Add README

This commit is contained in:
kalmarek 2017-09-10 15:10:55 +02:00
parent ee11e32030
commit 08f36ad3ff

86
README.md Normal file
View File

@ -0,0 +1,86 @@
This repository contains code for computations in [Certifying Numerical Estimates of Spectral Gaps](https://arxiv.org/abs/1703.09680).
# Installing
To run the code You need `julia-v0.5` (should work on `v0.6`, but with warnings).
You also need to install julia packages: `Nemo-v0.6.3`, `ArgParse`. To do so in `julia`'s REPL run:
```julia
Pkg.update()
Pkg.add("Nemo")
Pkg.add("ArgParse")
```
Then clone the main repository of `Groups.jl`, `GroupRings.jl` and `PropertyT.jl`:
```julia
Pkg.clone("https://git.wmi.amu.edu.pl/kalmar/Groups.jl.git")
Pkg.clone("https://git.wmi.amu.edu.pl/kalmar/GroupRings.jl.git")
Pkg.clone("https://git.wmi.amu.edu.pl/kalmar/PropertyT.jl.git")
Pkg.resolve()
```
This should resolve all dependencies (e.g. install `JuMP`, `SCS`, `IntervalArithmetic`, `JLD`, `Memento`). Exit julia and finally clone this repository:
```shell
git clone https://git.wmi.amu.edu.pl/kalmar/GroupsWithPropertyT.git
cd GroupswithPropertyT
```
# Running
To check that $\Delta^2-\lambda\Delta$ is not decomposable to a sum of hermitian squares of elements in the ball of radius $2$ in $SL(2,7)$ run
```shell
julia SL.jl -N 2 -p 7 --radius 2 --iterations 100000
```
(~30 seconds, depending on hardware). The monotonous decreasing $\lambda$ during the optimisation is in column `pri obj` (or `dua obj`) of `solver.log`.
Compare this to
```shell
julia SL.jl -N 2 -p 7 --radius 3 --iterations 100000
```
which finds $\lambda \geq 0.5857$ and decomposes $\Delta^2-\lambda\Delta$ into sum of $47$ hermitian squares in less than 20 seconds (including certification).
If You see in the output (or in `full.log`) that the upper end of the interval where $\lVert\Delta^2 - \lambda\Delta - \sum{\xi_i}^*\xi_i\rVert_1$ belongs to is too large (resulting in positive `Floating point distance`, but negative `The Augmentation-projected actual distance`), decrease the `--tol` parameter, e.g.
```
julia SL.jl -N 2 -p 7 --radius 3 --iterations 100000 --tol 1e-9
```
to achieve a better estimate (the residuals $\ell_1$-norm should be around $\|B_d(e))\|*tol$)
# Help
```shell
julia SL.jl --help
usage: SL.jl [--tol TOL] [--iterations ITERATIONS]
[--upper-bound UPPER-BOUND] [--cpus CPUS] [-N N] [-p P]
[--radius RADIUS] [-h]
optional arguments:
--tol TOL set numerical tolerance for the SDP solver
(type: Float64, default: 1.0e-6)
--iterations ITERATIONS
set maximal number of iterations for the SDP
solver (default: 20000) (type: Int64, default:
50000)
--upper-bound UPPER-BOUND
Set an upper bound for the spectral gap (type:
Float64, default: Inf)
--cpus CPUS Set number of cpus used by solver (type:
Int64)
-N N Consider elementary matrices EL(N) (type:
Int64, default: 2)
-p P Matrices over field of p-elements (p=0 => over
ZZ) (type: Int64, default: 0)
--radius RADIUS Radius of ball B_r(e,S) to find solution over
(type: Int64, default: 2)
-h, --help show this help message and exit
```
# Specific version of the article
To checkout the specific versions of packages used for [Certifying Numerical Estimates of Spectral Gaps](https://arxiv.org/abs/1703.09680) run (inside the cloned `GroupswithPropertyT`)
```shell
git checkout 1703.09680v1
```
Unfortunately: You need to link `~/.julia/v0.5/GroupRings` to `~/.julia/v0.5/GroupAlgebras` due to change in the name of the package. Then run in `julia`
```julia
Pkg.checkout("Groups", "1703.09680v1")
Pkg.checkout("GroupRings", "1703.09680v1")
Pkg.checkout("PropertyT", "1703.09680v1")
Pkg.resolve()
```