---
title: Dynamic Selective Mapping (DSLM)
url: https://www.emergentmind.com/topics/dynamic-selective-mapping-dslm
type: topic
---

# Dynamic Selective Mapping (DSLM)

Searching arXiv for the provided DSLM-related papers and closely related context papers.
Dynamic Selective Mapping (DSLM) is used in two different ways in the provided literature. In "Envelope Control Enabled Probabilistic Shaping for Peak Power Constrained IM DD Systems" [2507.18149], DSLM is a **memory-dependent bit-to-symbol mapping scheme** for \(M\)-PAM in **peak-power-constrained (PPC) IM–DD systems**. In "Mapping the Unseen: Unified Promptable Panoptic Mapping with Dynamic Labeling using Foundation Models" [2405.02162], DSLM is **not a term used in the paper**, but UPPM is described as **essentially a concrete realization of that idea** in the sense of “only map what I care about, when I ask for it, and allow new classes on the fly.”

## 1. Terminological scope

The two usages differ in domain, objective, and implementation, but both are explicitly framed around **dynamic** and **selective** operations rather than a fixed mapping or fixed label set.

| Domain | Operational meaning | Representative source |
|---|---|---|
| PPC IM–DD optical links | A **dynamic selective mapping (DSLM) mechanism at the transmitter** enabling an **untypical bit-to-symbol mapping** in which the current symbol is determined by the current bits pattern and by previously generated symbols within a specified memory length | [2507.18149] |
| Promptable panoptic mapping | DSLM is **not a term used in the paper**, but UPPM is **essentially a concrete realization of that idea**: “only map what I care about, when I ask for it, and allow new classes on the fly” | [2405.02162] |

In the optical paper, DSLM is the formal name of a transmitter-side finite-state mapping layer. In the mapping paper, the phrase is an interpretive label applied to a system whose semantics are **selected and updated dynamically through prompts**. This suggests that the expression denotes a family resemblance rather than a single canonical technique.

## 2. DSLM in peak-power-constrained IM–DD systems

The optical formulation is motivated by short-reach **intensity-modulation/direct-detection (IM–DD)** links **without optical amplifiers**, under a **peak optical (equivalently electrical) power** constraint. The model is
\[
X \ge 0, \quad \text{optical field } \mathcal{X} = \sqrt{X},
\]
\[
\max[X] \le P_{\text{peak}},
\]
and after fiber and photodiode,
\[
Y = X + Z,
\]
with AWGN \(Z\) dominated by receiver thermal noise etc. The paper treats this as a **peak-power-constrained (PPC) IM–DD channel**. In practice, the link also exhibits **memory effects** due to electrical and optical bandwidth limitations, modulator nonlinearity, fiber dispersion at high baud rates, and overshoot or ringing of analog components. These effects cause waveform overshoot beyond the nominal peak limit, increased PAPR, and pattern-dependent distortion. The paper therefore argues that conventional **probabilistic shaping (PS)**, which optimizes a **memoryless input PMF**, is not sufficient in high-rate PPC IM–DD, because the dominant impairments are often **memory-induced distortions associated with the signal envelope** rather than the memoryless mutual information [2507.18149].

DSLM is introduced as an **indirect PS** mechanism. Rather than directly specifying a target PMF and then matching to it, the transmitter modifies the **bit-to-symbol mapping rule** based on **past output symbols** in order to suppress “bad” temporal patterns, reduce envelope excursions, and indirectly realize a non-uniform effective symbol distribution. The resulting output PMF is therefore described as a **by-product of envelope control**.

## 3. Mapping rule, envelope metric, and finite-state structure

For an \(M\)-PAM system, a conventional Gray mapper uses
\[
S_k = \text{Gray}\big(b_{km-m+1},\dots,b_{km}\big),
\]
so \(S_k\) depends only on the current bits. In DSLM, the symbol depends on the current bit pattern and on the **previous \(L-1\)** transmitted symbols. With
\[
\mathbf{S}_{k-1}^{(L-1)} = [S_{k-L+1},\dots,S_{k-1}],
\]
the mapping becomes
\[
S_k = f\big(b_{km-m+1},\dots,b_{km};\, S_{k-L+1},\dots,S_{k-1}\big).
\]
The paper emphasizes three consequences: the mapping is **context-dependent**, multiple constellation points may effectively share the same bit label because of remapping, and symbol observation alone does not uniquely determine bits.

