Discrete Hartley Transform¶
A real, finite cas-kernel transform encodes a real sequence in N real coefficients and uses the same kernel for inversion up to scale.
Core Idea¶
The discrete Hartley transform maps \(N\) real samples to \(N\) real coefficients using \(\operatorname{cas}\theta=\cos\theta+\sin\theta\). Under an unnormalized forward convention, \(H[k]=\sum_nx[n]\operatorname{cas}(2\pi kn/N)\) and the same sum recovers \(x[n]\) after division by \(N\). Bracewell's 1983 original article put the \(1/N\) factor on the forward direction instead; the conventions are equivalent but must not be mixed.[^ref-a7a52b8e2915]
Scope of Application¶
The real coefficients contain the same information as a complex DFT of real input: with \(F[k]=\sum_nx[n]e^{-2\pi i kn/N}\), its real part is \((H[k]+H[-k])/2\) and imaginary part is \((H[-k]-H[k])/2\) (indices modulo \(N\)). Bracewell proposed real-arithmetic use in convolution and two-dimensional images, but no universal speed advantage follows.[^ref-a7a52b8e2915]
Clarity¶
The DHT is not a cosine-only transform, continuous Hartley integral, or a DFT's real part alone. Its odd Hartley component preserves information corresponding to the DFT imaginary part. It is self-reciprocal up to scale, not literally unscaled in both directions. Generic convolution requires a parity-aware mixed Hartley formula; pointwise multiplication is a special case when a convolving sequence is even.
Manages Complexity¶
For the constructed \(N=4\) sequences \([1,0,0,0]\) and \([0,1,0,0]\), the unnormalized DHTs are \([1,1,1,1]\) and \([1,1,-1,-1]\) respectively. Applying the same cas transform again and dividing by four recovers the inputs. The second case shows how real coefficients retain information about a shift rather than collapsing to a mere real-part projection.
Abstract Reasoning¶
Bracewell's original convolution discussion distinguishes general even/odd mixed terms from an even-filter special case where the odd filter part vanishes and simple Hartley-domain multiplication works.[^ref-a7a52b8e2915] Thus the transform's real arithmetic and shared kernel simplify representation, while composition for an arbitrary filter may be more complicated than the familiar DFT product.
Knowledge Transfer¶
Equivalent representations can retain information while changing which operations are simple. For this named transform, the cas kernel, modular even/odd relation and matching normalization are indispensable; a generic “change basis” slogan would not distinguish it from other transforms. Live Linear Operator is the strict domain-specific parent; the DFT is related, not the necessary genus.
[^ref-a7a52b8e2915]: R. N. Bracewell, “Discrete Hartley transform,” Journal of the Optical Society of America 73 (1983), 1832–1835, original full-text mirror. https://gwern.net/doc/math/1983-bracewell.pdf
Relationships to Other Abstractions¶
Current abstraction Discrete Hartley Transform Domain-specific
Parents (1) — more general patterns this builds on
-
Discrete Hartley Transform is a kind of Linear Operator Domain-specific
The DHT is a finite real linear operator with a particular cas-kernel matrix and inverse convention.
Hierarchy path (1) — routes to 1 parentless root
- Discrete Hartley Transform → Linear Operator → Mathematical Operator → Function (Mapping)
Neighborhood in Abstraction Space¶
Discrete Hartley Transform sits in a sparse region of the domain-specific corpus (85th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (2551 abstractions)
Nearest neighbors
- Regularization by spectral filtering — 0.82
- Lifting Scheme — 0.82
- Blind deconvolution — 0.82
- Upsampling — 0.81
- Riesz potential — 0.81
Computed from structural-signature embeddings · 2026-10-08