Skip to content
TheChatArchiveThe Chat Archive
Geminigemini-2.5-pro-preview-05-06· GoogleHugging Face LMArena Human Preference 140K

infinity norm is the max. That means, the higher the norm (2, 3, ...), the more it…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Geminigemini-2.5-flash’s answer.
  1. User

    infinity norm is the max. That means, the higher the norm (2, 3, ...), the more it privileges the largest component?

  2. Gemini Assistant

    Yes, your intuition is spot on!

    Let's break it down:

    1. L-p norm definition: The L-p norm of a vector x = (x_1, x_2, ..., x_n) is defined as: ||x||_p = (|x_1|^p + |x_2|^p + ... + |x_n|^p)^(1/p)

    2. The effect of the exponent p:

      • When you raise each component |x_i| to the power p, the components with larger absolute values become disproportionately larger compared to the smaller components.
        • Example: Vector [1, 5]
        • For p=2: 1^2 = 1, 5^2 = 25. The ratio is 25:1.
        • For p=3: 1^3 = 1, 5^3 = 125. The ratio is 125:1.
        • For p=10: 1^10 = 1, 5^10 = 9,765,625. The ratio is almost 10 million to 1.
      • As p increases, the sum (|x_1|^p + |x_2|^p + ... + |x_n|^p) becomes overwhelmingly dominated by the term corresponding to the largest absolute component. Let |x_max| be that largest component. The sum will be approximately |x_max|^p.
    3. The effect of the (1/p) root:

      • When you take the p-th root of this sum, (approximately |x_max|^p)^(1/p), you get approximately |x_max|.
    4. The limit: It's a mathematical fact that the L-infinity norm (max norm) is the limit of the L-p norms as p approaches infinity: lim (p -> infinity) ||x||_p = ||x||_infinity = max(|x_i|)

    So, you are correct:

    • L1 norm (p=1): sum(|x_i|). All components contribute equally (additively) according to their magnitude. No special privilege to the largest.
    • L2 norm (p=2): sqrt(sum(|x_i|^2)). The familiar Euclidean distance. Larger components contribute more due to squaring, but it's a balanced contribution.
    • As p increases (p=3, 4, ...): The process of raising to the p-th power increasingly emphasizes the largest component. The p-th root then scales it back, and in the limit, only the largest component's magnitude remains.

    Therefore, higher values of p in the L-p norm indeed mean that the norm is increasingly "sensitive" to or "privileges" the single component with the largest absolute value, eventually becoming just that maximum value for p = infinity.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.