# The Simplex in $\mathbb{R}^n$ is convex

Problem: Show that $S:= \lbrace v \in \mathbb{R}^m \mid v=\displaystyle \sum_{j=1}^n a_j v_j, \text{ with } a_1, \dots , a_m \in [0,1], \ \sum_{j=1}^m a_j=1 \rbrace$ the Simplex of $\mathbb{R}^n$ with vertices $v_1, \dots , v_m$ is convex.

So I have to verify the following definition

Definition (convex): $M \subset V$ where $V$ is a real or complex Vectorspace is called convex if $\forall (a,b) \in M^2$ and $\forall \lambda \in \mathbb{R}$ such that $0 \leq \lambda \leq 1$ the following always holds $\lambda a + (1-\lambda)b \in M$

My approach: Geometrically this property is very intuitive, for instance the 3-Simplex or Tetrahedron is of course convex, I will never be able to construct a line that somehow leaves the defined region. I do however fail when it comes to show this property in an analytic approach.

Let $(v,v^*) \in S^2$ and $\lambda \in [0,1]$, I then need to verify that: $$\lambda v+ (1-\lambda)v^* \in S \iff \lambda \sum_{j=1}^n a_j v_j + (1-\lambda) \sum_{j=1}^m a_j^* v_j^* \in S$$ What I can see is that for sure $\lambda \cdot a_j \in [0,1]$ and similarly $(1-\lambda) a_j^* \in [0,1]$ . But that is already the only thing I can grasp from this problem. Do I need to work with the norm and try to find an upper bound for the scalars? I would appreciate some hints on how to continue/get started on this problem

• The vertices of the simplex are the same for $v$ and $v^\ast$, so $$\lambda v + (1-\lambda)v^\ast = \lambda\sum_{j=1}^n a_jv_j + (1-\lambda)\sum_{j=1}^n a_j^\ast v_j.$$ Can you see that you can group some terms to get what you need? – Daniel Fischer Apr 16 '14 at 16:33
• @DanielFischer Intuitively I would merge the scalars into one and then say that $\lambda v + (1- \lambda)v^*= \sum_{j=1}^n b_jv_j$, then I would still have to verify that $\sum_{j=1}^n b_j=1$right? But since the vertices are the same, the same argument should hold for the scalars I presume. – Spaced Apr 16 '14 at 16:38
• The scalars are (generally) different, but you have a formula for $b_j$, what is that? – Daniel Fischer Apr 16 '14 at 16:44
• $\lambda \sum_{j=1}^n a_j v_j + (1-\lambda) \sum_{j=1}^n a_j^* v_j = \sum_{j=1}^n ( \lambda a_j + (1-\lambda) a_j^*) v_j \implies b_j = \lambda a_j + (1-\lambda) a_j^*$. Summing these up leads to $\sum_{j=1}^n b_j = \lambda \sum_{j=1}^n a_j + (1-\lambda) \sum_{j=1}^n a_j^* = \lambda + 1 - \lambda = 1$ Thanks @DanielFischer, your answers/comments never fail to enlighten me. – Spaced Apr 16 '14 at 16:49

Let $u = \sum_k a_k v_k, v = \sum_k b_k v_k$, and $\lambda \in [0,1]$. We have $\sum_k a_k = \sum_k b_k = 1$ and $a_k,b_k \ge 0$.
Let $w = \lambda u + (1-\lambda) v = \sum_k (\lambda a_k + (1-\lambda) b_k) v_k$.
We have $\sum_k (\lambda a_k + (1-\lambda) b_k) = \lambda \sum_k a_k + (1-\lambda) \sum_k b_k = 1$, and $(\lambda a_k + (1-\lambda) b_k) \ge 0$ (where the latter follows because $\lambda, 1-\lambda, a_k,b_k$ are non-negative).
Hence $w \in S$.