eofunc_n
Computes empirical orthogonal functions (EOFs, aka: Principal Component Analysis), given an index that specifies which dimension contains the number of observations.
Available in version 6.4.0 and later.
Prototype
function eofunc_n ( data : numeric, neval : integer, optEOF : logical, dim [1] : integer ) return_val : numeric
Arguments
dataA multidimensional array in which the dim index specifies the dimension that contains the number of observations. Generally, this is the time dimension. Commonly, the data array contains anomalies from some base climatology, however, this is not required.
nevalA scalar integer that specifies the number of eigenvalues and eigenvectors to be returned. This is usually less than or equal to the minimum number of observations or number of variables.
optEOFA logical variable to which various optional arguments may be assigned as attributes. These optional arguments alter the default behavior of the function. Must be set to True prior to setting the attributes which are assigned using the @ operator:

optEOF = True optEOF@jopt = 1
optEOF@jopt = 1: use correlation matrix to compute EOFs. The default is to use a covariance matrix (optEOF@jopt = 0). 
optEOF = True optEOF@pcrit = 85
optEOF@pcrit = %: a float value that indicates the percentage of nonmissing points that must exist at any single point in order to be included in the calculations. The default is 50%. Points that contain more than 'pcrit' missing values will be excluded from the computations.
The dimension index of data that represents the dimension containing the number of observations. Generally, this is the time dimension.
Return value
A multidimensional array containing normalized EOFs. The returned array will be of the same size as data with the 'time' dimension removed and an additional leftmost dimension of the same size as neval added. Double if data is double, float otherwise.
The return variable will have associated with it the following attributes:
 eval: a onedimensional array of size neval that contains the eigenvalues.
 pcvar: a onedimensional float array of size neval equal to the percent variance associated with each eigenvalue.
 pcrit: The same value and type of optEOF@pcrit if the user changed the default.
 matrix: A string indicating the type of matrix used, "correlation" or "covariance".
 method: A string indicating if the input array, data, was/wasnot transposed for the purpose of computing the eigenvalues and eigenvectors. The string can have two values: "transpose" or "no transpose"
 eval_transpose: This attribute is returned only if method="transpose". eval_transpose will contain the eigenvalues of the tranposed covariance matrix. These eigenvalues are then scaled such that they are consistent with the original input data. As of version 4.3.0, the scaled eigenvalues are returned as eval.
print(return_val@pcvar) print(return_val@pcrit) print(return_val@sig) ; v6.3.0 onward
Description
This function is identical to eofunc, except it has an extra dim argument that allows you to specify which dimension index is the "time" dimension. This keeps you from having to unnecessarily reorder the data to force "time" to be the rightmost dimension.
This function computes Empirical Orthogonal Functions (EOFs) via a covariance matrix or, optionally, via a correlation matrix. This is also known as Principal Component Analysis or Eigen Analysis. The EOFs are calculated using LAPACK's "dspevx" routine. Missing values are ignored when computing the covariance or correlation matrix. The returned values are normalized such that the sum of squares for each EOF pattern equals one. To denormalize the returned EOFs multiply by the square root of the associated eigenvalue (aka, the singular value).
Most commonly, the input data consists of anomalies.
This function differs from the deprecated eofcov and eofcor functions in that it may transpose the input data array prior to computing the EOFs. If data is transposed, a linear transformation is applied to the EOFs of the transposed array prior to returning. The reason for using this approach is computational efficiency.
Comments on weighting observations
Generally, when performing an EOF analysis on observations over the globe or a portion of the globe, the values are weighted prior to calculating. This is usually required to account for the convergence of the meridions (area weighting) which lessens the impact of highlatitude grid points that represent a small area of the globe. Most frequently, the square root of the cosine of the latitude is used to compute the area weight. The square root is used to create a covariance matrix that reflects the area of each matrix element. If weighted in this manner, the resulting covariance values will include quantities calculated via:
[x*sqrt(cos(lat(x)))]*[y*sqrt(cos(lat(y)))] = x*y*sqrt(cos(lat(x)))*sqrt(cos(lat(y)))Note that the covariance of a grid point with itself yields standard cosine weighting:
[x*sqrt(cos(lat(x)))]*[x*sqrt(cos(lat(x)))] = x^2 * cos(lat(x)).Note on standard EOF analysis
Conventional EOF analysis yields patterns and time series which are both orthogonal. The derived patterns are a function of the domain. However, the EOF procedure is strictly mathematical (not statistical) and is not based upon physics. The results may produce patterns that are similar to physical modes within the the system. However, physical meaning is dependent on your interpretation of the mathematical result.
Note on signs of EOF analysis (conributed by Andrew Dawson, UEA)
EOFs are eigenvectors of the covariance matrix formed from the input data. Since an eigenvector can be multiplied by any scalar and still remain an eigenvector, the sign is arbitrary. In a mathematical sense the sign of an eigenvector is rather unimportant. This is why the EOF analysis may yield different signed EOFs for slightly different inputs. Sign only becomes an issue when you wish to interpret the physical meaning (if any) of an eigenvector.
