The Regularized Fast Hartley Transform

Low Complexity Parallel Computation

of the FHT in One and Multiple
Dimensions 2nd Edition Jones Keith
John
Visit to download the full and correct content document:
https://ebookmeta.com/product/the-regularized-fast-hartley-transform-low-complexity-
parallel-computation-of-the-fht-in-one-and-multiple-dimensions-2nd-edition-jones-keit
h-john/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

The Regularized Fast Hartley Transform Low Complexity

Parallel Computation of the FHT in One and Multiple
Dimensions 2nd Edition Keith John Jones

https://ebookmeta.com/product/the-regularized-fast-hartley-
transform-low-complexity-parallel-computation-of-the-fht-in-one-
and-multiple-dimensions-2nd-edition-keith-john-jones/

The Risk Management Handbook: A Practical Guide to

Managing the Multiple Dimensions of Risk, 2nd Edition
David Hillson

https://ebookmeta.com/product/the-risk-management-handbook-a-
practical-guide-to-managing-the-multiple-dimensions-of-risk-2nd-
edition-david-hillson/

Leadership 2nd Edition Keith Grint Owain Smolovic Jones

https://ebookmeta.com/product/leadership-2nd-edition-keith-grint-
owain-smolovic-jones/

Parallel Scientific Computation: A Structured Approach

Using BSP 2nd Edition Rob H. Bisseling

https://ebookmeta.com/product/parallel-scientific-computation-a-
structured-approach-using-bsp-2nd-edition-rob-h-bisseling/
Economic Dynamics: Theory and Computation 2nd Edition
John Stachurski

https://ebookmeta.com/product/economic-dynamics-theory-and-
computation-2nd-edition-john-stachurski/

Multiple Systems: Complexity and Coherence in

Ecosystems, Collective Behavior, and Social Systems 1st
Edition Gianfranco Minati

https://ebookmeta.com/product/multiple-systems-complexity-and-
coherence-in-ecosystems-collective-behavior-and-social-
systems-1st-edition-gianfranco-minati/

The Riesz Transform of Codimension Smaller Than One and

the Wolff Energy 1st Edition Benjamin Jaye

https://ebookmeta.com/product/the-riesz-transform-of-codimension-
smaller-than-one-and-the-wolff-energy-1st-edition-benjamin-jaye/

Numerical Infinities and Infinitesimals in Optimization

Emergence Complexity and Computation 43 Yaroslav D.
Sergeyev (Editor)

https://ebookmeta.com/product/numerical-infinities-and-
infinitesimals-in-optimization-emergence-complexity-and-
computation-43-yaroslav-d-sergeyev-editor/

Fast Facts Multiple Sclerosis 5th Edition Stella E

Hughes Gabrielle Macaron

https://ebookmeta.com/product/fast-facts-multiple-sclerosis-5th-
edition-stella-e-hughes-gabrielle-macaron/
Keith John Jones

The Regularized
Fast Hartley
Transform
Low-Complexity Parallel Computation
of the FHT in One and Multiple
Dimensions
Second Edition
The Regularized Fast Hartley Transform
Keith John Jones

The Regularized Fast Hartley

Transform
Low-Complexity Parallel Computation of the
FHT in One and Multiple Dimensions

Second Edition
Keith John Jones
Wyke Technologies Ltd.
Weymouth, Dorset, UK

ISBN 978-3-030-68244-6 ISBN 978-3-030-68245-3 (eBook)

https://doi.org/10.1007/978-3-030-68245-3

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2010, 2022
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

Most real-world spectrum analysis problems involve the computation of the real-
data discrete Fourier transform (DFT), a unitary transform that maps elements of the
linear (or vector) space of real-valued N-tuples, RN, to elements of its complex-
valued counterpart, CN. The computation is conventionally carried out via a ‘real-
from-complex’ strategy using a complex-data version of the familiar fast Fourier
transform (FFT), the generic name given to the class of fast recursive algorithms
used for the efficient computation of the DFT. Such algorithms are typically derived
by exploiting the property of symmetry, whether it exists in just the transform kernel
or, in certain circumstances, in the input data and/or output data as well. When the
input data to the DFT is real-valued, for example, the resulting output data is in the
form of a Hermitian (or conjugate)-symmetric frequency spectrum which may be
exploited to some advantage in terms of reduced arithmetic-complexity.
To make effective use of a complex-data FFT, however, via the chosen real-from-
complex strategy, the input data to the DFT must first be converted from elements of
the linear space RN to those of CN. The reason for choosing the computational
domain of real-data problems such as this to be CN, rather than RN, is due in part
to the fact that manufacturers of computing equipment have invested so heavily in
producing digital signal processing (DSP) devices built around the design of the fast
complex-data multiplier-and-accumulator (MAC). This is an arithmetic unit that’s
ideally suited to the implementation of the radix-2 butterfly, which is the computa-
tional engine used for carrying out the repetitive arithmetic operations required by
the complex-data version of the radix-2 FFT.
The net result of such a strategy is that the problem of computing the real-data
DFT is effectively modified so as to match an existing complex-data solution, rather
than a solution being sought that matches the actual problem needing to be solved –
which is the approach that’s been adopted in this book.
The accessibility of the increasingly powerful field-programmable gate array
(FPGA) and application-specific integrated circuit (ASIC) technologies is now
giving DSP design engineers far greater control, however, over the type of algorithm
that may be used in the building of high-performance DSP systems, so that more

v
vi Preface

appropriate algorithmically specialized solutions to the real-data DFT may be

actively sought and exploited to some advantage with implementations based upon
the use of these silicon-based technologies. These technologies facilitate the use of
both multiple arithmetic units – such as those based upon the fast multiplier and/or
the CORDIC phase rotator – and multiple banks of fast memory in order to enhance
the performance of key signal processing algorithms, such as the FFT, via their
parallel computation.
The first part of the book, after providing the background information necessary
for a better understanding of both the problems to be addressed and of the proposed
solutions, concerns itself with the design of a new and highly parallel formulation of
the fast Hartley transform (FHT) which is to be used, in turn, for the efficient
computation of the real-data DFT – where both transforms are restricted to the
one-dimensional (1-D) case – which would, in turn, enable it to be used for those
DSP-based problems commonly addressed via the FFT. The FHT is the generic
name given to the class of fast recursive algorithms used for the efficient computa-
tion of the discrete Hartley transform (DHT) – a bilateral and orthogonal transform
and close relative of the DFT that possesses many of the same properties – which, for
the processing of real-valued data, has attractions over the complex-data FFT in
terms of reduced arithmetic and memory requirements. Its bilateral property means
that it may be straightforwardly applied to the transformation from Hartley-space to
data-space as well as from data-space to Hartley-space, thus making it equally
applicable to the computation of both the forward and the inverse DFT algorithms
and an attractive option for carrying out filtering-type operations with real-
valued data.
A drawback, however, of conventional FHT algorithms lies in the lack of
regularity (as relates to the algorithm structure and which equates to the amount of
repetition and symmetry present in the design) arising from the need for two sizes of
butterfly – and thus for two separate butterfly designs – single-sized and double-
sized for efficient fixed-radix formulations where, for a radix ‘R’ algorithm, a single-
sized butterfly produces R outputs from R inputs whilst a double-sized butterfly
produces 2R outputs from 2R inputs. A generic version of the double-sized butterfly,
to be referred to as the generic double butterfly, has therefore been sought for the
radix-4 factorization of the FHT that might overcome the problem in an elegant
fashion, where the resulting single-design solution, to be referred to as the regular-
ized FHT, would be required to lend itself to an efficient implementation with
parallel computing technology – as typified by the FPGA and the ASIC.
Consideration has been given in the design process to the fact that when produc-
ing electronic equipment, whether for commercial or military use, great emphasis is
inevitably placed upon minimizing the unit cost so that the design engineer is seldom
blessed with the option of using the latest state-of-the-art device technology. The
most common situation encountered is one where the expectation is to use the
smallest (and thus least expensive) device that’s capable of yielding solutions able
to meet the desired performance objectives, which means using devices that are often
one or more generations behind the latest specification. As a result, there are
situations where there would be great merit in having designs that are not totally
Preface vii

reliant on the increasing availability of large quantities of expensive embedded

resources, such as the fast multipliers and fast memory provided by the manufac-
turers of the latest silicon-based devices, but are sufficiently flexible as to yield
efficient implementations in silicon even when such resources are scarce.
To help address the problem, several versions of a processing element (PE) – as
required for the low-complexity parallel computation of the generic double
butterfly – have been sought which are each required to be a simple variation of
the same basic design and each compatible with a single-PE computing architecture.
This would enable parallel solutions to be defined for the radix-4 factorization of the
FHT and the real-data FFT that are resource-efficient (whereby just a single PE may
be used by the transform to carry out the repetitive arithmetic operations required by
each and every instance of the large double butterfly), scalable (which refers to the
ease with which the solution may be modified in order to accommodate increasing or
decreasing transform sizes) and possess universal application (in that each new
application would necessitate minimal re-design effort and costs). Such solutions
would enable the use of the available silicon resources to be optimized so as to
maximize the achievable computational density – that is, the throughput per unit area
of silicon – and, in so doing, to match the computational density of the most
advanced commercially available complex-data FFTs (which are invariably based
upon a multi-PE computing architecture) for potentially just a fraction of the silicon
resources.
A further requirement of any design is that it should be able to cater for a range of
resource-constrained environments, as might be encountered in applications typified
by that of mobile communications, for example, where a small battery may be the
only source of power supply for long periods of time so that power-efficiency would
have to be a key requirement in the design of any such solution. Such a design
process inevitably involves particular resources being consumed and traded off, one
against another, this being most simply expressed in the general terms of a trade-off
of the power requirement (which would have to satisfy some pre-defined constraint)
against the space-complexity (through the memory and arithmetic components –
which typically exist as embedded resources on an FPGA – and the programmable
logic) and the time-complexity (through either the latency or the update time –
although for a single-PE architecture the two parameters are identical – which would
be constrained by the data set refresh rate).
The choice of which particular computing device to use for carrying out a
comparative analysis of the various design options to be considered in the book
has not been regarded as being of relevance to the results obtained, as the intention
has been that the attractions of the proposed solutions should be considered to be
valid regardless of the specific device onto which they are mapped – that is, a ‘good’
design should be device-independent. The author is well aware, however, that the
intellectual investment made in achieving such a design may seem to fly in the face
of current wisdom, whereby the need for good engineering design and practice is
often dispensed with through the adoption of ever-more costly and powerful (and
power consuming) computing devices offering seemingly endless quantities of
embedded resources for overcoming design issues.
viii Preface