The sequence-selection rule is based on an **envelope evaluation function** over a length-\(L\) symbol window
\[
\mathbf{S} = [S_{k-L+1},\dots,S_k].
\]
Without CSI, the main heuristic is
\[
f(\mathbf{S}) = -10 \cdot \log_{10}\Big\{\mathrm{var}(\mathbf{S}) \cdot \max\big[\lvert \mathbf{S}\rvert\big]\Big\}. \tag{15}
\]
Large variance and large peak amplitude produce a low quality score, so patterns with large peaks and strong fluctuations are penalized. If CSI is available, the paper suggests
\[
f(\mathbf{S}) = \mathbf{S} * h, \tag{16}
\]
where \(h\) is the estimated impulse response. For modulation alphabet
\[
\mathcal{X} = \{-M+1, -M+3, \dots, M-1\},
\]
all \(M^L\) length-\(L\) patterns are evaluated and sorted. A **forbidden ratio** \(\gamma \in [0,1]\) partitions them into \(\mathcal{F}\), the worst \(\gamma\) fraction, and \(\mathcal{F}^c\), the allowed fraction.

The transmitter then applies a staged remapping rule. It first computes a preferred Gray symbol \(S_k^{\text{Gray}}\) and checks whether the tentative sequence
\[
\mathbf{S}_k^{\text{tent}} = [S_{k-L+1},\dots,S_{k-1}, S_k^{\text{Gray}}]
\]
is allowed. If so, it transmits \(S_k^{\text{Gray}}\). If not, it forms the candidate set
\[
\mathcal{A} = \{a \in \mathcal{X} : [S_{k-L+1},\dots,S_{k-1}, a] \in \mathcal{F}^c\}.
\]
Among allowed candidates, the selection is refined by minimum **Hamming distance** to the Gray target,
\[
\mathcal{C} = \{c \in \mathcal{A} : d_H(c, S_k^{\text{Gray}}) = \min_{x \in \mathcal{A}} d_H(x, S_k^{\text{Gray}})\},
\]
then by maximum local envelope quality,
\[
\mathcal{D} = \{d \in \mathcal{C} : f([S_{k-L+1},\dots,S_{k-1}, d]) = \max_{x \in \mathcal{C}} f([S_{k-L+1},\dots,S_{k-1}, x])\},
\]
and finally by minimum Euclidean distance,
\[
\mathcal{G} = \{g \in \mathcal{D} : \lvert g - S_k^{\text{Gray}} \rvert = \min_{x \in \mathcal{D}} \lvert x - S_k^{\text{Gray}} \rvert\}.
\]
If multiple elements remain, one is chosen uniformly at random. This rule directly suppresses patterns such as \([+7,-7,+7,-7,\dots]\) in PAM8, curbs overshoot in the analog channel response, and tends to produce **more low-pass spectral content**.

## 4. Trellis-based detection, PAS compatibility, and measured gains

Because DSLM introduces memory and ambiguity, the receiver is not a conventional symbol-by-symbol demapper. For memory length \(L\), the trellis state at stage \(i\) is
\[
\mathbf{s}_i = [S_{i-L+2},\dots,S_i],
\]
with
\[
K = M^{L-1}
\]
states. In a conventional memoryless mapper each state has \(M\) outgoing and \(M\) incoming transitions, whereas in DSLM some transitions are eliminated because they generate forbidden patterns. The receiver therefore runs a **BCJR (MAP) sequence detector** on the constrained trellis, with forbidden transitions excluded from the recursion. The observation model in the branch metric includes
\[
p(Y_i|S_i) \propto \exp\left(-\frac{(Y_i - S_i)^2}{2\sigma^2}\right),
\]
together with the a priori probability of the bit pattern from the decoder. To reduce complexity, the paper uses a **modified M-BCJR** within a **turbo equalization** loop with an LDPC decoder; only the strongest states are retained at each stage [2507.18149].

The same paper places DSLM after a **probabilistic amplitude shaping (PAS)** chain:
\[
\text{DM} \;\to\; \text{PAS} \;\to\; B^{-1}(\cdot) \;\to\; \text{interleaver} \;\to\; \text{DSLM} \;\to\; S_k.
\]
PAS determines the base information rate and average entropy, while DSLM provides envelope control and indirectly perturbs both the PMF and the temporal pattern statistics. For coded modulation the paper uses
\[
\text{NGMI} = 1 - \frac{H(X|Y)}{m},
\]
with \(m=\log_2 M\).

The reported measurements are specific. For **uniform PAM8** with \(L=5\) and \(\gamma=0.33\), the entropy drop is only \(\approx 0.03\) bits/symbol, from \(3.00\) to \(2.97\), and the measured NGMI shows **~1 dB improvement** in required ROP for a given NGMI. With LDPC of rate \(2/3\) and **11 turbo iterations**, **Typical PAM8** is error-free at about \(-5.5\) dBm, while **DSLM PAM8** is error-free at about \(-6.5\) dBm, yielding **~1 dB receiver-sensitivity improvement**. For **MB-PS PAM8** with entropy \(H=2.8\), DSLM with \(L=5\) and \(\gamma=0.72\) gives a **~0.7 dB** better ROP threshold for error-free operation than MB-PS alone. The paper also notes that the DSLM output spectrum is more low-pass, eye diagrams show less fluctuation, and the symbol histogram shifts toward a **cap-shaped** distribution.

