FMIP: Inertial Positioning & Optimization

• FMIP is a multi-domain acronym with key applications in indoor positioning, mixed-integer programming, inverse problems, and federated data imputation.
• In indoor positioning, FMIP uses foot-mounted IMUs with WiFi fingerprints and particle filtering to correct drift and align multi-user trajectories.
• In optimization and privacy contexts, FMIP underpins multimodal flow matching for MILP, foundation flow-matching priors, and federated Markov imputation for clinical data.

Searching arXiv for “FMIP” to ground the article in current usage. FMIP is a domain-dependent acronym rather than a single settled term. In indoor positioning, it denotes Foot-mounted Inertial Positioning, a pedestrian dead-reckoning backbone built around a foot-mounted IMU and often coupled to WiFi fingerprints to control drift (Gu et al., 2017). In mathematical optimization, it denotes “FMIP: Multimodal Flow Matching for Mixed Integer Linear Programming,” a generative framework that models the joint distribution of integer and continuous MILP solutions (Li et al., 31 Jul 2025). The acronym also appears in related discussions of foundation flow-matching priors for inverse problems and, in some federated or privacy-preserving contexts, alongside closely neighboring names such as Federated Markov Imputation, Membership Inference Privacy, and FTMCFE-IP rather than a single canonical FMIP term (Wan et al., 20 Nov 2025); (Düsing et al., 25 Sep 2025); (Izzo et al., 2022); (Zhang et al., 17 Oct 2025).

1. Terminological scope

The literature exhibits at least two direct, explicit expansions of FMIP. The 2017 indoor-positioning paper “Pedestrian Positioning Using WiFi Fingerprints and a Foot-mounted Inertial Sensor” uses FMIP as Foot-mounted Inertial Positioning and treats it as an encapsulated step-wise inertial navigation module that outputs position increments or step parameters such as $\Delta L_k$ and $\Delta \theta_k$ (Gu et al., 2017). The 2025 optimization paper “FMIP: Multimodal Flow Matching for Mixed Integer Linear Programming” uses the same acronym for a multimodal generative model in mixed discrete-continuous solution space $\mathbb{Z}^q \times \mathbb{R}^{n-q}$ (Li et al., 31 Jul 2025).

Other papers in the corpus show neighboring or overloaded usage. One paper explicitly states that “FMIP in this paper stands for Federated Markov Imputation,” although its title abbreviates the method as FMI (Düsing et al., 25 Sep 2025). Another paper states that “Provable Membership Inference Privacy” does not define FMIP as a separate term and instead centers on MIP (Izzo et al., 2022). A federated-learning cryptography paper is presented as an interpretation of FMIP through the proposed Flexible Threshold Multi-Client Functional Encryption for Inner Product (FTMCFE-IP) construction rather than through an acronym introduced by the paper itself (Zhang et al., 17 Oct 2025). This domain sensitivity makes contextual disambiguation essential.

2. Foot-mounted inertial positioning in indoor localization

In indoor positioning, FMIP is a PDR system in which a foot-mounted inertial measurement unit (IMU) measures specific force and angular rate, processes the data in real time inside a foot module, and outputs step-wise position increments rather than raw IMU measurements. The 2017 WiFi-assisted formulation treats FMIP, including ZUPTs, as an encapsulated black-box solution: the rest of the system receives a new estimate of position or position increment each time a step is detected (Gu et al., 2017).

At the step level, the nominal update is written as

$$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$

To represent uncertainty, each particle perturbs the FMIP output by additive noise,

$$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$

which yields a trajectory-valued particle filter whose weights are updated from RSS-space consistency rather than from a prebuilt radio map. The method assumes that small distance in RSS space typically implies small distance in coordinate space, uses a normalized Euclidean distance on RSS vectors with missing AP values set to $-110$ dBm, and uses loop-closure-style reweighting to suppress trajectories inconsistent with past signal patterns. The principal motivation is that FMIP errors remain “accumulative and unbounded in the long term” even though ZUPTs reduce error growth from cubic in time to linear in time (Gu et al., 2017).

