Question about Secure Multi-Party Computation

I am doing research about Non-interactive Secure Multi-party Computation and encounter a dilemma that I am not quite sure if it is possible and wonder if there are better thoughts that could help.

Situation:

There are $n$ parties ($i \in [n]$) and a dealer. The dealer has a number $U$ and weights $w_i (i \in [n])$ and parties have their inputs $x_i$. Now the parties would like to calculate $U - \sum_{i \in [n]} w_i x_i$ under the limitation:

1. only the parties know $x_i$
2. only the dealer knows $w_i$
3. only the dealer can send messages to parties but parties cannot send messages to the dealer.

I've been thinking of ways like FHE or Secret sharing that seem fine but actually are not.

• Is there any other way that might work or any consideration that might help?
• Or is there any way that we can prove that it is impossible to realize a protocol for this situation?

Edited:

As @poncho mentioned in the comment, it seems impossible if the parties cannot send messages to anybody. So I'd like to change the situation a little bit (an Evaluator is added) and hope there might be other thoughts:

Situation:

There are $n$ parties ($i \in [n]$), a dealer, and an evaluator. Dealers have a number $U$ and weights $w_i (i \in [n])$ and parties have their inputs $x_i$. Now the evaluator would like to calculate $U - \sum_{i \in [n]} w_i x_i$ under the limitation:

1. only the parties know $x_i$
2. only the dealer knows $w_i$
3. Dealer can send messages to parties and evaluator. (single way)
4. Parties can send messages to the evaluator (single way).
5. The evaluator and parties may collude.

Under the new situation, is there any way that might be a helper under a fully robust secure model?

Great thanks to everyone that stops and takes a look.