I have a dataset with 1000 samples, 1 timestep, and 1 feature (1000,1,1). I see a lot of discussions where they create subsequences (split one timestep column into multiple by using some X number - ex: (100,10,1)) and use that as input to LSTM autoencoder and argue that LSTMs/autoencoder can only learn when the data has multiple timesteps. But, I get good results with just 1 timestep column. What is the correct input here? Is it incorrect if I use univariate (1000,1,1) input?