---
title: Harmonic Mean Density Fusion
url: https://www.emergentmind.com/topics/harmonic-mean-density-hmd-fusion
type: topic
---

# Harmonic Mean Density Fusion

Harmonic Mean Density (HMD) fusion is an interpolation-based probabilistic fusion rule designed for robust combination of multiple local probability density functions (pdfs)—notably in distributed and track fusion scenarios where unknown cross-correlations or mismatches between local densities make conventional fusion methods inconsistent or overly conservative. HMD fusion is grounded in the power mean (Hölder mean) formalism at exponent $\alpha = -1$, and features a reciprocal formulation that prioritizes suppression of spurious or over-confident peaks, avoids double-counting of priors, and produces tight and consistent fused posteriors in both Gaussian and non-Gaussian contexts. Its mathematically principled foundations, computational tractability for Gaussian and Gaussian mixture models, and empirically demonstrated error reductions have led to its adoption and further development in distributed estimation, consensus tracking, and multisensor inference architectures [2412.06716], [2511.03130], [2202.11633], [2412.06725].

## 1. Mathematical Formulation

Given $N$ local densities $f_i(x)$ with weights $w_i$ summing to one, the harmonic mean density is defined as
\[
f_{HMD}(x) = \left( \sum_{i=1}^N w_i\, f_i(x)^{-1} \right)^{-1}
\]
with normalization constant $Z_{HMD} = \int \left( \sum_{i=1}^N w_i\,f_i(x)^{-1} \right)^{-1} dx$, yielding the normalized fused density. For the two-density, two-weight case (with $\omega \in [0,1]$),
\[
f_{HMD}(x) = \frac{f_1(x) f_2(x)}{(1-\omega) f_1(x) + \omega f_2(x)}  \Bigg/  \int \frac{f_1(u) f_2(u)}{(1-\omega) f_1(u) + \omega f_2(u)} du
\]
For more than two densities, a recursive structure or generalized weighted sum is used; for three densities with weights $(\omega_1, \omega_2, \omega_3)$, recursion is realized as
\[
\mathcal{M}^h_{\omega_1,\omega_2,\omega_3}(f_1, f_2, f_3) = \mathcal{M}^h_{\alpha,\omega_3}\left(\mathcal{M}^h_{\omega_1,\omega_2}(f_1,f_2),\,f_3\right)
\]
with $\alpha = \omega_1 / (\omega_1 + \omega_2)$. This reciprocal structure imposes “veto” behavior: the fused density vanishes wherever any local density vanishes, strongly penalizing isolated high-likelihood spikes.

## 2. Theoretical Properties

HMD fusion arises from multiple foundational perspectives:

- **Axiomatic Characterization**: HMD is the unique member of the power-mean ($\alpha$-mean) family at $\alpha = -1$ satisfying joint likelihood commutation (Likelihood Principle), weak setwise function property (WSFP), and unanimity preservation, but not strong linearity [2202.11633].
- **Optimization Principle**: It uniquely minimizes the weighted sum of Pearson $\chi^2$ divergences,
  \[
  J[g] = \frac{1}{N}\sum_{i=1}^N D_P(g \Vert f_i) = \frac{1}{2N}\sum_{i=1}^N \int \frac{(g(x)-f_i(x))^2}{f_i(x)} dx
  \]
  over normalized, nonnegative densities $g$, yielding $f_{HMD}(x)$ as a variational solution [2412.06725], [2202.11633].
- **Supra-Bayesian Interpretation**: In supra-Bayesian pooling of posteriors for linear-Gaussian models under certain priors, HMD fusion emerges as the consensus estimate for uncorrelated or weakly correlated noise models [2202.11633].
- **Support and Shape**: HMD satisfies the tight bound
  \[
  \min_i f_i(x) \le f_{HMD}(x) \le \max_i f_i(x)
  \]
  for all $x$, preserving the support intersection of all input densities and producing robustly conservative interpolation.
- **Divergence and Consistency**: For all $i, j$, $\mathrm{D}_{KL}(f_i \Vert f_{HMD}) \le \mathrm{D}_{KL}(f_i \Vert f_j)$, i.e., the fused density is always closer (in KL sense) to each local density than any two local densities are to each other. The fused density is nondecreasing in any argument, monotonic in all inputs, and consistent (in the sense of Normalized Estimation Error Squared tracking) on empirical benchmarks [2412.06716].

## 3. Extension to Gaussian and Mixture Densities

For univariate or multivariate Gaussian inputs, HMD fusion supports closed-form implementation with mean and covariance updates:

Given $p_i(x) = \mathcal{N}(x;\,\mu_i, \Sigma_i)$, the denominator mixture $(1-\omega)p_1 + \omega p_2$ is approximated by a single Gaussian matched in moments, with
\[
\mu_{eq}  = (1-\omega)\mu_1 + \omega \mu_2
\]
\[
\Sigma_{eq} = (1-\omega)[\Sigma_1 + (\mu_{eq} - \mu_1)(\mu_{eq} - \mu_1)^T] + \omega[\Sigma_2 + (\mu_{eq} - \mu_2)(\mu_{eq} - \mu_2)^T]
\]
The fused covariance and mean are given by
\[
\Sigma_{HMD} = (\Sigma_1^{-1} + \Sigma_2^{-1} - \Sigma_{eq}^{-1})^{-1}
\]
\[
\mu_{HMD} = \Sigma_{HMD} (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2 - \Sigma_{eq}^{-1}\mu_{eq})
\]
Generalization to $N$ densities yields analogous, recursively computed expressions, with explicit formulas available via leave-one-out Gaussian products and moment matching [2511.03130].