Thus, in order to meet the desired design objectives subject to the stated con-
straints, a single-PE architecture has been sought for the parallel computation of the
generic double butterfly and of the resulting regularized FHT that would yield
attractive solutions, particularly when implemented with parallel computing tech-
nology, that would be resource-efficient, scalable and device-independent (so that
they would not be dependent upon the specific characteristics of any particular
device, being able to exploit whatever resources happen to be available on the target
device). A high computational throughput has been sought by exploiting a combi-
nation of parallel processing techniques, these including the use of both pipelined
and single-instruction multiple-data (SIMD) processing and the adoption and exploi-
tation of partitioned memory for the parallel storage and retrieval of both the data and
the trigonometric coefficients (as defined by the transform kernel).
The above issues have already been successfully addressed and discussed in some
depth in a previous Springer book from 2010: The Regularized Fast Hartley
Transform: Optimal Formulation of Real-Data Fast Fourier Transform for
Silicon-Based Implementation in Resource-Constrained Environments (Signals
and Communication Technology Series). A new feature of the present edition of
the book – which is an updated and expanded version of the previous edition –
involves the search for attractive solutions for the parallel computation of the multi-
dimensional (m-D) DHT – and, in particular, of a separable version referred to as the
m-D SDHT – and equivalently, via the relationship of their kernels, of the m-D real-
data DFT. The aim has been to use the regularized FHT as a building block in
producing attractive parallel solutions that would exploit the already proven benefits
of the regularized FHT and which, like those solutions for the 1-D case, would be
resource-efficient, scalable and device-independent, being thus able to optimize the
use of the available silicon resources so as to maximize the achievable computational
density. The adoption of the regularized FHT for the provision of such solutions to
the 2-D and 3-D problems, in particular, would enable it to be beneficially used as a
key component in the design of systems for the real-time processing of 2-D and 3-D
images, respectively, as well as that of conventional 1-D signals.

Weymouth, UK Keith John Jones

Audience

The book is aimed at practising DSP engineers, academics, researchers and students
from engineering, computer science and mathematics backgrounds with an interest
in the design and implementation of sequential and parallel algorithms and, in
particular, of the FHT and the real-data FFT for both 1-D and m-D cases. It is
intended, in particular, to provide the reader with the tools necessary to both
understand and implement the new formulations with a choice of simple design
variations that offer clear implementational attractions/advantages, both theoretical
and practical, when compared to more conventional solutions based upon the
adoption of the familiar complex-data FFT.

ix
Acknowledgements

I would like to thank my wife Deborah for suggesting and encouraging the work as I
would otherwise have been forced to engage more fully with my DIY and gardening
duties – with my skills, like my enthusiasm, leaving much to be desired. Thanks also
to our aging cat, Titus, who has spent many an hour sat with me eating biscuits and
staring at an unchanging computer screen!

xi
Contents

Part I The Discrete Fourier and Hartley Transforms

1 Background to Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 The DFT and Its Efficient Computation . . . . . . . . . . . . . . . . . . 4
1.3 Twentieth-Century Developments of the FFT . . . . . . . . . . . . . . 6
1.4 The DHT and Its Relation to the DFT . . . . . . . . . . . . . . . . . . . 8
1.5 Attractions of Computing the Real-Data DFT via the FHT . . . . . 10
1.6 Modern Hardware-Based Parallel Computing Technologies . . . . 11
1.7 Hardware-Based Arithmetic Units . . . . . . . . . . . . . . . . . . . . . . 12
1.8 Performance Metrics and Constraints . . . . . . . . . . . . . . . . . . . . 13
1.9 Key Parameters, Definitions and Notation . . . . . . . . . . . . . . . . . 15
1.10 Organization of Monograph . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2 The Real-Data Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . 23
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 Real-Data FFT Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.1 The Bergland Algorithm . . . . . . . . . . . . . . . . . . . . . . . 25
2.2.2 The Bruun Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Real-From-Complex Strategies . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.1 Computation of Real-Data DFT via
Complex-Data FFT . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.2 Computation of Two Real-Data DFTs via
Complex-Data FFT . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.3 Computation of Real-Data DFT via Half-Length
Complex-Data FFT . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Data Reordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

xiii
xiv Contents

3 The Discrete Hartley Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2 Orthogonality of DHT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3 Decomposition into Even and Odd Components . . . . . . . . . . . . 37
3.4 Connecting Relations Between DFT and DHT . . . . . . . . . . . . . 38
3.4.1 Real-Data DFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.4.2 Complex-Data DFT . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.5 Fundamental Theorems for DFT and DHT . . . . . . . . . . . . . . . . 40
3.5.1 Reversal Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5.2 Addition Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.5.3 Shift Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5.4 Convolution Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5.5 Product Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.5.6 Autocorrelation Theorem . . . . . . . . . . . . . . . . . . . . . . 44
3.5.7 First Derivative Theorem . . . . . . . . . . . . . . . . . . . . . . 44
3.5.8 Second Derivative Theorem . . . . . . . . . . . . . . . . . . . . 44
3.5.9 Summary of Theorems and Related Properties . . . . . . . 45
3.6 Fast Solutions to DHT – The FHT Algorithm . . . . . . . . . . . . . . 46
3.7 Accuracy Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

Part II The Regularized Fast Hartley Transform

4 Derivation of Regularized Formulation of Fast Hartley
Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Derivation of the Conventional Radix-4 Butterfly Equations . . . 53
4.3 Single-to-Double Conversion of Radix-4 Butterfly Equations . . . 57
4.4 Radix-4 Factorization of the FHT . . . . . . . . . . . . . . . . . . . . . . . 58
4.5 Closed-Form Expression for Generic Radix-4 Double
Butterfly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.5.1 Twelve-Multiplier Version of Generic Double
Butterfly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.5.2 Nine-Multiplier Version of Generic Double
Butterfly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.6 Trigonometric Coefficient Storage, Retrieval and Generation . . . 69
4.6.1 Minimum-Arithmetic Addressing Scheme . . . . . . . . . . 70
4.6.2 Minimum-Memory Addressing Scheme . . . . . . . . . . . . 70
4.6.3 Trigonometric Coefficient Generation
via Trigonometric Identities . . . . . . . . . . . . . . . . . . . . 71
4.7 Comparative Complexity Analysis with Existing
FFT Designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.8 Scaling Considerations for Fixed-Point Implementation . . . . . . . 74
4.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Contents xv

5 Design Strategy for Silicon-Based Implementation

of Regularized Fast Hartley Transform . . . . . . . . . . . . . . . . . . . . . . 79
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2 The Fundamental Properties of FPGA and ASIC Devices . . . . . 80
5.3 Low-Power Design Techniques . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3.1 Clock Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3.2 Silicon Area . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3.3 Switching Frequency . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.4 Proposed Hardware Design Strategy . . . . . . . . . . . . . . . . . . . . . 86
5.4.1 Scalability of Design . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.4.2 Partitioned-Memory Processing . . . . . . . . . . . . . . . . . . 87
5.4.3 Flexibility of Design . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.5 Constraints on Available Resources . . . . . . . . . . . . . . . . . . . . . 88
5.6 Assessing the Resource Requirements . . . . . . . . . . . . . . . . . . . 89
5.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6 Architecture for Silicon-Based Implementation
of Regularized Fast Hartley Transform . . . . . . . . . . . . . . . . . . . . . . 93
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.2 Single-PE Versus Multi-PE Architectures . . . . . . . . . . . . . . . . . 94
6.3 Conflict-Free Parallel Memory Addressing Schemes . . . . . . . . . 96
6.3.1 Parallel Storage and Retrieval of Data . . . . . . . . . . . . . 96
6.3.2 Parallel Storage, Retrieval and Generation of
Trigonometric Coefficients . . . . . . . . . . . . . . . . . . . . . 102
6.4 Design of Pipelined PE for Single-PE Recursive Architecture . . 105
6.4.1 Parallel Computation of Generic Double Butterfly . . . . 107
6.4.2 Space-Complexity Considerations . . . . . . . . . . . . . . . . 109
6.4.3 Time-Complexity Considerations . . . . . . . . . . . . . . . . 110
6.5 Performance and Requirements Analysis of FPGA
Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.6 Derivation of Range of Validity for Regularized FHT . . . . . . . . 114
6.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7 Design of CORDIC-Based Processing Element for Regularized
Fast Hartley Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.2 Accuracy Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.3 Fast Multiplier Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.4 CORDIC Arithmetic Approach . . . . . . . . . . . . . . . . . . . . . . . . 122
7.4.1 CORDIC Formulation of Complex Multiplier . . . . . . . 122
7.4.2 Parallel Formulation of CORDIC-Based PE . . . . . . . . . 124
7.4.3 Discussion of CORDIC-Based Solution . . . . . . . . . . . . 124
7.4.4 Logic Requirement of CORDIC-Based PE . . . . . . . . . . 128
xvi Contents

7.5 Comparative Analysis of PE Designs . . . . . . . . . . . . . . . . . . . . 129

7.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

Part III Applications of Regularized Fast Hartley Transform

8 Derivation of Radix-2 Real-Data Fast Fourier Transform
Algorithms Using Regularized Fast Hartley Transform . . . . . . . . . 137
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
8.2 Computation of Real-Data DFT via Two Half-Length
Regularized FHTs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.2.1 Derivation of Radix-2 Algorithm via
Double-Resolution Approach . . . . . . . . . . . . . . . . . . . 139
8.2.2 Implementation of Double-Resolution Algorithm . . . . . 145
8.3 Computation of Real-Data DFT via One Double-Length
Regularized FHT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
8.3.1 Derivation of Radix-2 Algorithm via
Half-Resolution Approach . . . . . . . . . . . . . . . . . . . . . 153
8.3.2 Implementation of Half-Resolution Algorithm . . . . . . . 154
8.4 Comparative Complexity Analysis with Standard
Radix-2 FFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
9 Computation of Common DSP-Based Functions
Using Regularized Fast Hartley Transform . . . . . . . . . . . . . . . . . . . 161
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9.2 Fast Transform-Space Convolution and Correlation . . . . . . . . . . 162
9.3 Up-Sampling and Differentiation of Real-Valued Signal . . . . . . 163
9.3.1 Up-Sampling via Hartley-Space . . . . . . . . . . . . . . . . . 164
9.3.2 Differentiation via Hartley-Space . . . . . . . . . . . . . . . . . 165
9.3.3 Combined Up-Sampling and Differentiation . . . . . . . . . 165
9.4 Correlation of Two Arbitrary Signals . . . . . . . . . . . . . . . . . . . . 166
9.4.1 Computation of Complex-Data Correlation
via Real-Data Correlation . . . . . . . . . . . . . . . . . . . . . . 167
9.4.2 Cross-Correlation of Two Finite-Length Data Sets . . . . 169
9.4.3 Auto-Correlation: Finite-Length Against
Infinite-Length Data Sets . . . . . . . . . . . . . . . . . . . . . . 170
9.4.4 Cross-Correlation: Infinite-Length Against
Infinite-Length Data Sets . . . . . . . . . . . . . . . . . . . . . . 172
9.4.5 Combining Functions in Hartley-Space . . . . . . . . . . . . 175
9.5 Channelization of Real-Valued Signal . . . . . . . . . . . . . . . . . . . 175
9.5.1 Single Channel: Fast Hartley-Space Convolution . . . . . 176
9.5.2 Multiple Channels: Conventional Polyphase
DFT Filter Bank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
Contents xvii

