Multi-Dimensional Hyperspace Vectors

• Multi-Dimensional Hyperspace Vectors are high-dimensional representations that encode symbols, features, and sequences using HDC and VSA frameworks.
• They employ core algebraic operations—bundling, binding, and permutation—to combine random and structured encodings in binary, bipolar, and real-valued forms.
• MDHVs optimize accuracy and efficiency in applications such as bioinformatics, image retrieval, and graph-based localization while managing resource trade-offs.

Searching arXiv for the cited MDHV and hyperdimensional computing papers to ground the article in current literature. arxiv_search(query="(Aygun et al., 2023) OR (Stock et al., 2024) OR (Neubert et al., 2021) OR (Kazemi et al., 2021) OR (Clarkson et al., 2023) OR (Basaklar et al., 2021) OR (Gufran et al., 15 Jul 2025)", max_results=10) Multi-Dimensional Hyperspace Vector (MDHV) denotes a high-dimensional representation used in Hyperdimensional Computing (HDC) and related Vector Symbolic Architecture (VSA) systems. In the HDC literature, MDHVs are often treated as synonymous with hypervectors or high-dimensional vectors: single points in a very high-dimensional space used to encode symbols, features, sequences, graphs, signals, or whole data instances. Their defining operational properties are near-orthogonality under random initialization, concentration of similarity statistics in large dimension, robustness to partial corruption, and compatibility with a small algebra of bundling, binding, and permutation. In more recent task-specific systems, the term also acquires a narrower meaning, referring to a deliberately constructed composite representation rather than merely an atomic random vector (Aygun et al., 2023, Stock et al., 2024, Gufran et al., 15 Jul 2025).

1. Formal definition and representational domains

In its most standard form, an MDHV is a vector of dimension $d$ or $D$, typically in the range $1\text{K}$ to $10\text{K}$, with other works reporting $N \approx 10{,}000$ and $D=10{,}000$–$50{,}000$ depending on the application and representational regime (Aygun et al., 2023, Stock et al., 2024, Neubert et al., 2021). The basic domains are:

$$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$

with real-valued MDHVs often normalized so that $\|\mathbf{h}\|=1$ (Aygun et al., 2023). A bipolar MDHV is commonly generated by sampling each coordinate independently from the Rademacher distribution,

$P(h_i=+1)=P(h_i=-1)=\tfrac12,$

and the same independence principle is used for binary or real-valued atomic vectors (Stock et al., 2024). In multi-bit HDC, the representation is generalized further to integer-valued MDHVs,

$D$0

with $D$1 in the MIMHD formulation (Kazemi et al., 2021).

The literature distinguishes several functional roles for MDHVs. Atomic or symbol vectors represent primitive entities such as amino acids, nucleotides, or discrete symbols. Level hypervectors encode quantization bins of continuous features. Sample hypervectors aggregate feature-wise encodings for one datum, and class encoders or class memories accumulate sample hypervectors over training instances (Stock et al., 2024, Basaklar et al., 2021). This usage indicates that “MDHV” refers less to one fixed datatype than to a family of high-dimensional encodings that share common algebraic and statistical behavior.

A recurring misconception is that MDHVs are necessarily binary and random. The surveyed literature explicitly includes binary, bipolar, real-valued, sparse, quasi-random, and multi-bit constructions, and it also distinguishes uncorrelated atomic vectors from structured correlated vectors such as level encodings and sequence encodings (Aygun et al., 2023, Kazemi et al., 2021).

2. Core algebra: bundling, binding, and permutation

The operational core of MDHV-based systems consists of three transformations: bundling, binding, and permutation. These are the same primitives that organize the broader VSA literature (Neubert et al., 2021, Clarkson et al., 2023).

Bundling, or superposition, combines several vectors so that the result remains similar to each constituent. In the bipolar case, the standard construction is

$D$2

or equivalently

$D$3

In binary space, majority thresholding plays the analogous role; in real-valued space, bundling may remain an average or a rescaled sum (Aygun et al., 2023, Stock et al., 2024, Neubert et al., 2021).