A recurrent misconception is that DSLM is simply the optical **Selective Mapping (SLM)** technique known from OFDM. The paper distinguishes them explicitly: classic SLM is block-level, requires side information, and does not change the mapping between bits and constellation points, whereas DSLM is **symbol-by-symbol**, **state-dependent**, sends **no explicit side information**, and is decoded by a trellis consistent with the allowed and forbidden pattern rules.

## 5. UPPM as a concrete realization of dynamic selective mapping

In the robotics paper, **Dynamic Selective Mapping (DSLM)** is introduced only as an interpretive lens: **it is not a term used in the paper**, but UPPM is presented as **essentially a concrete realization of that idea**. The target behavior is stated directly: **“only map what I care about, when I ask for it, and allow new classes on the fly.”** The paper argues that conventional semantic and panoptic mapping systems use **fixed, closed label sets**, rely on pre-trained 2D semantic or instance segmentation networks, and cannot handle **novel or long-tail categories** without retraining or re-annotation. UPPM replaces this with a **prompt-driven, open-vocabulary, dynamically labeled** panoptic map [2405.02162].

UPPM takes posed **RGB-D** frames and builds a **panoptic 3D map** using a **multi-resolution, multi-TSDF** representation with **object-centric submaps**. Its semantic layer is built from foundation models: **Tag2Text** generates captions and tags, **Honnibal’s POS tagger** and the **WordNet lemmatizer** extract nouns, noun phrases, and adjectives, **MPNet + COCO-Stuff STS** map open-vocabulary labels to a parent COCO class by cosine similarity, **Grounding-DINO** performs open-set detection from curated label prompts, and **SAM** produces instance masks. The paper’s **Dynamic Labeling** mechanism then unifies diverse textual descriptions such as “table”, “small dining table”, and “wooden table” into a **single internal category ID** while preserving all textual variants.

This pipeline operationalizes promptability and selectivity in several ways. Automatic label generation during mapping uses Tag2Text-derived captions and tags as queries for Grounding-DINO. User prompts support later querying and interaction, including prompts of increasing specificity such as those illustrated for a “small round wooden table.” New textual descriptions can be integrated and linked to existing objects, because the **3D geometry is decoupled from semantics** and semantics can be updated **at mapping time or even post hoc**. The paper explicitly describes this as **geometry-first**, with semantics **prompt-driven and late-bound**.

The system’s **Unified Semantics** are central. Open-vocabulary labels are embedded and matched to COCO-Stuff classes using
\[
\hat{c} = \argmax_{c' \in C} \text{sim}(E(c), E(c')),
\]
with cosine similarity
\[
\text{sim}(\mathbf{u}, \mathbf{v}) =
\frac{\mathbf{u} \cdot \mathbf{v}}{\|\mathbf{u}\|\,\|\mathbf{v}\|}.
\]
This yields a parent COCO class and size estimate, enabling a stable unified category ID across frames even when the visible text labels vary. The paper therefore presents UPPM as a panoptic mapping system whose semantics are **selected and updated dynamically through prompts**, and explicitly states that this aligns well with the DSLM concept.

## 6. Limitations, trade-offs, and cross-domain significance

The two literatures attach the same phrase to different engineering problems, and each emphasizes a different trade-off structure. In the optical setting, larger memory length \(L\) improves control over longer patterns but increases complexity roughly with \(M^L\) when designing forbidden sets, while the receiver state space scales as \(M^{L-1}\). Higher forbidden ratio \(\gamma\) gives stronger envelope control but also increases the **DSLM-induced BER** before FEC, so \(\gamma\) must be chosen so that the combined channel and remapping errors remain correctable by the turbo equalizer and FEC. The paper describes transmitter implementation as essentially LUT-based and feasible in high-speed digital logic, but receiver **M-BCJR** and turbo iterations increase computation and latency [2507.18149].

In the mapping setting, the principal limitations are different. UPPM depends on **Tag2Text, Grounding-DINO, SAM**, and related modules, so noisy captions, detection errors, and image quality degradation propagate into the 3D map. The paper highlights sensitivity to motion blur and noisy images in **RIO**, a trade-off between open-set flexibility and closed-set precision, and the absence of explicit integration with planning or control. It also notes that foundation models are computationally heavy and that full online deployment may not meet strict real-time constraints. The proposed mitigations are **postprocessing (POS tagging, lemmatization, STS, NMS)** and offline or incremental semantic updates rather than full map reconstruction [2405.02162].

Taken together, the two usages show that **dynamic selective mapping** can denote either a **finite-state, pattern-aware mapping layer** that controls short-term envelope statistics in PPC IM–DD, or a **promptable, dynamically labeled panoptic map** in which semantics are added, merged, or refined on demand. This suggests a shared abstraction: the mapping is not fixed in advance, but conditioned on context—previous symbols in one case, natural-language prompts and stored semantic embeddings in the other.

Source: https://www.emergentmind.com/topics/dynamic-selective-mapping-dslm