MA-MPL: Disambiguating Domain-Specific Uses

• MA-MPL is a context-dependent acronym representing distinct innovations in photonics (multimode architecture with MPLC), multi-agent online coordination, and time-series clustering.
• In photonics, MA-MPL describes programmable spatial-mode interfacing via multiple phase masks, achieving efficient mode conversion and broadband operation in the telecom C-band.
• In multi-agent optimization and time-series analysis, MA-MPL denotes parameter-free decentralized algorithms and robust maximum pseudolikelihood estimators that ensure computational stability and effective approximation guarantees.

MA-MPL is not a single standardized term. In current technical usage across the supplied literature, it denotes several unrelated constructs: a multimode architecture built on multi-plane light conversion for bridging free-space and guided optical modes, a parameter-free algorithm for multi-agent online coordination, and a maximum-pseudolikelihood methodology for mixture-of-autoregressions time-series clustering. Closely related acronyms—MPL, mPL, AmPL, and MMAP-based matrix-analytic modelling—appear in neighboring literatures with different meanings, so domain-specific expansion is necessary for unambiguous interpretation (Stranden et al., 2 Dec 2025, Zhang et al., 26 Sep 2025, Nguyen et al., 2016).

1. Disambiguation across research domains

The term is best understood as a context-dependent acronym rather than a unique technical object. The principal usages in the supplied corpus are summarized below.

Usage	Domain	Representative source
Multimode architecture built on multi-plane light conversion	Integrated and free-space photonics	(Stranden et al., 2 Dec 2025)
Multi-Agent Meta-Policy Learning	Multi-agent online optimization	(Zhang et al., 26 Sep 2025)
Maximum pseudolikelihood for mixture of autoregressions	Model-based time-series clustering	(Nguyen et al., 2016)
Closely related but distinct acronyms: MPL, mPL, AmPL, MMAP-based matrix-analytic modelling	Information extraction, privacy, reliability	(2505.16107, Chen et al., 1 May 2026, Ruiz-Castro et al., 2024)

The photonics usage is the most explicit instance in which the label itself is tied to “a multimode architecture built on multi-plane light conversion (MPLC).” A second fully explicit usage appears in online optimization, where MA-MPL names the parameter-free algorithm introduced alongside MA-SPL. A third usage associates MA-MPL with maximum pseudolikelihood estimation for mixture-of-autoregressions models. The remaining acronym family is adjacent rather than identical, but it materially contributes to ambiguity because all of the terms are active in recent arXiv literature.

2. MA-MPL in multimode photonic interfaces

In photonics, MA-MPL denotes a multimode architecture built on multi-plane light conversion: a passive, programmable interface connecting free-space spatial modes to on-chip multimode silicon waveguide modes across the telecom C-band. MPLC is formulated as a sequence of phase masks separated by free-space propagation, implementing an approximately unitary mapping between an input basis of orthogonal modes and an output basis of orthogonal modes. In the reported realization, the device uses four phase modulation planes displayed on different regions of a single reflective liquid-crystal spatial light modulator; the beam is bounced four times between the SLM and a mirror, and the transformation is optimized so that $LG_{00}\rightarrow TE_{00}$, $LG_{10}\rightarrow TE_{10}$, and $LG_{20}\rightarrow TE_{20}$ for the first three TE modes of a 26.9 µm-wide silicon rib waveguide. Experimentally, the interface yields approximately 65% overlap for each mode at the MPLC output before the chip, visibility of approximately 90%, overall power coupling efficiency of 10–15%, and after propagation through the chip about 86% overlap for $TE_{00}$ and about 65% for $TE_{10}$ and $TE_{20}$, with overall visibility of approximately 75%; the broadband response is measured from 1528 to 1568 nm and optimized using four wavelengths from 1540 to 1570 nm (Stranden et al., 2 Dec 2025).

Within the same photonic family, a related interpretation treats MA-MPL as a mode-addressable MPLC device for multimode fibre. That formulation emphasizes programmability through an SLM and computer-generated hologram optimization, so that the same MPLC can launch arbitrary subsets of LP modes or Schmidt modes, adapt its transfer function as fibre conditions change, and support quasi-SISO operation by channel diagonalisation. In that setting, direct-search optimization can improve average mode extinction ratio by as much as 15 dB at the expense of insertion-loss deterioration of less than 3 dB, while preserving explicit control of the MER/IL trade-off (Rothe et al., 2024).

