through a simplex. Second, the name does not refer to logarithmic terms in the surrogate objective; it refers to logarithmic qubit scaling.

## 3. Quantum ansatz and logarithmic-qubit structure

In the original quantum formulation, the variational objective is
\[
\theta^*=\underset{\theta\in \mathbb{R}^n}{\arg\min}\; C(\theta),
\]
with
\[
C(\theta)=-2^{N-2}\bra{\Psi(\theta)}\hat{L}\ket{\Psi(\theta)},
\qquad
\ket{\Psi(\theta)} = \frac{1}{n}\sum_{i=0}^{n-1}f(\theta_i)\ket{i},
\]
and
\[
N=\lceil \log_2 n\rceil.
\]
The state uses one amplitude per original variable index \(i\), but the amplitudes are constrained to equal magnitude and differ only by phase. It can also be written as
\[
\ket{\Psi(\theta)}= U(\theta)H^{\otimes N}\ket{0}^{\otimes N},
\]
where \(U(\theta)\) implements the diagonal phases on top of the uniform superposition [2604.12925].

This is the structural source of the quantum efficiency claim. Since the \(n\) logical variables are indexed by computational basis states \(\ket{i}\), the method uses only
\[
N=\lceil \log_2 n\rceil
\]
qubits to represent \(n\) amplitudes. The formulation explicitly states that this yields “an exponential reduction in the number of qubits and a quadratic reduction in quantum circuit depth” compared to standard encodings such as QAOA’s computational-basis encoding [2604.12925]. A related analysis states that QAOA needs \(n\) qubits for a QUBO with \(n\) variables, whereas LogQ needs only \(\lceil \log_2 n\rceil\), and that the LogQ circuit uses about \(2^N\approx n\) CNOTs, versus \(p(n^2-n)\) for \(p\)-layer QAOA [2507.08489].

The cost operator is expanded as
\[
\hat{L}=\dfrac{1}{n}\sum\limits_{k=1}^{n^2}\mathrm{Tr}(J_k Q)J_k,
\]
where the \(J_k\) are the \(n^2=4^N\) tensor products of \(N\) Pauli matrices and identities. Because \(Q\) is symmetric, at most \((n^2+n)/2\) components and the same number of Pauli-basis measurements are needed. This Pauli decomposition and measurement burden later becomes a central target of the classical reformulation [2604.12925].

## 4. Exact classical reformulation as a non-linear continuous relaxation

The central reformulation result is that the same phase parameterization can be written directly as a classical optimization problem:
\[
\theta^*=\underset{\theta \in \mathbb{R}^n}{\arg\min}\;-\frac{1}{2}\sum_{i,j=1}^n f^\dagger(\theta_i) Q_{ij} f(\theta_j),
\]
subject to
\[
|f(\theta_i)|=1,
\qquad
f(\theta_i^*)\in\{-1,1\}.
\]
This is a continuous relaxation because the \(\theta_i\) are continuous, but it is not a relaxation into a real interval. It is a relaxation into the unit circle in \(\mathbb C\), and the modulus constraint enforces a strong analogue of \(|s_i|=1\) [2604.12925].

Substituting
\[
f(\theta_i)=e^{-i\pi R(\theta_i)}
\]
and using symmetry of \(Q\), the paper derives the real-valued surrogate
\[
\frac{1}{2}\sum_{i,j=1}^n f^\dagger(\theta_i) Q_{ij} f(\theta_j)
=
\sum_{i>j=1}^n \cos\!\big(\pi(R(\theta_i)-R(\theta_j))\big) Q_{ij} + ctnt.
\]
Equivalently, the discrete spin products are replaced by cosine couplings of phase differences:
\[
s_i s_j
\quad\leadsto\quad
\cos\!\big(\pi(R(\theta_i)-R(\theta_j))\big).
\]
The analogy emphasized in the formulation is
\[
s^{(\mathrm{lin})}_i s^{(\mathrm{lin})}_j\in[-1,1]
\quad\leftrightarrow\quad
\cos\!\big(\pi(R(\theta_i)-R(\theta_j))\big)\in[-1,1].
\]
This is the final classical surrogate induced by the LogQ parameterization [2604.12925].

The connection between the quantum and classical views is exact at the level of parameterization. Both use the same \(\theta_i\), the same \(R(\theta_i)\), and the same map \(f(\theta_i)=e^{-i\pi R(\theta_i)}\). The quantum version evaluates the cost as an expectation value of a Pauli-decomposed operator on a logarithmic-qubit state; the classical version analytically collapses that expectation into a deterministic nonlinear objective involving pairwise cosines. The practical consequences stated in the source are the elimination of Pauli decomposition and the bypassing of measurement overhead [2604.12925].

The practical algorithm is correspondingly simple in outline: choose \(R\), initialize \(\theta\in\mathbb R^n\), optimize the quantum or classical cost over \(\theta\), then decode each optimized coordinate using \(R(\theta_i^*)\in\{0,1\}\) and map to \(s_i^*\in\{-1,1\}\). No initialization rule is specified, and no optimizer pseudocode is given for the 2026 reformulation paper.