9.6 Distortion-Free Multi-Carrier Communications . . . . . . . . . . . . . 183

9.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186

Part IV The Multi-dimensional Discrete Hartley Transform

10 Parallel Reordering and Transfer of Data Between Partitioned
Memories of Discrete Hartley Transform for 1-D
and m-D Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
10.2 Memory Mappings of Regularized FHT . . . . . . . . . . . . . . . . . . 192
10.3 Requirements for Parallel Reordering and Transfer of Data . . . . 193
10.4 Sequential Construction of Reordered Data Sets . . . . . . . . . . . . 197
10.5 Parallelization of Data Set Construction Process . . . . . . . . . . . . 199
10.6 Parallel Transfer of Reordered Data Sets . . . . . . . . . . . . . . . . . . 201
10.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
11 Architectures for Silicon-Based Implementation
of m-D Discrete Hartley Transform Using Regularized
Fast Hartley Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
11.2 Separable Version of 2-D DHT . . . . . . . . . . . . . . . . . . . . . . . . 209
11.2.1 Two-Stage Formulation of 2-D SDHT . . . . . . . . . . . . . 209
11.2.2 Hartley-Space Filtering of 2-D Data Sets . . . . . . . . . . . 211
11.2.3 Relationship Between 2-D SDHT and 2-D DFT . . . . . . 212
11.3 Architectures for 2-D SDHT . . . . . . . . . . . . . . . . . . . . . . . . . . 213
11.3.1 Single-FHT Recursive Architecture . . . . . . . . . . . . . . . 214
11.3.2 Two-FHT Pipelined Architecture . . . . . . . . . . . . . . . . . 216
11.3.3 Relative Merits of Proposed Architectures . . . . . . . . . . 218
11.4 Complexity Analysis of 2-D SDHT . . . . . . . . . . . . . . . . . . . . . 220
11.4.1 Complexity Summary for Regularized FHT . . . . . . . . . 221
11.4.2 Space-Complexity of 2-D Solutions . . . . . . . . . . . . . . . 221
11.4.3 Time-Complexity of 2-D Solutions . . . . . . . . . . . . . . . 222
11.4.4 Computational Density of 2-D Solutions . . . . . . . . . . . 225
11.4.5 Comparative Complexity of 2-D Solutions . . . . . . . . . . 225
11.4.6 Relative Start-up Delays and Update Times
of 2-D Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
11.4.7 Application of 2-D SDHT to Filtering of 2-D
Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
11.4.8 Application of 2-D SDHT to Computation
of 2-D Real-Data DFT . . . . . . . . . . . . . . . . . . . . . . . . 231
11.5 Generalization of 2-D Solutions to Processing of m-D
Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
11.5.1 Space and Time Complexities of m-D Solutions . . . . . . 234
xviii Contents

11.5.2 Comparative Complexity of M-D Solutions . . . . . . . . . 235

11.5.3 Relative Start-up Delays and Update Times
of m-D Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
11.6 Constraints on Achieving and Maintaining Real-Time
Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
11.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242

Part V Results of Research

12 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
12.1 Outline of Problems Addressed . . . . . . . . . . . . . . . . . . . . . . . . 247
12.2 Summary of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
12.3 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252

Appendix A: Computer Programme for Regularized Fast Hartley

Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
A.2 Description of Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
A.2.1 Control Routine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
A.2.2 Generic Double Butterfly Routines . . . . . . . . . . . . . . . . 254
A.2.3 Address Generation and Data Reordering Routines . . . . 254
A.2.4 Data Memory Retrieval and Updating Routine . . . . . . . 255
A.2.5 Trigonometric Coefficient Generation Routines . . . . . . . 255
A.2.6 Look-Up Table Generation Routines . . . . . . . . . . . . . . . 256
A.2.7 FHT-to-FFT Conversion Routine . . . . . . . . . . . . . . . . . 256
A.3 Brief Guide to Running the Programme . . . . . . . . . . . . . . . . . . . 256
A.4 Available Scaling Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Appendix B: Source Code for Regularized Fast Hartley Transform . . . . 261
B.1 Listings for Main Programme and Signal Generation Routine . . . 261
B.2 Listings for Preprocessing Functions . . . . . . . . . . . . . . . . . . . . . 273
B.3 Listings for Processing Functions . . . . . . . . . . . . . . . . . . . . . . . 277
Appendix C: MATLAB Code for Parallel Reordering of Data
via Dibit-Reversal Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
C.1 Listing for MATLAB Data Reordering Program . . . . . . . . . . . . . 311
C.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
About the Author

Keith John Jones is a Chartered Mathematician (CMath) and Fellow of the Insti-
tute of Mathematics & its Applications (FIMA), in the UK, having obtained a BSc
Honours degree in Mathematics from the University of London in 1974 as an
external student, a MSc in Applicable Mathematics from Cranfield Institute of
Technology in 1977, and a PhD in Computer Science from Birkbeck College,
University of London, in 1992, again as an external student. The PhD was awarded
primarily for research into the design of novel systolic processor array architectures
for the parallel computation of the discrete Fourier transform (DFT). Dr Jones has
subsequently published widely in the signal processing and sensor array processing
fields, having a particular interest in the application of number theory, algebra and
non-standard arithmetic techniques to the design of low-complexity algorithms and
circuits for efficient implementation with suitably defined parallel computing archi-
tectures. Dr Jones, who also holds a number of patents in these fields, has been
regularly named in both the Who’s Who in Science and Engineering and the
Dictionary of International Biography (otherwise known as the ‘Cambridge Blue
Book’) since 2008.

xix
 Part I
The Discrete Fourier and Hartley
Transforms
Chapter 1
Background to Research

1.1 Introduction

The subject of spectrum or harmonic analysis started in earnest with the work of
Joseph Fourier (1768–1830), who asserted and proved that an arbitrary function
could be represented via a suitable transformation as a sum of trigonometric func-
tions [6]. It seems likely, however, that such ideas were already common knowledge
amongst European mathematicians by the time Fourier appeared on the scene,
mainly through the earlier work of Joseph-Louis Lagrange (1736–1813) and
Leonhard Euler (1707–1783), with the first appearance of the discrete version of
this transformation, the discrete Fourier transform (DFT) [45, 48], dating back to
Euler’s investigations of sound propagation in elastic media in 1750 and to the
astronomical work of Alexis Claude Clairaut (1713–1765) in 1754 [26]. The DFT is
now widely used in many branches of science, playing in particular a central role in
the field of digital signal processing (DSP) [45, 48], enabling digital signals –
namely, those that have been both sampled and quantized – to be viewed in the
frequency domain where, compared to the time domain, the information contained in
the signal may often be more easily extracted and/or visualized, or where many
common DSP functions, such as that of the finite impulse response (FIR) filter or the
matched filter [45, 48], may be more easily or efficiently carried out.
The monograph is primarily concerned with the problem of computing the real-
valued discrete Hartley transform (DHT) and equivalently, via the relationship of
their kernels, of the above-mentioned DFT, initially in just one dimension (1-D) but
later to be extended to that of multiple dimensions (m-D). Solutions are achieved via
the application of various factorization techniques and are targeted at implementa-
tion with silicon-based parallel computing equipment – as typified by the field-
programmable gate array (FPGA) and the application-specific integrated circuit
(ASIC) technologies [40] – bearing in mind the size and power constraints relevant
to the particular field of interest. With mobile communications, for example, a small
battery may be the only source of power supply for long periods of time so that

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 3

K. J. Jones, The Regularized Fast Hartley Transform,
https://doi.org/10.1007/978-3-030-68245-3_1
4 1 Background to Research

power efficiency would have to be a key requirement of any such solution. Through
the adoption of the DHT, the monograph looks also to exploit the fact that the
measurement data, as with many real-world problems, is real-valued in nature, with
each sample of data thus belonging to R, the field of real numbers [4], although the
restriction to fixed-point implementations effectively limits the range of interest still
further to that of Z, the commutative ring of integers [4].

1.2 The DFT and Its Efficient Computation

Turning firstly to its definition, the DFT is a unitary transform [19], which, for the
1-D case of N input/output samples, may be expressed in normalized form via the
equation:

1 X
N 1
X ðF Þ ½k ¼ pffiffiffiffi x½n:W nk
N, k ¼ 0, 1, . . . , N  1 ð1:1Þ
N n¼0

where the input/output data vectors belong to CN, the linear (or vector) space of
complex-valued N-tuples [4]. This equation may be rewritten in matrix-vector form
as

