Direct Point Estimator (DPE)

• Direct Point Estimator (DPE) is a framework that directly computes point estimates of latent variables from raw data without relying on intermediate surrogate measurements.
• It is applied in GNSS, wireless localization, DPP kernel estimation, and deep learning for real-time regression, showcasing versatility across signal processing and control domains.
• DPE methods integrate robust cost functions, multipath mitigation, and efficient optimization strategies to enhance accuracy and manage noise and bias in complex environments.

A Direct Point Estimator (DPE) is a methodological framework that produces a point estimate of a latent variable or a model parameter directly from observed data, eschewing intermediate measurements or surrogate variables. DPEs have emerged as a central paradigm in inverse problems spanning signal processing, wireless localization, probabilistic modeling, and data-driven control. In modern applications, DPEs are implemented as maximum-likelihood estimators, explicit moment-based kernel reconstructors, or deep neural parametric regressors, operating across modalities from GNSS and 5G localization to determinantal point processes and real-time robotics.

1. Core Principles and Formulations

The defining feature of a DPE is its directness—the estimator operates on raw data and outputs the target quantity (position, kernel, or landing point) in a single computational block, typically by maximizing a domain-specific cost function. In GNSS and wireless positioning, the DPE evaluates the cross-ambiguity function (CAF) or likelihood over navigation-domain variables (PVT) and yields the optimal solution as the maximizer (Vicenzo et al., 2024, Huang et al., 24 Jul 2025, Li et al., 2023). In statistical modeling, the DPE (for instance, in determinantal point processes) reconstructs the kernel matrix by directly replacing theoretical moments with their empirical counterparts, guaranteeing statistical consistency and circumventing identification ambiguities (Gouriéroux et al., 20 May 2025).

The generic DPE workflow for maximum-likelihood-based positioning is:

• For each candidate value of the latent variable (e.g., PVT vector in GNSS), synthesize the predicted signal or kernel.
• Evaluate the joint cost function (e.g., sum of squared CAFs, negative log-likelihood).
• (Optionally) Use a grid search or differentiable optimizer to maximize (or minimize) the cost.
• Output the maximizer as the point estimate.

2. DPE in GNSS and Wireless Localization

Direct Point Estimation in satellite navigation inverts the conventional two-step paradigm by treating the navigation variables (position p, velocity v, time t) as primary unknowns, generating local signal replicas for candidate hypotheses, and correlating directly with the received baseband signal. The CAF for each satellite encodes delay and Doppler dependencies, forming a multidimensional likelihood surface over the PVT space (Vicenzo et al., 2024, Huang et al., 24 Jul 2025).

GNSS DPE Workflow:

• For each candidate PVT $\mathbf{y}_j = [\mathbf{p}_j^T, t_{r,j}]^T$, predict code delay $\Delta\tau_{i}(\mathbf{y}_j)$ and Doppler $\Delta f_{i}(\mathbf{y}_j)$ for each satellite.
• Generate replica vectors and compute correlation:

$$J(\mathbf{y}_j) = \sum_{i=1}^M \left| \mathbf{x}^H \mathbf{c}_i(\mathbf{y}_j) \right|^2$$

• Estimate:

$$\widehat{\mathbf{y}} = \arg \max_{\mathbf{y}_j} J(\mathbf{y}_j)$$

• Optionally, employ hierarchical or refined grid search for computational efficiency.

In cellular and 5G/OFDM positioning, DPE generalizes to non-GNSS signals by jointly fitting user state, channel gains, and multipath delays in a single maximum-likelihood navigation domain problem (Li et al., 4 Aug 2025, Li et al., 25 Feb 2025). The DPE cost incorporates all available symbols and subcarriers, directly marginalizing or concentrating out nuisance channel coefficients:

$$\widehat{\mathbf{p}} = \arg \max_{\mathbf{p}} \sum_{m=1}^M \| S_m^H(\mathbf{p}) y_m \|^2$$

yielding robust performance in NLoS and multipath-rich environments, with empirical RMSE ≤ 6 m at 10 dB SNR and outperformance versus OTDoA by over 95% in NLoS scenarios (Li et al., 25 Feb 2025).

3. Geometric Error Propagation, Multipath, and Bias

DPE error sources decompose naturally into random (noise variance) and systematic (multipath bias) components. Multipath induces deterministic bias via off-centering of correlation peaks that geometrically projects into PVT errors—quantified by the Satellite Circular Multipath Bias (SCMB) model (Huang et al., 24 Jul 2025). The induced range bias for satellite m is:

$\delta\rho_m = c\Delta\tau_m \sec\phi_m$

where $\phi_m$ is satellite elevation. The locus of possible intersection points is a circle in the ENU plane, tangent to each satellite’s bias-projected “center line.” Final bias for two satellites with azimuths $\theta_i$, $\theta_j$ and delays $\Delta\tau_{i}(\mathbf{y}_j)$0, $\Delta\tau_{i}(\mathbf{y}_j)$1 is:

$\Delta\tau_{i}(\mathbf{y}_j)$2

Bias magnitude grows with satellite elevation (due to increased secant projection), while lower-elevation satellites suffer more from increased noise variance. Thus, DPE performance in multipath-dominated regions is maximized by balancing high- and low-elevation satellites for optimal geometry (Huang et al., 24 Jul 2025).