## 5. Trainability and gradient-compliant parameterizations

A major issue in the earlier LogQ literature is that trainability depends strongly on the choice of \(R\). In the 2025 development focused on improving the scheme, the original parameterization is the step function
\[
R^{(0)}(\theta_z)=
\begin{cases}
0 & \text{if } \theta_z \in [0,\pi[,\\
1 & \text{if } \theta_z \in [\pi,2\pi].
\end{cases}
\]
This is ideal from an encoding standpoint because it directly produces binary outputs, but it has
\[
R'(\theta_z)=0
\]
almost everywhere, so gradients vanish almost everywhere and the scheme is “not optimizable using gradient-inspired methods.” Earlier implementations therefore relied on genetic algorithms and other evolutionary or global methods [2507.08489].

The first smoothing considered is the logistic sigmoid
\[
R^{(1)}_\lambda(\theta_z)=\frac{1}{1+e^{\lambda(\pi-\theta_z)}}.
\]
This introduces continuous gradients, but the source analysis states that large \(\lambda\) recreates broad plateaus while small \(\lambda\) softens the encoding too much. The proposed replacement is the “distorted sigmoid”
\[
R^{(2)}_{\lambda,\kappa}(\theta_z)= \text{sgm}_\lambda(\pi-\theta_z)\times\text{sgm}_{-\lambda}(\kappa\pi+2\pi-\theta_z) +\text{sgm}_{\lambda}(\kappa\pi-\theta_z),
\]
with
\[
0\le\kappa<\gamma\le 1.
\]
The optimization domain is extended from \([0,2\pi]\) to
\[
[-\gamma\pi,(2+\gamma)\pi],
\]
with the numerical choice \(\gamma=0.6\), and in practice \(\kappa=0.2\) is used [2507.08489].

This trainability analysis is directly relevant to the later classical surrogate. The 2026 paper does not print an explicit gradient formula for the classical objective, but it makes clear that derivatives involve chain-rule factors \(R'(\theta_i)\) and phase-difference terms. A plausible implication is that the same design logic for \(R\) persists after dequantization: if \(R'\) vanishes frequently away from the binary endpoints, the classical surrogate will also exhibit poor gradient signal [2604.12925].

The 2025 paper packages the improved trainable scheme as LogQ-grad. Its practical recipe is: use \(R^{(2)}_{\lambda,\kappa}\) with \(\kappa=0.2\), restrict \(\theta_z\in[-0.6\pi,2.6\pi]\), start with \(\lambda=5\) or \(6\), use Cobyla with a large initial trust-region parameter \(\text{rhobeg}\approx 3\), decrease \(\text{rhobeg}\) over the run, perform a post-processing stage with \(\lambda=30\), and use multistart initialization. On MaxCut instances, the reported objective values are lower than those of the original GA-based LogQ: for \(n=50\), \(-238\) versus \(-219\); for \(n=128\), \(-1410\) versus \(-1325\); and for \(n=256\), \(-5383\) versus \(-5149\). The same source states that converged solutions satisfy the intended binary condition,
\[
R(\theta^*)=0 \text{ or } 1 \quad \text{only},
\]
for all coordinates [2507.08489].

## 6. Interpretation, limitations, and nomenclature

LogQ is best classified as a non-linear phase relaxation of binary variables. It is not an LP relaxation, not an SDP or vector relaxation, and not a conventional variational ansatz in which \(n\) qubits directly represent \(n\) bits. Its defining feature is that the binary structure is encoded through unit-modulus phases associated with amplitudes of a \(\log n\)-qubit register, and that this same encoding survives intact when the quantum machinery is removed [2604.12925].

The practical advantages stated for the method are logarithmically fewer qubits,
\[
N=\lceil \log_2 n\rceil,
\]
reduced circuit depth versus standard encodings, elimination of Pauli decomposition after classical reformulation, elimination of measurement overhead, and compatibility with gradient-inspired optimization if \(R\) is chosen well. The corresponding limitations are equally explicit: the method remains heuristic; success depends strongly on the choice of \(R(\cdot)\); the landscape may still contain plateaus or local minima, ideally only “small barrier” ones; and the papers do not establish guarantees of global optimality or approximation ratios [2604.12925].

The literature also contains several unrelated uses of similar terminology. “LogQuant” denotes a 2-bit KV-cache compression method whose “log” component is a base-2 log-distributed token-retention rule, not a formal method named LogQ [2503.19950]. In large-scale retrieval, “logQ correction” denotes a sampled-softmax importance-sampling adjustment that subtracts \(\log Q(d)\) or \(\log Q'(d)\) from sampled negative logits and, in its refined form, excludes the positive item from the sampled denominator [2507.09331]. These usages are terminological collisions rather than variants of the phase-based QUBO parameterization.

In concise form, LogQ parameterization denotes the substitution
\[
s_i \mapsto f(\theta_i)=e^{-i\pi R(\theta_i)},
\]
with \(R(\theta_i)\to\{0,1\}\) and \(f(\theta_i)\to\pm1\), together with a logarithmic-qubit amplitude encoding in the quantum formulation and an exactly corresponding cosine-coupled nonlinear surrogate in the classical formulation. Its significance lies less in any single optimizer than in the structural claim that a quantum-inspired phase encoding can define a distinct continuous relaxation of QUBO variables that remains meaningful even after the quantum layer is removed [2604.12925].

Source: https://www.emergentmind.com/topics/logq-parameterization