ðF Þ
X N ¼ F NN :xN ð1:2Þ

where FNN, the N  N complex-valued Fourier matrix, is the matrix representation

of the transform kernel whose elements derive from the term:
pffiffiffiffiffiffiffiffiffi
W N ¼ exp ði2π=N Þ, i¼ 1, ð1:3Þ

the primitive N’th complex root of unity [4, 41, 43]. The Fourier matrix is clearly
conjugate-symmetric about its leading diagonal and therefore equal to its own
conjugate-transpose. The unitary nature of the DFT means, in addition, that the
inverse of the Fourier matrix is equal to its conjugate-transpose, whilst its columns
form an orthogonal basis [6, 7, 19] – similarly, a transform is said to be orthogonal
when the inverse of the transform matrix is equal simply to its transpose, as is the
case with any real-valued kernel.
Note that the multiplication of any power of the term WN by any number
belonging to C, the field of complex numbers [4], results in a simple phase shift of
that complex number – the amplitude or magnitude remains unchanged. This
suggests that the dominant operation of any fast solution to the DFT will be that
of phase rotation which, with the right choice of arithmetic unit, could be exploited
to some advantage. The direct computation of the N-point DFT, however, as defined
above in Eq. 1.1, involves O(N2) arithmetic operations, so that many of the early
scientific problems involving the DFT could not be seriously attacked without access
1.2 The DFT and Its Efficient Computation 5

to fast algorithms for its efficient solution, where the key to the design of such
algorithms is the identification and exploitation of the property of symmetry [54, 58],
whether it exists in just the transform kernel or, in certain circumstances, in the input
data and/or output data as well.
One early area of activity with such transforms involved astronomical calcula-
tions, and in the early part of the nineteenth century the great Carl Friedrich Gauss
(1777–1855) used the DFT for the interpolation of asteroidal orbits from a finite set
of equally spaced observations [26]. He developed a fast two-factor algorithm for its
computation that was identical to that described in 1965 by James Cooley and John
Tukey [14] – as with many of Gauss’s greatest ideas, however, the algorithm was
never published outside of his collected works and only then in an obscure Latin
form. This algorithm, which for a transform length of N ¼ N1  N2 involves just
O((N1 + N2)  N ) arithmetic operations, was probably the first member of the class
of fast recursive algorithms now commonly referred to as the fast Fourier transform
(FFT) [5, 6, 10, 14, 19, 44], which is unquestionably the most ubiquitous algorithm
in use today for the analysis or manipulation of digital data. In fact, Gauss is known
to have first used the above-mentioned two-factor FFT algorithm for the solution of
the DFT as far back as 1805, the same year that Admiral Nelson routed the French
fleet at the Battle of Trafalgar – interestingly, Fourier served in Napoleon
Bonaparte’s army from 1798 to 1801, during its invasion of Egypt, acting as
scientific advisor.
Although the DFT, as defined above, allows for both the input and output data
sets to be complex-valued (i.e. possessing both amplitude and phase), many real-
world spectrum analysis problems, including those addressed by Gauss, involve
only real-valued (i.e. possessing amplitude only) input data, so that there is a genuine
need for the identification of a subset of the class of FFT algorithms that are able to
exploit this fact – bearing in mind that the use of real-valued data leads to a
Hermitian (or conjugate) -symmetric frequency spectrum:

complexdata FFT ) exploitation of kernel symmetry,

whilst

realdata FFT ) exploitation of kernel symmetry þ spectral symmetries,

with the exploitation of symmetry in the transform kernel being typically achieved
by invoking the property of periodicity and of the shift theorem, as will be discussed
later in Chap. 3 of the monograph.
There is a requirement, in particular, for the development of real-data FFT
algorithms which retain the regularity – as relates to the algorithm structure – of
their complex-data counterparts as regular algorithms lend themselves more natu-
rally to an efficient implementation. Regularity, which equates to the amount of
repetition and symmetry present in the design, is most straightforwardly achieved
through the adoption of fixed-radix formulations, such as with the familiar radix-2
and radix-4 algorithms [10, 13], as this essentially reduces the FFT design to that of a
6 1 Background to Research

single fixed-radix butterfly, which is the computational engine used for carrying out
the repetitive arithmetic operations required by the fixed-radix algorithm. Note that
with such a formulation, the radix actually corresponds to the size of the resulting
butterfly (in terms of the number of inputs/outputs), although in Chap. 8 it is seen
how a DFT, whose length is a power of two (a radix-2 integer) but not a power of
four (a radix-4 integer), may be solved by means of a highly optimized radix-4
butterfly.
An additional attraction of fixed-radix FFT formulations, which for an arbitrary
radix R decomposes an N-point DFT into logRN temporal stages each comprising
N/R radix-R butterflies, is that they lend themselves naturally to a parallel solution.
Such decompositions may be defined over either (1) the spatial domain, facilitating
its mapping onto a single-instruction multiple-data (SIMD) computing architecture
[1], whereby the same set of operations (e.g. those corresponding to the radix-R
butterfly) may be carried out simultaneously on multiple sets of data stored within
the same memory, or (2) the temporal domain, facilitating its mapping, via the
technique of pipelining [1], onto a systolic computing architecture [1, 35], whereby
all stages of the systolic array (referred to hereafter simply as a ‘pipeline’ or
‘computational pipeline’) operate simultaneously on different temporal stages of
the computation and each of the pipeline’s stages communicates with only its nearest
neighbours. A parallel solution may also be defined over both the spatial and the
temporal domains which would involve a computational pipeline where each of its
stages involves the simultaneous execution of multiple arithmetic operations via
SIMD processing – such an architecture being often referred to in the computing
literature as ‘parallel-pipelined’. Parallel decompositions such as these suggest that
the fixed-radix FFT would lend itself naturally to an efficient implementation with
one of the increasingly more accessible/affordable parallel computing technologies.

1.3 Twentieth-Century Developments of the FFT

As far as modern-day developments in FFT design are concerned, it is the names of

Cooley and Tukey that are invariably mentioned first in any historical account, but
this does not really do justice to the many contributors from the first half of the
twentieth century whose work was simply not picked up on, or appreciated, at the
time of their development or publication. The prime reason for such a situation was
the lack of a suitable technology for their efficient implementation, this remaining
the case until the advent of the semiconductor technology of the 1960s.
Early pioneering work was carried out by the German mathematician Carl Runge
[49], who in 1903 recognized that the periodicity of the DFT kernel could be
exploited to enable the computation of a 2N-point DFT to be expressed in terms of
the computation of two N-point DFTs, this factorization technique being subse-
quently referred to as the doubling algorithm. The Cooley-Tukey algorithm, which
does not rely on any specific factorization of the transform length, may thus be
viewed as a simple generalization of this algorithm, as the successive application of
1.3 Twentieth-Century Developments of the FFT 7

the doubling algorithm leads straightforwardly to the radix-2 version of the Cooley-
Tukey algorithm. Runge’s influential work was subsequently picked up and popu-
larized in publications by Karl Stumpff [55] in 1939 and Gordon Danielson and
Cornelius Lanczos [15] in 1942, each in turn making contributions of their own to
the subject. Danielson and Lanczos, for example, produced reduced-complexity
solutions to the DFT through the exploitation of symmetries in the transform kernel,
whilst Stumpff discussed versions of both the doubling algorithm and the analogous
tripling algorithm, whereby a 3N-point DFT is expressed in terms of the computation
of three N-point DFTs.
All of the techniques developed, including those of more recent origin such as the
nesting algorithm of Shmuel Winograd [60] and the split-radix algorithm of Pierre
Duhamel [17], rely upon the ‘divide-and-conquer’ [34] principle, whereby the
computation of a composite-length DFT is broken down into that of a number of
smaller DFTs where the small-DFT lengths correspond to the multiplicative factors
of the original transform length. Depending upon the particular factorization of the
transform length, this process may be repeated in a recursive fashion on the increas-
ingly smaller DFTs.
When the lengths of the small DFTs have common factors, as encountered with
the familiar fixed-radix formulations, there will be a need for the intermediate results
occurring between the successive stages of small DFTs to be modified by elements
of the Fourier matrix, these terms being commonly referred to in the FFT literature as
‘twiddle factors’. When the algorithm in question is a fixed-radix algorithm of the
decimation-in-time (DIT) type [10], whereby the set of data-space samples is
decomposed into successively smaller subsequences, the twiddle factors are applied
to the inputs to the butterflies, whereas when the fixed-radix algorithm is of the
decimation-in-frequency (DIF) type [10], whereby the set of transform-space sam-
ples is decomposed into successively smaller subsequences, the twiddle factors are
applied to the outputs to the butterflies.
Note, however, that when the lengths of the small DFTs have no common factors
at all – that is, when they are relatively prime [4, 41] – the need for the application of
the twiddle factors disappears as each twiddle factor becomes equal to one. This
particular result was made possible through the development of a new number-
theoretic data reordering scheme in 1958 by the statistician Jack Good [22], the
scheme being based upon the ubiquitous Chinese remainder theorem (CRT) [41, 43,
44] – which for the interest of those readers of a more mathematical disposition
provides a means of obtaining a unique solution to a set of simultaneous linear
congruences – whose origins supposedly date back to the first century A.D. [16].
Also, it should be noted that in the FFT literature, the class of fast algorithms
based upon the decomposition of a composite-length DFT into smaller DFTs whose
lengths have common factors – such as the Cooley-Tukey algorithm – is often
referred to as the common factor algorithm [41, 43, 44], whereas the class of fast
algorithms based upon the decomposition of a composite-length DFT into smaller
DFTs whose lengths are relatively prime is often referred to as the prime factor
algorithm [41, 43, 44].
8 1 Background to Research

Before moving on from this brief historical discussion, it is worth returning to the
last name mentioned, namely, that of Jack Good, as his background is a particularly
interesting one for anyone with an interest in the history of computing. During World
War II, Good served at Bletchley Park in Buckinghamshire, England, working
alongside Alan Turing [27] on, amongst other things, the decryption of messages
produced by the Enigma machine [21] – as used by the German armed forces. At the
same time, and on the same site, a team of engineers under the leadership of Tom
Flowers [21] – all seconded from the Post Office Research Establishment at Dollis
Hill in North London – were, unbeknown to the outside world (and remaining so for
several decades), developing the world’s first electronic computer, the Colossus
[21], under the supervision of Turing and Cambridge mathematician Max Newman
(Turing’s former supervisor from his student days at Cambridge University and
future colleague at Manchester University [38] where work was to be undertaken on
the development of both hardware and software for the Mark I computer [37]). The
Colossus was built primarily to automate various essential code-breaking tasks such
as the cracking of the Lorenz code used by Adolf Hitler to communicate with his
generals and was the first serious device – albeit a very large and a very specialized
one – on the path towards our current state of technology whereby entire signal
processing systems may be mapped onto a single silicon chip.

