I have 35 tumor and 4 normal samples. I'm using DESeq2 for differential analysis. Differential analysis between tumor and normal gave only two upregulated genes which could be due to statistical power. So, I'm interested in selection of random samples from tumor condition and do differential analysis with that and repeat the process `n` times. 

I have a matrix with genes as rows and samples as columns. Columns 1-35 are tumor and 36-39 are normal samples.

    nb.replicates <- 10
    samples.Normal <- sample(36:39, replace=FALSE)
    set.seed(123)
    
    ## Random sampling of the Tumor
    samples.Tumor <- sample(c(1:35), size=nb.replicates, replace=FALSE)
    samples.Tumor
    
    selected.samples <- c(samples.Normal, samples.Tumor)

So, with the above code I repeated differential analysis `n` number of times. I have different number of differential expressed genes with each analysis. 

Now, from all the analysis should I merge and consider only common genes as differentially expressed genes for the whole cohort?

In each analysis it is 4 normal vs and 10 random tumor samples.

                baseMean        log2FoldChange    lfcSE       stat    pvalue    padj
    AL357060.1    8.50582           6.1871       1.67335    3.54023    0.0003    0.03245

In another analysis it is 4 normal vs another 10 random tumor I see the same gene differentially expressed but with different values as results

                  baseMean     log2FoldChange       lfcSE        stat    pvalue    padj
    AL357060.1    10.58937424    6.552371044      1.6296950    3.85921    0.00011    0.02642

There are many genes with different results in different analysis so, which one should I consider? 

You will not increase power by reducing the sample size in one group.

You may get different sets of genes, and sometimes more than you get in the full analysis, but this is induced by your random process of sub-sampling. It's not an increase in power per se but just random fluctuations. So to answer "Now, from all the analysis should I merge and consider only common genes as differentially expressed genes for the whole cohort?": no, I would not recommend to do this random subsampling procedure.

I would recommend to instead increase the FDR cutoff and only look at the full dataset.