Efficient analysis of photoplethysmography (PPG) signals is essential for wearable monitoring, but it remains unclear how much physiological information is preserved in low-dimensional representations especially when DL architectures like autoencoders are used. The latent space of an autoencoder determines how much physiological information is preserved when compressing a PPG signal. We trained a multi-scale convolutional autoencoder with temporal attention on 30-second PPG segments from the MIMIC-III finger oximeter dataset (N=5,074). We systematically varied the bottleneck dimension (d=8 to 72) and loss function (MSE vs. MAE). Generalizability was evaluated on an internal test set and an external wrist-worn dataset (Simband, N=292). Performance was assessed via reconstruction MSE and stratified results by heart rhythm (NSR, AF, PAC/PVC), heart rate and patient was analyzed. Reconstruction error improved consistently with larger bottleneck dimensions, with MSE decreasing from 0.0415 at d=8 to 0.0117 at d=72 on MIMIC-III. However, performance gains gradually diminished beyond d=48. For the best performing model, AF segments showed significantly higher error that PAC/PVC (p_corrected = 0.040) on the MIMIC test set and across both datasets patient-level variability in reconstruction quality dropped substantially as dimension increased (inter-subject standard deviation from 0.0102 at d=8 to 0.0019 at d=72). Reconstruction quality improves with larger bottleneck dimensions, though gains diminish beyond d=48. Smaller dimensions produce higher error and greater patient-level variability. While larger bottlenecks improve consistency across heterogeneous subgroups, the additional benefit at higher dimensions must be weighed against the cost of reduced compression.