---
title: Covariant Approximation Averaging in Lattice QCD
url: https://www.emergentmind.com/topics/covariant-approximation-averaging-caa
type: topic
---

# Covariant Approximation Averaging in Lattice QCD

Covariant Approximation Averaging (CAA) is a symmetry-based error reduction strategy in lattice QCD that constructs improved stochastic estimators for correlation functions by combining a low-cost, symmetry-covariant approximation with a single unbiased correction from an exact computation. The main principle is to leverage exact lattice symmetries (such as translations) to generate multiple low-cost evaluations of an approximate observable, average them, and remove all bias via a single correction, yielding a statistically improved and unbiased estimator applicable to two-point, three-point, and more complex correlation functions. CAA is particularly effective in applications requiring all-to-all quark propagators and has been realized as all-mode averaging (AMA) in large-scale lattice computations [1402.0244][1212.5542][2111.15138].

## 1. Formal Definition and Theoretical Framework

Let $S$ represent the exact quark propagator and $\mathcal{O}[S]$ any derived observable (e.g., a hadron $n$-point function). CAA constructs an improved estimator as follows. Define a group $G$ of lattice symmetries (typically site translations), and an approximation $S^{(\rm appx)}$ (commonly a low-mode truncated or a loose-tolerance CG solution) such that
\[
S = S^{(\rm appx)} + (S - S^{(\rm appx)}).
\]
The observable splits accordingly. For each $g \in G$, $\mathcal{O}^{(\rm appx),g}$ is the observable constructed from $S^{(\rm appx)}$ at the $g$-shifted configuration or source. Defining
\[
\mathcal{O}_G^{(\rm appx)} = \frac{1}{|G|} \sum_{g\in G} \mathcal{O}^{(\rm appx),g}, \qquad \mathcal{O}^{(\rm rest)} = \mathcal{O}[S] - \mathcal{O}[S^{(\rm appx)}],
\]
the CAA/AMA estimator is
\[
\mathcal{O}_{\rm CAA} = \mathcal{O}^{(\rm rest)} + \mathcal{O}_G^{(\rm appx)}.
\]
By the covariance of the lattice action and approximation, CAA is strictly unbiased:
\[
\langle \mathcal{O}_{\rm CAA} \rangle = \langle \mathcal{O}[S] \rangle.
\]
This construction reduces variance by exploiting the statistical independence among symmetry-related approximations while correcting any remaining bias from a single exact computation.

## 2. Methodology and Implementation

The canonical implementation uses translation symmetry on the lattice. The workflow proceeds as follows:

1. **Eigenmode Computation**: Compute $N_\lambda$ low eigenmodes $\{\psi_k, \lambda_k\}$ of the Hermitian Dirac kernel $H=\gamma_5 D$. Construct the low-mode part
   \[
   S_{\rm low}(x,y) = \sum_{k=1}^{N_\lambda} \lambda_k^{-1} \psi_k(x)\psi_k^\dagger(y).
   \]
2. **Approximate Propagator Construction**: For each symmetry $g\in G$ (e.g., $|G|=32$ or 64 spatial translations or shifts),
   - Generate $S^{(\rm appx),g}$ by injecting a shifted source and applying a low-cost approximation (e.g., low-mode truncation or relaxed CG, possibly combined with deflation), reusing precomputed modes.
   - Construct the approximate observable $\mathcal{O}^{(\rm appx),g}$.
3. **Exact Correction**: At a reference point (typically a single source location), compute the full exact observable $\mathcal{O}[S]$.
4. **Estimator Assembly**: The improved observable is the mean over all approximations plus the difference between the exact and its approximation at the reference:
   \[
   \mathcal{O}_{\rm CAA} = \mathcal{O}_G^{(\rm appx)} + \left(\mathcal{O}[S] - \mathcal{O}[S^{(\rm appx)}]\right).
   \]

A typical pseudocode (as explicitly described in [2111.15138]):

```
1. Compute {λ_i, v_i(x)} for i = 1..N_eig (Lanczos)
2. Form D_low^{-1}(x,y) = Σ_i (1/λ_i) v_i(x) v_i†(y)
3. For each translation g_j in G (j = 1..N_ave):
     a. Let sink position = g_j·x₀
     b. Construct O_app_j via one-end trick + sequential, using D_low^-1 at sink
4. O_app_avg = sum of O_app_j over G / N_ave
5. Compute O_exact at x₀ (full all-to-all, including noise for high-modes as needed)
6. CAA estimator: O_CAA = O_app_avg + [O_exact(x₀) - O_app(x₀)]
```
In large-scale applications such as the HAL QCD $\rho$-resonance study, $N_{\text{eig}}=300$ and $N_{\text{ave}}=64$ were typical values [2111.15138].

