diff --git a/deflate.c b/deflate.c
index 7f421e4..2d638ca 100644
--- a/deflate.c
+++ b/deflate.c
@@ -674,36 +674,50 @@
 }
 
 /* =========================================================================
- * For the default windowBits of 15 and memLevel of 8, this function returns
- * a close to exact, as well as small, upper bound on the compressed size.
- * They are coded as constants here for a reason--if the #define's are
- * changed, then this function needs to be changed as well.  The return
- * value for 15 and 8 only works for those exact settings.
+ * For the default windowBits of 15 and memLevel of 8, this function returns a
+ * close to exact, as well as small, upper bound on the compressed size. This
+ * is an expansion of ~0.03%, plus a small constant.
  *
- * For any setting other than those defaults for windowBits and memLevel,
- * the value returned is a conservative worst case for the maximum expansion
- * resulting from using fixed blocks instead of stored blocks, which deflate
- * can emit on compressed data for some combinations of the parameters.
+ * For any setting other than those defaults for windowBits and memLevel, one
+ * of two worst case bounds is returned. This is at most an expansion of ~4% or
+ * ~13%, plus a small constant.
  *
- * This function could be more sophisticated to provide closer upper bounds for
- * every combination of windowBits and memLevel.  But even the conservative
- * upper bound of about 14% expansion does not seem onerous for output buffer
- * allocation.
+ * Both the 0.03% and 4% derive from the overhead of stored blocks. The first
+ * one is for stored blocks of 16383 bytes (memLevel == 8), whereas the second
+ * is for stored blocks of 127 bytes (the worst case memLevel == 1). The
+ * expansion results from five bytes of header for each stored block.
+ *
+ * The larger expansion of 13% results from a window size less than or equal to
+ * the symbols buffer size (windowBits <= memLevel + 7). In that case some of
+ * the data being compressed may have slid out of the sliding window, impeding
+ * a stored block from being emitted. Then the only choice is a fixed or
+ * dynamic block, where a fixed block limits the maximum expansion to 9 bits
+ * per 8-bit byte, plus 10 bits for every block. The smallest block size for
+ * which this can occur is 255 (memLevel == 2).
+ *
+ * Shifts are used to approximate divisions, for speed.
  */
 uLong ZEXPORT deflateBound(strm, sourceLen)
     z_streamp strm;
     uLong sourceLen;
 {
     deflate_state *s;
-    uLong complen, wraplen;
+    uLong fixedlen, storelen, wraplen;
 
-    /* conservative upper bound for compressed data */
-    complen = sourceLen +
-              ((sourceLen + 7) >> 3) + ((sourceLen + 63) >> 6) + 5;
+    /* upper bound for fixed blocks with 9-bit literals and length 255
+       (memLevel == 2, which is the lowest that may not use stored blocks) --
+       ~13% overhead plus a small constant */
+    fixedlen = sourceLen + (sourceLen >> 3) + (sourceLen >> 8) +
+               (sourceLen >> 9) + 4;
 
-    /* if can't get parameters, return conservative bound plus zlib wrapper */
+    /* upper bound for stored blocks with length 127 (memLevel == 1) --
+       ~4% overhead plus a small constant */
+    storelen = sourceLen + (sourceLen >> 5) + (sourceLen >> 7) +
+               (sourceLen >> 11) + 7;
+
+    /* if can't get parameters, return larger bound plus a zlib wrapper */
     if (deflateStateCheck(strm))
-        return complen + 6;
+        return (fixedlen > storelen ? fixedlen : storelen) + 6;
 
     /* compute wrapper length */
     s = strm->state;
@@ -740,11 +754,12 @@
         wraplen = 6;
     }
 
-    /* if not default parameters, return conservative bound */
+    /* if not default parameters, return one of the conservative bounds */
     if (s->w_bits != 15 || s->hash_bits != 8 + 7)
-        return complen + wraplen;
+        return (s->w_bits <= s->hash_bits ? fixedlen : storelen) + wraplen;
 
-    /* default settings: return tight bound for that case */
+    /* default settings: return tight bound for that case -- ~0.03% overhead
+       plus a small constant */
     return sourceLen + (sourceLen >> 12) + (sourceLen >> 14) +
            (sourceLen >> 25) + 13 - 6 + wraplen;
 }