Matrix Chernoff Bounds for sums of independent random variables
Summary
In this talk, we proved a matrix-version of a simple Chernoff bound. The scalar version which we will generalize is the following.
Scalar Chernoff
Let $X = \sum_i \eps_ia_i $ where $a_i \in \R$ are fixed, and, $\eps_i$ are $\pm 1$ Rademacher random variables. Then, \[ \prob{\sum_i \eps_ia_i > t} ~\leq~ e^{\frac{-t^2}{2\sigma^2}},\;\; \sigma^2 = \ex{X^2} =\; \sum_i a_i^2 \] Let’s look at a standard proof of this to see where it breaks down. For a single Rademacher variable, one can show that $\ex{e^{\theta \eps a}} = \mathrm{cosh}(\theta a) \leq e^{\frac{\theta^2 a^2}{2}}$. The goal is to break the sum on $n$ such variables as a product of these. Fix $\theta > 0\in \R$. \[ \begin{align} \prob{X > t} ~&=~ \prob{e^{\theta X} > e^{\theta t}}\;\;\small{\text{ (Monotonicity of exponential)} }\\ &\leq e^{-\theta t}\,\ex{ e^{\theta X}} \;\;\small{\text{ (Markov)} }\\ ~&=~ e^{-\theta t}\,\ex{ \prod_i e^{\theta \eps_i a_i}} \;\; \small{(e^{a+b} = e^a\cdot e^b)} \\ ~&=~ e^{-\theta t}\,\prod_i\ex{ e^{\theta \eps_i a_i}} \;\; \small{\text{ (Independence)} } \\ ~&\leq~ e^{-\theta t}\,\prod_i e^{\frac{\theta^2 a_i^2}{2}} \;\; \small{\text{ (Plugging in the 1-variable bound )} }\\ ~&\leq~ e^{\frac{-t^2}{2(\sum_i a_i^2)}} \;\; \small{\text{ (Optimizing for }\theta) } \end{align} \]
Matrix Chernoff
We will generalize the above setup as follows. Let $X = \sum_i \eps_iA_i $ where $A_i$ are fixed Hermitian matrices, and, $\eps_i$ are $\pm 1$ Rademacher random variables. Then, \[ \prob{\lambda_{\max}(X) > t} ~\leq~ e^{\frac{-t^2}{2\sigma^2}} \]
We will give two proofs each of which give a slightly different $\sigma$. The first question is to make sense of exponentials of matrix valued random variables. This, fortunately, is easy.
Lifting functions to matrices
Let $f: I \to \R$ be a function defined on an interval $I$. Let $A = Q\Lambda Q^* $ be an Hermitian matrix such that all eigenvalues of $A$ lies in $I$, i.e., $\mathrm{Spec}(A)\subseteq I$. Then, we can define $f(A) := Qf(\Lambda) Q^*$, where $f(\Lambda)$ is obtained by applying $f$ entry-wise to the diagonal matrix $\Lambda$, i.e., $f(\Lambda) = \mathrm{diag}(f(\lambda_1), \cdots, f(\lambda_d))$.
Thus, we can talk of $e^A$ for any matrix $A$, of $\log (B)$ for any positive-definite matrix $B$. In particular, $\log (\sum_i \alpha_i e^{B_i})$ is well-defined for $\alpha_i > 0$ as $e^{B_i}$ are all positive-definite and the set of PD matrices forms a (open) cone.
Mimicking the scalar proof
One can repeat the scalar argument for the first two steps but it is not clear how to handle the term, $\ex{\lambda_{\max}(e^{\theta X})}$. Ideally, we would like to have $\prod_i \ex{\lambda_\max (e^{\theta \eps_i A_i})}$. The key difficulty is that $e^{A+B}\neq e^Ae^B$ for matrices. This can be resolved (by at least) two approaches.
\[ \begin{align} \small{[\text{Naive Hope}]}\;\;\; &\ex{\lambda_{\max}(e^{\theta X})} ~=~ \ex{\lambda_{\max}\left(\prod_i e^{\theta \eps_i A_i}\right)} ~\leq~ \prod_i \ex{\lambda_\max (e^{\theta \eps_i A_i})} \\ \small{[\text{Ahlswede-Winter}]}\;\;\; &\ex{\lambda_{\max}(e^{\theta X})} ~\leq~ \ex{\tr(e^{\theta X})} ~\leq~ d\prod_i \ex{\lambda_\max (e^{\theta \eps_i A_i})}\\ \small{[\text{Tropp}]}\;\;\; &\ex{\lambda_{\max}(e^{\theta X})} ~\leq~ \ex{\tr(e^{\theta X})} ~\leq~ \tr\, \mathrm{exp}\left( \sum_i \log \ex{(e^{\theta \eps_i A_i})}\right) \\ \end{align} \]
Using the trace inequalities
[Ahlswede-Winter] By definition of the matrix exponential, $\lambda_\max(e^A) = e^{\lambda_\max(A)}$. Thus, treating $\lambda_\max(A)$ as a scalar random variable, we can plug it into the scalar inequality we used earlier, $\ex{e^{\theta \eps a}} \leq e^{\frac{\theta^2 a^2}{2}}$. Thus, we get, $\ex{\lambda_\max (e^{\theta \eps_i A_i})} \leq e^{\frac{\theta^2 \lambda_\max(A_i)^2}{2}}$. It is now exactly like the scalar Chernoff and we get a variance term of $\sum_i \lambda_\max(A_i)^2$.
