This work presents the first comprehensive study of structure formation at the peak epoch of cosmic star formation over $1.4\leq z \leq 3.6$ in the COSMOS field, including the most massive high redshift galaxy proto-clusters at that era. We apply the extended COSMIC BIRTH algorithm to account for a multi-tracer and multi-survey Bayesian analysis at Lagrangian initial cosmic times. Combining the data of five different spectroscopic redshift surveys (zCOSMOS-deep, VUDS, MOSDEF, ZFIRE, and FMOS-COSMOS), we show that the corresponding unbiased primordial density fields can be inferred, if a proper survey completeness computation from the parent photometric catalogs, and a precise treatment of the non-linear and non-local evolution on the light-cone is taken into account, including (i) gravitational matter displacements, (ii) peculiar velocities, and (iii) galaxy bias. The reconstructions reveal a holistic view on the known proto-clusters in the COSMOS field and the growth of the cosmic web towards lower redshifts. The inferred distant dark matter density fields concurrently with other probes like tomographic reconstructions of the intergalactic medium will explore the interplay of gas and dark matter and are ideally suited to study structure formation at high redshifts in the light of upcoming deep surveys.

We present COSMIC BIRTH: COSMological Initial Conditions from Bayesian Inference Reconstructions with THeoretical models: an algorithm to reconstruct the primordial and evolved cosmic density fields from galaxy surveys on the light-cone. The displacement and peculiar velocity fields are obtained from forward modelling at different redshift snapshots given some initial cosmic density field within a Gibbs-sampling scheme. This allows us to map galaxies, observed in a light-cone, to a single high redshift and hereby provide tracers and the corresponding survey completeness in Lagrangian space including phase-space mapping. These Lagrangian tracers in turn permit us to efficiently obtain the primordial density field, making the COSMIC BIRTH code general to any structure formation model. Our tests are restricted for the time being to Augmented Lagrangian Perturbation theory. We show how to robustly compute the non-linear Lagrangian bias from clustering measurements in a numerical way, enabling us to get unbiased dark matter field reconstructions at initial cosmic times. We also show that we can greatly recover the information of the dark matter field from the galaxy distribution based on a detailed simulation. Novel key ingredients to this approach are a higher order Hamiltonian sampling technique and a non-diagonal Hamiltonian mass-matrix. This technique could be used to study the Eulerian galaxy bias from galaxy surveys and could become an ideal baryon acoustic reconstruction technique. In summary, this method represents a general reconstruction technique, including in a self-consistent way a survey mask, non-linear and non-local bias and redshift space distortions, with an efficiency about 20 times superior to previous comparable methods.