Improved upper limits on the 21-cm signal power spectrum at z=17.0 and z=20.3 from an optimal field observed with NenuFAR

We report the deepest upper limits to date on the power spectrum of the 21-cm signal during the Cosmic Dawn (redshifts: $z>15$), using four nights of observations with NenuFAR. The limits are derived from two redshift bins, centred at $z=20.3$ and $z=17.0$, with integration times of 26.1 h and 23.6 h, from observations of an optimal target field chosen to minimise sidelobe leakage from bright sources. Our analysis incorporates improvements to the data processing pipeline, particularly in subtracting strong radio sources in the primary beam sidelobes and mitigating low-level radio frequency interference, yielding a 50-fold reduction in the excess variance compared to a previous analysis of the north celestial pole field. At $z=20.3$, we achieve a best $2\sigma$ upper limit of $\Delta^{2}_{21}<4.6 \times 10^5 \, \textrm{mK}^{2}$ at $k=0.038$ $h\, \mathrm{cMpc}^{-1}$, while at $z=17.0$, the best limit is $\Delta^{2}_{21}<5.0 \times 10^6 \, \textrm{mK}^{2}$ at $k=0.041$ $h\, \mathrm{cMpc}^{-1}$. These are the strongest constraints on the 21-cm power spectrum at the respective redshifts, with the limit at $z = 20.3$ being deeper by more than an order of magnitude over all previous Cosmic Dawn power spectrum limits. Comparison against simulated exotic 21-cm signals shows that while the $z=20.3$ limits begin to exclude the most extreme models predicting signals stronger than the EDGES detection, an order-of-magnitude improvement would constrain signals compatible with EDGES. A coherence analysis reveals that the excess variance is largely incoherent across nights for the $z=20.3$ redshift bin, suggesting that deeper integrations could yield significantly stronger constraints on the 21-cm signal from the Cosmic Dawn.