Binding forms an association that is intended to be dissimilar to either input while remaining algebraically tractable. In bipolar or real-valued regimes, binding is usually element-wise multiplication,

$D$4

and in binary regimes it is realized as XOR,

$D$5

Circular convolution is another established binding operator,

$D$6

or, in the biological-data review’s notation,

$D$7

with approximate inversion through circular correlation (Aygun et al., 2023, Stock et al., 2024). The reversible character of binding underlies the claim that HDC admits white-box-style explainability and tractable decomposition of complex representations (Stock et al., 2024).

Permutation injects order, role, or positional information by fixed reindexing of coordinates. A one-step circular shift is a standard example,

$D$8

or, in alternate notation, $D$9 (Aygun et al., 2023, Stock et al., 2024). Powers of the permutation operator, such as $1\text{K}$0 or $1\text{K}$1, are used to encode position in sequences, temporal order in signals, or role distinctions in structured symbolic representations.

These operators are simple enough to map efficiently onto bit-parallel hardware. The literature specifically notes XOR, bit-count, thresholding, and similar primitives as suitable for CPUs, GPUs, and FPGAs with minimal energy per operation (Stock et al., 2024). This suggests that the practical identity of an MDHV is inseparable from the operator algebra defined over it.

3. Encoding methodologies

MDHV encoding methods divide broadly into uncorrelated random generation, sparse generation, structured correlated encoding, and learned or projected embeddings (Aygun et al., 2023).

The canonical random construction samples each coordinate independently. For bipolar vectors, one formulation is

$1\text{K}$2

and an equivalent form is

$1\text{K}$3

Sparse variants replace the dense $1\text{K}$4 regime by vectors with many zeros, for example

$1\text{K}$5

or the “random indexing” case in which $1\text{K}$6 with only a small fraction of nonzero entries (Aygun et al., 2023).

Structured encoding deliberately introduces correlations. In level-based or record encoding for a quantized value $1\text{K}$7, one chooses a random base trace $1\text{K}$8 and then generates successive levels by controlled bit flips,

$1\text{K}$9

so that adjacent levels remain correlated (Aygun et al., 2023). Ordered sequences use N-gram or permutation encoding,

$10\text{K}$0

and biological sequence models further combine permutation and binding to encode $10\text{K}$1-mers,

$10\text{K}$2

Graphs and phylogenies can be encoded by binding endpoint vectors and bundling all edges, for example

$10\text{K}$3

These constructions show that an MDHV is not merely a random codeword but the output of a compositional encoding grammar (Stock et al., 2024).

A more explicitly learning-oriented workflow projects a real-valued input $10\text{K}$4 into hyperspace via an item-memory matrix $10\text{K}$5. With

$10\text{K}$6

or a quasi-random alternative such as Sobol or VDC, one computes

$10\text{K}$7

Class memories are then accumulated by superposition,

$10\text{K}$8

and inference reduces to nearest-prototype comparison using cosine similarity (Aygun et al., 2023). In image descriptor aggregation, the corresponding fusion formula is

$10\text{K}$9

which combines feature vectors and position encodings into one MDHV of fixed dimensionality (Neubert et al., 2021).

4. Similarity, orthogonality, and capacity

The mathematical utility of MDHVs arises from concentration phenomena in very high dimensions. For two MDHVs $N \approx 10{,}000$0 and $N \approx 10{,}000$1, the standard similarity measure in bipolar or real-valued spaces is cosine similarity,

$N \approx 10{,}000$2

For random bipolar vectors, the survey reports

$N \approx 10{,}000$3

and gives a near-orthogonality threshold

$N \approx 10{,}000$4

Binary settings use normalized Hamming similarity, Hamming distance, or Jaccard similarity for sparse binary vectors (Aygun et al., 2023, Stock et al., 2024).

The bioinformatics review characterizes this regime as a “blessing of dimensionality”: random vectors are nearly orthogonal; element-wise statistics concentrate with variance $N \approx 10{,}000$5; and corrupting up to $N \approx 10{,}000$6 bits for small $N \approx 10{,}000$7 leaves the vector closer to its original than to an unrelated vector (Stock et al., 2024). The same concentration underlies robustness after bundling and the detectability of weak correlations introduced by structured encoding.

