*4.1. Transmission Results*

We first measure the performance of our transparent fiber-optical and THz wireless 2 × 2 MIMO transmission system for the back-to-back (BtB) case, i.e., without fiber and wireless distance transmission. Figure 4 gives the measured BER of X- and Y-polarization versus different input power into each AIPM. At the 3.8 × <sup>10</sup>−<sup>3</sup> HD-FEC limit, the THzwave carrier frequency at 500 GHz for the BtB case has a successfully transmitted range from 28 GBaud to 36 GBaud. The insets show that the QPSK constellation points can be demodulated well at 28 GBaud. However, the BER performance degrades at high transmission rates due to the limitations of the receiver LNA bandwidth.

**Figure 4.** BER versus input power into each AIPM for the BtB case without fiber and wireless transmission. (**a**) X-polarization; (**b**) Y-polarization.

Then, we measure the BER versus the input power into each AIPM over one span of 20-km SSMF and 3-m wireless distance, as shown in Figure 5. At the 3.8 × <sup>10</sup>−<sup>3</sup> HD-FEC limit, the THz-wave carrier frequency at 500 GHz for the fiber and wireless transmission can also be successfully transmitted. However, the transmission performance is not ideal at 36 GBaud. Figure 6 gives the electrical spectrum of the 24/28/32/36 GBaud QPSK IF signal with the corresponding bandwidth (BW). The signals are more damaged when the transmission speed increases. Moreover, we can see that the required bandwidth becomes limited as the transmission rate increases, especially at the 36 GBaud rate with 36.36 GHz bandwidth. To protect the AIPMs, the maximum input optical power is set at 13.5 dBm. A THz-wave carrier frequency at 500 GHz can be successfully transmitted at the 3.8 × <sup>10</sup>−<sup>3</sup> HD-FEC threshold. The best BER performance occurs for the lower transmission rates. The insets also show that the constellation points can be demodulated well. In order to further improve the performance for higher transmission rates, the J-DBN equalizer based on the conventional DSP algorithm is introduced.

**Figure 5.** BER versus input power into each AIPM over one span of 20-km SSMF and 3-m wireless transmission. (**a**) X-polarization; (**b**) Y-polarization.

**Figure 6.** Electrical spectrum of the received QPSK IF signal: (**a**) 24 GBaud signal with 24.24 GHz BW; (**b**) 28 GBaud signal with 28.28 GHz BW; (**c**) 32 GBaud signal with 32.32 GHz BW; (**d**) 36 GBaud signal with 36.36 GHz BW.

Here, the training data are only used for the DBN-1 adaptive equalizer in our proposed J-DBN equalizer since the DBN-2 blind equalizer and optimization is a self-recovering equalization method without the aid of a training sequence. It indicates that the J-DBN equalizer has good training accuracy and satisfactory tracking speed. Figure 7 illustrates the BER of 36 GBaud QPSK signals versus the input optical power into each AIPM and SNR over one span of 20-km SSMF and 3-m wireless distance; it can be found that increasing the optical power can help to improve the BER performance due to the larger SNR. Next, we compare four equalization schemes (including *Opt*. 1–*Opt*. 4) with a 53-tap CMA equalizer, 37-tap DD-LMS equalizer, DBN-1 equalizer with 25 cells, and DBN-2 equalizer with 50 cells. From the comparison between *Opt*. 1 and *Opt*. 4, it can be concluded that the required power under HD-FEC (3.8 × <sup>10</sup><sup>−</sup>3) utilizing the J-DBN scheme is close to 12.6 dBm, which increases by almost up to 0.2 dB in receiver sensitivity and 0.8 dB in SNR gain compared with the conventional scheme. Moreover, *Opt*. 2 and *Opt*. 3 also obtain a slight performance improvement after the DBN-1 equalizer or DBN-2 equalizer. When the input power into the AIPM is 12.8 dBm, the illustrations depict the constellation diagrams of QPSK symbols after recovery. The received QPSK symbol constellation before the receiver DSP chain is also given. Furthermore, we compared the constellation diagrams of the *Opt*. 1 and *Opt*. 4 schemes. The illustrations (i) and (ii) also show that the J-DBN equalizer can reduce the residual error after convergence and visually improve the nonlinear decision capacity, which is suitable for the nonlinear channel balance.

**Figure 7.** BER versus input power into each AIPM over one span of 20-km SSMF and 3-m wireless transmission for different proposed equalizers. (**a**) X-polarization. (**b**) Y-polarization.