The same year, a second paper moved from single-trajectory correction to multi-user alignment and radio-map construction by embedding FMIP in a graph-SLAM framework (Gu et al., 2017). There the node state is

$$\text{pos}_k^{(i)} = (x_k^{(i)}, y_k^{(i)}, \theta_k^{(i)}),$$

and the step-wise inertial recurrence is

$$\begin{bmatrix} x_k \ y_k \ \theta_k \end{bmatrix} = \begin{bmatrix} x_{k-1} \ y_{k-1} \ \theta_{k-1} \end{bmatrix} + \begin{bmatrix} \Delta x_{k-1}\cos(\theta_{k-1}) - \Delta y_{k-1}\sin(\theta_{k-1}) \ \Delta x_{k-1}\sin(\theta_{k-1}) + \Delta y_{k-1}\cos(\theta_{k-1}) \ \Delta \theta_{k-1} \end{bmatrix}.$$

IMU edges encode pairwise inertial constraints, while WiFi edges encode the premise that vicinity in RSS space corresponds statistically to vicinity in coordinate space. By minimizing the combined error, the system aligns trajectories from different users into a common frame and produces a crowd-sourced radio map.

The reported empirical results define the practical significance of this FMIP line. In the particle-filter paper, a trajectory of approximately $1855$ m over approximately $24$ min produced raw FMIP mean errors of $\Delta \theta_k$0 at the first revisit and $\Delta \theta_k$1 at the second revisit, while the proposed WiFi-assisted approach reduced those to $\Delta \theta_k$2 and $\Delta \theta_k$3, respectively, with runtimes of $\Delta \theta_k$4 s and $\Delta \theta_k$5 s versus $\Delta \theta_k$6 s and $\Delta \theta_k$7 s for the Gaussian-process-based comparison (Gu et al., 2017). In the graph-SLAM paper, five long trajectories with total walking length about $\Delta \theta_k$8 km yielded a maximum landmark error of about $\Delta \theta_k$9 m and mean error $\mathbb{Z}^q \times \mathbb{R}^{n-q}$0 m after optimization, whereas raw FMIP errors reached about $\mathbb{Z}^q \times \mathbb{R}^{n-q}$1 m; in the straight-corridor alignment experiment the mean heading alignment error was about $\mathbb{Z}^q \times \mathbb{R}^{n-q}$2 with standard deviation about $\mathbb{Z}^q \times \mathbb{R}^{n-q}$3 (Gu et al., 2017).

3. FMIP as multimodal flow matching for MILP

In optimization, FMIP denotes a graph-based generative model for Mixed-Integer Linear Programming that learns a joint distribution over integer and continuous solutions and then samples high-quality candidates guided by objective value and constraint satisfaction (Li et al., 31 Jul 2025). The underlying MILP is written as

$\mathbb{Z}^q \times \mathbb{R}^{n-q}$4

The central modeling claim is that existing GNN-based heuristics usually predict only integer variables and therefore fail to capture integer–continuous coupling and do not model a distribution of high-quality mixed solutions.

FMIP addresses this by representing a MILP instance as a tripartite graph with integer-variable nodes, continuous-variable nodes, and constraint nodes, and by applying multimodal flow matching to a mixed solution space. The model conditions on a noisy state $\mathbb{Z}^q \times \mathbb{R}^{n-q}$5 and predicts both denoised continuous variables and a distribution over clean integer assignments. Its training objective is

$\mathbb{Z}^q \times \mathbb{R}^{n-q}$6

The paper couples this generative model with decision-focused guidance via

$\mathbb{Z}^q \times \mathbb{R}^{n-q}$7

so sampling is biased toward low objective value and low violation.

The architecture uses an extension of Bi-GCN to tripartite graphs, specifically a 12-layer residual GCN. The reported evaluation covers seven standard MILP benchmarks and four downstream heuristics: Neural Diving, Predict-and-Search, PMVB, and Apollo-MILP. Across those settings, FMIP improves solution quality by 50.04% on average over GNN-based predictive baselines, and guidance contributes around 15 percentage points of additional improvement over unguided FMIP in the reported ablations (Li et al., 31 Jul 2025). This suggests that, in this literature, FMIP names not merely a predictor but a learned transport mechanism over the mixed feasible-solution manifold.

4. FMIP and foundation flow-matching priors for inverse problems

A related usage associates FMIP with the use of foundation flow-matching models as priors for inverse problems, later reworked by the FMPlug framework (Wan et al., 20 Nov 2025). The starting point is the generative-prior formulation

$\mathbb{Z}^q \times \mathbb{R}^{n-q}$8

where a foundation FM model supplies the prior through a generator $\mathbb{Z}^q \times \mathbb{R}^{n-q}$9.

The paper’s diagnosis is that naïvely plugging a foundation FM prior into this optimization is substantially weaker than domain-specific FM priors and can even trail an untrained Deep Image Prior. The proposed remedy has two components. The first is an instance-guided, time-dependent warm-start based on the FM probability path,

$$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$0

and the corresponding optimization

$$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$1

The second is a sharp Gaussianity regularization that constrains $$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$2 to the thin shell

$$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$3

implemented by projection rather than by a soft chi-square penalty.

