How are electrodes, magnetometers or gradiometers described?
Sensor locations are described by the elec or grad field in the data object. These definitions of the sensors can contain fewer or more channels that present in the data, i.e., you can have bipolar EOG channels that do not have a unique position on the scalp, but you can also have reference gradiometers in the MEG system that do not have a signal attached to them.
The definition of EEG, ECoG and iEEG electrodes
As of September 23, 2011 we updated the description of how the sensors are defined in FieldTrip. The electrode definition contains the following field
elec.label % cellarray of length N with the label of each channel
elec.elecpos % Mx3 matrix with the cartesian coordinates of each electrode
elec.chanpos % Nx3 matrix with the cartesian coordinates of each channel
Note that there is typically a onetoone match between electrodes and channels, but in principle channels and electrodes can refer to different entities. In the context of EEG, one may consider a setup containing bipolar derivations, in which each ‘channel’ represents the voltage difference between a pair of electrodes. Consequently, the number of channels N then is different from the number of electrodes M. An additional field is needed in the elecstructure
elec.tra % NxM matrix with the weight of each electrode into each channel
to tell FieldTrip how to combine the electrodes into channels. This array can be stored as a sparse array and it also allows to set the position of the reference electrode in unipolar recordings. In case elec.tra is not provided, the forward and inverse calculations will be performed assuming an average reference over all electrodes.
The EEG potential is in first instance computed on the locations in elec.elecpos, and when applicable combined using elec.tra. The elec.chanpos field is used e.g. for visualization and determining neighbours.
The definition of MEG sensors
The gradiometer definition generally consists of multiple coils per channel, e.g. two coils for a 1st order axial gradiometer, in which the orientation of the coils is opposite. Each coil is described separately and one large matrix (grad.tra: can be sparse) has to be given that defines how the forward computed field is combined over the coils to generate the output of each channel. The gradiometer definition consists of the following fields as of September 23, 201
grad.coilpos % Mx3 matrix with the position of each coil
grad.coilori % Mx3 matrix with the orientation of each coil
grad.tra % NxM matrix with the weight of each coil into each channel
grad.label % cellarray of length N with the channel label
grad.chanpos % Nx3 matrix with the position of each channel
grad.chanori % Nx3 matrix with the orientation of each channel.
The channel ‘orientation’ is needed for synthetic gradient computation for axial gradiometer or magnetometer systems. If you don’t know what it means and need to construct your own grad structure, please set it to nan(N,3).
MEG forward computations are performed for each grad.coilpos and grad.coilori, and subsequently combined using grad.tra. Although they are called “coils”, you can better think of them as “field digitization points”.
By default a first order gradiometer is described by 2 “coils”, but you could use more digitization points to get a more accurate forward model.
The old electrode and gradiometer structure
The old electrode definition contained the following field
elec.pnt % Mx3 matrix with the position of each electrode
elec.label
The old gradiometer definition contained the following field
grad.pnt % Mx3 matrix with the position of each coil
grad.ori % Mx3 matrix with the orientation of each coil
grad.label
grad.tra
The upgrade from this to the current representation is motivated by the fact that the relevant information that is needed from the grad/elec structure is different for different analysis/visualization step

for displaying purposes, usually the channels are the entities of relevance. Also, in the context of finding neighbours to a given channel (for clustering, or synthetic gradient computation, or interpolation as in scalpcurrentdensity or channelrepair), the channels are the relevant entities.

for forward and inverse modelling purposes, the sensing elements, i.e. the electrodes or coils are of relevance.
Originally, FieldTrip relied on the fact that the channel positions can be recovered from the electrode/coil positions by looking into the tramatrix, because the tramatrix specifies which electrode/coil contributes to which channel. However, FieldTrip supports increasingly complicated tramatrices that for example include balancing coefficients (obtained through ft_denoise_synthetic, or ft_denoise_pca), projectedout spatial topographies (obtained through a sequence of ft_componentanalysis and ft_rejectcomponent), or synthetic planar gradients (obtained through ft_megplanar). With these increasingly complicated tramatrices, recovery of the channel positions from the coil/electrode positions is not straightforward and sometimes impossible. We decided to make the distinction between channels on the one hand, and electrodes/coils on the other hand explicit in the code.
Some additional notes on the ‘tra’matrix
The tramatrix is a very important piece of information that needs to be taken into account when building forward models (leadfields) for the sensor data in a given data structure. When building a forward model, we compute the magnetic/electric field distribution at the described sensors/electrodes in the data, given a known dipolar source. If the sensor data has been manipulated in any way  e.g. by creating higher order synthetic gradients by using additional information from the reference coils (as can be done with CTF MEG data with ft_denoise_synthetic, or with the custom CTF software), by using adaptive weights estimated from the data (as can be done with 4Ddata, using custom software or ft_denoise_pca), or also when removing spatial topographies from the sensor data (using a combination of ft_componentanalysis and ft_rejectcomponent)  the corresponding leadfields need to be manipulated in the same way, to keep the forward model consistent with the data. In FieldTrip this is achieved with the tramatrix.
As the tramatrix provides the recipe of how the individual electrodes/coils relate to the individual channels in the data structure, it is updated automatically upon manipulation of the sensor data in the following function
ft_denoise_synthetic ft_denoise_pca ft_componentanalysis ft_rejectcomponent
Algorithmically, the tramatrix is used as a left multiplier of the unbalanced lead field (i.e. the leadfield that represents the magnetic field distribution at the location of the original magnetometer coils) in the following way: lf_balanced = grad.tra * lf_unbalanced. For example, to obtain a firstorder axial gradiometer, each row in the tramatrix contains two ‘1’s (assuming the orientation of the top and bottom coils to be opposite), indicating that the modelled field estimated at the top and bottom coil of a gradiometer should be summed to obtain a model of the axial gradients. In order to obtain synthetic higher order gradients, the columns in the tramatrix that correspond to the reference coils will have nonzero values, reflecting the ‘balancing’ coefficients.
In summary, if you are doing fancy things with your data, and later on want to do source reconstruction, it’s not safe to just take any gradstructure to construct your leadfields with. For obvious reasons, not only do the positions of the coils need to be the same as the ones used during the measurement. Additionally, the tramatrix needs to reflect all the manipulations you have applied to the data that you wish to sourcereconstruct.