The relation between the stellar mass and the star formation rate characterizes how the instantaneous star formation is determined by the galaxy past star formation history and by the growth of the dark matter structures. We deconstruct the M-SFR plane by measuring the specific SFR functions in several stellar mass bins from z=0.2 out to z=1.4. Our analysis is primary based on a MIPS 24$\mu m$ selected catalogue combining the COSMOS and GOODS surveys. We estimate the SFR by combining mid- and far-infrared data for 20500 galaxies. The sSFR functions are derived in four stellar mass bins within the range 9.5<log(M/Msun)<11.5. First, we demonstrate the importance of taking into account selection effects when studying the M-SFR relation. Secondly, we find a mass-dependent evolution of the median sSFR with redshift varying as $sSFR \propto (1+z)^{b}$, with $b$ increasing from $b=2.88$ to $b=3.78$ between $M=10^{9.75}Msun$ and $M=10^{11.1}Msun$, respectively. At low masses, this evolution is consistent with the cosmological accretion rate and predictions from semi-analytical models (SAM). This agreement breaks down for more massive galaxies showing the need for a more comprehensive description of the star-formation history in massive galaxies. Third, we obtain that the shape of the sSFR function is invariant with time at z<1.4 but depends on the mass. We observe a broadening of the sSFR function ranging from 0.28 dex at $M=10^{9.75}Msun$ to 0.46 dex at $M=10^{11.1}Msun$. Such increase in the scatter of the M-SFR relation suggests an increasing diversity of SFHs as the stellar mass increases. Finally, we find a gradual decline of the sSFR with mass as $log(sSFR) \propto -0.17M$. We discuss the numerous physical processes, as gas exhaustion in hot gas halos or secular evolution, which can gradually reduce the sSFR and increase the SFH diversity.

We estimate the galaxy stellar mass function and stellar mass density for star-forming and quiescent galaxies with 0.2<z<4. We construct a deep K<24 sample of 220000 galaxies selected using the UltraVISTA DR1 data release. Our analysis is based on precise 30-band photometric redshifts. By comparing these photometric redshifts with 10800 spectroscopic redshifts from the zCOSMOS bright and faint surveys, we find a precision of sigma(dz/(1+z))=0.008 at i<22.5 and sigma(dz/(1+zs))=0.03 at 1.5<z<4. We derive the stellar mass function and correct for the Eddington bias. We find a mass-dependent evolution of the global and star-forming populations. This mass-dependent evolution is a direct consequence of the star formation being quenched in galaxies more massive than M>10^10.7Msun. For the mass function of the quiescent galaxies, we do not find any significant evolution of the high-mass end at z<1; however we observe a clear flattening of the faint-end slope. From z~3 to z~1, the density of quiescent galaxies increases over the entire mass range. Their comoving stellar mass density increases by 1.6 dex between z~3 and z~1 and by less than 0.2dex at z<1. We infer the star formation history from the mass density evolution and we find an excellent agreement with instantaneous star formation rate measurements at z<1.5, while we find differences of 0.2dex at z>1.5 consistent with the expected uncertainties. We also develop a new method to infer the specific star formation rate from the mass function of star-forming galaxies. We find that the specific star formation rate of 10^10Msun galaxies increases continuously in the redshift range 1<z<4. Finally, we compare our results with a semi-analytical model and find that these models overestimate the density of low mass quiescent galaxies by an order of magnitude, while the density of low-mass star-forming galaxies is successfully reproduced.