[Tropp] We need two more facts.
- Firstly, a matrix version of the 1-variable inequality. This is given by $\log \ex{e^{\theta \eps_i A_i}} \preceq \frac{\theta^2 A_i^2}{2}$. Here the order being used is the Loewner order ($A\preceq B$ iff $B-A$ is PSD). The proof is analogous to the scalar proof and is given in Tropp’s book [Lemma 4.6.3].
- The fact that trace-exponential is monotone in the sense that if $A\preceq B$, $\tr\, e^A \leq \tr\, e^B$. This is easy to establish by using the fact that $A\preceq B$ implies $\lambda_i(A)\leq \lambda_i(B)$ which itself follows by Courant-Fischer.
Now, we are ready. \[ \begin{align} \log \ex{e^\theta \eps_i A_i} &~\preceq~ \frac{\theta^2 A_i^2}{2}\;\;\; \small{\text{[Fact 1]}}\\ \sum_i \log \ex{e^\theta \eps_i A_i} &~\preceq~ \frac{\theta^2 \sum_i A_i^2}{2}\\ \tr \; \mathrm{exp}\left(\sum_i \log \ex{e^\theta \eps_i A_i}\right) &~\leq~ \tr \; \mathrm{exp}\left(\frac{\theta^2 \sum_i A_i^2}{2}\right) \;\; \; \small{\text{[Fact 2]}}\\ &~\leq~ d\, \lambda_\max \left( \mathrm{exp}\left( \frac{\theta^2 \sum_i A_i^2}{2}\right)\right)\\ &~=~ d\, \mathrm{exp}\left( \frac{\theta^2 \lambda_\max\left(\sum_i A_i^2\right)}{2}\right). \end{align} \] This, gives us a variance term of $\lambda_\max\left(\sum_i A_i^2\right) $ which is better than the earlier one of $\sum_i \lambda_\max\left(A_i\right)^2$ by up to a factor of $d$. This matters as we have the factor of $d$ in the exponent.
Deriving the trace inequalities
[Golden-Thompson] Ahlswede–Winter use the following sequence of inequalities, \[\tr(e^{A+B}) \leq \tr(e^A e^B) \leq \lambda_\max(e^A)\tr(e^B).\]
The first inequality is called the Golden-Thompson inequality which can be derived from the Lie-Trotter formula which says that although $e^{A+B}\neq e^Ae^B$, this holds in the limit. Formally, $e^{A+B} = \lim_{n\to \infty} \left(e^{A/n}e^{B/n} \right)^{n}$. However, the Golden-Thompson inequality is false for three or more matrices. Thus, it cannot be applied directly. To overcome this, Ahlswede–Winter use the second inequality, a proof of which can be found in Harvey’s notes. We now show the derivation, \[ \begin{align} \ex{\tr(e^{\theta X + 0)})} ~=~ \ex{\tr(e^{\theta \sum_i \eps_i A_i + 0})} ~&\leq~ \ex{\lambda_\max(e^{\eps_1A_1}) \tr(e^{\theta \sum_{i=2}^n \eps_i A_i+ 0})}\\ ~&\leq~ \ex{\prod_{i} \lambda_\max(e^{\eps_iA_i}) \tr(e^{0})} \\ ~&\leq~ d \prod_i \ex{ \lambda_\max(e^{\eps_iA_i})} \end{align} \]
[Lieb’s Concavity] Tropp’s insight is that one must instead work with the cumulant generating function, $\log e^{X}$. The advantage of this POV is that this approach generalizes to a much more general settings. Moreover, as we have seen, it gives tighter bounds. The key is the following result of Lieb.
One now only needs to apply Jensen’s to get the trace inequality, \[ \begin{align} \mathbb{E}\Ex{\eps_n}{\tr\, e^{\sum_{i<n} \eps_i A_i + \eps_n A_n } } ~&=~ \mathbb{E}\Ex{\eps_n}{\tr\, e^{\sum_{i<n} \eps_i A_i+ \log (e^{\eps_nA_n})} }\\ ~&\leq~ \mathbb{E}\, \tr\, e^{\sum_{i<n} \eps_i A_i +\log \Ex{\eps_n}{e^{\eps_nA_n}}} \;\;\small{\text{(Jensen’s)}}\\ ~&\leq~ \tr\, e^{\sum_{i} \log \ex{e^{\eps_i A_i}}} \;\;\small{\text{(Repeating the above steps for all variables)}} \end{align} \]
Resources
Tropp, Joel A. “An Introduction to Matrix Concentration Inequalities.” arXiv.
Lecture Notes on Ahlswede-Winter Inequality by Nicholas Harvey.
Garg, Ankit, Lee, Yin T., Song, Zhao, and Nikhil Srivastava. “A Matrix Expander Chernoff Bound.” arXiv.
Talk by Joel Tropp at [Youtube Link]