Random Mapping and Projection (RMRP)

• RMRP is a class of random embeddings that transforms high-dimensional data into lower-dimensional representations while approximately preserving key geometric properties.
• It encompasses both linear methods like Johnson–Lindenstrauss projections and nonlinear approaches such as random Fourier features, with broad applications in machine learning and optimization.
• Structured and tensorized RMRP variants offer computational efficiency and enable physical implementations, including photonic systems for ultra-high-speed data processing.

Searching arXiv for the specified papers and closely related work on random projection, tensorized projection, manifold embeddings, optimization, and photonic implementations. Random Mapping and Random Projection (RMRP) denotes a class of randomized embeddings used to construct lower-dimensional or otherwise tractable representations of high-dimensional data. In its broadest form, a random mapping is a randomly drawn function $f:\mathbb R^d\to\mathbb R^k$; random projection is the linear special case $f(x)=Px$, where $P$ is sampled from a prescribed ensemble. The central idea is that a suitably chosen random map can preserve salient geometry—most notably pairwise Euclidean distances—while sharply reducing storage, arithmetic cost, or downstream model complexity. Within the literature represented here, RMRP spans classical Johnson–Lindenstrauss embeddings, sparse and structured sketches, tensorized constructions, neural-network front ends, optimization methods, and integrated photonic implementations (Nabil, 2017, Miri, 2021).

1. Definitions and conceptual scope

A random mapping is any potentially non-linear function

$f:\mathbb R^d\longrightarrow\mathbb R^k$

drawn according to some probability measure over a family of candidate functions. In general, such mappings need not preserve inner products or distances exactly; instead, they may provide unbiased or low-variance estimates of kernel values, dot-products, or norms. Random Fourier features for kernel approximation and count-sketch mappings for sparse recovery are standard examples. A random projection is a special case in which $f$ is linear: one draws a random matrix $P\in\mathbb R^{k\times d}$ from a specified distribution and defines $f(x)=Px$. Every random projection is therefore a random mapping, but not every random mapping is a random projection (Nabil, 2017).

This distinction is conceptually important because much of the theory that made RMRP foundational in machine learning and numerical linear algebra depends on linearity. In the survey literature, random mapping is often used as a broad umbrella term, while random projection refers specifically to linear, distance-preserving dimension reduction. The same surveys also emphasize a complementary point: the projection matrix is usually generated without considering the intrinsic structure of the original data, which explains both the method’s broad applicability and its characteristic distortion trade-offs (Xie et al., 2017).

2. Geometric foundations: Johnson–Lindenstrauss, concentration, and manifold embeddings

The classical theoretical cornerstone of RMRP is the Johnson–Lindenstrauss lemma. In a standard form, for any finite set $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$ and any $0<\varepsilon<1$, there exists a linear map $f:\mathbb R^d\to\mathbb R^k$ with

$f(x)=Px$0

such that all pairwise distances are preserved up to a factor $f(x)=Px$1: $f(x)=Px$2 For Gaussian random projection, one takes $f(x)=Px$3 with $f(x)=Px$4 having independent standard Gaussian entries. The standard proof fixes a unit vector $f(x)=Px$5, observes that $f(x)=Px$6 is $f(x)=Px$7 with $f(x)=Px$8 degrees of freedom, applies concentration of measure, and then uses a union bound over all $f(x)=Px$9 pairwise differences (Nabil, 2017).

The same concentration template reappears in closely related settings. For sparse signals, the Restricted Isometry Property is presented as a direct analogue of the JL lemma: a random matrix $P$0 with Gaussian or Bernoulli entries satisfies

$P$1

with $P$2. In neural-network theory this stable embedding of sparse inputs and low-dimensional manifolds is used to argue that approximation complexity depends primarily on sparsity or intrinsic manifold dimension, and only weakly on ambient dimension (Cai et al., 2018).

For smooth manifolds, the finite-point JL lemma is not the whole story. One representative analysis considers a smooth Gaussian random manifold $P$3 of intrinsic dimension $P$4, dimensionless volume

$P$5

and failure probability $P$6. In the regime $P$7, $P$8, and $P$9, an approximate high-probability bound takes the form

$f:\mathbb R^d\longrightarrow\mathbb R^k$0

This replaces the finite-set cardinality term $f:\mathbb R^d\longrightarrow\mathbb R^k$1 by explicit dependence on intrinsic dimension, ambient dimension, and a volume parameter that counts correlation cells (Lahiri et al., 2016).

