Packet　loss　greatly　influences　the　overall　performance　of　network　applications.　A　novel　statistical　model　for　network　packet　transmission　probability　is　proposed　to　study　statistical　inference　of　network　link-level　performance　in　network　tomography.　Under　the　assumptions　that　link-level　packet　transmission　probability　is　spatially　and　temporally　independent,　the　statistical　model　adopts　cumulant　generating　function　(CGF)　to　analyze　path-level　packet　transmission　probability　from　end-to-end　unicast　measurements　and　link-level　packet　transmission　probability　is　determined　by　the　method　of　moments.　Due　to　the　sum　of　network　packet　transmission　probability　and　loss　probability　being　one,　link-level　packet　loss　information　can　be　indirectly　obtained.　Simulation　experiments　demonstrate　that　the　proposed　model　can　accurately　infer　network　link-level　packet　transmission　probability　information,　obtain　the　order　relation　of　packet　transmission　probability　among　all　links,　and　locate　bottleneck　link　according　to　Chernoff　bound.　Copyright　©2009　Binary　Information　Press.