The photonic usage sits in a broader MPLC lineage. An early efficient, mode-selective spatial mode multiplexer based on MPLC reported a typical 3-mode multiplexer with mode selectivity better than $-23$ dB and total insertion efficiency of $-4.1$ dB across the full C-band, with optical-coating improvements projected to increase efficiency to $-2.4$ dB (Labroille et al., 2014). A later 45-mode MPLC spatial multiplexer and demultiplexer saturating all the modes of a standard 50 µm core graded-index OM2 multimode fiber reported average 4 dB insertion loss and $-28$ dB cross-talk across the C band, while exploiting a separable Hermite-Gaussian basis to reduce the number of reflections (Bade et al., 2018). This suggests that, in photonics, MA-MPL is best read as a systems-level label for programmable, approximately unitary spatial-mode interfacing rather than as a single device architecture.

3. MA-MPL as Multi-Agent Meta-Policy Learning

In multi-agent online optimization, MA-MPL stands for Multi-Agent Meta-Policy Learning. It is introduced for the multi-agent online coordination problem, where a set of agents connected by an undirected communication graph chooses actions online under partition constraints, and performance is measured by dynamic $LG_{10}\rightarrow TE_{10}$0-regret relative to the per-round combinatorial optimum. MA-MPL is the second of two algorithms in the paper: MA-SPL attains the target approximation ratios but relies on unknown structural parameters such as the diminishing-return ratio $LG_{10}\rightarrow TE_{10}$1 or the submodularity-ratio pair $LG_{10}\rightarrow TE_{10}$2, whereas MA-MPL is entirely parameter-free and maintains the same approximation ratio (Zhang et al., 26 Sep 2025).

The central technical device is the policy-based continuous extension. For each agent $LG_{10}\rightarrow TE_{10}$3, a policy vector $LG_{10}\rightarrow TE_{10}$4 specifies a distribution over the private action set $LG_{10}\rightarrow TE_{10}$5, and the extension

$LG_{10}\rightarrow TE_{10}$6

equals the expected value of the set function under independent sampling from those policies. Its stated advantage over the multilinear extension is a lossless rounding scheme for any set function, because each agent samples at most one action and the expected discrete utility equals the continuous objective. That property is then used to handle weakly submodular objectives, not only submodular ones.

Algorithmically, MA-MPL maintains multiple online linear oracles per agent and implements a decentralized Meta-Frank-Wolfe procedure with communication, consensus-style coordinatewise maxima, action sampling from the learned policies, and batch gradient estimation based on local marginal oracles. The theoretical guarantees are explicit. For monotone $LG_{10}\rightarrow TE_{10}$7-weakly DR-submodular objectives, the approximation factor is $LG_{10}\rightarrow TE_{10}$8; for monotone $LG_{10}\rightarrow TE_{10}$9-weakly submodular objectives, it is

$LG_{20}\rightarrow TE_{20}$0

With $LG_{20}\rightarrow TE_{20}$1 and $LG_{20}\rightarrow TE_{20}$2, the expected dynamic regret satisfies

$LG_{20}\rightarrow TE_{20}$3

where $LG_{20}\rightarrow TE_{20}$4 is the graph diameter and $LG_{20}\rightarrow TE_{20}$5 is the deviation of the optimal sequence. The paper also states the corresponding communication and query costs, $LG_{20}\rightarrow TE_{20}$6 and $LG_{20}\rightarrow TE_{20}$7, and experimentally reports competitive performance in multi-target tracking and EKF-based Bayesian A-optimal design settings (Zhang et al., 26 Sep 2025).

4. MA-MPL as maximum pseudolikelihood for mixture autoregressions

A separate usage associates MA-MPL with maximum pseudolikelihood estimation for mixture of autoregressions models used in model-based clustering of time series. In that setting, $LG_{20}\rightarrow TE_{20}$8 univariate series $LG_{20}\rightarrow TE_{20}$9 are assumed to arise from $TE_{00}$0 latent AR($TE_{00}$1) components with Gaussian innovations. Full maximum-likelihood estimation becomes numerically problematic because the component density for a long series is a product of many Gaussian conditional densities, so underflow occurs as $TE_{00}$2 increases. The proposed remedy is to maximize a pseudolikelihood built from one-step conditional densities,

$TE_{00}$3

rather than the full joint likelihood (Nguyen et al., 2016).

