Rssa (version 1.1)

fossa: Nested Filter-adjusted O-SSA decomposition

Description

Perform nested decomposition by Filter-adjusted O-SSA (FOSSA).

Usage

# S3 method for ssa
fossa(x, nested.groups, filter = c(-1, 1), gamma = Inf, normalize = TRUE, ...)

Value

Object of class `ossa'. The field `ossa.set' contains the vector of indices of elementary components used in Filter-adjusted O-SSA (that is, used in nested.groups).

Arguments

x

SSA object holding SSA decomposition

nested.groups

vector of numbers of eigentriples from full decomposition for nested decomposition. The argument is coerced to a vector, if necessary

filter

numeric vector or array of reversed impulse response (IR) coefficients for filter adjustment or list of such vectors or arrays

gamma

weight of filter adjustment. See `Details' and `References'

normalize

logical, whether to normalize left decomposition vectors before filtering

...

additional arguments passed to decompose routines

Details

See Golyandina N. and Shlemov A. (2015) and Section 2.5 in Golyanina et al (2018) for full details in the 1D case and p.250-252 from the same book for an example in the 2D case.

Briefly, FOSSA serves for decomposition of series components that are mixed due to equal contributions of their elementary components, e.g. of sinusoids with equal amplitudes or of complex-form trend and periodics. FOSSA performs a new decomposition of a part of the ssa-object, which is given by a set of eigentriples. Note that eigentriples that do not belong to the chosen set are not changed.

In particular, Filter-adjusted O-SSA performs a nested decomposition specified by a number of eigentriples via Oblique SSA with a specific inner product in the row space: $$% \langle x, y \rangle = (x, y) + \gamma^2(\Phi(x), \Phi(y)), $$ where \((\cdot, \cdot)\) denotes conventional inner product and `\(\Phi\)' is linear filtration which is specified by filter argument.

The default value of \(\Phi\) corresponds to sequential differences, that is, to derivation. Such version of Filter-adjusted O-SSA is called `DerivSSA'. See `References' for more details.

filter argument

For 1D-SSA, Toeplitz-SSA and MSSA: Filter can be given by a vector or a list of vectors. Each vector corresponds to reversed IR for a filter, these filters are applied independently and their results are stacked such that the matrix \([X:\Phi_1(X):\Phi_2(X)]\) is decomposed.

For 2D-SSA: the following variants are possible: (1) a list of vectors. Each vector corresponds to reversed IR for a filter. Each filter is applied to different dimensions, the first to columns, the second to rows, and the results are stacked. (2) single vector. Given vector corresponds to one-dimensional filter applied to both dimensions, the same as list of two equal vectors. (3) a list of matrices, where each matrix provides 2d filter coefficients and the results are stacked. (4) single matrix. Given matrix corresponds to two-dimensional filter applied once, the same as list of one matrix.

For nD-SSA: the same as for 2D-SSA, a list of vectors for filters by directions, single vector, a list of arrays (matroids) for nD filters or single array.

Normalization

Let us explain for the 1D case. Let \(X\) be the reconstructed matrix, corresponding to the selected eigentriples \(\{(\sigma_i,U_i,V_i)\}\), \(\Psi(X)\) is the matrix, where the filter is applied to each row of \(X\).

Then normalize = FALSE (Algorithm 2.9 or 2.10 in Golyandina et al (2018)) corresponds to finding the basis in the column space of \(X\) by means of the SVD of \([X, \Psi(X)]\), while normalize = TRUE (by default, see Algorithm 2.11 in Golyandina et al (2018)) corresponds to finding the basis by the SVD of \([V, \Phi(V)]\), where the rows of matrix \(V\) are \(V_i\). The value by default TRUE guaranties that the contributions of sine waves will be ordered by decreasing of frequencies, although can slightly worsen the weak separability

References

Golyandina N., Korobeynikov A., Zhigljavsky A. (2018): Singular Spectrum Analysis with R. Use R!. Springer, Berlin, Heidelberg.

Golyandina N. and Shlemov A. (2015): Variations of Singular Spectrum Analysis for separability improvement: non-orthogonal decompositions of time series, Statistics and Its Interface. Vol.8, No 3, P.277-294. https://arxiv.org/abs/1308.4022

See Also

Rssa for an overview of the package, as well as, iossa.

