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.

Version: v1 · 2026-10-03 · History
Domain-specific #: 13154
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomain: Digital Signal Processing → Mathematics
Aliases: DHT

Core Idea¶

The discrete Hartley transform (DHT) sends a real length-\(N\) sequence \(x[n]\) to a real length-\(N\) sequence \(H[k]\) using the kernel \(\operatorname{cas}\theta=\cos\theta+\sin\theta\). Under the unnormalized-forward convention used here,

\[H[k]=\sum_{n=0}^{N-1}x[n]\operatorname{cas}(2\pi kn/N),\qquad x[n]=\frac1N\sum_{k=0}^{N-1}H[k]\operatorname{cas}(2\pi kn/N).\]

Thus one cas-kernel operation performs both directions, with a factor \(1/N\) on inversion.^[1] Bracewell's 1983 original article places \(1/N\) on the forward transform and not on its inverse. These are equivalent normalization choices; mixing them within one calculation is not.

For real input, the DHT carries the same information as a discrete Fourier transform (DFT), but it is not merely the DFT's real part. If \(F[k]=\sum_nx[n]e^{-2\pi i kn/N}\), then \(\Re F[k]=(H[k]+H[-k])/2\) and \(\Im F[k]=(H[-k]-H[k])/2\), with indices modulo \(N\).^[1] The antisymmetric Hartley part preserves the sine/imaginary information that would otherwise be lost.

Structural Signature¶

Finite real input: \(N\) ordered real samples provide the domain.
Real cas kernel: cosine plus sine mixes the even and odd spectral contributions.
Real output: \(N\) real coefficients retain enough information to reconstruct \(x\).
Reciprocal inverse: the same kernel applies again, with the matching \(N\) scale.
Parity-aware relation: even and odd Hartley parts recover complex DFT components.
Conditional convolution simplification: componentwise multiplication is straightforward for an even filter; generic convolution needs a mixed rule.

Sig role-phrases: ordered real samples; cosine-plus-sine kernel; real coefficient array; same-kernel inverse with scale; even/odd spectral recovery; parity-qualified convolution.

What It Is Not¶

The DHT is not a discrete cosine transform, which omits the sine contribution, nor a continuous Hartley integral, which does not take a finite \(N\)-point sequence. It is not the DFT itself: the DFT outputs complex coefficients under the stated convention, whereas the DHT stores the same real-input information in a different real basis. It is also not an assertion that every convolution becomes elementwise Hartley multiplication. Bracewell's general convolution rule combines even/odd or reversed-index terms; the simple product appears in a special even-filter setting.^[1]

Scope of Application¶

Bracewell proposed the finite transform as a real-arithmetic alternative in numerical spectral work and discussed convolution and two-dimensional images.^[1]^[2] Those uses follow from the exact transform and its algebra, not from a universal runtime advantage. A modern complex FFT can be highly optimized, and particular hardware, data shape, and operation determine which implementation is faster. Here the entry is about the transform's mathematical identity, not a performance promise or a particular fast Hartley algorithm.

Clarity¶

Record the normalization before comparing formulas. With the convention above, a unit impulse at sample zero transforms to the all-ones vector, and applying the same unnormalized transform again produces \(N\) times that impulse. Bracewell's printed Eq. 5 instead scales the forward output by \(1/N\), giving the corresponding reverse placement of the factor. Also keep the modular negative index: \(H[-k]\) means \(H[(N-k)\bmod N]\). The difference between \(H[k]\) and \(H[-k]\) is exactly where asymmetric input stores information analogous to DFT phase.

Manages Complexity¶

The DHT avoids representing a real sequence's spectrum as \(N\) complex values when \(N\) real values suffice. Its shared forward/inverse kernel can simplify a library interface. This economy is not free of algebraic obligations: DFT-domain convolution uses complex componentwise multiplication, whereas Hartley convolution usually needs the even/odd mirrored mixture. Bracewell notes that an even convolving sequence has vanishing odd Hartley part, so its special case simplifies to real componentwise products and another transform.^[1] This is a real operational boundary, not a claim that the transforms differ in retained information.

Abstract Reasoning¶

Write \(H_e[k]=(H[k]+H[-k])/2\) and \(H_o[k]=(H[k]-H[-k])/2\). With the negative-exponential DFT convention, \(F[k]=H_e[k]-iH_o[k]\). Therefore the real Hartley array is sufficient for reconstructing complex Fourier information of real input. This is a coordinate conversion, not a lossy projection. The inverse relation follows from orthogonality of the finite cas matrix; using it twice returns \(N x\) under our convention. If the odd part is discarded, however, asymmetric signals can no longer be distinguished by their Fourier imaginary part.