Capacity analysis makes these intuitions explicit. One summary bound in the hypervector-encoding survey states that reliable retrieval with error rate $N \approx 10{,}000$8 requires

$N \approx 10{,}000$9

The VSA capacity analysis sharpens this for specific symbolic tasks. For set membership in MAP-I or MAP-B, the required dimension satisfies

$D=10{,}000$0

while for estimating set-intersection size in MAP-I with additive or relative tolerance $D=10{,}000$1,

$D=10{,}000$2

For Bloom-filter-like sparse binary VSAs, the analysis yields the familiar membership scaling

$D=10{,}000$3

and also provides intersection-size bounds for Bloom and Counting-Bloom variants (Clarkson et al., 2023).

These results are significant because they connect MDHVs to two established analytical traditions: Johnson–Lindenstrauss-type random sketching in the dense real or bipolar case, and Bloom-filter or count-min-sketch reasoning in the sparse binary case (Clarkson et al., 2023). A common misconception is that “higher dimension is always better.” The literature instead treats $D=10{,}000$4 as a trade-off parameter: larger $D=10{,}000$5 improves orthogonality and retrieval reliability but increases area, power, storage, and latency (Aygun et al., 2023, Basaklar et al., 2021).

5. Multi-bit, hardware-aware, and optimized MDHV design

Although early HDC formulations emphasize binary or bipolar vectors, recent work treats MDHV design itself as an optimization target. Two lines are especially prominent: multi-bit in-memory HDC and dimensionality/robustness co-optimization for edge deployment (Kazemi et al., 2021, Basaklar et al., 2021).

In the MIMHD framework, each feature value $D=10{,}000$6 is first quantized by

$D=10{,}000$7

and each quantization level is assigned a random MDHV $D=10{,}000$8. Base hypervectors $D=10{,}000$9 encode feature positions, and an $50{,}000$0-dimensional feature vector $50{,}000$1 is mapped to

$50{,}000$2

After accumulation, each component is saturated or uniformly quantized back to $50{,}000$3 bits. Binding is implemented as element-wise multiplication modulo $50{,}000$4,

$50{,}000$5

and similarity can be computed by quantized cosine similarity or by a conductance-based FeFET-MCAM distance metric (Kazemi et al., 2021).

The hardware realization uses FeFET crossbar arrays for multiply-and-add encoding and FeFET multi-bit content-addressable memories for associative search. To accommodate the mismatch between cosine-trained class memories and the MCAM search metric, the authors introduce Hardware-Aware Retraining (HWART), updating high-precision class accumulators according to

$50{,}000$6

followed by quantization after each mini-batch (Kazemi et al., 2021). For $50{,}000$7 over six datasets, the reported average accuracies are $50{,}000$8 for binary, $50{,}000$9 for 2-bit, $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$0 for 3-bit, and $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$1 for 8-bit cosine-based reference. At the same dimensionality, the reported energy improvements over a GPU are approximately $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$2, $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$3, and $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$4 for 1-, 2-, and 3-bit regimes, with latency speedups of approximately $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$5, $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$6, and $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$7, respectively; 3-bit MIMHD is further reported as $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$8 more energy-efficient and $$\mathbf{h}\in\{0,1\}^d,\qquad \mathbf{h}\in\{+1,-1\}^d,\qquad \mathbf{h}\in\mathbb{R}^d,$$9 faster than a binary HDC accelerator at equal accuracy (Kazemi et al., 2021).

A different optimization perspective appears in edge-device hypervector design. There, level hypervectors $\|\mathbf{h}\|=1$0 encode feature quantization levels, sample hypervectors are

$\|\mathbf{h}\|=1$1

and class encoders are

$\|\mathbf{h}\|=1$2

The design variable is a matrix $\|\mathbf{h}\|=1$3 whose entries $\|\mathbf{h}\|=1$4 specify how many bits are flipped between adjacent levels, under the constraint