You should approach the interpretation of EOFs by looking at both the EOF pattern and the associated time series together. For example, consider an EOF of sea surface temperature. If your EOF has a positive centre and the associated time series is increasing, then you will interpret this centre as a warming signal. If your EOF had come out the other sign (ie. a negative centre), then the associated time series would also be the opposite sign and you would still interpret the centre as a warming signal.
In essence, the sign flip does not change the physical interpretation of the result. Hence, it is up to you to choose which sign to associate with your EOF patterns for visualisation (remembering that any sign change to an EOF must be applied to the associated time series also). Usually you would simply adjust the sign so that all your EOF patterns with the same physical interpretation also look the same.
If desired, EOF spatial patterns may be tested for orthogonality by using the dot product:
d01 = sum(eof(0,:,:)*eof(1,:,:)) d12 = sum(eof(1,:,:)*eof(2,:,:)) d02 = sum(eof(0,:,:)*eof(2,:,:)) print("d01="+d01+" d12="+d12+" d02="+d02) ; may be +/ 1e8Use eofunc_n_Wrap if retention of metadata is desired.
References:
Quadrelli, Roberta, Christopher S. Bretherton, John M. Wallace, 2005: On Sampling Errors in Empirical Orthogonal Functions. J. Climate, 18, 37043710 North, G. R., T. L. Bell, R. F. Cahalan, and F. J. Moeng, Sampling errors in the estimation of empirical orthogonal functions, Mon. Wea. Rev., 110, 699706, 1982. Dawson, A.: EOF Analysis
Acknowledgement: The code used is a modified version of David Pierce's fortran code.
See Also
eofunc_Wrap, eofunc_north, eofunc, eofunc_n_Wrap, eofunc_ts, eofunc_ts_n, eofunc_ts_Wrap, eofunc_ts_n_Wrap, eofunc_varimax, eof2data, eof2data_n
Examples
In the following, the attribute pcvar can be output via:
print(ev@pcvar) ; 1D vector of length "neval"
This attribute could also be used in graphics. For example, it could be used in a title.
title = "%=" + ev@pcvar(1)
sprintf can be used to format the title more precisely:
title = "%=" + sprintf("%5.2f", ev@pcvar(1) )Example 1
Let x be twodimensional with dimensions "variables" (size = nvar) and "time". Commonly, 'x' contains anomalies.
neval = 3 ; calculate 3 EOFs out of 7 ev = eofunc_n(x,neval,False,1) ; ev(neval,nvar) ; print the percentvariance explained print("% var="+ ev@pcvar) option = True option@jopt = 1 ; use correlation matrix ev_cor = eofunc_n(x,neval,option,1) ; ev_cor(neval,nvar) ; print the percentvariance explained print("% var="+ ev_cor@pcvar)Example 2
Let x be threedimensional with dimensions of time, lat, lon:
neval = nvar ; calculate all EOFs ev = eofunc_n(y,neval,False,0) ; 0=index of time dimension ; ev(neval,nlat,nlon) ; denormalize the EOFs [units same as data] do ne=0,neval1 ev(ne,:,:) = ev(ne,:,:)*sqrt( ev@eval(ne) ) end doExample 3
Let z be fourdimensional with dimensions time, lev, lat, lon:
neval = 3 ; calculate 3 EOFs out of klev*nlat*mlon ev = eofunc_n(z,neval,False,0) ; ev will be dimensioned neval, level, lat, lonExample 4
Calculate the EOFs at every other lat/lon point:
neval = 5 ; calculate 5 EOFs out of nlat*mlon ev = eofunc_n(z(:,::2,::2),neval,False,0) ; ev(neval,nlat/2,mlon/2)Example 5
Let z be fourdimensional with dimensions level, lat, lon, time. Calculate the EOFs at one specified level:
kl = 3 ; specify level neval = 8 ; calculate 8 EOFs out of nlat*mlon ev = eofunc_n(z(kl,:,:,:),neval,False,3) ; ev will be dimensioned neval, lat, lonExample 6
Let z be fourdimensional with dimensions time, lev, lat, lon. Calculate EOFs on one specified level:
kl = 3 ; specify level neval = 8 ; calculate 8 EOFs out of nlat*mlon ev = eofunc_n(z(:,kl,:,:),neval,False,0) ; ev will be dimensioned neval, lat, lonExample 7
Areaweight the data prior to calculation. Let p be threedimensional with dimensions time x lat x lon. The array lat contains the latitudes.
; calculate the weights using the square root of the cosine of the latitude and ; also convert degrees to radians wgt = sqrt(cos(lat*0.01745329)) pw = p ; create an array with metadata ; weight each point prior to calculation. ; conform is used to make wgt the same size as pt pw = p*conform(p, wgt, 1) optEOF = True optEOF@pcrit = 80 evec = eofunc_n(pw,neval,optEOF,0) ; denormalize the EOFs ; print the % variance explained do ne=0,neval1 evec(ne,:,:) = evec(ne,:,:)*sqrt( evec@eval(ne) ) ; units same as data print("% var="+ evec@pcvar(ne) ) end do