The empirical record is framed as a rescue of foundation FM priors. On AFHQ-Cat Gaussian blur, D-Flow (FD) attains PSNR $$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$4 and SSIM $$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$5, whereas FMPlug (FD) attains PSNR $$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$6 and SSIM $$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$7; PnP-Flow (DS) is reported at PSNR $$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$8 and SSIM $$\text{pos}_k = \text{pos}_{k-1} + \Delta L_k \begin{bmatrix} \cos \theta_k\ \sin \theta_k \end{bmatrix}, \qquad \theta_k = \theta_{k-1} + \Delta \theta_k.$$9 (Wan et al., 20 Nov 2025). On the few-shot scientific inverse problems, FMPlug reaches 31.83 dB / 0.97 SSIM on LIS and 23.26 dB / 0.48 SSIM on MRI $$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$0, substantially above D-Flow with a foundation prior. The paper’s broader implication is that the limitations of foundation FM priors are not purely representational; they also depend on whether inference respects the geometry of the Gaussian training shell and the time structure of the FM path.

5. Federated and clinical-data usage

One paper explicitly uses FMIP for Federated Markov Imputation, a privacy-preserving temporal-imputation protocol for multi-center ICU federated learning (Düsing et al., 25 Sep 2025). Its core object is a global first-order Markov transition model built without sharing raw patient data. Each ICU computes local transition counts $$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$1 over discretized feature bins, secure aggregation produces

$$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$2

and the server normalizes to obtain

$$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$3

That federated transition matrix is then used locally for temporal imputation. For a single missing state with observed neighbors, the imputed bin is

$$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$4

and for longer gaps the paper states that the most probable path is inferred recursively.

The experimental setting uses MIMIC-IV, 28,610 patients, 7 ICUs, 25 clinically relevant features, and the first 6 hours after admission aggregated into six 1h windows. Under the Regular setting, mean AUC rises from 0.8634 for Local Mean and 0.8710 for LMI to 0.8878 for FMI. Under the Irregular setting, where some ICUs are assigned 2h or 3h intervals, Local Mean drops to mean AUC 0.7961 and LMI is reported as n/a, while FMI reaches 0.8629. Particularly large gains are reported for MICU/SICU (3h), from 0.5958 to 0.8038, and NSICU (3h), from 0.5180 to 0.8503 (Düsing et al., 25 Sep 2025). In this strand of the literature, FMIP names a federated imputation mechanism rather than an optimization or localization framework.

6. Adjacent but non-identical acronyms

Two further literatures are relevant because they are often adjacent to FMIP in abbreviation space but are not identical to it. The first is Membership Inference Privacy (MIP). The paper “Provable Membership Inference Privacy” explicitly states that it does not define FMIP as a separate term; its central object is $$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$5-MIP, under which an attacker’s success probability is bounded by

$$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$6

For $$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$7, this becomes a $$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$8 bound, and the paper proves that $$\Delta L_k^{(i)} = \Delta L_k + \delta L_k^{(i)}, \qquad \Delta \theta_k^{(i)} = \Delta \theta_k + \delta \theta_k^{(i)},$$9-DP implies MIP with

$-110$0

It also gives a generic wrapper that adds norm-based noise calibrated to central moments rather than to worst-case sensitivity, thereby obtaining $-110$1-MIP with less randomness than DP in the reported constructions (Izzo et al., 2022).

The second is Flexible Threshold Multi-client Functional Encryption for Inner Product (FTMCFE-IP), a federated-learning cryptographic scheme that is presented in the supplied material as an interpretation of FMIP rather than as a paper-defined expansion (Zhang et al., 17 Oct 2025). It supports non-interactive setup, per-round threshold selection, and client dropout. The decryption condition is thresholded: if $-110$2, decryption aborts; if $-110$3, the scheme recovers $-110$4. The implementation reports, for example, Dec times of 3819 ms for $-110$5, 7297 ms for $-110$6, and 7628 ms for $-110$7, with setup and encryption substantially cheaper than some prior threshold MCFE constructions (Zhang et al., 17 Oct 2025).

Taken together, these papers show that FMIP is best treated as a contextual label. In localization it refers to a drifting but infrastructure-free inertial backbone; in MILP it denotes a multimodal generative solver heuristic; in inverse problems it is tied to foundation flow-matching priors; and in federated or privacy-preserving settings it borders on, overlaps with, or is explicitly contrasted against FMI, MIP, and FTMCFE-IP. The literature therefore supports a plural, domain-indexed reading of the acronym rather than a single universal definition.