Continuous non-invasive monitoring of blood pressure is essential for cardiovascular risk patients. This article presents the sparse characterization of Photoplethysmogram (PPG) using K-SVD technique for beat-to-beat blood pressure estimation. The relative changes in the blood volume that is represented by PPG is influenced by the pressure changes in the peripheral circulation. Owing to the anomalous nature of the time domain features obtained from ECG and PPG, we propose the use of sparse representation as means of feature extraction method. The sparse features generated from K-SVD processed dictionary that can approximate the shape of the PPG signal were taken as the features to predict the BP values. The proposed system is evaluated using the Multi-Parameter Intelligent Monitoring for Intensive Care (MIMIC-II) database. The proposed method is compared with the baseline system that employs time domain features for BP prediction. The Mean Absolute Error (MAE) and Root mean square Error (RMSE) between the predicted BP and ground truth BP were chosen as the performance measures. The system achieved the error measure of (MAE ± RMSE) 5.06 ± 6.27 mmHg for systolic BP and 2.99 ± 3.93 mm Hg for Diastolic BP. Further, the comparison studies suggests that the proposed system outperforms the baseline system with an overall reduction in MAE and RMSE by 20.55% and 22.14% respectively for systolic BP and 28.15% and 23.48% respectively for Diastolic BP. Hence, the sparse representation of PPG can be successfully utilized for the beat-to-beat prediction of BP.