Examples

# Separation of two mixed sine-waves with equal amplitudes
N <- 150
L <- 70
omega1 <- 1/15
omega2 <- 1/10

v <- sin(2*pi*omega1 * (1:N)) + sin(2*pi*omega2 * (1:N))
s <- ssa(v, L)
fs <- fossa(s, nested.groups = 1:4, gamma = 100)

# Rssa does most of the plots via lattice
ws <- plot(wcor(s, groups = 1:4))
wfs <- plot(wcor(fs, groups = 1:4))
plot(ws, split = c(1, 1, 2, 1), more = TRUE)
plot(wfs, split = c(2, 1, 2, 1), more = FALSE)

opar <- par(mfrow = c(2, 1))
plot(reconstruct(s, groups = list(1:2, 3:4)))
plot(reconstruct(fs, groups = list(1:2, 3:4)))
par(opar)

# Real-life example: Australian Wine Sales
# \donttest{
data(AustralianWine)
s <- ssa(AustralianWine[1:120, "Fortified"], L = 60)
fs <- fossa(s, nested.groups = list(6:7, 8:9, 10:11), gamma = 10)

plot(reconstruct(fs, groups = list(6:7, 8:9, 10:11)))
plot(wcor(s, groups = 6:11))
plot(wcor(fs, groups = 6:11))

# Real life example: improving of strong separability
data(USUnemployment)
unempl.male <- USUnemployment[, "MALE"]
s <- ssa(unempl.male)
fs <- fossa(s, nested.groups = 1:13, gamma = 1000)

# Comparison of reconstructions
rec <- reconstruct(s, groups = list(c(1:4, 7:11), c(5:6, 12:13)))
frec <- reconstruct(fs, groups <- list(5:13, 1:4))
# Trends
matplot(data.frame(frec$F1, rec$F1, unempl.male), type= 'l',
        col=c("red","blue","black"), lty=c(1,1,2))
# Seasonalities
matplot(data.frame(frec$F2, rec$F2), type = 'l', col=c("red","blue"), lty=c(1,1))

# W-cor matrices before and after FOSSA
ws <- plot(wcor(s, groups = 1:30), grid = 14)
wfs <- plot(wcor(fs, groups = 1:30), grid = 14)
plot(ws, split = c(1, 1, 2, 1), more = TRUE)
plot(wfs, split = c(2, 1, 2, 1), more = FALSE)

# Eigenvectors before and after FOSSA
plot(s, type = "vectors", idx = 1:13)
plot(fs, type = "vectors", idx = 1:13)

# 2D plots of periodic eigenvectors before and after FOSSA
plot(s, type = "paired", idx = c(5, 12))
plot(fs, type = "paired", idx = c(1, 3))

# Compare FOSSA with and without normalize
N <- 150
L <- 70
omega1 <- 1/15
omega2 <- 1/10

v <- 3*sin(2*pi*omega1 * (1:N)) + 2*sin(2*pi*omega2 * (1:N))
s <- ssa(v, L)
fs <- fossa(s, nested.groups = 1:4, gamma = 100)
fs.norm <- fossa(s, nested.groups = 1:4, gamma = 100, normalize = TRUE)
opar <- par(mfrow = c(2, 1))
plot(reconstruct(fs, groups = list(1:2, 3:4)))
plot(reconstruct(fs.norm, groups = list(1:2, 3:4)))
par(opar)

# 2D example
data(Mars)
s <- ssa(Mars)
plot(s, "vectors", idx = 1:50)
plot(s, "series", idx = 1:50)
fs <- fossa(s, nested.groups = 1:50, gamma = Inf)
plot(fs, "vectors", idx = 1:14)
plot(fs, "series", groups = 1:13)

# Filters example, extracting horizontal and vertical stripes
data(Mars)
s <- ssa(Mars)
fs.hor <- fossa(s, nested.groups = 1:50, gamma = Inf,
                filter = list(c(-1, 1), c(1)))
plot(fs.hor, "vectors", idx = 1:14)
plot(fs.hor, "series", groups = 1:13)
fs.ver <- fossa(s, nested.groups = 1:50, gamma = Inf,
                filter = list(c(1), c(-1, 1)))
plot(fs.ver, "vectors", idx = 1:14)
plot(fs.ver, "series", groups = 1:13)
# }