3. Projection ensembles, structured transforms, and computational trade-offs

The most elementary construction is the dense Gaussian projector, whose entries are sampled as $f:\mathbb R^d\longrightarrow\mathbb R^k$2. Its appeal is theoretical cleanliness: it gives the simplest JL proof and exact subspace embedding, but it is dense and requires $f:\mathbb R^d\longrightarrow\mathbb R^k$3 storage and $f:\mathbb R^d\longrightarrow\mathbb R^k$4 time per projected vector. Achlioptas-style discrete projections replace Gaussians by sparse integer-valued entries such as

$f:\mathbb R^d\longrightarrow\mathbb R^k$5

so that rescaling by $f:\mathbb R^d\longrightarrow\mathbb R^k$6 retains JL guarantees while reducing computation to integer add/subtract operations. Sparse subspace embeddings and count-sketch variants push this further by using $f:\mathbb R^d\longrightarrow\mathbb R^k$7 nonzeros per column, yielding $f:\mathbb R^d\longrightarrow\mathbb R^k$8 application time for sparse vectors (Nabil, 2017).

Survey treatments typically distinguish four broad families: dense Gaussian, Achlioptas-type discrete projections, sparse projections of the Li–König–Wright form, and structured fast transforms such as the Subsampled Randomized Hadamard Transform. The computational motivation is straightforward. Dense multiplication costs $f:\mathbb R^d\longrightarrow\mathbb R^k$9 for $f$0 points in $f$1 dimensions, whereas sparse or structured maps reduce either arithmetic or storage. SRHT uses a signed Hadamard transform and row sampling; Count-Sketch uses exactly one nonzero per row and can be applied in $f$2 time. The recurring trade-off is that denser matrices usually yield better concentration constants, while sparser or more structured constructions sacrifice some statistical sharpness to gain speed, streaming compatibility, or memory efficiency (Xie et al., 2017, Wójcik, 2018).

At the algorithmic level, the generic workflow is remarkably stable across constructions: choose a projection distribution, sample the matrix or its implicit structure, compute $f$3 for each datum, and optionally normalize by $f$4 or orthonormalize rows. This simplicity is one reason RMRP functions as a reusable primitive in large-scale data analysis, where its role is often to shrink the dimension before more problem-specific computation is applied.

4. Tensorized and structured-domain RMRP

When the ambient object is a high-order tensor rather than a flat vector, naively vectorizing the data destroys multilinear structure and makes the ambient dimension exponential in order. Tensorized random projections address this by representing the random projector itself in compressed tensor format. In the Tensor-Train (TT) construction, each projected coordinate is an inner product with a random TT tensor. For TT-Gaussian and TT-Rademacher maps, the expected isometry statement is exact: $f$5 and one obtains the variance bound

$f$6

A corresponding JL-type guarantee holds when

$f$7

Within this framework, TT-based projections are substantially superior to CP-based projections in required sketch size, because the CP bound grows like $f$8, whereas the TT factor $f$9 can be driven toward $P\in\mathbb R^{k\times d}$0 by increasing $P\in\mathbb R^{k\times d}$1 (Rakhshan et al., 2020, Rakhshan et al., 2021).

The distribution of the compressed cores need not be Gaussian. A TT-Rademacher projector draws each core entry from $P\in\mathbb R^{k\times d}$2 with probability $P\in\mathbb R^{k\times d}$3, rescales appropriately, and retains the same expectation and variance bounds as TT-Gaussian maps. Empirically, tensorized Rademacher projection is reported to outperform tensorized Gaussian projection on synthetic data, while also being faster because the random sketch uses sign flips rather than floating-point multiplications (Rakhshan et al., 2021).