Knowledge Transfer¶

The transferable lesson inside signal processing is to distinguish the information carried by a representation from the arithmetic it makes convenient. DHT and DFT can encode the same real data while favoring different primitive operations. The exact cas kernel, modular parity relation, and normalization are indispensable here; replacing them with “convert data to another view” would erase the mathematical identity. Any general representation-change prime is a separate question, not a necessary parent inferred from similarity to the live DFT page.

Examples¶

Two constructed four-sample signals. Let \(N=4\). For \(x=[1,0,0,0]\), every cas argument is zero and \(H=[1,1,1,1]\). Shift the impulse one sample to \(x=[0,1,0,0]\); evaluating cas at \(0,\pi/2,\pi,3\pi/2\) gives \(H=[1,1,-1,-1]\). A second DHT divided by four reconstructs each respective input. Mapped back: the ordered real samples feed the cas kernel, which yields real spectra; the shifted impulse changes the odd/parity structure without losing invertibility. These arrays are direct calculations from Bracewell's formula, not arrays printed in his paper.^[1]

Bracewell's convolution setting. The 1983 article derives a general DHT convolution theorem with an even/odd mixture. It then notes that when one convolving sequence is even, its odd part vanishes and the transformed convolution reduces to a pointwise product; he proposes real-transform filtering and discusses extension to two-dimensional image data.^[1] Mapped back: the signal and even filter are finite real sequences, DHT coefficients replace complex DFT arithmetic, and the evenness condition is what makes the simple product valid. Without that condition, copying this procedure to an arbitrary filter would produce the wrong result.

Structural Tensions¶

There is an arithmetic-versus-composition tradeoff in the original proposed use. Real coefficients and a reciprocal routine can reduce representational and implementation burden, yet a generic convolution needs parity-aware cross terms rather than the DFT's single complex product. In the even-filter special case the extra algebra collapses. Diagnostic: before claiming a Hartley-domain pointwise filter, check whether the filter is even under the chosen cyclic indexing; if not, use the full convolution identity. This is a choice of representation for a task, not a theorem that DHT is universally faster.

Structural–Framed Character¶

The transform is predominantly structural: for specified \(N\), input, kernel and normalization, its coefficients and inverse are algebraically fixed. Practitioners choose the normalization and implementation, and Bracewell's engineering context gave prominence to real arithmetic and image/convolution use; neither choice changes the identity of the cas transform. Vocabulary travels between mathematics, signal processing and imaging because the same finite kernel can operate on samples or image axes. It is recognized genuinely when the real cas relation and invertible scaling survive. Calling any “real Fourier-like” method a DHT, or importing simple DFT convolution multiplication without its parity condition, is a superficial import. Its character: an exact finite real spectral representation with convention-dependent scaling and task-dependent computational advantages.

Structural Core vs. Domain Accent¶

The skeleton is an invertible finite linear transform whose matrix entries are \(\operatorname{cas}(2\pi kn/N)\) and whose inverse uses that matrix again up to scale. The domain-bound mechanism is real sampled signals and modular frequency indices; the even/odd split relates it to complex DFT coefficients. Image filtering, fast algorithms and whether multiplication counts favor a machine are accents, not definitional. The named entry fails the prime bar because removing the cas kernel, index symmetry and \(N\)-point inverse leaves only generic “re-encode a signal,” which cannot distinguish DHT from DFT or DCT. The live Linear Operator is the verified strict domain-specific parent; the DFT is related and information-equivalent for real data but not its necessary genus.

This entry is a kind of Linear Operator.

The live Linear Operator is the strict domain-specific parent: for each N the DHT is a linear map on real N-vectors. The live discrete Fourier transform is a related equivalent representation for real input, and discrete-time Fourier transform differs by its frequency-domain construction for sequences rather than this particular finite array. Equivalence alone is not a DAG relation.

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 (85^th 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

Not to Be Confused With¶

Discrete Fourier transform: complex exponential kernel and complex coefficients under the stated convention.
Discrete cosine transform: cosine-only basis, different boundary conventions and data relation.
Continuous Hartley transform: integral, not the \(N\)-point discrete sum.
Fast Hartley transform: an algorithm for computing a DHT, not the transform's definition.
Generic pointwise Hartley-domain convolution: missing even/odd mixed terms unless an appropriate even-filter special case applies.

References¶

[1] R. N. Bracewell, “Discrete Hartley transform,” Journal of the Optical Society of America 73 (1983), 1832–1835, especially Eqs. 5–6 and the even/odd and convolution sections. Original article full-text mirror. https://gwern.net/doc/math/1983-bracewell.pdf registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g

[2] R. N. Bracewell, “Discrete Hartley transform”, Journal of the Optical Society of America 73 (1983), 1832–1835, publisher record for the same article. registry ↩