1.4 The DHT and Its Relation to the DFT

A close relative of the Fourier transform is that of the Hartley transform, as

introduced by Ralph Hartley (1890–1970) in 1942 for the analysis of transient and
steady-state transmission problems [25]. The discrete-time version of this bilateral
and orthogonal transform is referred to as the DHT [8], which, for the 1-D case of
N input/output samples, may be expressed in normalized form via the equation:

1 X
N 1
X ðH Þ ½k ¼ pffiffiffiffi x½n:casð2πnk=N Þ k ¼ 0, 1, . . . , N  1 ð1:4Þ
N n¼0

where the input/output data vectors belong to RN, the linear space of real-valued N-
tuples, and the transform kernel – which may be represented when using matrix-
vector terminology by means of the symmetric Hartley matrix – is expressed in terms
of the ‘cas’ function:

casð2πnk=N Þ  cos ð2πnk=N Þ þ sin ð2πnk=N Þ: ð1:5Þ

1.4 The DHT and Its Relation to the DFT 9

Note that as the elements of the Hartley matrix are all real-valued, the DHT is
orthogonal (although unitary as well, given that R ⊂ C), with the columns of the
matrix forming an orthogonal basis.
Unlike the DFT, the DHT has no natural interpretation as a frequency spectrum,
although the discrete version of the power spectral density (PSD) may be determined
directly from the DHT coefficients. Its most natural use is as:
1. A bilateral transform that satisfies the circular convolution theorem (CCT), so that
it may be used for both the forward and the inverse transformations of a
transform-based solution for the filtering of real-valued data (in both one and
multiple dimensions)
2. A means for computing the DFT so that fast solutions to the DHT – which are
referred to generically as the fast Hartley transform (FHT) [7–9, 52] – have
become increasingly popular as an alternative to the FFT for the efficient com-
putation of the DFT
The FHT is particularly attractive for when the input data to the DFT is real-
valued, its applicability being made possible by the fact that all of the familiar
properties associated with the DFT, such as the CCT and the shift theorem, are
also applicable to the DHT (as will be discussed in Chap. 3), and that the complex-
valued DFT output set and real-valued DHT output set may each be simply obtained,
one from the other. To see the truth of this, note that the equality
   nk 
casð2πnk=N Þ ¼ Re W nk
N  Im W N ð1:6Þ

(where ‘Re’ stands for the real component and ‘Im’ for the imaginary component)
relates the kernels of the two transformations, both of which are periodic with a
period of 2π radians. As a result
   
X ðH Þ ½k  ¼ Re X ðF Þ ½k  Im X ðFÞ ½k  , ð1:7Þ

which expresses the DHT output in terms of the DFT output, whilst
   
Re X ðFÞ ½k ¼ 1=2 X ðH Þ ½N  k þ X ðH Þ ½k ð1:8Þ

and
   
Im X ðFÞ ½k  ¼ 1=2 X ðH Þ ½N  k  X ðH Þ ½k  , ð1:9Þ

which express the real and imaginary components of the DFT output, respectively, in
terms of the DHT output.
10 1 Background to Research

1.5 Attractions of Computing the Real-Data DFT via

the FHT

Although applicable to the computation of the DFT for both real-valued and
complex-valued data, the major computational advantage of the FHT over the
FFT, as implied above, lies in the processing of real-valued data. As most real-
world spectrum analysis problems involve only real-valued data, significant perfor-
mance gains may be obtained by using the FHT without any great loss of generality.
This is evidenced by the fact that if one computes the complex-data FFT of an N-
point real-valued data set, the result will be 2N real-valued (or, equivalently,
N complex-valued or N pairs of real-valued) samples, one half of which are redun-
dant. The FHT, on the other hand, will produce just N real-valued outputs, from
which the required N/2 complex-valued DFT outputs may be straightforwardly
obtained, thereby requiring only one half as many arithmetic operations and one
half the memory requirement for storage of the input/output data. The reduced
memory requirement is particularly relevant when the transform is large and the
available resources are limited, as might be encountered in applications typified by
that of mobile communications.
The traditional approach to computing the DFT has been to use a complex-data
solution, regardless of the nature of the data, this often entailing the initial conver-
sion of real-valued data to complex-valued data via a wideband digital down-
conversion (DDC) process or through the adoption of a ‘real-from-complex’ strategy
whereby two real-data DFTs are computed simultaneously via one full-length
complex-data FFT [51] or where one real-data DFT is computed via one half-
length complex-data FFT [13, 51]. Each of the real-from-complex solutions, how-
ever, involves a computational overhead when compared to the more direct approach
of a real-data FFT in terms of increased memory, increased processing delay to allow
for the possible acquisition/processing of pairs of data sets and additional packing/
unpacking complexity. With the DDC approach, the integrity of the information
content of short-duration signals may also be compromised through the introduction
of the filtering operation.
The reason for such a situation is due in part to the fact that manufacturers of
computing equipment have invested so heavily in producing DSP devices built
around the fast complex-data multiplier-and-accumulator (MAC). This is an arith-
metic unit ideally suited to the implementation of the radix-2 butterfly, which is the
computational engine used for carrying out the repetitive arithmetic operations
required by the complex-data version of the radix-2 FFT. The net result is that the
problem of computing the real-data DFT is effectively modified so as to match an
existing complex-data solution, rather than a solution being sought that matches the
actual problem needing to be solved – which is the approach to be adopted here.
1.6 Modern Hardware-Based Parallel Computing Technologies 11

It should be noted that specialized FFT algorithms [2, 11, 17, 18, 20, 36, 42, 53,
56] do however exist for dealing with the case of real-valued data. Such algorithms
compare favourably, in terms of their arithmetic and memory requirements, with
those of the FHT, but suffer in terms of a loss of regularity and reduced flexibility in
that different algorithms are often required for the computation of the forward and
the inverse DFT algorithms. Clearly, in applications requiring transform-space
processing followed by a return to data-space, this could prove something of a
disadvantage, particularly when compared to the adoption of a bilateral transform,
such as the DHT, which may be straightforwardly applied to the transformation from
Hartley-space to data-space as well as from data-space to Hartley-space, making it
thus equally applicable to the computation of both the forward and the inverse DFT
algorithms – the bilateral property of the DHT means that its definitions for the two
directions, up to a possible scaling factor, are identical.
A drawback of conventional FHT algorithms [7–9, 52], however, lies in the
need for two sizes of butterfly – and thus for two separate butterfly designs – for
fixed-radix formulations where, for a radix ‘R’ algorithm, a single-sized butterfly
produces R outputs from R inputs and a double-sized butterfly produces 2R outputs
from 2R inputs. A generic version of the double-sized butterfly, referred to here-
after as the generic double butterfly [28], is therefore developed in this monograph
for the radix-4 factorization of the FHT which overcomes the problem in an elegant
fashion. The resulting radix-4 FHT, referred to hereafter as the regularized FHT
[28], will be shown to lend itself naturally to an efficient implementation with
parallel computing technology.

1.6 Modern Hardware-Based Parallel Computing

Technologies

The type of high-performance parallel computing equipment referred to above is

typified by the increasingly accessible FPGA and ASIC technologies which now
give design engineers far greater flexibility and control over the type of algorithm
that may be used in the building of high-performance DSP systems, so that more
appropriate hardware solutions to the real-data FFT may be actively sought and
exploited to some advantage with these silicon-based technologies. With such
technologies, however, it is no longer adequate to view the complexity of the FFT
purely in terms of arithmetic operation counts, as has conventionally been done, as
there is now the facility to use both multiple arithmetic units – such as those based
upon the fast multiplier – and multiple banks of fast memory in order to enhance the
FFT performance via its parallel computation. As a result, a whole new set of
constraints has arisen relating to the design of ‘efficient’ FFT algorithms.
12 1 Background to Research

With the recent and explosive growth of wireless technology, and in particular
that of mobile communications, algorithms are now being designed subject to new
and often conflicting performance criteria whereby the ideal is to simultaneously:
1. Maximize the throughput
2. Satisfy a timing constraint, relating either to the ‘latency’, which is defined as the
elapsed time involved in the production of an output data set from its input data
set, or the ‘update time’, which is defined as the elapsed time between the arrival
of the latest input data set and the subsequent production of the latest output data
set (noting that for a 1-D block-based solution these two timing parameters are
identical)
3. Minimize the required silicon resources (and thus the cost of implementation)
4. Keep the power consumption to within the available power budget
The task, therefore, is to find those solutions that are best able to deal with the
trade-offs that inevitably need to be made if the above objectives are to be adequately
addressed. Note that the difference between the latency and the update time is most
evident and easy to visualize when the algorithm is implemented via a computational
pipeline, as the latency then corresponds to the elapsed time across the entire length
of the pipeline whereas the update time corresponds to the elapsed time across a
single stage of the pipeline.
Such trade-offs are considered in some considerable detail for the silicon-based
implementations of the regularized FHT discussed in this monograph, bearing in
mind the aim of achieving resource-efficient and power-efficient solutions for the
parallel computation [1, 3, 23] of the DHT and equivalently, via the relationship of
their kernels, of the real-data DFT, for the processing of both 1-D data, to be
discussed in Chaps. 4, 5, 6 and 7, and m-D data, to be discussed in Chaps. 10 and 11.
As a final observation, the adoption of the FHT for wireless communications
technology would seem to be particularly apt, given the contribution made by the
originator of the Hartley transform (albeit the continuous-time rather than the
discrete-time version) to the foundation of information theory, where the Shannon-
Hartley theorem [46] helped to establish Shannon’s idea of channel capacity
[46, 50]. The theorem simply states that if the amount of digital data or information
transmitted over a given communication channel is less than the channel capacity,
then error-free communication may be achieved, whereas if it exceeds that capacity,
then errors in transmission will always occur no matter how well the communication
equipment is designed.

1.7 Hardware-Based Arithmetic Units

When producing electronic equipment, whether for commercial or military use, great
emphasis is inevitably placed upon minimizing the unit cost so that one is seldom
blessed with the option of using the latest state-of-the-art device technology. The
1.8 Performance Metrics and Constraints 13

most common situation encountered is one where the expectation is to use the
smallest (and thus least expensive) device that’s capable of yielding solutions able
to meet the desired performance objectives, which means using devices that are often
one or more generations behind the latest specification. As a result, there are
situations where there would be great merit in having designs that are not totally
reliant on the increasing availability of large quantities of expensive embedded
resources, such as the fast multipliers and fast memory provided by the manufac-
turers of the latest silicon-based devices, but are sufficiently flexible as to yield
efficient implementations in silicon even when such resources are scarce.
One way of achieving such flexibility with the regularized FHT would be through
the design of a processing element (PE) for the computation of the generic double
butterfly that minimizes or perhaps even avoids the need for fast multipliers, or fast
memory, or both, according to the availability of the resources on the target
computing device. Despite the increased use of the hardware-based computing
technologies, however, there is still a strong reliance upon the use of software-
based techniques for the design of the arithmetic unit. These techniques, as typified
by the familiar fast multiplier, are relatively inflexible in terms of the precision they
offer and, although increasingly more power-efficient, tend to be expensive in terms
of silicon resources.
There are a number of hardware-based arithmetic techniques available, however,
such as the shift-and-add techniques, as typified by the Co-Ordinate Rotation DIgital
Computer (CORDIC) arithmetic [57] unit and the look-up table (LUT) techniques,
as typified by the distributed arithmetic (DA) [59] unit, which date back to the DSP
revolution of the mid-twentieth century but nevertheless still offer great attractions
for use with the new hardware-based technologies. The CORDIC arithmetic unit, for
example, which may be used to carry out in an optimal fashion the operation of
phase rotation – the key operation involved in the computation of the DFT – may be
implemented by means of a computational structure whose form may range from
fully sequential to fully parallel, with the update time of the CORDIC operation
decreasing linearly with increasing parallelism. The application of the CORDIC
technique to the computation of the regularized FHT is considered in this monograph
for its ability both to minimize the memory requirement and to yield a flexible-
precision solution to the problem of computing the real-data DFT.