Not every compressed tensor format is equally suitable. An important negative result is that the Matrix Product Operator variant is not a Johnson–Lindenstrauss transform: although $P\in\mathbb R^{k\times d}$4, the variance contains a term that does not decrease as $P\in\mathbb R^{k\times d}$5, so the vanishing-variance criterion fails. This establishes a sharp distinction between compression schemes that merely reduce storage and those that preserve the probabilistic geometry needed for JL-style embeddings (Rakhshan et al., 2021).

A separate low-memory line is Tensor Random Projection (TRP), which forms the projector as a Khatri–Rao product

$P\in\mathbb R^{k\times d}$6

TRP reduces storage from $P\in\mathbb R^{k\times d}$7 to $P\in\mathbb R^{k\times d}$8, preserves expected norms, and admits a reduced-variance extension $P\in\mathbb R^{k\times d}$9. For $f(x)=Px$0, a non-asymptotic tail bound yields $f(x)=Px$1, which is weaker than the optimal JL logarithmic dependence but is accompanied by substantial empirical savings in memory and query cost (Sun et al., 2021).

5. Applications in machine learning, numerical linear algebra, optimization, and deep networks

The most classical applications are geometric. After projection into $f(x)=Px$2, one can run $f(x)=Px$3-means or hierarchical clustering with lower per-iteration cost, use one-bit or angular random projections inside locality-sensitive hashing, and train linear SVMs or logistic regression in projected space. Compressed sensing uses essentially the same principle under the RIP formulation. These uses explain why random projection is often described as a unifying mathematical basis underlying disparate machine-learning algorithms (Nabil, 2017).

In numerical linear algebra, RMRP functions as a sketching mechanism. Randomized low-rank approximation draws a random test matrix $f(x)=Px$4, forms $f(x)=Px$5, orthonormalizes to a basis $f(x)=Px$6, and computes a reduced SVD of $f(x)=Px$7. In kernel methods, random Fourier features replace an expensive kernel matrix by Monte Carlo features

$f(x)=Px$8

so that $f(x)=Px$9 with variance $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$0. These two strands illustrate the broader RMRP theme: randomized embeddings can either reduce dimension directly or lift data into an explicit randomized feature space that makes subsequent linear methods practical (Advani et al., 2020).

Optimization provides a different extension. In linear programming, sub-Gaussian random projectors preserve not only Euclidean distances but also distance to a cone. This leads to a probabilistic algorithm that replaces the original equality system $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$1 by the projected system $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$2, solves the smaller LP, and then retrieves an approximate solution. Under the paper’s assumptions, the optimal values satisfy

$V=\{v_1,\dots,v_n\}\subset\mathbb R^d$3

with high probability when $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$4. The accompanying experiments report practical speedups on large dense problems and successful application to $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$5-based error-correction decoding (Vu et al., 2017).

Deep learning uses RMRP in two distinct ways. One line treats random projection as an explicit architectural layer: a fixed or learnable RP front end reduces extremely high-dimensional sparse input before the trainable network begins. Another line uses RP distributions for parameter initialization. In theoretical work, fixed random projection layers are used to show that approximation complexity for sparse or manifold-structured data depends on the compressed dimension $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$6, not directly on the ambient dimension $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$7. In empirical work, RP layers make it possible to train on data with as many as tens of millions of input features and training examples, and SRHT-based initialization improves very deep residual networks relative to classical He initialization on several benchmarks (Cai et al., 2018, Wójcik, 2018).

6. Physical and system-level realizations

A notable physical realization of random projection is implemented by light propagation in disordered photonic lattices. If $V=\{v_1,\dots,v_n\}\subset\mathbb R^d$8 denotes the complex modal amplitudes in a waveguide array, then

$V=\{v_1,\dots,v_n\}\subset\mathbb R^d$9

with $0<\varepsilon<1$0 real-symmetric, diagonal disorder $0<\varepsilon<1$1, and nearest-neighbor evanescent couplings $0<\varepsilon<1$2. In the diffusive regime—intermediate diagonal disorder $0<\varepsilon<1$3 with $0<\varepsilon<1$4 of order the uniform coupling $0<\varepsilon<1$5, and sufficiently long propagation length $0<\varepsilon<1$6—the entries of $0<\varepsilon<1$7 become statistically independent complex Gaussians with variance scaling approximately as $0<\varepsilon<1$8. Selecting $0<\varepsilon<1$9 output channels by a projector $f:\mathbb R^d\to\mathbb R^k$0 yields the embedding $f:\mathbb R^d\to\mathbb R^k$1, and the normalized map satisfies the Johnson–Lindenstrauss guarantee for $f:\mathbb R^d\to\mathbb R^k$2 input points when $f:\mathbb R^d\to\mathbb R^k$3 (Miri, 2021).

