Microhistological recognition of food particles in vole stomachs was based on leaf epiderm morphology. The shape of epidermal cells is taxon-specific, and several additional features, such as trichomes, hairs and characteristics of cells surrounding stomata, can be used for species identification [25, 35]. A photography-guide (Soininen & Nielsen, unpublished data) of epidermis of all vascular plants recorded at the sampling area (Ravolainen et al. unpublished data) was prepared using a method modified from Carrière . Dry plant samples were soaked overnight, scraped to reveal the epidermis and bleached with household bleach to clear the tissue of chlorophyll. Hard leaves were first boiled in table vinegar to soften the mesophyll tissue. Microphotographs (40×) were taken of abaxial and adaxial leaf side and leaf edge of all plants. Additional photographs were taken of stems and seeds of certain species of special interest. In addition to the specifically prepared epidermis photographs, photographs and microscopy slides of arctic plant epidermis, received from C. Hübner and E. Bjørkevoll, were used to aid identification.
A method modified from Hansson  was used for microscopy analysis. After taking subsamples for DNA analysis, stomach contents were filtered to > 0.16 mm and > 0.56 mm fractions. These were bleached with approximately 2 mL of household bleach for approximately 1/2 hour. One sample per fraction was analyzed, mounting a droplet of it on a microscopy slide. The frequency of occurrence of food items was recorded by light microscope (40×), by counting 25 hits on identifiable material along a measure grid. When approximately 95% of fragments were unidentifiable, the slide was discarded. For four individuals, no slide with adequate amount of identifiable material could be made. Therefore, they were discarded from the microhistological analysis.
Two samples were analyzed for each individual. Total DNA was extracted from about 10 mg of sample with the DNeasy Tissue Kit (Qiagen GmbH, Hilden, Germany), following the manufacturer's instructions. The DNA extracts were recovered in a total volume of 300 μ L. Mock extractions without samples were systematically performed to monitor possible contaminations.
DNA amplifications were carried out in a final volume of 25 μ L, using 2.5 μ L of DNA extract as template. The amplification mixture contained 1 U of AmpliTaq® Gold DNA Polymerase (Applied Biosystems, Foster City, CA), 10 mM Tris-HCl, 50 mM KCl, 2 mM of MgCl2, 0.2 mM of each dNTP, 0.1 μ M of each primer, and 0.005 mg of bovine serum albumin (BSA, Roche Diagnostic, Basel, Switzerland). The mixture was denatured at 95°C for 10 min, followed by 35 cycles of 30 s at 95°C, and 30 s at 55°C; the elongation was removed in order to reduce the +A artifact [37, 38]. Samples were amplified with using the universal primers g and h described by Taberlet et al. . The addition of a specific tag on the 5' end allowed an assignement of sequences to the respective samples. After amplification all samples were pooled for the pyrosequencing run. Each sample was recognized by a specific five bases long tag with at least two differences between tags for a better assignation of sequences to samples during bioinformatic segregation of sequences.
PCR products were purified using the MinElute PCR purification kit (Qiagen GmbH, Hilden, Germany). DNA quantification was carried out using the BioAnalyzer (Agilent Technologies, Inc., Santa Clara, CA). Taking these concentrations into account PCR products were pooled leading to equal amounts per sample. Large-scale pyrosequencing was carried out using GS FLX sequencer (Roche, Basel, Switzerland) following the manufacturer's instructions.
The first step of analyzing the output of the pyrosequencing consisted of sorting the different sequences according to the tag present on the 5' end of the primers. Thus, for each sample (each stomach content), a new file was generated, containing all the sequences having the relevant tag. Then, these sequences were analyzed to determine the diet. To limit the influence of sequence errors , only sequences that were present more than three times were considered in the subsequent analyses.
The sequences were compared to a database of 842 species representing all widespread and/or ecologically important taxa of the arctic flora (GenBank accession number GQ244527 to GQ245667) (Sønstebø et al: A minimalist DNA barcoding approach for reconstructing past Arctic vegetation and climate, submitted). It was developed by sequencing the whole chloroplast trn L (UAA) intron of these species using primer pair designed by Taberlet et al. , and following the protocol described and evaluated in Taberlet et al. . In the database a total of 33,5% of species and 77,1% of genera could be identified by the P6 loop. All families were unambiguously identified (Sønstebø et al: A minimalist DNA barcoding approach for reconstructing past Arctic vegetation and climate, submitted). When sequences were not fully identified using the arctic plant database, they were compared with sequences retrieved from GenBank, using ecoPCR ; http://www.grenoble.prabi.fr/trac/ecoPCR. The taxon was assigned to each sequence in a dataset by similarity assessment with a reference database using FASTA  algorithm. A FASTA alignment was retrieved if there was at least 98% of identity between query and database sequences and 100% of query coverage. If two or more taxa could be assigned with the same score for a given sequence, we assigned this sequence to the higher taxonomic level that included both taxa. This method resulted in some sequenced taxa being assigned to the rank of genus or family.
Chimeric sequences are a well know problem when amplifying a mixture of homologous genes, and it is impossible to avoid their formation . But if two unrelated taxa compose the chimeric sequence the resulting sequence is not taken into account because for taxon identification the FASTA alignments is retrieved only if the sequence have 100% of query coverage with the reference sequence. If two related taxa compose the chimeric sequence, this sequence is assigned to the higher taxonomic level that included both taxa (e.g. genus, family, order, etc.).
Results of the two methods were compared in two ways. First, the taxonomic resolution obtained by the two methods was examined by comparing the relative frequencies of hits (microscopy) and sequences (DNA) at different taxonomic levels. For this comparison, the four individuals with no microscopy data were excluded also from DNA dataset.
Second, the relative frequencies of food items in diets were compared between the methods. Taxonomic adjustments were first made to make the results from the two methods comparable. Nomenclature from species to family level follows Lid & Lid  and Elven , and for higher taxonomy Judd et al. . Several genera are represented only by one species in the study area and sequences assigned to these were therefore attributed to the respective species (e.g. Arctous alpinus, Bistorta vivipara, Chamaepericlymenum suecicum, Rumex acetosa). Similar adjustments were done at other taxonomic levels (e.g. Salicaceae to Salix and Ranunculales to Ranunculaceae). Then, proportions of food items estimated at the level of individual voles were averaged (mean ± standard deviation) across species and sampling season for both methods.