1.8 Performance Metrics and Constraints

Having introduced and defined the algorithms of interest in this introductory chapter,
namely, the DFT and its close relation the DHT, as well as discussing very briefly the
various types of computing architecture and technology available for the implemen-
tation of their fast solutions, via the FFT and the FHT, respectively, it is now worth
devoting a little time to considering the type of performance metrics and constraints
Another random document with
no related content on Scribd:
degrees outen the way. You can use your compass if you want to but
I’m goin’ by the blinkin’ Sun and the bloomin’ North Star, I am.”
All that day as they were mushing on Jack kept tab on his
compass and Bill kept his eye on the sun and while they both firmly
believed they were headed right, the compass, by which the mariner
pushes boldly forward, steering always as it directs, knowing it will
not send him astray, had the boys worked up into something that
very nearly approached a nervous state of mind.
All the time they were on the march that afternoon the going was
very much heavier than it had been on the No Name River, for they
had to break the trail as they went along. Jack kept wondering what
had come over the compasses that so persistently made them point
east instead of north.
When they had established camp that night they were still
discussing the frivolous peculiarities of compasses which enabled
them to point east when they were on top-o’-the-world with the same
degree of freedom that they pointed north when they were used on
the rim-o’-the-world.
The weather was crisp and cold and the air as thin and clear as
crystal. Bill, who had lost faith in the instrument that is the symbol of
unerring accuracy, stood forth in the night, looking more like some
barbarian of the glacial age than a pampered boy of the gas-house
district and viewed the twinkling lights in the bowl of the heavens. He
called Jack and indicating the North Star with his finger said:
“Either that star is wrong and our compasses are right or the other
way about, but ’tween you and me, Bud, I’ll bank on the North Star
every time and dish the compasses.”
“I know exactly where the trouble comes in, Bill; funny I couldn’t
have thought of it before,” said Jack, brightening up as though his
brain-cells had decohered. “The North Star and the compasses are
both right. You know that the magnetic north pole and the true, or
geographic, north pole are not in the same place.
“In fact the magnetic pole is way south of the true pole—let me
see, if I remember rightly it is pretty close to the meridian which is
one hundred degrees west of Greenwich and on the sixty-eighth
parallel, and is, consequently, nearly twenty degrees south of the
geographic pole. This is the reason, then, our compasses point to
the east instead of to the north; the only thing we don’t want to forget
to do is to allow for this difference.”
“Right you are, Jack,” Bill made answer, for of all times that his
admiration for his partner welled in his breast it was when the latter
explained what he called “this high-brow stuff.” “Say if I had a brain
like yourn I wouldn’t be up here seekin’ moosehide sacks o’ gold, I’d
be back there in little ole Noo York on Wall Street shovelin’ it into
vaults; that’s what I’d be doin’.”
Having disposed of the vexatious problem of the North Pole Bill
again took an interest in his compass and began figuring out how
many points this way or that way they would have to go to get so
many points the other side of somewhere else. Bill didn’t know it but
up there in the cold, cold North he was developing his gray matter,
for he was thinking and this is the only process by which it can be
done.
And so for the next three days they kept steadily onward over
tundras, on streams, through wooded lands, up hills and down dales
and always north by east. Nor did the boys feel a bit lonesome here
in these vast stretches of the sub-Arctic ice and snow and the great,
grim solitude of nature but this may be accounted for in virtue of
there being hardly ever a minute but that they were kept on the jump
doing something for either themselves or the dogs.
Neither were they without companions for the dogs were the most
wonderful company ever. They showed the most amazing
intelligence, particularly ’Frisco and Sate, and Bill was not far from
the truth when he said “they’re human and that’s all there is to it.”
And in very truth so it seemed, for whatever they wanted to do or
say, they knew precisely how to go about it, or to make themselves
understood.
“We still have another day’s journey before us,” Jack announced
as they made their last temporary camp, and they were, indeed,
getting pretty close to the end of the rainbow, for they were even
then in the land of the Yeehats, which was the land of their golden
hopes.
But to Bill, instead of there being more gold the farther north they
went, the snowscape grew more desolate and forbidding, for he was
better acquainted with a semi-torrid climate than he was with a
wholly frigid one, and to him the outlook was far from alluring. Jack
who had spent nine months in the Arctic didn’t mind it a little bit. He
had the makings in him of a polar explorer.
Harking back to that July morning when Jack had unfolded the
fascinating story of gold in moosehide sacks to him in his apartment,
and now looking out upon the snow-veiled land as far as his eye
could reach Bill again began to wonder if, after all, it wasn’t a fairy
tale told by a writer of fiction, or, more likely, a hoax perpetrated by
the early miners on the tenderfeet who pestered them with
questions.
“What I’d like to know is if this metal is really up here,” he finally
said to Jack, “why haven’t men like Jack McQuesten, Doc Marling,
Sam Stoneback and all the other old timers who have lived in Ilasker
ever since gold was discovered, searched for and found this
treasure.”
Jack smiled cynically—that is, as cynically as a boy can smile.
“You might just as reasonably ask me why the head door-keeper
of the Stock Exchange has not made a fortune on the floor—he’s on
the ground too you know. Or why is it a boot-black sometimes
becomes a millionaire, or a girl from Tin Can Alley rises out of the
depths and is crowned a queen?” Jack argued.
“Or Bill Adams, of Claremont Avenoo, seekin’ the yellow metal in
the shadow o’ the North Pole,” Bill commented and then he added,
“I’m gettin’ to be some poet like Mr. Service, what say, Jack?”
“Yes, this beautiful Northland will make a poet of anybody. But
were the bootblack and the alley wench destined to do and become
what they did do and did become?” Jack went on.
“Is it because they thought their way up, or is the element of
chance responsible for it all? Perhaps it is like pemmican, due to a
little of everything mixed together. These are things for you to think
about, Bill.”
Bill was thinking but he couldn’t think fast enough to keep up with
Jack’s line of talk, though he had the satisfaction of knowing what his
partner was driving at and this was more than he was sometimes
able to do.
 “It sounds to me, Jack,” he finally said, “but I’m hopin’ as how
you’re right. I wouldn’t take any stock in it comin’ from any one else
’ceptin’ yourself. Your hunches from the time I first knowed you has
got the weegie board locked in a vault. An’ consekently I’m sayin’ as
how I take it your hunch inkubator is in just as good workin’ order
and reliable here in Ilasker, as it was down in Mexico.”
“Now you’re talking sense,” said Jack, throwing out his chest, only
it couldn’t be noticed from the exterior because his caribou coat was
so big it covered up his abnormal expansion. “And see here, Bill, you
want to cut out this ‘it sounds to me’ stuff. I’m not exactly what you
call a Christian Scientist but we’ll never find the pot of gold if you’re
going to keep doubting it all the time.”
This little talk gave Bill some food for thought too, and he resolved
that let come what may he would never show any signs of its
“sounding to him” again.
Along in the late afternoon of the next day they came to a river
and Jack proclaimed that they had at last reached the end of their
long trip.
“This is the Big Black River all right and if I haven’t missed my
guess we are about ten miles below the Arctic Circle and fifteen or
twenty miles west of the International boundary line. Put her there,
old pard, we’re in the land of the Yeehats at last!”
“With nary a Yeehat in sight,” said Bill as they grasped hands, “but
I’m goin’ to keep my rifle handy if it’s all the same to you.”
Then came the work of building their winter quarters which was to
be a log cabin of one room about twelve feet wide and fourteen feet
long. There were plenty of trees about, the chief kind being Alaska
spruce, and owing to its abundance in the more northern parts of
Alaska it is used for work of every description, such as cabins,
mining timber, firewood, sleds, etc.
The first thing to be done was to fell the trees and they began by
sawing them down with their crosscut saw. Bill said he would rather
chop them down and that he could do it easier and quicker than both
of them could do it together with the saw. While this work was in
progress the dogs grew restless on account of their inactivity and
enlivened things up every now and then with a fight; then Jack would
go among them, like Daniel in the lion’s den, and use the butt-end of
his whip handle on them until they broke apart.
“I’ll give you muts something to do that will take the fight out of
you,” he told them, and he did, for as Bill felled each tree his
pardner, as he had now begun to call him, lashed a rope round an
end and hitching the dogs to it put them to doing work the like of
which none of them had ever done before.
And pull? Why, boy, they pulled so hard that their muscles looked
as if they would break through their hides. After he had broken out a
log and was ready to start Jack would give his long whip a
tremendous crack and yell mush! when every dog did his duty and
they liked it too.
It was a never ending source of wonder to the boys that these
animals liked to work. And yet under the influence of kind treatment
they were very affectionate, especially the malamutes, though none
of them showed it in a way at all like dogs that live in the lap of
luxury. Neither would it do to pet one of them to the exclusion of the
others else there would be a terrific fight going on in an instant for
they were fearfully jealous, and would not tolerate the slightest show
of partiality.
“I’ve got one o’ them high-brow ideas, Jack; I’ve been thinkin’ and
thinkin’ as I’ve watched these huskies, and after what you told me
about the way the dogs acted on the front over there in France, I’ve
conclooded they’ve got human brains just the same as you and me.
They could talk if they wants to but they just pretend they can’t so
they won’t have to argy with a feller. They’re just like them furriners
in Noo York, they can savvy anything they wanter and anything they
don’t wanter savvy—why they don’t.”
“Then you believe in reincarnation,” said Jack.
“Reindarnation!” was Bill’s near echo. “I might believe in it if I knew
what it is, but not knowin’ I cannot say.”
Then Jack explained how some folks, including about four
hundred million in India, believed that the souls of animals, when
they died, passed on into the bodies of people. This was all easy
enough for Jack to tell about but when Bill wanted to know what Jack
meant by soul his partner had no small time telling him about it in a
way that he could understand.
“It sounds reasonable,” declared Bill, “and I would believe in this
reindarnation thing only these dogs are so much decenter than most
people.”
 “‘I’VE CONCLOODED THEY’VE GOT HUMAN BRAINS JUST THE SAME AS
YOU AND ME.’”

