Abstract: In vivo 31P magnetic resonance spectra of 16 isolated dog brains were studied by using a 9.4-T wide-bore superconducting magnet. The observed Pi peak had an irregular shape, which implied that it represented more than one single homogeneous pool of Pi. To evaluate our ability to discriminate between single and multiple peaks and determine peak areas, we designed studies of simulated 31Pi spectra with the signal-to-noise (S/N) ratios ranging from ∞ to 4.4 with reference to the simulated Pi peak. For the analysis we used computer programs with a linear prediction algorithm (NMR-Fit) and a Marquardt–Levenberg nonlinear curve-fit algorithm (Peak-Fit). When the simulated data had very high S/N levels, both methods located the peak centers precisely; however, the Marquardt-Levenberg algorithm (M-L algorithm) was the more reliable at low S/N levels. The linear prediction method was poor at determining peak areas; at comparable S/N levels, the M-L algorithm determined all peak areas relatively accurately. Application of the M-L algorithm to the individual experimental in vivo dog brain data resolved the Pi peak into seven or more separate components. A composite spectrum obtained by averaging all spectral data from six of the brains with normal O2 utilization was fitted using the M-L algorithm. The results suggested that there were eight significant peaks with the following chemical shifts: 4.07, 4.29, 4.45, 4.62, 4.75, 4.84, 4.99, and 5.17 parts per million (ppm). Although linear prediction demonstrated the presence of only three peaks, all corresponded to values obtained using the M-L algorithm. The peak indicating a compartment at 5.17 ppm (pH 7.34) was assigned to venous pH on the basis of direct simultaneous electrode-based measurements. On the basis of earlier electrode studies of brain compartmental pH, the peaks at 4.99 ppm (pH 7.16) and 4.84 ppm (pH 7.04) were thought to represent interstitial fluid and the astrocyte cytoplasm, respectively.