The photonic setting also makes transport regime explicit. For $f:\mathbb R^d\to\mathbb R^k$4, propagation is ballistic; for $f:\mathbb R^d\to\mathbb R^k$5, impulse responses spread over the full array and $f:\mathbb R^d\to\mathbb R^k$6 become i.i.d. $f:\mathbb R^d\to\mathbb R^k$7; for $f:\mathbb R^d\to\mathbb R^k$8, Anderson localization produces sparse, non-Gaussian transfer matrices. This identifies intermediate disorder, rather than maximal disorder, as the regime relevant to near-isometric optical embedding. The proposed implementations include femtosecond-laser-written waveguide arrays in glass and silicon-photonics lithography, with input coupling by a spatial-light modulator or phase-and-amplitude mask, empirical full mixing at $f:\mathbb R^d\to\mathbb R^k$9, and representative numerical or experimental values such as distance distortion $f(x)=Px$00 for $f(x)=Px$01, footprint of a few $f(x)=Px$02, $f(x)=Px$03 ps transit time, and throughput $f(x)=Px$04 vectors/s (Miri, 2021).

A more recent system-level use of the term RMRP appears in mobile robotics, where a random mapping

$f(x)=Px$05

is followed by a sparse Achlioptas projection $f(x)=Px$06 to form the linear parametric map

$f(x)=Px$07

This framework is coupled to a Residual Energy Preservation Theorem and then integrated into a perception-aware trajectory planner. In that system, occupancy is modeled as a classifier $f(x)=Px$08, ESDF as a regressor $f(x)=Px$09, and both admit closed-form gradients through $f(x)=Px$10. The same construction is used to unify grid and ESDF maps for UAV planning and to model terrain for UGV planning (Nie et al., 12 Jul 2025).

7. Limitations, misconceptions, and evolving terminology

Several recurrent misconceptions are corrected by the literature itself. First, random projection is not synonymous with random mapping: the former is linear and enjoys JL-style near-isometry, whereas the latter may be nonlinear and may preserve only selected statistics. Second, data-independence is both strength and weakness. Because the projection matrix is usually sampled without regard to class labels, manifolds, or other intrinsic structure, RMRP is simple, parallelizable, and broadly applicable, but it can also incur relatively high distortion compared with data-adaptive methods (Nabil, 2017, Xie et al., 2017).

Third, reduced memory does not automatically imply JL behavior. TRP has strong storage advantages but currently weaker non-asymptotic dependence in the proved $f(x)=Px$11 bound than dense JL maps. MPO-based tensor projections preserve expected norm but fail the vanishing-variance criterion, and therefore are not well-suited random projection maps in the JL sense. Photonic realizations offer constant-time physical multiplication and ultra-high bandwidth, but their randomness is fixed after fabrication, moderate fabrication tolerances and temperature drift can bias the distribution away from perfectly i.i.d. Gaussians, and readout noise may limit precision (Sun et al., 2021, Rakhshan et al., 2021, Miri, 2021).

The acronym itself is not stable across all adjacent literatures. “Random Iteration and Projection Method” denotes a different framework based on random orbits of nonexpansive maps, including random orthogonal projections onto affine subspaces, and is unrelated to JL-style dimensionality reduction despite the overlapping initials (Leśniak, 2013). Conversely, some recent works retain the projection theme but modify the underlying random ensemble. “Free Random Projection,” for example, constructs orthogonal projections from products of independent Haar matrices indexed by words in a free group, thereby inducing a hierarchical bias that improves in-context reinforcement learning on several benchmarks. This suggests that contemporary RMRP research is no longer confined to dense Gaussian or sparse sign matrices, but is expanding toward structured randomness tailored to specific statistical or dynamical biases (Hayase et al., 9 Apr 2025).