And so they worked and talked and talked and worked and
another month slipped by before they got their log cabin done. The
way Bill could swing an ax made Jack envious and while building the
cabin was the hardest of hard work, both of these youngsters got a
lot of pleasure seeing it go up log by log. And when it was all done
they were as proud of it as any millionaire who ever built a mansion
on Fifth Avenue.
And furniture! They made mission furniture, table, chairs and all
the accessories of home, the like of which no missionary in the heart
of lightest Africa ever set eyes upon. And comfortable! With a
rousing fire, ham and Alaska strawberries, coffee and biscuits that
Jack made so well (I didn’t say so light) they were as comfortable as
a husky after a double ration of dried fish, fast asleep under the
snow.
“I’m thinkin’ we’ve got to get out and kill some fresh meat,”
suggested Bill after a meal in which the spirit of Sing Nook was
present, i.e., when the strawberries came on as usual.
“I thought you declared that Alaska strawberries were every whit
as good as the spaghetti we used to get at The Black Cat back in
New York, when we thought we were a couple of highflyers,” Jack
laughed.
“Oh, for a dish of spaghetti,” sighed Bill, and then he came back
with this statement: “Ilasker strawberries are all right but after you’ve
et them for thirty or forty meals you get a lee-tle tired of them and
pine for a young oyster, in a bowl of cracker soup, or a couple of
fried eggs—one fried on one side and one on the other—or even a
steak from a hoof of a panhandle longhorn.”
“I move that to-morrow we begin ‘prospecting’,” Jack said, paying
no attention to Bill’s likes and dislikes. “We’ve been away now for
over three months and all we’ve got to show for it is an outlay of
more than a thousand dollars, these two mighty good dog teams, our
cabin and the fun we’re having.”
 “Then let’s go to it,” Bill said.
“We’ll strike out across the river and go due north; then every trip
we make we’ll veer round five points until we’ve boxed the
compass.”
 CHAPTER VIII
ON THE ARCTIC CIRCLE
While the boys did not expect to be gone longer than a week, or
ten days at the most, on any one spoke of their prospecting wheel,
and carried good grub to last them for this length of time, they
nevertheless took the precaution to stock up with enough alcohol,
compressed tea, hard tack and pemmican for themselves, and dried
fish for the dogs, to stave off starvation for a month in the event of
meeting with an accident, getting stormbound, or wanting to make a
longer stay.
With a team apiece of seven dogs and a load of only a hundred
and fifty pounds it was possible for them to ride on their sleds a good
deal of the time. But this does not mean that they could very often
actually sit on them but the way they did it was to stand on the rear
ends of the runners and hold on to the handle bars.
The night before they made their first trip out they packed their
traveling mess-gear, which consisted of a collapsible stove and
alcohol for fuel, grub and the few other necessary things of their
outfit, on the sleds, so that they could make a start the next morning
at day-break.
They crossed the Big Black River and drove due north over the
tundra (a Russian word, pronounced toon-dra,) which is a rolling
prairie, without any trees on it; the soil is black and soft, or muck as it
is called, and on it both mosses and lichens grow. They drove due
north and in the course of time Bill announced that according to the
sun, his watch and his stomach there should be a period of rest and
of eating. According to Jack’s calculations they had made about
twelve miles and were moreover right then on the Arctic Circle.
“After we gets through with the eats, Jack, I wants you to edicate
me on this Arctic Circle thing,” said Bill as he threw the dogs their
fish.
Jack was busy opening the thermos bottles of hot tea and getting
out the sandwiches.
“What do you want to know about it?” he asked absent-mindedly,
for he was not a little bit interested in this at the particular moment.
“I wants to know why is the Arctic Circle, and everything else
about the bloomin’ thing. The way I’ve doped it out it is like a
meridian or the equator, that is, it’s a line that you can’t see and yet
it’s there or here just the same. I’m settin’ on it and I know it but I
can’t prove it, As man to man, now, I’m askin’ you what is it?” asked
Bill with great earnestness.
Jack looked at him and laughed.
“You asked a question and then answered it yourself in the next
breath. You’ve said all there is to say about it except that it’s a circle
running round the North Pole like an ostrich feather on a lady’s hat,
only, different from the latter, it extends on all sides of the pole to
latitude sixty-six degrees and thirty-two minutes north.”
“But why is it?” persisted Bill.
Jack thought a moment.
“The chief reason the Arctic Circle is so called is because it is the
circle below which the sun does not drop in mid-summer. If we were
here on the Arctic Circle in summer we’d see the sun at midnight just
above the horizon, and the farther north a person goes in summer
the higher he will see the sun above the horizon at midnight. Lots of
tourists come up here every summer just to take a look at the
midnight sun, and the natives call them sunners.”
“An’ we won’t get to see it then?” kicked Bill; “it’s just my luck. If it
’ud be rainin’ soup I’d be standin’ out in it with a fork.”
“We’re not up here to see the sun at midnight,” Jack came back at
him, “we’re lucky if we get a glimpse of it at noon. What we’re up
here for is to get the yellow stuff.”
“Oh yes, I kinda lost sight o’ the bloomin’ gold for a minute,” was
Bill’s reply.
It was great sport, now that their loads were light, for the young
drivers to flourish their whips and crack them in the dry air, while the
dogs, fed-up, fresh and eager, raced along, with tinkling bells where
the going was good, as though they were making a dash for the
pole. The boys and their outfit would have made a capital movie, but
there wasn’t a cinematograph camera nearer than Skagway on the
south or St. Michaels on the west.
At this time of the year, the period of daylight on the Arctic Circle
is very short and as darkness came on they pulled up on the banks
of a stream to make camp.
“This must be the Rat River,” said Bill.
“It is, but it certainly isn’t much at this point. We’re close to its
head waters though and that accounts for it. It empties into the
Porcupine River about sixty or seventy miles west of here. It might
be worth our while to make a survey up and down the river for a few
miles, so to-morrow let’s go down stream.”
They had not gone more than five miles the next morning when
their attention was attracted by a huge fire a couple of hundred feet
back of the north bank and they drove up to see what was going on.
“Bet it’s the Yeehats barbecuin’ a caribou,” suggested Bill who
was dying by inches for the want of a caribou steak.
“Look again,” said Jack, and then Bill saw the winter diggings of
some miners, three all told, one white man and two Indians, busy
with picks and shovels.
“Lookin’ for our gold,” was Bill’s idea of it.
“More likely they are mining for some on their own account. A
great deal of placer mining is done up here in the winter—has to be
done in winter as a matter of fact—because the ground is so low and
wet that they can’t do any digging in the summer time, for the hole
fills up with water as fast as the dirt is thrown out.
“The way they work it according to what Rip Stoneback told me, is
like this. The miner cuts all the fire-wood he can in the summer,
which isn’t a great deal as it is so scarce in these parts, and builds
his sluice-box; then when winter sets in and it begins to freeze, he
clears the moss off of a small patch. On this clearing he builds a fire
and keeps it going until the ground is thawed down a foot or so when
he digs it out; then he builds another fire, digs out the thawed ground
and repeats the operation until he has sunk a shaft through the muck
and gravel to bed-rock.
 “Now between the gravel and bed-rock is a layer of gold-bearing
dirt called pay-streak and this is hoisted to the surface by means of a
windlass on the ends of whose rope are spliced a couple of buckets;
and this windlass, of course, sets over the shaft. Usually two men go
down in the shaft and pick the frozen pay streak from the ground.
The shafts vary in depth from fifteen to forty feet depending on what
part of the country the mine is located.
“The third man stays on top to draw up the buckets and with a
wheel-barrow wheels the gold-bearing dirt back and dumps it in a
pile where it will be in no danger of getting washed away by the
melting snows when spring comes. In the spring when water is
plentiful the fun begins for then the clean-up takes place and the
men who were as poor as Indian dogs all winter wax rich and take
their winnings back to civilization where they can be separated from
it.
“The clean-up means that the color-bearing dirt is shoveled into
the sluice-box, that is, a trough without ends, into which the miner
has contrived to keep a steady stream of water running. The water
washes away the dirt and leaves the free gold just as it does in the
more primitive method of panning.”
The miners were as glad to see the boys as the latter were to see
them, yes even more so. They immediately knocked off all work and
there was a regular “chin-fest,” as Bill called it, from that time on.
They made the boys stay to supper and improvised bunks in their
cabin for them to sleep on. After Art Jennings, who, as you will
gather from his name, was the lone white man, had heard the news
of the outside world they talked about three other things only, the first
of which was gold, the second gold and the third gold.
“This placer minin’ is altogether too slow a game for me,”
remarked Bill when they were on their way again. “What I wants is to
see moosehide sacks of it piled up like cordwood, I do.”
“Well there are moosehide sacks of it cached right here in
Yeehatville on the Circle. From the Pacific Ocean on up to the Arctic
Ocean there’s gold. In every stream and river, as well as the land
between them, this precious metal is found in either particles or in
nuggets. Take the Klondike! it’s not much larger than the Rat River
here and yet so much gold was found there its name became known
all over the world. Every river in Alaska and the Yukon, I suppose, is
just as rich but you don’t hear much about them because the
Klondike was the first and so outshone all the rest. We’ll get ours yet,
don’t worry,” said Jack hopefully.
Each trip the boys made from their base of supplies took them
from one to two weeks. Their marches in and out were usually made
in a couple of days and when they had worked away from their
permanent base as far as they wanted to go they would set up a
temporary camp.
If the weather was not too severe, that is to say below zero, they
pitched their tent, but when it got to twenty, forty or sixty below, or a
blizzard struck them as it frequently did in mid-winter, they made a
better camp by cutting out blocks of snow and piling them up into a
dome-shaped shelter like the igloo of the Eskimo, but which Bill, who
always persisted in nick-naming everything that was new to him,
called a butter-dish.
Building a snow igloo was a simple matter after they had put up a
couple, and the boys got it down to such a fine point that they could
do the complete job in two or three hours. Of course this was largely
the result of Jack’s experience in the Arctic which enabled him to go
about it in the right way. He had brought his saw-knife with him for
this express purpose. This useful tool is about eighteen inches long
and one and three-fourths inches wide and while one of the edges of
it is sharp like a knife the other edge has teeth cut in it like a saw.
With this saw-knife Jack or Bill would saw out the hard frozen
snow into blocks which for the lower layers of the igloo measured
about two feet in length and eighteen inches wide and high; as the
upper layers were reached they used smaller and smaller blocks.
Finally when all of the snow-blocks but one were laid up and the