For Gaussian mixture models (GMMs), pairwise products of mixture components generate a new mixture which is divided by a moment-matched denominator Gaussian, followed by pruning and merging steps to control mixture cardinality [2412.06716], [2412.06725].

## 4. Algorithmic Implementation

Efficient and scalable implementations target both closed-form (Gaussian) and sampling-based (mixture/nonparametric) scenarios:

- **Gaussian Case**: Sequentially compute, for each node, moment-matched denominator and apply Sherman–Morrison or blockwise inversion to obtain fused covariance/mean as above. Recommended weights can be uniform or computed by optimizing a criterion over symmetric KL divergence between the fused and local densities [2511.03130].
- **Mixture Densities**: For each fusion step, compute all pairwise products for the numerator, approximate the denominator as a single Gaussian, divide, and normalize. Merge/prune mixture components to handle complexity growth.
- **Sampling-based (HMD-S algorithm)**: For arbitrary densities, perform self-normalized importance sampling using local densities as proposal distributions. For each sample, compute weights proportional to the ratio of the other densities to the denominator, then normalize. This approach requires no additional proposal and generalizes naturally to GMM inputs [2412.06725].
- **Consensus/Distributed Fusion**: Each agent computes and broadcasts local estimates and covariance; global weights may be optimized at each time step via numerical minimization of a KL-divergence–based cost function [2511.03130].

## 5. Comparative Analysis: HMD, Naive Fusion, and Covariance Intersection

| Method                    | Consistency       | Covariance Tightness    | Computational Structure   |
| ------------------------- | ---------------- | ---------------------- | ------------------------ |
| Naive/Kalman Fusion       | Over-confident   | Often singular         | Product of densities     |
| Covariance Intersection   | Consistent, conservative | Loose (deflated)    | Geometric mean           |
| Harmonic Mean Density     | Consistent, robust | Tighter but valid     | Reciprocal (harmonic) mean |

- **Naive Fusion** applies direct multiplication of posteriors, leading to over-confident covariances due to double-counting of common prior information. For Gaussians: $\Sigma_N = (\Sigma_1^{-1} + \Sigma_2^{-1})^{-1}$, which often leads to divergence [2412.06716].
- **Covariance Intersection (CI)/Geometric Mean Density (GMD)**, $f_{CI}\propto f_1^\omega f_2^{1-\omega}$, provides a consistent fusion strategy under unknown correlations but is recognized to be conservative, with expanded covariances: $\Sigma_{CI} = (\omega \Sigma_1^{-1} + (1-\omega)\Sigma_2^{-1})^{-1}$.
- **HMD** explicitly subtracts an information-theoretic denominator term and produces $\Sigma_{HMD} = (\Sigma_1^{-1} + \Sigma_2^{-1} - \Sigma_{eq}^{-1})^{-1}$, resulting in covariances that are always at least as tight as those of CI or the inverse covariance intersection (ICI), but with consistency validated in both NEES and RMSE metrics [2412.06716], [2412.06725].

## 6. Applications and Empirical Performance

Distributed target tracking, sensor consensus, and multi-sensor scenario benchmarks demonstrate that HMD fusion offers superior trade-offs in real-world tasks:

- **3D and 2D Target Tracking**: In both nearly constant velocity (NCV) and maneuvering multi-model (IMM) regimes, HMD outperforms CI, often showing 20–50% RMSE improvement and maintaining NEES tightly within 95% confidence bands [2412.06716].
- **Underwater Consensus Tracking**: Weighted HMD fusion, with KL-divergence–optimized weights, achieves minimum average RMSE and drives track divergence to zero, sustaining ANEES within confidence bounds [2511.03130].
- **Real-world Multi-Target Fusion**: Across multi-radar and multi-track fusion with clutter and non-Gaussian densities, HMD-GA yields consistently lower positional RMSE and NEES near theoretical means, dominating both CI and naive fusion for most targets [2412.06725].

## 7. Limitations, Computational Considerations, and Future Directions

HMD fusion’s primary limitations include the need for accurate approximation of the denominator mixture in non-Gaussian settings, and the complex optimization of fusion weights when targeting further covariance control or entropy minimization. Normalization requires one-dimensional or higher-order integration, though sampling-based implementations relieve the need for explicit proposal distributions. A plausible implication is that advances in mixture reduction and moment-matching techniques, or adaptive and decentralized weight optimization, will further enhance HMD’s applicability to high-dimensional and resource-constrained systems [2412.06716], [2511.03130], [2412.06725].

---

References:
- "On Pooling-Based Track Fusion Strategies: Harmonic Mean Density" [2412.06716]
- "Consensus Tracking of an Underwater Vehicle Using Weighted Harmonic Mean Density" [2511.03130]
- "Fusion of Probability Density Functions" [2202.11633]
- "Harmonic Mean Density Fusion in Distributed Tracking: Performance and Comparison" [2412.06725]

Source: https://www.emergentmind.com/topics/harmonic-mean-density-hmd-fusion