The resulting estimator is computed by an EM algorithm. In the E-step, time-point-level responsibilities $TE_{00}$4 are computed from the conditional densities. In the M-step, the mixture proportions are updated from averaged responsibilities, and the component-specific autoregressive parameters and innovation variances are updated by weighted least squares and weighted residual-variance formulas. The supplied description states that the maximum pseudolikelihood estimator is consistent and that simulations show performance comparable to ML where ML is feasible, while remaining numerically stable for long time series where ML becomes impractical.

The methodology is illustrated on resting-state fMRI data, where many voxel-wise BOLD time series must be clustered under strong temporal dependence. In that application, MA-MPL is presented as scalable to large $TE_{00}$5 and moderate-to-large $TE_{00}$6, with clustering based on posterior probabilities aggregated over time points. A plausible implication is that, in this statistical usage, MA-MPL emphasizes computational robustness and tractable inference rather than any connection to the photonic or multi-agent meanings of the same label.

5. Neighboring acronym families and near-collisions

The ambiguity around MA-MPL is intensified by several nearby acronyms that are active in unrelated literatures. In information extraction, MPL stands for Multiple Programming Languages with LLMs for Information Extraction. That framework reformulates IE as code-style generation, introduces a compact function-prompt with virtual running, and uses Python, C++, and Java jointly during supervised fine-tuning. Reported results include average F1 improvements over Python-only or prior code-prompt baselines, such as 77.6 average F1 for an MPL configuration with LLaMA3-8B versus 76.5 for GoLLIE-34B, and an approximately 20% reduction in prompt length for the function-prompt relative to class prompts (2505.16107).

In privacy-preserving machine learning, mPL denotes metric-normalized posterior leakage, and AmPL denotes Adaptive mPL. Here the focus is not pseudolikelihood or policy learning, but an attacker-aligned, distance-calibrated measure of posterior-odds shift under joint observation. The paper proves that, for single or independent releases, uniformly bounding mPL is equivalent to metric differential privacy, then introduces probabilistically bounded mPL and a trust-and-verify framework in which learned attackers audit privacy leakage and perturbation parameters are adapted accordingly (Chen et al., 1 May 2026).

In reliability and queueing-style stochastic modelling, the supplied description explicitly states that a paper on a complex redundant multi-state system with preventive maintenance and a multiple-vacation repairperson “does not use the label ‘MA-MPL’ explicitly,” but is “a prototypical matrix-analytic / MAP-type model” whose constructions “map directly to what is usually meant by MA-MPL (matrix-analytic multi-phase/point process) modelling.” The underlying representation is an MMAP with marked arrivals,

$TE_{00}$7

combined with PH distributions and matrix-analytic computation of transient and stationary measures (Ruiz-Castro et al., 2024).

These neighboring uses are not interchangeable. The acronym collision is structural: each literature binds the letters to a different expansion and an unrelated mathematical framework.

6. Interpretation, usage conventions, and technical significance

Taken together, the supplied literature indicates that MA-MPL functions as a local acronym whose meaning is fixed only by disciplinary context. In photonics it refers to multimode architectures or mode-addressable systems built on MPLC and concerns approximately unitary spatial transformations, mode selectivity, insertion loss, crosstalk, and broadband operation. In online optimization it denotes a specific parameter-free decentralized algorithm whose identity is inseparable from the policy-based continuous extension and weak-submodularity approximation guarantees. In time-series statistics it denotes maximum pseudolikelihood machinery for MoAR estimation, with EM-based computation and consistency guarantees. Closely adjacent literatures use MPL, mPL, AmPL, or MMAP-based matrix-analytic language for still different purposes.

This suggests that accurate expansion on first use is not optional. In technical writing, “MA-MPL” without expansion is liable to be underdetermined even for expert readers, because the same string now indexes at least three non-overlapping research programs. A plausible editorial rule is therefore to bind the acronym immediately to its local expansion and, where ambiguity is likely, to cite the defining paper directly—for example, the multimode MPLC interface in photonics (Stranden et al., 2 Dec 2025), Multi-Agent Meta-Policy Learning in online coordination (Zhang et al., 26 Sep 2025), or maximum pseudolikelihood for mixture autoregressions (Nguyen et al., 2016).