igloo was as hemispherical as the half of a ball, the last block, which
they beveled on four sides, was set in the center and this held all of
the other blocks out like the keystone of a bridge.
They made these snow igloos about six feet in diameter on the
inside of the base so that they could lie down comfortably. To get into
the igloo they left one of the snow blocks of the first layer out and
through this hole they also took in the grub they needed, the alcohol
stoves and the sleeping bags. To close the hole it was only
necessary to push in the snow block when they were pretty well
housed in.
What, then, with their fur clothing, a log house at their permanent
base and these snow igloos at the ends of their trips, they were able
to keep quite comfortable. Nearly every one who has never put in a
winter in the Arctic, or sub-Arctic, regions seems to think that the
extreme cold is a thing to be feared, but it isn’t if one has the right
kind of clothes, enough food and if, when outside of the shelter, he
does not stop but keeps right on going or working. But the long hours
of darkness often get to be mighty monotonous.
Being boys, however, nothing could chill their ardor or cast a
gloom on their spirits for any length of time and they were always
ready for a frolic. Thus it was when they were sledding on streams
where the ice was good they had some great races. Each contended
that his team was the swiftest that ever pulled a sled and this
difference of opinion invariably led to a challenge to prove it.
The dogs entered into the spirit of the races with as much zest as
their young masters and when they were abreast and the signal to
go was given, the whips cracked and the dogs jumped to get first
place. Onward they dashed with an ease and grace that made them
seem more like rubber balls bouncing along low on the course, than
four-footed animals whose business it was to work.
But the spirit of sport was strangely strong in these living, vibrant
creatures and as they fairly flew along over the course they voiced
their joy by short howls and yelps when they were in the lead or their
anguish by whines and cries when they dropped behind.
Jack was, perhaps, a better driver than Bill but in his own heart he
gave the credit to his team when they won and win they nearly
always did. Bill was a good “sport” though and never got “sore” when
he lost a race; he always took the blame on himself for his poor
driving and nothing could shake his belief that his was the fastest
team, bar none, in all Alaska.
There were a few times though when Bill’s team won. One of
these rare occasions was when a snowshoe rabbit ran from a bank
onto the ice intending to cross to the other side; finding himself in
front of a terrible pack of running dogs or wolves, he knew not which,
that were bent on catching him, instead of going on across to safety
he ran straightaway ahead of them.
Sate, Bill’s lead-dog, spotted him first and he ran as he had never
run before; the dogs of his team felt this super-burst of speed on his
part and as the rabbit paced him, so he paced them with the highly
gratifying result, to Bill, that his team jumped ahead of Jack’s by a
length. The boys urged their teams on with their “yow-yows,” and the
bells jingled joyously while the wild race was on.
The dogs of both teams had forgotten that there were such things
as a trace or trail, while the boys had lost sight of the treasure they
were seeking and let nothing impede their mad flight toward
destruction. At the end of a quarter of a mile Bill’s team was nearly
three lengths ahead of Jack’s and he felt the race well won. His dogs
had lost all interest in the race, indeed, they did not know they were
racing for it was the rabbit they were after now. Then little snowshoe
fooled them, for he made a sharp turn and ran up the bank.
Sate likewise turned as sharp as the high speed he was making
would allow; the team swerved abruptly, slipped and slid for half-a-
dozen yards, the sled upset and everything was piled up in a heap.
Jack’s team shot by them like an arrow and they ran for another
quarter of a mile before he could stop them in their mad flight. When
he got back he had to admit that Bill’s team had won the race but it
cost them an hour’s work to make good the damage done. There
was no more racing that day.
“You see, Jack, as I always told you, my team is faster than yourn
and all it needs to show speed is a rabbit for a pace maker,” was
Bill’s comment as he picked himself up.
In their goings and comings they ran across all sorts of wild
animal life from the little lemmings, a mouse-like animal with short
ears and tail, which looks like a miniature yellow rabbit, to the giant
moose. In between these two extremes they saw squirrels,
snowshoe rabbits, red and black foxes, lynxes, gray wolves and
caribou. They had also seen the tracks of bears, for the species of
bear that live in the sub-Arctic regions does not hibernate.
They often shot squirrel, rabbit and ptarmigan (pronounced tar’-
mi-gan), a bird of the grouse order, and these served as dishes of
great delicacy for the boys, as well as giving the dogs a welcome
change from dried fish. Bill declared it to be the open season for
bagging some big game and Jack agreed that they must. But it is
hard to seek cached treasure and be big game hunters at the same
time.
Once while they were moving leisurely along after a satisfying
dinner and they were talking about hunting the caribou, moose and
bear, the tables were suddenly turned on them when they became
the hunted prey of wild beasts, for a pack of famished wolves had
scented them out and were headed straight for them.
Pell-mell came the lean, long-legged beasts with ears erect, ribs
bulging out of their loose skins, tails drooping and starved to
desperation. Instantly the boys halted their teams and had barely
time enough to cut the dogs out of their traces before the pack was
upon them. The dogs knew they were in for a fight to the death and
braced themselves for it, while the boys drew their revolvers and
stood on their sleds ready for the attack.
In less than a minute the wolves were upon them and the fight
was on. The dogs met the onslaught with the strength and courage
the wolves lacked; and in between pistol shots, each of which picked
off a wolf, the dogs snapped in two the legs, and broke the necks of
their savage ancestors with a crunch of their powerful jaws, or
opened their bellies, which let their entrails half out, or severed the
jugular veins when streams of blood spurted forth from the rips made
by merciless fangs.
But the dogs suffered too, for often three or even four wolves
would fight a single one and in this unequal struggle he would go
down unless his master took a hand and evened up numbers by a
few well-placed bullets. Nor was it easy for the boys to shoot the
wolves, for the fight was so fast and furious it was well-nigh
impossible at times to send a piece of cold lead into their miserable
carcasses without the danger of hitting their dogs.
One of the curious things was that when a wolf got hold of the
harness on a dog it mistook it for brute substance instead of inert
leather and it would bite it viciously and shake it furiously without
getting the living response that it had the right to expect.
When the number of wolves had been brought down to twice that
of the dogs, they knew they were beaten and the moment this
happened their courage failed them and those that were left with
strength enough to take to their heels slunk quickly away.
An examination of the dogs showed that far from coming out of
the fight unscathed every one of them was in a bad way and, still
more sad to relate, Jennie and Prince, two of the outside dogs of
Jack’s team, had to be shot to put them out of their misery. As the
dogs were so badly off and the harness cut up and chewed to pieces
the boys had to make camp on the spot.
They dressed the wounds of the dogs as well as they could and
gave them half-a-can of pemmican apiece—a food that the dogs
liked above all else. While the dogs laid down and rested and nursed
their hurts, their masters built an igloo, for they couldn’t tell when
they would be able to move on. While the igloo was going up there
was nothing but kind words and praise for the dogs and it could be
seen by the looks in their eyes and the expressions on their faces
that they knew every word which was said to and about them, and
enjoyed and appreciated it all. As Bill saw them now he was more
thoroughly convinced than ever that these particular dogs were
endowed with human brains and not just common dog brains.
“I always told you my team could outrun yourn and you’ll have to
admit they out-fought yourn too,” said Bill boastfully after the gloom
had somewhat worn off.
“I don’t see how you make that out,” Jack flared up.
“Well, two of your dogs will never mush again pullin’ a sled after
them here on earth—though they may haul a little red cart with
angels in it when they go tearin’ along the trails o’ heaven.”
“That’s no argument at all,” returned Jack soberly, “and you can’t
get away with it either. Why, I saw ’Frisco rip the throats open of one
wolf after another when four of them were at him at once. Prince and
Jennie went down in a fluke—in a fluke I tell you—and that is the
only reason they lost out.’”
“This is soitenly tough luck,” said Bill as he was going over the
wounds of the dogs before they turned in.
“And I’m two dogs short,” moaned Jack, “though I’m mighty glad
they were not the malamutes.”
 “Never youse mind, Buddy. I’ll give youse one of mine and we’ll
still be even.”
“I don’t want any of your dogs, Bill, I’ll just drive my five dogs
along until we strike an Indian village or some camp and then I’ll buy
a couple of Siwashes. But I’m sure sorry to lose Prince and Jennie
for they were a couple of dandy dogs to say the least.”
Just the same when Bill had fixed the harness and hitched up the
dogs preparatory to making a fresh start, Jack saw with grim
pleasure that the teams were even and that Bill’s best dog, next to
Sate his leader, was in the traces of his team.
Jack didn’t say anything about it then but he made up his mind
that when he went ’round the world on a pleasure jaunt, or anywhere
else, Bill could go with him however crude his speech, and rough his
manner.
They limped back to their base of supplies and stayed there for a
week until the dogs got into shape again.
 CHAPTER IX
THE LAND OF THE YEEHATS
On the various trips they had made from their base of supplies on
the Big Black River the boys had kept a sharp lookout for marks or
signs or other visual evidence which might indicate in some way the
location of the treasure they sought. Jack’s hunch was responsible
for his belief that so great a store of gold would not, in fact could not,
have been abandoned without some clew which would serve as a
key to its recovery.
They often dug off the snow from a pile of dirt which they thought
might cover the sacks of gold; as wood was frequently hard to get,
they couldn’t thaw it out and, consequently, had to work like “niggers”
with their picks and shovels to penetrate it. And to what purpose?—
usually only to find it was the dump of some discarded mine. But a
gold seeker wots not of either hardship or work if his efforts give
promise of bringing about the desired result. And they hoped great
hopes.
Again they would find a cache (pronounced cash) but it was not of
the kind that is formed of a hole in the ground, or a cavity under a
pile of stones, but a box-like structure erected on poles set in the
ground. Some of the better ones had notched logs which served as
steps and these were set up at an angle on one side so that access
to the cache could be made with greater ease and lesser agility.
These caches were used by prospectors and miners who
transported their outfits on their backs, or hauled them on sleds, and
who had to double back on the trail time and time again before they
got to their journey’s end.
In nearly all of these caches the stores were of ancient vintage, a
few of them dating back to the pioneers of ’94 or perhaps a little
later, and those who made the caches never returned to claim their