Symbol-Level Monte Carlo Simulation

• Symbol-level Monte Carlo simulations are techniques to numerically estimate symbol error rates (SER) by evaluating multidimensional Gaussian integrals over error regions.
• The ALOE method employs multiple importance sampling with truncated Gaussian proposals to focus on critical error events, greatly reducing estimator variance.
• Empirical results demonstrate that ALOE achieves orders-of-magnitude improved accuracy over naive Monte Carlo, especially in high SNR and non-standard lattice constellations.

Symbol-level Monte Carlo simulations are central to the numerical estimation of symbol error rates (SER) in advanced digital communication systems, especially for two-dimensional constellations formed by non-square or hexagonal lattices. Estimating SERs typically requires evaluating multi-dimensional integrals that are intractable analytically, leading to reliance on Monte Carlo (MC) methods. However, standard MC is often computationally inefficient, particularly at high signal-to-noise ratios (SNRs), motivating the adoption of multiple importance sampling (MIS) strategies such as the ALOE (“At Least One rare Event”) technique for vastly accelerating convergence and obtaining unbiased estimates with dramatically less computational effort (Elvira et al., 2019).

1. Mathematical Formulation of Symbol Error Rate Estimation

Let $\{s_1,\dots,s_M\} \subset \mathbb{R}^2$ denote the constellation points, with equal-probability transmission. For a transmitted symbol $s_m$, the received vector under additive white Gaussian noise (AWGN) is $x = s_m + n$, where $n \sim \mathcal{N}(0, \sigma^2 I_2)$. The Voronoi region (decision region) for $s_m$ is given as a convex polytope:

$$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$

which is the intersection of $K_m$ half-spaces. The symbol error probability conditioned on $s_m$ is

$$p_m = P \{ x \notin R_m \mid s_m \} = \int_{\mathbb{R}^2} 1_{\mathbb{R}^2 \setminus R_m}(x)\ \pi_m(x)\ dx,$$

where $\pi_m(x) = \mathcal{N}(x; s_m, \sigma^2 I_2)$ is the Gaussian density. The overall SER is then

$s_m$0

The error region $s_m$1 can be written as the union of $s_m$2 half-spaces $s_m$3, i.e., $s_m$4 (Elvira et al., 2019).

2. Multiple Importance Sampling with the ALOE Technique

ALOE applies multiple importance sampling (MIS) by constructing a proposal distribution as a mixture of $s_m$5 components, each focused on a different half-space error event:

$s_m$6

where $s_m$7 is a one-dimensional Gaussian tail integral, and $s_m$8 is the standard Gaussian tail function. Each component receives a mixture weight $s_m$9, where $x = s_m + n$0 forms a union-bound on $x = s_m + n$1.

The full proposal for symbol $x = s_m + n$2 is the mixture

$x = s_m + n$3

Sampling from this proposal focuses computational effort on the error region, giving rise to much lower estimator variance at high SNR.

3. Algorithmic Steps and Estimator Properties

The ALOE procedure is as follows:

1. Precompute for each facet ($x = s_m + n$4):
   • Hyperplane normals $x = s_m + n$5 and thresholds $x = s_m + n$6
   • Tail probabilities $x = s_m + n$7
   • Mixture weights $x = s_m + n$8
2. For each sample ($x = s_m + n$9 per symbol $n \sim \mathcal{N}(0, \sigma^2 I_2)$0):
   • Draw $n \sim \mathcal{N}(0, \sigma^2 I_2)$1 Categorical($n \sim \mathcal{N}(0, \sigma^2 I_2)$2)
   • Draw $n \sim \mathcal{N}(0, \sigma^2 I_2)$3 from the truncated Gaussian $n \sim \mathcal{N}(0, \sigma^2 I_2)$4 (sample $n \sim \mathcal{N}(0, \sigma^2 I_2)$5 with $n \sim \mathcal{N}(0, \sigma^2 I_2)$6)
   • Compute $n \sim \mathcal{N}(0, \sigma^2 I_2)$7, the number of half-spaces containing $n \sim \mathcal{N}(0, \sigma^2 I_2)$8
3. Weighted Estimation:

$n \sim \mathcal{N}(0, \sigma^2 I_2)$9

This estimator is unbiased, $s_m$0. The variance is tightly bounded:

$s_m$1

In contrast, naive MC's variance is $s_m$2, with a relative RMSE scaling as $s_m$3, implying MC requires $s_m$4 samples for order-one accuracy at low error rates, whereas ALOE’s variance falls rapidly as $s_m$5 at high SNR (Elvira et al., 2019).

4. Comparative Performance and Numerical Results

The method's empirical evaluation uses a 64-point "improper" lattice constellation (circularity $s_m$6), comparing:

• Naive Monte Carlo (MC) with $s_m$7
• Single-proposal IS with overdispersed Gaussians of variance $s_m$8 for $s_m$9
• ALOE MIS with $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$0

Over 200 independent repetitions and at high $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$1 (e.g., $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$2), ALOE achieves a relative error $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$3–$$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$4 times smaller than MC, indicating orders-of-magnitude fewer samples or runtime for the same relative accuracy. While the focus is on non-square lattices, similar gains are observed for hexagonal constellations and square QAM at high SNR, consistent with the underlying theory (Elvira et al., 2019).

5. Extension to General Constellations, Lattices, and Noise Models

ALOE applies to any two-dimensional constellation whose decision region is the complement of a convex polytope, i.e., an intersection of half-spaces. Extension to higher-dimensional lattices (such as those encountered in multiple-input multiple-output—MIMO—antenna arrays) is straightforward: the error region remains a union of half-spaces in $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$5, and the proposal mixture generalizes naturally, one truncated Gaussian per facet.

For non-Gaussian noise models (e.g., Laplacian or mixture distributions), the base density $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$6 in the proposal is replaced by the appropriate PDF, with the truncated proposal remaining the "base PDF restricted to $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$7." The same MIS bookkeeping applies, though the required tail integrals $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$8 must now be evaluated for the new base distribution. As SNR increases, $$R_m = \{ x \in \mathbb{R}^2 : a_{m,k}^T x < \beta_{m,k},\ k=1,\dots,K_m \},$$9 and ALOE’s variance approaches zero, reflecting that most errors occur through the nearest hyperplane. At low SNR, both naive MC and ALOE need only a modest number of samples, so the method is robust across all regimes (Elvira et al., 2019).

6. Principled Advantages and Theoretical Implications

ALOE transforms otherwise intractable two-dimensional SER integrals into the sum over $K_m$0 one-dimensional Gaussian (or more generally, noise-model-specific) tail integrals, using these for both proposal construction and estimator computation. The resulting estimator remains unbiased and its variance is always smaller than, and typically vastly outperforms, naive MC at moderate-to-high SNR. Importantly, by construction, every generated sample lies within the error region, ensuring that no simulated effort is wasted—each sample contributes positively to resolving low-probability error events. These properties indicate that symbol-level MC simulation via ALOE is particularly well-suited to regimes where rare-event analysis is essential, such as in evaluating deep-fade, high-SNR, or high-density lattice communication systems (Elvira et al., 2019).