## 3. Performance and Error Reduction

The statistical error of the CAA estimator is determined by the correlation coefficient $r$ between the exact and approximate observable:
\[
\mathrm{Var}[\mathcal{O}_{\rm CAA}] \simeq \mathrm{Var}[\mathcal{O}] [2(1-r) + |G|^{-1}]
\]
and
\[
\Delta_{\rm CAA} \simeq \Delta_{\rm orig} \sqrt{2(1-r) + 1/|G|}.
\]
When $r \to 1$ (highly correlated approximation), the error scales as $1/\sqrt{|G|}$, approaching the optimum for independent measurements at cost $\sim 1/|G|$ of the naive approach [1212.5542][1402.0244].

In the HAL QCD $\rho$-resonance application, this yielded roughly a tenfold reduction in statistical error: uncertainties on the leading-order potential at $r \approx 0.5$ fm decreased from $\mathcal{O}(20–30\%)$ to $\mathcal{O}(2–3\%)$ per time-slice, sufficient to enable the first stable N$^2$LO extraction in a $P$-wave channel [2111.15138]. Similarly, for nucleon two- and three-point functions, error-reduction factors close to $1/\sqrt{32}\approx 0.18$ were observed, with overall cost savings of $5$–$25\%$ compared to standard methods [1212.5542][1402.0244].

| Observable              | Standard Error | CAA/AMA Error Ratio | Normalized Cost Ratio |
|-------------------------|---------------|---------------------|----------------------|
| $m_N$                   | 156 MeV       | 0.17                | 0.04                 |
| $m_\pi$                 | 12 MeV        | 0.36                | 0.19                 |
| $m_V$                   | 176 MeV       | 0.20                | 0.06                 |

## 4. Applications in Lattice QCD

CAA is widely applicable in the context of all-to-all propagators, essential for multi-hadron and resonance studies with spatially extended operators and non-localities. Major applications include:

- **HAL QCD Method**: Used for high-precision, non-local potential extraction, especially in channels requiring the Nambu–Bethe–Salpeter (NBS) wave function with explicit spatial dependence. In the $\rho$-resonance calculation, CAA was combined with the one-end trick and sequential propagator techniques, enabling efficient construction of many spatially displaced correlators with minimal additional cost while rigorously correcting for all bias at a single point [2111.15138].
- **Hadron Structure**: CAA/AMA yields an order-of-magnitude reduction in cost-to-error for nucleon and meson two-point/three-point functions, form factors, and disconnected diagrams [1212.5542][1402.0244].
- **Broader Observables**: The method is compatible with a wide array of fermion discretizations (domain-wall, Wilson, twisted-mass, overlap) and can be extended to multi-current matrix elements and flavor singlet channels [1402.0244].

## 5. Advantages, Limitations, and Extensions

### Advantages

- **Unbiasedness**: CAA preserves the expectation value of the target observable exactly by construction under the relevant group symmetry [1402.0244].
- **Dramatic Variance Reduction**: Effective statistics are amplified by $|G|$ with only one exact measurement per configuration required.
- **Compatibility with Other Tricks**: Integrates seamlessly with the one-end trick and sequential propagator techniques, eliminating stochastic noise at the sink [2111.15138].
- **Reduced Solver Count**: Substantial decrease in the number of Dirac-operator inversions compared to fully stochastic all-to-all approaches.

### Limitations

- **Quality of Approximation**: The approximation $S^{({\rm appx})}$ must be highly correlated with $S$; otherwise, residual fluctuations can dominate the estimator's variance. The balance between low-mode computation ($N_{\text{eig}}$) and number of symmetry shifts ($|G|$) is critical to optimize performance [2111.15138][1402.0244].
- **Computation of Low Modes**: Calculating and storing many low eigenvectors incurs significant initial cost, albeit amortized over many measurements.
- **Symmetry Scope**: In channels with additional symmetry mixing (rotations, parity), $G$ may need to be expanded beyond simple translations.

### Extensions

- **Inclusion of Rotational/Reflection Symmetries**: Amplifies effective statistics beyond translations [2111.15138].
- **Combination with Hierarchical Probing, High-Mode Deflation**: Targets further reduction of residual variance.
- **Generalization to Disconnected Diagrams, Multi-current Matrix Elements, High-momentum Baryon Observables**: Demonstrated in nucleon and meson matrix element calculations [1212.5542][1402.0244].

## 6. Practical Considerations and Outlook

Implementation of CAA requires careful attention to covariance under lattice group actions; in practice, fixed iteration counts or randomized shifts can be used to eliminate rounding-induced biases [1402.0244]. The cost of low-mode computation becomes negligible as the scope of observable averaging grows.

In all tested scenarios—including $\rho$-resonance potentials, nucleon form factors, and meson correlators—CAA and its variants (notably AMA) enabled factors of $3$–$10$ statistical error reduction for fixed computational resources, establishing CAA as an essential tool in modern lattice QCD [1402.0244][1212.5542][2111.15138]. A plausible implication is that the broad applicability and systematic unbiasedness of CAA make it broadly relevant to forthcoming high-precision studies in hadronic physics, spectrum determination, and nuclear interactions.

Source: https://www.emergentmind.com/topics/covariant-approximation-averaging-caa