4. Algorithmic Extensions: Robustness, Multipath Mitigation, Plug-Ins

DPE frameworks have been enhanced by integrating robust statistics and multipath mitigation strategies. Robust Interference Mitigation (RIM) processing replaces the quadratic loss by robust M-estimators (e.g., Huber, Laplacian, Cauchy), down-weighting outlier samples in time or frequency, and recovers near-nominal performance under high jamming/interference (Li et al., 2023). MMT-DPE introduces explicit multipath residual signal modeling via cost function augmentation and employs techniques such as Weill's multipath model to correct code delays before feeding to the DPE module, substantially reducing error in urban/NLoS regimes (Vicenzo et al., 2024).

A modular MATLAB plug-in architecture enables seamless integration of DPE (and MMT-DPE) into conventional 2SP SDRs, facilitating practical adoption and comparative analysis in both simulation and real open-sky/urban environments (Vicenzo et al., 2024).

5. DPE in Probabilistic and Data-Driven Estimation

In statistical inference, DPEs are explicit moment-based estimators, as illustrated for the correlation kernel of a determinantal point process (DPP). Given empirical first, second, and (selected) third moment inclusion probabilities in the sample, the DPE for the DPP kernel $\Delta\tau_{i}(\mathbf{y}_j)$3 is uniquely solved (up to D-similarity) from the principal minors:

• Diagonal: $\Delta\tau_{i}(\mathbf{y}_j)$4
• Off-diagonal: $\Delta\tau_{i}(\mathbf{y}_j)$5
• Signs from triple-inclusion moments

The DPE estimator enjoys strong consistency, asymptotic normality at $\Delta\tau_{i}(\mathbf{y}_j)$6-rate, explicit finite-sample deviation bounds, and O($\Delta\tau_{i}(\mathbf{y}_j)$7) complexity, resolving D-similarity by a fixed sign convention. This estimator also provides high-quality initialization for maximum-likelihood refinement in high-dimensional DPP kernel learning (Gouriéroux et al., 20 May 2025).

In control, data-driven DPEs synthesize estimators for unknown LTI systems by parameterizing all system matrices consistent with the observed data and known noise bounds via a quadratic matrix inequality (QMI), then solving an H∞-synthesis SDP for robust estimator gains. This approach attains robust performance bounds without restrictive invertibility assumptions (Brändle et al., 2024).

6. Deep and Neural DPE Architectures

DPEs have been instantiated as deep neural regressors in high-velocity real-time inference tasks such as in-flight object catching. For instance, the DPE in the DIPP framework employs a two-stage architecture: a fully-connected encoder (or discriminative LSTM feature embedding), followed by a recurrent LSTM stack, directly mapping an input state sequence to a 2D landing point prediction (Nguyen et al., 18 Sep 2025). Training combines state reconstruction and direct impact-point (mean squared error) loss. DPE variants yield minimal prediction error—sub-5cm landing error for seen and ≈6.5cm for unseen objects—and substantially higher simulated catching success rates compared to sequential trajectory-based (autoregressive) regressors.

7. Implementation, Performance, and Practical Considerations

GNSS/5G/OFDM Positioning

• Typical grid search: candidate spacing ≈1 m in each coordinate and clock, multidimensional grid across target region.
• Non-coherent integration over 1–10 ms intervals, maximizing the joint cost function for each candidate PVT.
• Optimizations: precomputed correlation banks, hierarchical or GPU-parallelized search, robustified cost functions.
• Empirical results: GNSS DPE outperforms two-step (2SP) by 36–76% RMS error in urban/NLOS; 5G DPE achieves ≤6 m RMSE in Tsim Sha Tsui, with 95.24% improvement over OTDoA at the 90th percentile (Vicenzo et al., 2024, Li et al., 25 Feb 2025).
• Plug-in modularity: seamless integration with SDRs via modular MATLAB function insertion with minimal code changes.

DPP Kernel Estimation

• Complexity: O(T d^2), numerically stable with variance/regularization correction.
• Sample complexity: $\Delta\tau_{i}(\mathbf{y}_j)$8 for error $\Delta\tau_{i}(\mathbf{y}_j)$9 in sup-norm pseudodistance (Gouriéroux et al., 20 May 2025).
• Role: Standalone estimator and high-quality initialization for likelihood-based kernel refinement.

Neural DPE for Real-Time Regression

• Achieves lowest average impact-point error and simulation catching success, outperforming autoregression by ≈50% in RMSE (Nguyen et al., 18 Sep 2025).
• Efficient runtime, minimal overfitting due to architectural and augmentation design.

8. Summary Table: DPE Signatures Across Modalities

Domain	DPE Cost / Model	Optimization
GNSS / 5G / OFDM Positioning	Likelihood / CAF sum	Grid / parallel search
DPP Kernel Estimation	Moment-based mapping	Closed-form
LTI System Identification	QMI / SDP synthesis	Convex optimization
Deep Impact Point Prediction	End-to-end regressor	SGD/Adam, direct loss

Across diverse applications, the Direct Point Estimator paradigm serves as an explicit, statistically grounded, and algorithmically direct methodology to infer critical hidden variables from observation, achieving state-of-the-art performance and robust error bounds under adversarial or information-limited conditions. (Huang et al., 24 Jul 2025, Li et al., 2023, Vicenzo et al., 2024, Li et al., 25 Feb 2025, Li et al., 4 Aug 2025, Gouriéroux et al., 20 May 2025, Brändle et al., 2024, Nguyen et al., 18 Sep 2025)