Summary
The starting point of this research proposal is a recent result by the PI, making progress in a half century old,
notoriously open problem. In the mid 1960’s, Tukey and Cooley discovered the Fast Fourier Transform, an
algorithm for performing one of the most important linear transformations in science and engineering, the
(discrete) Fourier transform, in time complexity O(n log n).
In spite of its importance, a super-linear lower bound has been elusive for many years, with only very limited
results. Very recently the PI managed to show that, roughly speaking, a faster Fourier transform must result
in information loss, in the form of numerical accuracy. The result can be seen as a type of computational
uncertainty principle, whereby faster computation increases uncertainty in data. The mathematical argument
is established by defining a type of matrix quasi-entropy, generalizing Shannon’s measure of information
(entropy) to “quasi-probabilities” (which can be negative, more than 1, or even complex).
This result, which is not believed to be tight, does not close the book on Fourier complexity. More importantly,
the vision proposed by the PI here reaches far beyond Fourier computation. The computation-information
tradeoff underlying the result suggests a novel view of complexity theory as a whole. We can now revisit
some classic complexity theoretical problems with a fresh view. Examples of these problems include better
understanding of the complexity of polynomial multiplication, integer multiplication, auto-correlation and
cross-correlation computation, dimensionality reduction via the Fast Johnson-Linednstrauss Transform (FJLT;
also discovered and developed by the PI), large scale linear algebra (linear regression, Principal Component
Analysis - PCA, compressed sensing, matrix multiplication) as well as binary functions such as integer multiplication.
notoriously open problem. In the mid 1960’s, Tukey and Cooley discovered the Fast Fourier Transform, an
algorithm for performing one of the most important linear transformations in science and engineering, the
(discrete) Fourier transform, in time complexity O(n log n).
In spite of its importance, a super-linear lower bound has been elusive for many years, with only very limited
results. Very recently the PI managed to show that, roughly speaking, a faster Fourier transform must result
in information loss, in the form of numerical accuracy. The result can be seen as a type of computational
uncertainty principle, whereby faster computation increases uncertainty in data. The mathematical argument
is established by defining a type of matrix quasi-entropy, generalizing Shannon’s measure of information
(entropy) to “quasi-probabilities” (which can be negative, more than 1, or even complex).
This result, which is not believed to be tight, does not close the book on Fourier complexity. More importantly,
the vision proposed by the PI here reaches far beyond Fourier computation. The computation-information
tradeoff underlying the result suggests a novel view of complexity theory as a whole. We can now revisit
some classic complexity theoretical problems with a fresh view. Examples of these problems include better
understanding of the complexity of polynomial multiplication, integer multiplication, auto-correlation and
cross-correlation computation, dimensionality reduction via the Fast Johnson-Linednstrauss Transform (FJLT;
also discovered and developed by the PI), large scale linear algebra (linear regression, Principal Component
Analysis - PCA, compressed sensing, matrix multiplication) as well as binary functions such as integer multiplication.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/682203 |
| Start date: | 01-06-2016 |
| End date: | 31-05-2022 |
| Total budget - Public funding: | 1 515 801,00 Euro - 1 515 801,00 Euro |
Cordis data
Original description
The starting point of this research proposal is a recent result by the PI, making progress in a half century old,notoriously open problem. In the mid 1960’s, Tukey and Cooley discovered the Fast Fourier Transform, an
algorithm for performing one of the most important linear transformations in science and engineering, the
(discrete) Fourier transform, in time complexity O(n log n).
In spite of its importance, a super-linear lower bound has been elusive for many years, with only very limited
results. Very recently the PI managed to show that, roughly speaking, a faster Fourier transform must result
in information loss, in the form of numerical accuracy. The result can be seen as a type of computational
uncertainty principle, whereby faster computation increases uncertainty in data. The mathematical argument
is established by defining a type of matrix quasi-entropy, generalizing Shannon’s measure of information
(entropy) to “quasi-probabilities” (which can be negative, more than 1, or even complex).
This result, which is not believed to be tight, does not close the book on Fourier complexity. More importantly,
the vision proposed by the PI here reaches far beyond Fourier computation. The computation-information
tradeoff underlying the result suggests a novel view of complexity theory as a whole. We can now revisit
some classic complexity theoretical problems with a fresh view. Examples of these problems include better
understanding of the complexity of polynomial multiplication, integer multiplication, auto-correlation and
cross-correlation computation, dimensionality reduction via the Fast Johnson-Linednstrauss Transform (FJLT;
also discovered and developed by the PI), large scale linear algebra (linear regression, Principal Component
Analysis - PCA, compressed sensing, matrix multiplication) as well as binary functions such as integer multiplication.
Status
CLOSEDCall topic
ERC-CoG-2015Update Date
27-04-2024
Images
No images available.
Geographical location(s)