$\|\mathbf{h}\|=1$5

The optimization jointly maximizes weighted accuracy and minimizes average inter-class similarity, producing a Pareto front over accuracy, robustness, and resource usage (Basaklar et al., 2021). Reported outcomes include $\|\mathbf{h}\|=1$6–$\|\mathbf{h}\|=1$7 model-size reduction, $\|\mathbf{h}\|=1$8–$\|\mathbf{h}\|=1$9 dimension compression, $P(h_i=+1)=P(h_i=-1)=\tfrac12,$0–$P(h_i=+1)=P(h_i=-1)=\tfrac12,$1 inference-time speedup on an Odroid XU3, $P(h_i=+1)=P(h_i=-1)=\tfrac12,$2–$P(h_i=+1)=P(h_i=-1)=\tfrac12,$3 lower energy consumption, maintained or improved classification accuracy, and $P(h_i=+1)=P(h_i=-1)=\tfrac12,$4–$P(h_i=+1)=P(h_i=-1)=\tfrac12,$5 improvement in robustness $P(h_i=+1)=P(h_i=-1)=\tfrac12,$6 along Pareto-front points (Basaklar et al., 2021). This suggests that MDHV design can be treated as a non-convex, application-specific optimization problem rather than a fixed random initialization heuristic.

6. Application domains and the task-specific reinterpretation of MDHV

MDHVs have been used as a unifying representational framework across biological data, image retrieval, biosignals, and graph-based localization (Stock et al., 2024, Neubert et al., 2021, Gufran et al., 15 Jul 2025).

In biological sequence analysis, reference genomes, reads, proteins, phylogenies, and multimodal omics data can all be reduced to fixed-size MDHVs by sequence $P(h_i=+1)=P(h_i=-1)=\tfrac12,$7-mer encoding, edge binding in graphs, and bundling over structured relations. The review on biological data reports that tools such as Demeter, HDNA, GenieHD, BioHD, and HDGIM encode each reference genome by bundling $P(h_i=+1)=P(h_i=-1)=\tfrac12,$8-mer hypervectors into a single fixed-size MDHV, after which classification becomes a similarity search between read and reference MDHVs. It further reports that Demeter’s in-memory XOR/XNOR pipelines on memristive hardware achieved $P(h_i=+1)=P(h_i=-1)=\tfrac12,$9 speedup and $D$00 memory reduction over Kraken2/MetaCache with $D$01 accuracy loss; HyperSpec achieved $D$02 speed gains for mass-spectrometry clustering; and HDC processors on ultra-low-power wearable hardware achieved real-time seizure or gesture detection with state-of-the-art accuracy and orders-of-magnitude energy savings (Stock et al., 2024).

In computer vision and mobile robotics, MDHVs serve as a systematic aggregation space for image descriptors. Feature vectors and image positions are bound and bundled into a single holistic descriptor of the same dimensionality, and the reported place-recognition experiments show a $D$03 improvement in average performance compared to runner-up and $D$04 better worst-case performance (Neubert et al., 2021). This is significant because it demonstrates that MDHVs can combine learned descriptors with additional symbolic structure rather than merely replace conventional embeddings.

A more specialized reinterpretation appears in GATE for indoor localization. Here, the MDHV is not an atomic HDC hypervector but a composite node representation in a graph neural system. For a node $D$05 with fingerprint $D$06 and neighbors $D$07, scalar attention is defined by

$D$08

message passing by

$D$09

and the Attention Hyperspace Vector (AHV) by the $D$10 matrix with entries

$D$11

The final representation is

$D$12

The paper states that this construction mitigates the GNN oversquashing or blind-spot problem by retaining both a global neighbor summary and a full feature-wise tensor of neighbor influence, and then feeding the result as an “image-like” feature map into a 2-layer GCN (Gufran et al., 15 Jul 2025). In ablation across five real buildings and seven smartphones, the reported mean localization errors are $D$13 m for GATE-Full, $D$14 m for GATE-No-MSG, $D$15 m for GATE-No-AHV, and $D$16 m for GATE-No-MDHV, with GATE-Full also showing the lowest device-variance at $D$17 m and stability when up to $D$18 of RSS features are dropped (Gufran et al., 15 Jul 2025).

This task-specific usage clarifies an important terminological point. In most HDC and VSA work, “MDHV” names the general class of high-dimensional symbolic representations. In GATE, it names a particular concatenated representation built from raw features, message passing, and attention structure. The two usages are compatible, but they are not identical. A precise reading of the literature therefore treats MDHV as a general representational paradigm with multiple concrete instantiations rather than as one canonical vector format.