See also <ulink url="https://www.unicode.org/reports/tr10">Unicode Technical
     Standard 10</ulink> for more information on the terminology.
    </para>

    <para>
     To create a nondeterministic collation, specify the property
     <literal>deterministic = false</literal> to <command>CREATE
     COLLATION</command>, for example:
<programlisting>
CREATE COLLATION ndcoll (provider = icu, locale = 'und', deterministic = false);
</programlisting>
     This example would use the standard Unicode collation in a
     nondeterministic way.  In particular, this would allow strings in
     different normal forms to be compared correctly.  More interesting
     examples make use of the ICU customization facilities explained above.
     For example:
<programlisting>
CREATE COLLATION case_insensitive (provider = icu, locale = 'und-u-ks-level2', deterministic = false);
CREATE COLLATION ignore_accents (provider = icu, locale = 'und-u-ks-level1-kc-true', deterministic = false);
</programlisting>
    </para>

    <para>
     All standard and predefined collations are deterministic, all
     user-defined collations are deterministic by default.  While
     nondeterministic collations give a more <quote>correct</quote> behavior,
     especially when considering the full power of Unicode and its many
     special cases, they also have some drawbacks.  Foremost, their use leads
     to a performance penalty.  Note, in particular, that B-tree cannot use
     deduplication with indexes that use a nondeterministic collation.  Also,
     certain operations are not possible with nondeterministic collations,
     such as some pattern matching operations.  Therefore, they should be used
     only in cases where they are specifically wanted.
    </para>

    <tip>
     <para>
      To deal with text in different Unicode normalization forms, it is also
      an option to use the functions/expressions
      <function>normalize</function> and <literal>is normalized</literal> to
      preprocess or check the strings, instead of using nondeterministic
      collations.  There are different trade-offs for each approach.
     </para>
    </tip>
   </sect3>
  </sect2>

  <sect2 id="icu-custom-collations">
   <title>ICU Custom Collations</title>

   <para>
    ICU allows extensive control over collation behavior by defining new
    collations with collation settings as a part of the language tag. These
    settings can modify the collation order to suit a variety of needs. For
    instance:

<programlisting>
-- ignore differences in accents and case
CREATE COLLATION ignore_accent_case (provider = icu, deterministic = false, locale = 'und-u-ks-level1');
SELECT 'Å' = 'A' COLLATE ignore_accent_case; -- true
SELECT 'z' = 'Z' COLLATE ignore_accent_case; -- true

-- upper case letters sort before lower case.
CREATE COLLATION upper_first (provider = icu, locale = 'und-u-kf-upper');
SELECT 'B' &lt; 'b' COLLATE upper_first; -- true

-- treat digits numerically and ignore punctuation
CREATE COLLATION num_ignore_punct (provider = icu, deterministic = false, locale = 'und-u-ka-shifted-kn');
SELECT 'id-45' &lt; 'id-123' COLLATE num_ignore_punct; -- true
SELECT 'w;x*y-z' = 'wxyz' COLLATE num_ignore_punct; -- true
</programlisting>

    Many of the available options are described in <xref
    linkend="icu-collation-settings"/>, or see <xref
    linkend="icu-external-references"/> for more details.
   </para>

   <sect3 id="icu-collation-comparison-levels">
    <title>ICU Comparison Levels</title>

    <para>
     Comparison of two strings (collation) in ICU is determined by a
     multi-level process, where textual features are grouped into
     "levels". Treatment of each level is controlled by the <link
     linkend="icu-collation-settings-table">collation settings</link>. Higher
     levels correspond to finer textual features.
    </para>

    <para>
     <xref linkend="icu-collation-levels"/> shows which textual feature
     differences are considered significant when