I'm playing with BWA-MEM.

I've extracted a pair of reads properly paired (FLAG=2) with bwa-mem:

bwa mem human_g1k_v37.fasta NA12878_1.fastq.gz NA12878_2.fastq.gz | samtools view -S -f 3 -

SRR098401.2297993    99    20    123171    60    76M    =    123252    157    TGGCTGAGCCCAACCCCACCTAAGAAGGACAATAAAGATCTGTGTTCAGAGTCATACTGAATAGAGACTTCTGGAC    BEDFDGBEEFDBDDDGDEDGF@EFC?F?BCEGDDBEFEBFGHCH?BC@FCEADECDHFJC:B@D?HDG=C6A9==8    NM:i:0    AS:i:76    XS:i:0
SRR098401.2297993    147    20    123252    60    76M    =    123171    -157    AGAACCCACTGCCTCCTGATGAAGTCCCTACTGTTCACCCTTGCAGTTTTTATGCTCCTGGCCCAATTGGTCTCAG    <5>09>F:EBGDFDD?DDDEGDFDDF:HDCH8ACCGB:?EF>HGGGEF?DCCEHGC<GC@CCFFAAB>DC@D@EB>    NM:i:0    AS:i:76    XS:i:0

and I put the two reads in two fastq files:

@SRR098401.2297993
TGGCTGAGCCCAACCCCACCTAAGAAGGACAATAAAGATCTGTGTTCAGAGTCATACTGAATAGAGACTTCTGGAC
+
BEDFDGBEEFDBDDDGDEDGF@EFC?F?BCEGDDBEFEBFGHCH?BC@FCEADECDHFJC:B@D?HDG=C6A9==8

and

@SRR098401.2297993
AGAACCCACTGCCTCCTGATGAAGTCCCTACTGTTCACCCTTGCAGTTTTTATGCTCCTGGCCCAATTGGTCTCAG
+
<5>09>F:EBGDFDD?DDDEGDFDDF:HDCH8ACCGB:?EF>HGGGEF?DCCEHGC<GC@CCFFAAB>DC@D@EB>

now, If I re-align those reads with the ~same bwa-mem command

bwa mem  human_g1k_v37.fasta tmp1.fq tmp2.fq | grep -v -E "^@SQ" 2> /dev/null 

SRR098401.2297993    65    20    123171    60    76M    =    123252    82    TGGCTGAGCCCAACCCCACCTAAGAAGGACAATAAAGATCTGTGTTCAGAGTCATACTGAATAGAGACTTCTGGAC    BEDFDGBEEFDBDDDGDEDGF@EFC?F?BCEGDDBEFEBFGHCH?BC@FCEADECDHFJC:B@D?HDG=C6A9==8    NM:i:0    AS:i:76    XS:i:0
SRR098401.2297993    129    20    123252    60    76M    =    123171    -82    AGAACCCACTGCCTCCTGATGAAGTCCCTACTGTTCACCCTTGCAGTTTTTATGCTCCTGGCCCAATTGGTCTCAG    <5>09>F:EBGDFDD?DDDEGDFDDF:HDCH8ACCGB:?EF>HGGGEF?DCCEHGC<GC@CCFFAAB>DC@D@EB>    NM:i:0    AS:i:76    XS:i:0

The pair is no more flagged as "mapped in proper pair". Does bwa-mem needs a large number of pairs to set this flag ? how can I run some tests with a small amount of sequences ?